A bridge damping ratio identification method based on monitoring data and deep learning

Abstract

This invention provides a bridge damping ratio identification method based on monitoring data and deep learning. This method improves the U-Net model, a deep learning method that can perform semantic segmentation on images, making it suitable for processing time-series acceleration monitoring data. A semantic segmentation dataset for free decay response segments is constructed, and the improved U-Net model is trained, performance evaluated, and optimized. The optimized improved U-Net model is used to perform semantic segmentation on massive amounts of bridge acceleration monitoring data, extracting free decay response segments with ideal decay shapes. The damping ratio is then calculated using the exponential decay method. The bridge damping ratio identification method based on monitoring data and deep learning provided by this invention addresses the difficulty of efficiently screening free decay response segments in massive amounts of monitoring data. This allows the exponential decay method to be applied to bridge monitoring data, achieving a more reliable calculation of the damping ratio than other methods.

Description

Technical Field

[0001] The present invention relates to the field of bridge structure health monitoring and computer deep learning application technology, and in particular to a bridge damping ratio identification method based on monitoring data and deep learning. Background Art

[0002] The damping ratio characterizes the energy dissipation capacity of a bridge structure under dynamic loads and plays a key role in dynamic analysis for purposes such as bridge seismic resistance, wind resistance, and human-induced vibration control. Relevant specifications generally set the damping ratio for bridges between 0.003 and 0.05, depending on the structural material and excitation type. However, operational bridges are affected by a variety of factors, including the environment and operational loads, and the structural damping ratio does not remain constant. If the damping ratio values recommended by the specifications are used in dynamic analysis, the calculated dynamic response may deviate significantly from the actual situation, making it difficult to understand the structure's true vibration conditions for feedback in structural dynamic design or vibration control. Therefore, it is essential to reliably observe the damping ratio of actual bridge structures and accurately understand its long-term variations.

[0003] Because it is impossible to interrupt traffic for long periods of time to conduct targeted dynamic testing, the identification of the damping ratio of bridge structures requires the use of long-term acceleration monitoring data observed by a bridge structural health monitoring (SHM) system under environmental excitation, calculated using operational modal analysis. Specific algorithms include the half-power bandwidth method, random subspace method, eigensystem realization method, Hilbert-Huang transform, autoregressive method with external input, and frequency domain decomposition method. However, each of these algorithms has limitations when calculating the damping ratio from SHM data. For example, the half-power bandwidth method can produce large errors when the external excitation cannot be observed and has high requirements for frequency resolution. The random subspace method and eigensystem realization method assume that the environmental excitation is white noise, which is inconsistent with actual conditions. The Hilbert-Huang transform is an empirical signal processing method with inherent issues. As a result, the damping ratio identified using monitoring data often exhibits significant dispersion, making the results less reliable.

[0004] Numerous studies have confirmed that the exponential decay method, based on the free decay response of the structure, provides more accurate and reliable damping ratio calculations. However, the exponential decay method presupposes the presence of a free decay dynamic response in the observed data. Therefore, it is often used in dynamic tests that simulate pulse excitations, such as vehicle jumping or falling objects. Due to the need for traffic closures, the number of such tests is limited, insufficient to generate statistically significant long-term observations of the damping ratio. In fact, bridge SHM data also contain free decay response data, which can be generated by the excitation of a single vehicle passing through the bridge. For small-span bridges, when traffic volume is relatively low and speeds are high, there are many single-vehicle load conditions. This provides a large sample of free decay response data for the exponential decay method, enabling more reliable identification and long-term analysis of damping ratio patterns. However, the free decay response in SHM data can exhibit varying durations and patterns depending on vehicle speed, weight, and bridge length. Manual screening of available free decay response segments by professionals is feasible, but given the massive amount of monitoring data, this process is extremely time-consuming and labor-intensive, making it unfeasible. Therefore, how to automatically identify the free decay response segment suitable for damping ratio calculation in massive monitoring data and accurately locate its occurrence and end time is a key problem to be solved. It is very necessary to design a bridge damping ratio identification method based on monitoring data and deep learning. Summary of the Invention

[0005] The purpose of the present invention is to provide a bridge damping ratio identification method based on monitoring data and deep learning, which can automatically identify the free decay response segments suitable for damping ratio calculation in massive monitoring data, and then calculate the damping ratio of the bridge structure using the exponential decay method, providing a reliable data basis for the analysis of the long-term change law of the damping ratio.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A bridge damping ratio identification method based on monitoring data and deep learning includes the following steps:

[0008] Step 1: Collect bridge acceleration monitoring data, manually screen the free decay response segments with ideal decay shapes in the bridge acceleration monitoring data, and semantically annotate the time periods where they occur to construct a semantic segmentation dataset for the free decay response segments.

[0009] Step 2: Based on a deep neural network that performs image semantic segmentation in computer vision, a time series semantic segmentation model is constructed to identify free decay response segments from bridge acceleration monitoring data.

[0010] Step 3: Train, evaluate, and fine-tune the semantic segmentation model using the semantic segmentation dataset.

[0011] Step 4: Input the unlabeled acceleration monitoring data into the tuned semantic segmentation model to identify and locate the acceleration free decay response segment in the acceleration monitoring data;

[0012] Step 5: Correct the acceleration free decay response segment obtained by identification and positioning, pick up the peak and valley values in each vibration cycle of the free decay segment, and calculate the damping ratio using the exponential decay method.

