BP neural network-based concrete box girder bridge temperature time lag prediction and elimination method

Abstract

The invention discloses a BP neural network-based concrete box girder bridge temperature time-lag prediction and elimination method, and belongs to the bridge health detection technology, the method provides a temperature time-lag calculation index, constructs a BP neural network-based daily temperature difference and temperature time-lag coefficient relation model, can determine the temperature time-lag under different sunshine temperature differences, and provides a BP neural network-based concrete box girder bridge temperature time-lag prediction and elimination method. The problem that a conventional time domain translation method is not suitable for a long-term monitoring data sequence with large temperature fluctuation is solved, the correlation between bridge structure response monitoring data and structure temperature data is greatly improved, and support is provided for structure state diagnosis.

Description

Technical Field

[0001] The invention belongs to bridge health detection technology, and specifically relates to a temperature lag prediction and elimination method for a concrete box girder bridge based on a BP neural network. Background Art

[0002] Concrete box girder bridges have the advantages of high bending and torsional rigidity, good overall performance, strong bearing capacity, and mature construction technology. They have been widely used in national and provincial trunk bridges. However, with the increase in the service life of bridges, concrete bridges are long-term affected by environmental factors such as solar radiation, seasonal changes, and temperature changes. The temperature field inside and on the surface of the concrete structure presents an uneven distribution state. The internal secondary stress of the structure caused by temperature often exceeds the stress generated by other loads, which may lead to local cracking of concrete, instability of components and other safety and durability issues. Therefore, the impact of temperature effect on concrete box girder bridges needs to be taken seriously. Analysis of the impact of temperature on the structure generally requires a correlation analysis of temperature-structure response, and then diagnose the health status of the bridge structure based on the change in temperature. However, due to the existence of temperature lag effect, that is, the change in structural response caused by temperature change (such as beam end displacement, main beam displacement, cable force, etc.) lags behind the temperature change, and the correlation analysis between temperature and structural response is inaccurate, which can easily lead to omission and misreporting of bridge diseases.

[0003] Taking the displacement of concrete bridge ends as an example, due to the thermal expansion and contraction characteristics of reinforced concrete materials, an increase in temperature will cause the displacement of the beam ends to increase (expansion), and a decrease in temperature will cause the displacement of the beam ends to decrease (contraction). In areas where the temperature fluctuates greatly between day and night, the thermal expansion and contraction effects of concrete bridges may be more significant. If the temperature changes and the displacement of the beam ends changes abnormally, the expansion joint may be damaged.

[0004] Existing research mainly eliminates the temperature lag effect through the time domain translation method, that is, the difference between the structural response and the temperature peak and valley time is taken as the temperature lag, and the bridge structure response data is translated forward on the time scale. However, the time lag caused by different temperatures is not a fixed value. In long-term monitoring, the temperature lag is nonlinear. The method of translating a single time is suitable for short-term monitoring data series with small temperature fluctuations, but not for long-term monitoring data series with large temperature fluctuations. This makes the elimination of the temperature lag effect poor, resulting in poor fitting of the correlation between the structural response and the structural temperature. Summary of the invention

[0005] Purpose of the invention: To solve the problem that the temperature lag effect exists in the analysis of the correlation between structural response and temperature, and the existing temperature lag translation method is not suitable for short-term monitoring data series with small temperature fluctuations, especially the existing technology does not take into account that the temperature lag is nonlinear and is not suitable for long-term monitoring data series with large temperature fluctuations, and the effect of eliminating the temperature lag effect is poor, the present invention proposes a temperature lag prediction and elimination method for concrete box girder bridges based on BP neural network.

