Anchor ring steel strand cable force identification method based on neural network

Abstract

The invention discloses a ring anchor steel strand cable force identification method based on a neural network, and particularly relates to the technical field of structural health monitoring, vibration measurement is performed on a prestressed ring anchor steel strand, an actual measurement time domain signal is obtained and post-processing is performed, and a normalized and noise-reduced time frequency domain signal is obtained; the method comprises the following steps of: establishing and dividing data sets together with structural characteristic parameters of a steel strand, dividing the data sets into a training data set, a verification data set and a test data set in proportion, initializing a GRU network model, performing hyper-parameter optimization on the GRU model through SSA to obtain an SSA-GRU combined network prediction model for identifying the cable force value of the ring anchor prestressed steel strand, and predicting the cable force value of the ring anchor prestressed steel strand according to the SSA-GRU combined network prediction model. And the model identification accuracy is verified by a test data set.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and in particular to a method for identifying tension in a ring anchor steel strand based on a neural network; Background Art

[0002] The vibration frequency method is generally used to measure the cable tension of steel strands. That is, the steel strands are struck to generate vibrations, and the fundamental frequencies of each order are identified through the vibration signals of the steel strands. Then, the cable tension is calculated using the steel string vibration equation. The vibration frequency method itself has the defect of difficulty in ensuring accuracy due to the complex environment and the difference between calculation conditions and reality. It is even more limited in the application of ring anchor prestressed steel strand tension testing. Signal extraction may result in the loss of key information. The ring anchor test section is short in length and the boundary conditions are completely different from those of the inclined cable, resulting in large differences in vibration characteristics and deviations in test results. Steel strands may exhibit nonlinear behavior when subjected to prestress, which may make simple linear vibration models unable to accurately describe their dynamic characteristics. Therefore, a neural network-based ring anchor steel strand tension identification method is urgently needed to solve this problem. Summary of the Invention

[0003] In order to solve the problem that the vibration frequency method has defects in measuring the tension of ring-anchor prestressed steel strands, the method of the present invention can give a more accurate steel strand tension after learning through the GRU neural network model based on the vibration signal data and structural parameters of the steel strand, which can be used for real-time and accurate identification of the tension of ring-anchor prestressed steel strands.

[0004] The specific technical solutions provided are as follows: A method for identifying tension in a ring anchor strand based on a neural network specifically comprises the following steps: S1: Use vibration sensors to measure the vibration of different ring anchor prestressed steel strands and record their time domain signals and structural parameters; S2: Preprocess the collected vibration time domain signal, obtain the time-frequency domain signal through continuous wavelet transform, and perform normalization and noise reduction; S3: The processed time-frequency domain signals and the structural parameters of the ring anchor steel strand are combined as input data, and the corresponding cable force is used as the output layer. The data set is established and divided into training data set, validation data set and test data set according to the proportion; S4: Initialize the GRU network model, update the parameters of the GRU network model through SSA, and train the GRU network model; S5: Input the actual ring anchor steel strand structure data to be predicted into the trained GRU network model, and quickly obtain the predicted ring anchor steel strand tension from the GRU model; Preferably, step S1 includes: The vibration sensor is attached to the longer rope, 5-10 cm away from the bottom of the longer rope or 5-10 cm below the middle anchor; The longer cable end of the ring anchor model was struck at different tensions: 50kN, 100kN, 150kN, and 200kN. The vibration signal of the steel strand was recorded after each strike, ensuring that sufficient sample signals were recorded to cover different vibration modes. The structural parameters of the steel strand were also recorded, including the number of steel strands N, the lengths of the upper and lower steel strands L1 and L2, the weight of the anchor M, the linear density of the steel strand ρ, and the corresponding cable force. By changing the structural parameters, a large number of different ring-anchor model sample data are obtained; Preferably, step S2 specifically includes: S21: Select Morlet as the wavelet basis function , input the original vibration time domain signal x [n] ,length There are 1024 sampling points; ; in, is the center frequency, , is the time variable; S22: Perform continuous wavelet transform (CWT) in discrete form ; in: is the input signal, is the scale parameter, is the translation parameter, is the complex conjugate of the wavelet function, N is the signal length; S23: Normalize the coefficients after continuous wavelet transform CWT and scale them to the range of [0,1]. The Min-Max normalization formula is: ; in, is the normalized value, is each element in the CWT coefficient matrix, and are the maximum and minimum values ​​in the matrix respectively; S24: Soft threshold denoising is performed on the normalized CWT coefficients: ; in, is the normalized transform coefficient of the input, is the threshold value set; Preferably, step S3 specifically includes: By changing the structural parameters, including the anchor mass M, the number of steel strands N, and the lengths of the upper and lower steel strands L1 and L2, a large number of different ring anchor models were measured to generate a large amount of labeled sample data. The sample data was divided into a training data set (70%), a validation data set (15%), and a test data set (15%). Preferably, step S4 specifically includes: S41: Use Keras' Functional API to define the two input branches of the GRU network model: the time-frequency domain signal data branch and the structural feature branch. The processed time-frequency domain signal is 1024 samples long and has five structural features (number of strands N, lengths of upper and lower strands L1 and L2, anchor weight M, and strand density ρ). Define the time-frequency domain signal data input layer and the structural feature data input layer separately. S42: Processing time-frequency domain signal data: In the multi-layer GRU of the GRU network model, each GRU layer (with ReLU activation function) is followed by a Dropout layer to prevent overfitting, and finally an attention mechanism layer is added to extract key information; S43: Processing structural feature data: Use a fully connected layer (Dense layer) to process structural feature data and connect a Dropout layer to prevent overfitting; S44: Use the Concatenate layer to merge the context vector generated by the attention mechanism with the feature vector processed by the structural features. The output layer is a fully connected layer (Dense