[0006] Technical solution: A method for predicting and eliminating temperature lag of concrete box girder bridge based on BP neural network, the method steps include:

[0007] (1) Collect beam end displacement and structural temperature data, and align the beam end displacement and structural temperature data sequence time, specifically:

[0008] Compare the start collection time t_dis(1) and t_tmp(1) of the beam end displacement and the structural temperature, and select the maximum value max(t_dis(1), t_tmp(1)) as the common start collection time of the two monitoring items;

[0009] Compare the end acquisition time t_dis(end) and t_tmp(end) of the beam end displacement and the structural temperature, and select the minimum value min(t_dis(end), t_tmp(end)) as the common end acquisition time of the two monitoring items;

[0010] The beam end displacement and structure temperature collection times are aligned through the determined start collection time and end collection time;

[0011] (2) Perform sliding average processing on the beam end displacement and structural temperature data. The time interval of the sliding average processing is adjusted to solve the problem of inconsistent sampling frequency of beam end displacement and structural temperature, including translating the beam end displacement data series on the time scale;

[0012] (3) Perform cross-correlation analysis on the beam end displacement and structural temperature data series, including calculating the Pearson correlation coefficient, drawing the cross-correlation coefficient curve, finding the cycle time corresponding to the maximum correlation coefficient, and obtaining the temperature lag in a day;

[0013] By calculating the cross-correlation function of the beam end displacement and structure temperature data series, the time offset that maximizes the correlation between the two signals is found, and the offset represents the temperature lag;

[0014] Keep the beam end displacement value unchanged, and shift the beam end displacement data points on the time scale by the temperature lag value lag to eliminate the temperature lag every day;

[0015] (4) Record the highest temperature of the processing unit during the period of time , minimum temperature , Temperature hysteresis , calculate the temperature hysteresis coefficient Qi:

[0016] ,i=1,2,…,n

[0017]

[0018] (5) Counting daily temperature differences within a certain period of time and temperature lag coefficient Qi, construct the temperature lag data set, and convert the daily temperature difference As an independent variable, the temperature hysteresis coefficient Q i As the dependent variable, a three-layer BP neural network model was built;

[0019] Temperature difference between days and temperature hysteresis coefficient Q i The data set is divided into a training set and a test set, and the BP neural network is trained to obtain a temperature time-lag neural network model that can predict different daily temperature differences;

[0020] (6) The daily temperature difference on the target date Input neural network model to calculate and predict temperature lag Q on target date i .

[0021] Furthermore, step (3) is to determine the similarity between the two data series by cross-correlation analysis, which is used to evaluate the correlation between the structural temperature time series and the beam end displacement time series at different time delays;

[0022] The calculation of Pearson's correlation coefficient is:

[0023]

[0024] In the formula, represents the Pearson correlation coefficient, is the amount of time series data, , are the values ​​of the i-th sequence of time series variables X and Y, respectively. , is the mean of the sequence variables X and Y;

[0025] Furthermore, the BP neural network model constructed in step (5) of the method includes an input layer, a hidden layer and an output layer;

[0026] Information forward propagation: data is input into the model and calculated through the activation function;

[0027] Error back propagation: Calculate and update the weights and biases from the hidden layer to the input layer;

[0028] Model fitting output: The error function is used to measure the degree of model fitting. The mean square error is used as the error function in BP neural network training. The smaller the MSE value, the higher the accuracy of the prediction model. Finally, the updated weights, thresholds and biases are recalculated and iterated continuously. When the planned number of iterations or the optimal error expectation value is reached, the training is completed.

[0029] The method collects the beam end displacement and structure temperature data in units of days, and regards the data sequence within each day as a processing unit, including looping the data sequence within a processing unit;

[0030] For long-term data series of one year or more, the data series is cycled daily to obtain a 365-day time-lag series, and then the time lag of each day is subtracted to align the structural temperature and beam end displacement data for each day, thereby obtaining an accurate linear relationship.

[0031] Furthermore, for the forward propagation of information in the BP neural network model package, the data is first normalized and then input into the model, calculated through the activation function, and then the error is back-propagated to calculate and update the weights and biases from the hidden layer to the input layer;

[0032] The BP neural network model uses mean square error as the error function in BP neural network training. The smaller the MSE value, the higher the accuracy of the prediction model. Finally, the updated weights, thresholds and biases are recalculated and iterated continuously. When the planned number of iterations or the optimal error expected value is reached, the training is completed.

[0033] On the other hand, the present application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for predicting and eliminating temperature lag of a concrete box girder bridge based on a BP neural network as described above.

[0034] A computer storage medium stores a computer program. When the computer program is executed by a computer, the method for predicting and eliminating the temperature lag of a concrete box girder bridge based on a BP neural network as described above is implemented.