layer) that maps the merged feature vector to a numerical output as the cable force value. The activation function is linear. S45: Use the root mean square error (RMSE) indicator to comprehensively evaluate the performance of the GRU network model; ; in, is the true value, is the predicted value, is the sample size; Preferably, the steps of SSA for hyperparameter optimization of the GRU network model are as follows: The hyperparameters that define the optimization of the GRU network model include the number of hidden units, Dropout rate, learning rate, and optimizer; The Sobol sequence is used to generate an efficient and uniform initial sparrow population, the number of iterations and the ratio of predators to joiners are set, and the fitness is calculated and sorted. Simulate the foraging process, update the positions of the discoverer, joiner, and sentinel, and add adaptive weights to the discoverer position update. The formula is as follows:

[0005] in, is the number of iterations, ; is the upper limit of the number of iterations; represents the individual position in the search space; ; is a safe value and ; is the warning value and ; is a random number that obeys Gaussian distribution; equal Matrix of Calculate the fitness value and update the position of the sparrows. The optimization iterative process adopts the elite retention strategy: retain the current best fitness individual to enter the next generation; if the new sparrow is better than the historical best individual, the historical best individual is replaced; Set the fitness threshold to 0.95, the maximum number of iterations to 100, and the fitness change rate to 1e-6. If the algorithm iteration stop condition is met, output the global optimal solution and its related parameters. Otherwise, repeat the iteration until the iteration stop condition set by the algorithm is met. Preferably, step S5 specifically includes: Once the training is complete and passes verification and testing, it can be deployed in real-world applications to identify the cable tension of ring-anchor prestressed steel strands in real time. This involves real-time acquisition and processing of vibration signals, combined with the trained GRU network model, to quickly and accurately predict cable tension values. The present invention has the following advantages: 1. The present invention statistically analyzes a large amount of vibration test data of ring-anchor prestressed steel strands, establishes a data set based on preprocessed time-frequency domain signal characteristics and structural characteristic parameters, and proportionally divides the data set into a training data set, a validation data set, and a test data set. A ring-anchor prestressed steel strand tension prediction model is constructed using a neural network. The hyperparameters of the neural network are optimized and automatically selected using the SSA (Sparrow Search Optimization Algorithm), resulting in a mature ring-anchor prestressed steel strand tension prediction model. This model accurately predicts the cable tension of the ring-anchor prestressed steel strand, avoiding the errors that may occur in the fundamental frequency identification process of the vibration frequency method and eliminating the need to consider complex boundary conditions. This allows for more accurate cable tension values ​​to be obtained, achieving automatic optimization of the hyperparameters in the network model. The optimized network model can obtain accurate cable tension values ​​using time-domain signals and structural parameters, replacing the cumbersome calculation process of the traditional vibration frequency method. 2. In addition to being applicable to the tension identification of ring anchor steel strands, the present invention can also be used to identify the tension of special cables or prestressed steel strands in other building structures, thus having a certain degree of versatility. 3. The calculation method of the present invention is simple and has a fast calculation speed. Real-time measurement can be performed after the ring anchor is tensioned. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a test diagram of the ring anchor prestressed steel strand provided by the present invention; Figure 2 A diagram showing the structural parameters of the steel strand provided by the present invention; Figure 3 The time-frequency diagram provided by the present invention; Figure 4 This is a structural diagram of the GRU network prediction model provided by the present invention; DETAILED DESCRIPTION