[0035] Beneficial effects: The method described in the present invention proposes a temperature lag calculation index and constructs a relationship model between daily temperature difference and temperature lag coefficient based on BP neural network, which can clarify the temperature lag under different sunshine temperature differences, solves the problem that conventional time domain translation methods are not applicable to long-term monitoring data series with large temperature fluctuations, greatly improves the correlation between bridge structure response monitoring data and structural temperature data, and provides support for structural status diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the method of the present invention;

[0037] Figure 2 It is a schematic diagram of the three-layer BP neural network structure. DETAILED DESCRIPTION

[0038] In order to explain the technical solution provided by the present invention in detail, a further introduction is given below in conjunction with the accompanying drawings.

[0039] Combination Figure 1 , the present invention provides a method for predicting and eliminating temperature lag of concrete box girder bridges based on BP neural network. Taking the displacement-temperature correlation analysis of a bridge end in Huai'an as an example, it should be noted that since the beam end displacement sensor of a bridge in Huai'an is arranged as a connection between the main beam of the approach bridge and the main beam of the main bridge, the decrease in the value of the beam end displacement sensor reflects the elongation of the main beam, and the calculated regression relationship shows a negative correlation. At present, front-line technicians manually calculate the temperature lag of a certain day. On the one hand, the number of manual calculations is limited, the efficiency is low, and the lag is basically calculated by guesswork, which is highly subjective. On the other hand, the temperature lag is different every day. This method mainly includes the following methods for eliminating temperature lag:

[0040] S1. Take the data sequence of each day as the processing unit and execute 2-9 loop operations.

[0041] The loop operation is for long-term data of more than one year. The temperature lag is nonlinear. The method of translating a single time is not suitable for long-term monitoring data series with large temperature fluctuations. Therefore, here we loop 365 days a year to obtain a 365-day lag sequence, and then subtract the lag of each day to achieve the alignment of the structural temperature and beam end displacement data for each day, and then obtain an accurate linear relationship. The purpose of statistics for one year is that the seasonal and daily temperature differences between years are basically consistent, and the linear relationship of previous years can be used to infer future correlations.

[0042] S2, compare the start collection time t_dis(1) and t_tmp(1) of the beam end displacement and the structure temperature, and select the maximum value max(t_dis(1), t_tmp(1)) as the common start collection time of the two monitoring items;

[0043] S3, compare the end collection time t_dis(end) and t_tmp(end) of the beam end displacement and the structure temperature, and select the minimum value min(t_dis(end), t_tmp(end)) as the common end collection time of the two monitoring items;

[0044] S4, aligning the beam end displacement and the structure temperature acquisition time through steps S2 and S3;

[0045] S5. Perform sliding average processing on the beam end displacement and structural temperature at intervals of 5 minutes to solve the problem of inconsistency between the structural response and the structural temperature sampling frequency;

[0046] S6, with 1 minute as the initial cycle value, 2 hours as the final cycle value, and 1 minute as the increment, the beam end displacement data series is translated on the time scale;

[0047] S7. Perform cross-correlation analysis on the beam end displacement and structure temperature data series, calculate the Pearson correlation coefficient cor, draw the cross-correlation coefficient curve, and find the cycle time corresponding to the maximum correlation coefficient cor_max, which is the temperature lag in a day;

[0048] Cross-correlation analysis is a statistical method for determining the similarity between two data series. It is often used in time series analysis to evaluate the correlation between two data series at different time delays. In this case, it is the structural temperature time series and the beam end displacement time series.

[0049] It is further pointed out that the Pearson correlation coefficient is a method to measure the linear correlation between two variables. The linear correlation degree of the variables can be evaluated through the covariance matrix of the data. The Pearson correlation coefficient is:

[0050]

[0051] In the formula, r XY is the Pearson correlation coefficient, n represents the amount of time series data, Xi and Yi are the values ​​of the i-th sequence of the time series variables X and Y respectively, , is the mean of the sequence variables X and Y.

[0052] r XY is the Pearson correlation coefficient of two variables X and Y, ranging from -1 to 1. The larger the absolute value, the more correlated the two variables are. When the value is 0, the two variables are unrelated. If the correlation coefficient is greater than 0, the two variables are positively correlated. If the correlation coefficient is less than 0, the two variables are negatively correlated. The Pearson correlation coefficient gives the similarity of two time series data at different time offsets, which can help us find the time lag between two sets of time series data.

[0053] In the analysis of bridge beam end displacement and structural temperature, cross-correlation analysis can be used to determine the response lag time of beam end displacement to structural temperature change. By calculating the cross-correlation function of the beam end displacement and structural temperature data series, the time offset that maximizes the correlation between the two signals can be found. This offset represents the temperature lag. Specifically, starting from 0 minutes, with a step of 1 minute, the time delay of the beam end displacement data is gradually increased on the time scale until 2 hours. Each value on the horizontal axis represents the time to delay the beam end displacement data relative to the structural temperature data.

[0054] S8, keeping the beam end displacement value unchanged, shifting the beam end displacement collection data point on the time scale by the temperature lag value lag, so as to eliminate the temperature lag every day;

[0055] S9. Record the highest temperature of each day imax , minimum temperature t imin , temperature lag i , calculate the temperature hysteresis coefficient Q i :

[0056] , i=1,2,…,n(1)

[0057] (2)

[0058] It should be noted that in this step, the present invention constructs statistical indicators and temperature hysteresis coefficient Qi. In the unprocessed data stage, the linear relationship between beam end displacement and temperature hysteresis is not obvious, and it is impossible to assist in disease diagnosis based on their relationship, so cross-correlation analysis is required.

[0059] S10, statistics of daily temperature difference a for 365 days i and temperature hysteresis coefficient Q i , construct a temperature time-lag dataset, and convert the daily temperature difference a i As an independent variable, the temperature hysteresis coefficient Q i As the dependent variable, a three-layer BP neural network model is built.

[0060] S11, the daily temperature difference a i and temperature hysteresis coefficient Q i The data set is divided into a training set and a test set according to a ratio of 9:1. The BP neural network is trained to obtain a temperature lag neural network model that can predict different daily temperature differences.

[0061] S12, the daily temperature difference a of the target date i Input the neural network model, and the temperature lag Q of the target date can be calculated and predicted. i .

[0062] Among them, the Matlab code for aligning the beam end displacement and structural temperature acquisition time can be as follows:

[0063] t1 = max(t_dis(1),t_tmp(1));

[0064] t2 = min(t_dis(end),t_tmp(end));

[0065] ind_DIS = t_dis >= t1 & t_dis <= t2;

[0066] ind_TMP = t_tmp >= t1 & t_tmp <= t2;

[0067] t_dis = t_dis(ind_DIS);

[0068] dis = disf(ind_DIS);

[0069] t_tmp = t_tmp(ind_TMP);

[0070] tmp = tmp(ind_TMP);

[0071] In summary, in order to solve the problem of temperature lag effect existing in structural response-temperature correlation analysis, the existing temperature lag translation method is applicable to short-term monitoring data series with small temperature fluctuations, but not applicable to long-term monitoring data series with large temperature fluctuations, and the elimination effect of temperature lag effect is poor. The present invention proposes a temperature lag calculation index, constructs a relationship model between daily temperature difference and temperature lag coefficient based on BP neural network, which can clarify the temperature lag under different sunshine temperature differences, and realizes the calculation, elimination and prediction of nonlinear temperature lag by optimizing and improving the BP algorithm, solving the problem that conventional time domain translation method is not applicable to long-term monitoring data series with large temperature fluctuations, greatly improving the correlation between bridge structure response monitoring data and structural temperature data, providing support for structural status diagnosis, and thoroughly solving the problem of current subjective determination of temperature lag, which has great practicality.

[0072] In order to make the present invention better understood by those skilled in the art, the three-layer BP neural network model and its algorithm constructed by the present invention are further introduced.

[0073] The BP (Back Propagation, BP) neural network model is a multi-layer feedforward network trained by the error back propagation algorithm. The present invention can fully approximate any complex nonlinear relationship through a three-layer BP neural network, and has a high degree of self-learning and self-adaptive ability. Its basic idea as a temperature lag elimination is: taking the daily temperature difference ai as the independent variable and the temperature lag coefficient Qi as the dependent variable, the regression analysis method is used to fit the explicit expression of the independent variable with respect to the dependent variable.