[0007] The following describes the implementation of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention. like Figure 1-Figure 4 As shown, this embodiment provides a method for identifying the tension of a ring anchor steel strand based on a neural network, which specifically includes the following steps: S1: Use high-sensitivity vibration sensors to measure the vibration of different ring-anchor prestressed steel strands and record their time domain signals and structural parameters; S2: Preprocess the collected vibration time domain signal, obtain the time-frequency domain signal through continuous wavelet transform, and perform normalization and noise reduction; S3: The processed time-frequency domain signals and the structural parameters of the ring anchor steel strand are combined as input data, and the corresponding cable force is used as the output layer. The data set is established and divided into training data set, validation data set and test data set according to the proportion; S4: Initialize the GRU network model, update the parameters of the GRU network model through SSA, and train the GRU network model; S5: Input the actual ring anchor steel strand structure data to be predicted into the trained GRU network model, and quickly obtain the predicted ring anchor steel strand tension from the GRU model; As an embodiment, step S1 includes: The high-sensitivity vibration sensor is attached to the longer rope and is 5-10 cm from the bottom or 5-10 cm below the middle anchor; The longer cable end of the ring anchor model was struck at different tensions: 50kN, 100kN, 150kN, and 200kN. The vibration signal of the steel strand was recorded after each strike, ensuring that sufficient sample signals were recorded to cover different vibration modes. The structural parameters of the steel strand were also recorded, including the number of steel strands N, the lengths of the upper and lower steel strands L1 and L2, the weight of the anchor M, the linear density of the steel strand ρ, and the corresponding cable force. By changing the structural parameters, a large number of different ring-anchor model sample data are obtained; As an embodiment, step S2 specifically includes: S21: Select Morlet as the wavelet basis function , input the original vibration time domain signal x [n] ,length There are 1024 sampling points; ; in, is the center frequency, , is the time variable; S22: Perform continuous wavelet transform (CWT) in discrete form ; in: is the input signal, is the scale parameter, is the translation parameter, is the complex conjugate of the wavelet function, N is the signal length; S23: Normalize the coefficients after continuous wavelet transform CWT and scale them to the range of [0,1]. The Min-Max normalization formula is: ; in, is the normalized value, is each element in the CWT coefficient matrix, and are the maximum and minimum values ​​in the matrix respectively; S24: Soft threshold denoising is performed on the normalized CWT coefficients: ; in, is the normalized transform coefficient of the input, is the threshold value set; As an embodiment, step S3 specifically includes: By changing the structural parameters, including the anchor mass M, the number of steel strands N, and the lengths of the upper and lower steel strands L1 and L2, a large number of different ring anchor models were measured to generate a large amount of labeled sample data. The sample data was divided into a training data set (70%), a validation data set (15%), and a test data set (15%). As an embodiment, step S4 specifically includes: S41: Use Keras' Functional API to define the two input branches of the GRU network model: the time-frequency domain signal data branch and the structural feature branch. The processed time-frequency domain signal is 1024 samples long and has five structural features (number of strands N, lengths of upper and lower strands L1 and L2, anchor weight M, and strand density ρ). Define the time-frequency domain signal data input layer and the structural feature data input layer separately. S42: Processing time-frequency domain signal data: In the multi-layer GRU of the GRU network model, each GRU layer (with ReLU activation function) is followed by a Dropout layer to prevent overfitting, and finally an attention mechanism layer is added to extract key information; S43: Processing structural feature data: Use a fully connected layer (Dense layer) to process structural feature data and connect a Dropout layer to prevent overfitting; S44: Use the Concatenate layer to merge the context vector generated by the attention mechanism with the feature vector processed by the structural features. The output layer is a fully connected layer (Dense layer) that maps the merged feature vector to a numerical output as the cable force value. The activation function is linear. S45: Use the root mean square error (RMSE) indicator to comprehensively evaluate the performance of the GRU network model; ; in, is the true value, is the predicted value, is the sample size; As an implementation method, the steps of SSA for hyperparameter optimization of the GRU network model are as follows: Hyperparameters for defining GRU network model optimization include but are not limited to the number of hidden units, dropout rate, learning rate, and optimizer; The Sobol sequence is used to generate the initial sparrow population, the number of iterations and the ratio of predators to joiners are set, and the fitness is calculated and sorted; Simulate the foraging process, update the positions of the discoverer, joiner, and sentinel, and add adaptive weights to the discoverer position update. The formula is as follows: ; in, is the number of iterations, ; is the upper limit of the number of iterations; represents the individual position in the search space; ; is a safe value and ; is the warning value and ; is a random number that obeys Gaussian distribution; equal Matrix of Calculate the fitness value and update the position of the sparrows. The optimization iterative process adopts the elite retention strategy: retain the current best fitness individual to enter the next generation; if the new sparrow is better than the historical best individual, the historical best individual is replaced; Set the fitness threshold to 0.95, the maximum number of iterations to 100, and the fitness change rate to 1e-6. If the algorithm iteration stop condition is met, output the global optimal solution and its related parameters. Otherwise, repeat the iteration until the iteration stop condition set by the algorithm is met. As an embodiment, step S5 specifically includes: Once the training is complete and passes verification and testing, it can be deployed in real-world applications to identify the cable tension of ring-anchor prestressed steel strands in real time. This involves real-time acquisition and processing of vibration signals, combined with the trained GRU network model, to quickly and accurately predict cable tension values. As an implementation method, specific experimental scenarios include: The experiment formulates a tensioning system consisting of upper steel strands, lower steel strands and intermediate anchorage devices, such as Figure 1 As shown in Figure 1, the upper and lower anchors are used to simulate the fixed state of the prestressed steel strand in the actual concrete structure. The middle anchor connects the upper and lower steel strands as shown in Figure 1. Figure 2 As shown, the length ratio of the upper and lower cables L1:L2 is randomly selected in the range of 1:2-1:6, the lower cable is a longer cable and the length L2 is randomly selected in the range of 0.5m-2m, the number of anchor holes is 2N, and the experimental loading device adopts a microcomputer-controlled electro-hydraulic servo pressure testing machine. A piezoelectric sensor is used, which is adsorbed on either end of the lower steel strand. The steel strand is struck with a hammer to stimulate its vibration, and the vibration signal generated is collected and analyzed. The vibration signal is collected by a vibration test analyzer to obtain the steel strand vibration signal. The sampling frequency is about ten times the fundamental frequency, and the average number of times is 2-4 times, and the vibration time domain signal is obtained. After Morlet wavelet transform, the time-frequency domain signal is obtained as shown below. Figure 3 As shown in the figure, the actual cable tension of the steel strand is determined through the neural network.

[0008] Although the present invention has been described in detail above using general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made on the basis of the present invention; therefore, these modifications or improvements made without departing from the spirit of the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for identifying the tension of a ring anchor strand based on a neural network, characterized in that: include: Step S1: using a vibration sensor to measure the vibration of different ring anchor prestressed steel strands, and recording their time domain signals and structural parameters; Step S2: preprocessing the collected vibration time domain signal, obtaining the time-frequency domain signal through continuous wavelet transform, and performing normalization and noise reduction; Step S3: The processed time-frequency domain signal and the structural parameters of the ring anchor steel strand are combined as input data, and the corresponding cable force is used as the output layer. The data set is established and divided, and is divided into a training data set, a verification data set, and a test data set in proportion; Step S4: Initialize the GRU network model, update the parameters of the GRU network model through SSA, and train the GRU network model; Step S5: Input the ring-anchor steel strand structure data actually required to be predicted into the trained GRU network model, and quickly obtain the predicted ring-anchor steel strand tension from the GRU model.

2. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S1 includes: The vibration sensor is attached to the longer rope, and the vibration sensor is set at a position 5-10 cm from the bottom of the longer rope or 5-10 cm below the middle anchor; Apply different tensile forces to the ring anchor model, and knock the longer cable end at 50kN, 100kN, 150kN, and 200kN tensile forces respectively. Record the vibration signal of the steel strand after each knock, and ensure that the recorded sample signal covers different vibration modes; at the same time, record the structural parameters of the steel strand, including the number of steel strands N, the lengths of the upper and lower steel strands L1 and L2, the weight of the anchor M, the linear density of the steel strand ρ and its corresponding cable force; By changing the structural parameters, sample data of different ring anchor models are obtained.

3. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S2 includes: S21: Select Morlet as the wavelet basis function , input the original vibration time domain signal x [n] ,length There are 1024 sampling points; ； in, is the center frequency, , is the time variable; S22: Perform continuous wavelet transform CWT in discrete form; ； in: is the input signal, is the scale parameter, is the translation parameter, is the complex conjugate of the wavelet function, N is the signal length; S23: Normalize the coefficients after continuous wavelet transform CWT and scale them to the range of [0,1]. The Min-Max normalization formula is: ； in, is the normalized value, is each element in the CWT coefficient matrix, and are the maximum and minimum values ​​in the matrix respectively; S24: Soft threshold denoising is performed on the normalized continuous wavelet transform CWT coefficients: ； in, is the normalized transform coefficient of the input, The threshold to set.

4. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S3 includes: By changing the structural parameters, including the anchor mass M, the number of steel strands N, the lengths of the upper and lower steel strands L1 and L2, and the steel strand density ρ, a large number of different ring anchor models are measured to generate a large amount of labeled sample data, which are divided into 70% training data set, 15% validation data set and 15% test data set.

5. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S4 includes: S41: Use Keras' Functional API to define two input branches of the GRU network model: the time-frequency domain signal data branch and the structural feature branch. The length of the processed time-frequency domain signal is 1024 sampling points, and there are 5 structural features, which are the number of steel strands N, the lengths of the upper and lower steel strands L1 and L2, the weight of the anchor M, and the linear density of the steel strand ρ. The time-frequency domain signal data input layer and the structural characteristic data input layer are also defined. S42: Processing time-frequency domain signal data: In the multi-layer GRU of the GRU network model, each GRU layer is followed by a Dropout layer to prevent overfitting, and finally an attention mechanism layer is added to extract key information; S43: Processing structural feature data: Use the fully connected layer Dense layer to process the structural feature data and connect the Dropout layer to prevent overfitting; S44: Use the Concatenate layer to merge the context vector generated by the attention mechanism with the feature vector processed by the structural characteristics. The output layer is a fully connected layer. The Dense layer maps the merged feature vector to a numerical output as the cable force value. The activation function is linear. S45: Use the root mean square error RMSE indicator to comprehensively evaluate the performance of the GRU network model; ； in, is the true value, is the predicted value, is the sample size.

6. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S4, the steps of SSA for hyperparameter optimization of the GRU network model are as follows: The hyperparameters that define the optimization of the GRU network model include the number of hidden units, Dropout rate, learning rate, and optimizer; The Sobol sequence is used to generate the initial sparrow population, the number of iterations and the ratio of predators to joiners are set, and the fitness is calculated and sorted; Simulate the foraging process, update the positions of the discoverer, joiner, and guard, and add adaptive weights to the discoverer position update. The formula is as follows: ； in, is the number of iterations, ; is the upper limit of the number of iterations; represents the individual position in the search space; ; is a safe value and ; is the warning value and ; is a random number that follows a Gaussian distribution; equal Matrix of Calculate the fitness value and update the position of the sparrows. The optimization iterative process adopts the elite retention strategy: retain the current best fitness individual to enter the next generation; if the new sparrow is better than the best individual in history, replace the best individual in history; The fitness threshold is set to 0.95, the maximum number of iterations is set to 100, and the fitness change rate is set to 1e-6. If the algorithm iteration stop condition is met, the global optimal solution and its related parameters are output. Otherwise, the iteration is repeated until the iteration stop condition set by the algorithm is met.

7. The neural network-based method for identifying the tension of a ring anchor strand according to claim 1, characterized in that: Step S5 includes: Once the training is completed and passes verification and testing, it can be deployed in actual applications to identify the cable tension of the ring-anchor prestressed steel strand in real time: through real-time collection and processing of vibration signals, combined with the trained GRU network model, the cable tension value can be predicted quickly and accurately.

Images