[0074] The calculation process of the BP algorithm is first the forward propagation of information, inputting the data into the model and calculating through the activation function, followed by the back propagation of the error, calculating and updating the weights and biases from the hidden layer to the input layer. The error function is usually used to measure the degree of model fitting. The present invention adopts the mean square error (MSE) as the error function in the BP neural network training. The smaller the value of MSE, the higher the accuracy of the prediction model. Finally, the updated weights, thresholds and biases are recalculated, and iterated continuously. When the proposed number of iterations or the optimal error expectation value is reached, the training is completed. The model construction process is as follows:

[0075] Below Figure 2 The three-layer neural network structure shown is used to illustrate the calculation process of the BP algorithm.

[0076] (1) First, information is forward propagated. In order to avoid the long training time of the BP neural network, the inconsistent dimensions of the input data or the large gap between the input data, the input and output data need to be normalized according to the following formula:

[0077] (3)

[0078] represents any value in the training sample, represents the mean of the training samples, represents the variance of the training samples, Represents the normalized data of the original training samples.

[0079] So as to obtain the normalized input sample ,Depend on Figure 2 As we know, the first layer is the input layer, which contains three neurons and the bias term ; The second layer is the hidden layer, which contains two neurons and the bias term ; The third output layer has only one neuron , in the figure is the weight of the connection between layers. The activation function from the input layer to the hidden layer and from the hidden layer to the output layer takes sigmoid as an example. The formula is as follows:

[0080] (4)

[0081] The input of the hidden layer is calculated as:

[0082] (5)

[0083] (6)

[0084] The output of the hidden layer is calculated as:

[0085] (7)

[0086] (8)

[0087] The input of the output layer is calculated as

[0088] (9)

[0089] The output of the output layer is calculated as

[0090] (10)

[0091] (2) Then comes the error back propagation. The error function is usually used to measure the degree of model fitting. This paper uses the mean square error (MSE) as the error function in BP neural network training. The smaller the MSE value, the higher the accuracy of the prediction model.

[0092] (11)

[0093] represents the number of training samples, represents the neural network prediction data of the jth sample, Represents the true value of the jth sample, and the training error of this forward information propagation is:

[0094] (12)

[0095] Calculate the weight and bias updates from the output layer to the hidden layer:

[0096] (13)

[0097] (14)

[0098] in:

[0099] , ,

[0100] , ,

[0101] We can get:

[0102] (15)

[0103] (16)

[0104] The same can be said

[0105] (17)

[0106] Update weights and biases:

[0107] (18)

[0108] (19)

[0109] in It is the step size, which determines the length of each step in the direction of the negative gradient during the error back propagation process.

[0110] Calculate the weight and bias updates from the hidden layer to the input layer:

[0111] (20)

[0112] (twenty one)

[0113] in:

[0114] ,

[0115] , , ,

[0116] We can get:

[0117] (twenty two)

[0118] (twenty three)

[0119] Similarly, we can get:

[0120] (twenty four)

[0121] (25)

[0122] (26)

[0123] Update weights and biases:

[0124] (27)

[0125] (28)

[0126] This completes the back propagation of the error, and finally the updated weights, thresholds and biases are recalculated and iterated continuously. When the planned number of iterations or the optimal expected error value is reached, the training is completed.

[0127] For the application of the three-layer BP neural network model, when restoring the BP neural network prediction value to the actual value, the output data needs to be denormalized. The denormalization formula is: .

Claims

1. A method for predicting and eliminating temperature lag of concrete box girder bridge based on BP neural network, characterized in that: The method steps include: (1) Collect beam end displacement and structural temperature data, and align the beam end displacement and structural temperature data sequence time, specifically: Compare the start collection time t_dis(1) and t_tmp(1) of the beam end displacement and the structural temperature, and select the maximum value max(t_dis(1), t_tmp(1)) as the common start collection time of the two monitoring items; Compare the end acquisition time t_dis(end) and t_tmp(end) of the beam end displacement and the structural temperature, and select the minimum value min(t_dis(end), t_tmp(end)) as the common end acquisition time of the two monitoring items; The beam end displacement and structure temperature collection times are aligned through the determined start collection time and end collection time; (2) Perform sliding average processing on the beam end displacement and structural temperature data. The time interval of the sliding average processing is adjusted to solve the problem of inconsistent sampling frequency of beam end displacement and structural temperature, including translating the beam end displacement data series on the time scale; (3) Perform cross-correlation analysis on the beam end displacement and structural temperature data series, including calculating the Pearson correlation coefficient, drawing the cross-correlation coefficient curve, finding the cycle time corresponding to the maximum correlation coefficient, and obtaining the temperature lag in a day; By calculating the cross-correlation function of the beam end displacement and structure temperature data series, the time offset that maximizes the correlation between the two signals is found, and the offset represents the temperature lag; Keep the beam end displacement value unchanged, and shift the beam end displacement data points on the time scale by the temperature lag value lag to eliminate the temperature lag every day; (4) Record the highest temperature of the processing unit during the period of time , minimum temperature , Temperature hysteresis , calculate the temperature hysteresis coefficient Qi: ,i=1,2,…,n ； (5) Counting daily temperature differences within a certain period of time and temperature lag coefficient Qi, construct the temperature lag data set, and convert the daily temperature difference As an independent variable, the temperature hysteresis coefficient Q i As the dependent variable, a three-layer BP neural network model was built; Temperature difference between days and temperature hysteresis coefficient Q i The data set is divided into a training set and a test set, and the BP neural network is trained to obtain a temperature time-lag neural network model that can predict different daily temperature differences; (6) The daily temperature difference on the target date Input neural network model to calculate and predict temperature lag Q on target date i .

2. The method for predicting and eliminating temperature lag of concrete box girder bridge according to claim 1 is characterized in that: Step (3) is to determine the similarity between the two data series through cross-correlation analysis, which is used to evaluate the correlation between the structural temperature time series and the beam end displacement time series at different time delays; The calculation of Pearson's correlation coefficient is: ； In the formula, represents the Pearson correlation coefficient, is the amount of time series data, , are the values ​​of the i-th sequence of time series variables X and Y, respectively. , is the mean of the sequence variables X and Y.

3. The method for predicting and eliminating temperature lag of concrete box girder bridge according to claim 1 is characterized in that: The BP neural network model constructed in step (5) includes an input layer, a hidden layer and an output layer; Information forward propagation: data is input into the model and calculated through the activation function; Error back propagation: Calculate and update the weights and biases from the hidden layer to the input layer; Model fitting output: The error function is used to measure the degree of model fitting. The mean square error is used as the error function in BP neural network training. The smaller the MSE value, the higher the accuracy of the prediction model. Finally, the updated weights, thresholds and biases are recalculated and iterated continuously. When the planned number of iterations or the optimal error expectation value is reached, the training is completed.

4. The method for predicting and eliminating temperature lag of concrete box girder bridge according to claim 1, characterized in that: The method collects the beam end displacement and structure temperature data in units of days, and regards the data sequence within each day as a processing unit, including looping the data sequence within a processing unit; For long-term data series of one year or more, the data series is cycled daily to obtain a 365-day time-lag series, and then the corresponding time lag is subtracted to align the structural temperature and beam end displacement data for each day to obtain an accurate linear relationship.

5. The method for predicting and eliminating temperature lag of concrete box girder bridge according to claim 3 is characterized in that: For the forward propagation of information in the BP neural network model package, the data is first normalized and then input into the model, calculated through the activation function, and then the error is back-propagated to calculate and update the weights and biases from the hidden layer to the input layer; The BP neural network model uses mean square error as the error function in BP neural network training. The smaller the MSE value, the higher the accuracy of the prediction model. Finally, the updated weights, thresholds and biases are recalculated and iterated continuously. When the planned number of iterations or the optimal error expected value is reached, the training is completed.

6. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the temperature lag prediction and elimination method of concrete box girder bridge based on BP neural network as described in any one of claims 1 to 5.

7. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a method for predicting and eliminating temperature lag of a concrete box girder bridge based on a BP neural network as described in any one of claims 1 to 5 is implemented.

Images