[0013] Optionally, in step 1, bridge acceleration monitoring data is collected, and free decay response segments with ideal decay shapes are manually screened from the bridge acceleration monitoring data. The time periods in which they occur are semantically annotated to construct a semantic segmentation dataset for the acceleration free decay response segments. Specifically,

[0014] Bridge acceleration monitoring data was collected. For each acceleration time series data in the bridge acceleration monitoring data, label time series data with the same timestamp was generated. The initial labels of all moments were set to 0. The free decay segments in the acceleration time series data were manually screened, and the label values of the initial labels of the time periods corresponding to the free decay segments were modified to 1. The acceleration time series data were standardized so that their mean was 0 and their standard deviation was 1. The standardized data were split into training set, validation set, and test set in a ratio of 0.8:0.1:0.1.

[0015] Optionally, based on a deep neural network that performs image semantic segmentation in computer vision, a time series semantic segmentation model is constructed to identify free decay response segments from bridge acceleration monitoring data. Specifically:

[0016] Improve the U-Net deep neural network model for performing image semantic segmentation tasks in computer vision. Replace the two-dimensional convolutional layer in the U-Net model with a one-dimensional convolutional layer, and change the input layer length. Replace the ReLU activation function in the U-Net model with the LeakyReLU activation function. Modify the sizes of the encoder and decoder convolutional layers in the U-Net model so that the jump connections between the symmetrically positioned convolutional layers of the encoder and decoder can be implemented by direct splicing. Modify the length and number of channels of the output layer of the U-Net model, and change the output layer activation function to Sigmoid.

[0017] Optionally, train, evaluate, and fine-tune the semantic segmentation model using a semantic segmentation dataset. Specifically:

[0018] The semantic segmentation model is trained and verified using the training set and validation set. The intersection-over-union (IoU) is defined as the performance metric of the semantic segmentation model, and the performance of the model is evaluated using the test set. The convolution kernel size of the semantic segmentation model is optimized.

[0019] Optionally, unlabeled acceleration monitoring data is input into the tuned semantic segmentation model to identify and locate the acceleration free decay response segment in the acceleration monitoring data, specifically:

[0020] The unlabeled acceleration monitoring data is standardized and then input into the tuned semantic segmentation model to obtain the probability value of each data point in the acceleration monitoring data belonging to the free decay response segment. A judgment threshold is set to identify data segments with probability values greater than the judgment threshold as free decay response segments. The acceleration monitoring data input into the semantic segmentation model is denormalized. All obtained free decay response segments are screened and data segments with sufficient sampling length are retained.

[0021] Optionally, the acceleration free decay response segment obtained by identification and positioning is corrected, the peak and valley values in each vibration cycle of the free decay segment are picked up, and the damping ratio is calculated using the exponential decay method, specifically:

[0022] The acquired acceleration free decay response segment is subjected to trend removal and offset correction. The data segment at the tail of the free decay response segment, which may contain high-frequency noise, is truncated. The acceleration amplitudes at the peaks and troughs of the remaining time periods are extracted and their natural logarithms are calculated. The natural logarithms are fitted, and the logarithmic decay rate is calculated from the slope of the fitted line. The damping ratio of the bridge structure is calculated using the exponential decay formula.

[0023] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the bridge damping ratio identification method based on monitoring data and deep learning provided by the present invention first collects bridge acceleration monitoring data, manually screens the free decay response segments with ideal decay shapes in the bridge acceleration monitoring data, and semantically annotates the time periods in which they occur, constructs a semantic segmentation dataset for the acceleration free decay response segments, and constructs a semantic segmentation model that can identify the free decay response segments from the bridge acceleration monitoring data based on a deep neural network that performs image semantic segmentation tasks in computer vision. The semantic segmentation model is trained, evaluated and tuned through the semantic segmentation dataset, and the unlabeled acceleration monitoring data is input into the tuned semantic segmentation model to identify the acceleration in the acceleration monitoring data. The free decay response segment is identified and located, the obtained acceleration free decay response segment is corrected, the peak and valley values in each vibration cycle of the free decay segment are picked up, and the damping ratio is calculated by the exponential decay method. This method performs semantic segmentation on the acceleration monitoring data collected by the bridge SHM system through the improved U-Net model, and can obtain the free decay response segment with a relatively ideal attenuation shape. Then, the exponential decay method can be used to calculate the damping ratio of the bridge structure, which solves the problem that human power cannot efficiently identify and extract the acceleration free decay response segment under massive monitoring data. The exponential decay method can be applied to bridge monitoring data, realizing a more reliable calculation of the damping ratio than other methods, and laying a reliable data foundation for further understanding the long-term change law of the damping ratio of the bridge structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is a flow chart of a bridge damping ratio identification method based on monitoring data and deep learning according to an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the free decay response section marked in the acceleration monitoring data;

[0027] Figure 3 Schematic diagram of the semantic segmentation model;

[0028] Figure 4 This is a schematic diagram of the acceleration free decay response segment data identified by the semantic segmentation model;

[0029] Figure 5This is a schematic diagram of the steps for calculating the damping ratio using the exponential decay method for the extracted free decay response segment;

[0030] Figure 6 The first-order damping ratio time history for 45 weeks was calculated from the monitoring data of Z24 Bridge. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] The purpose of the present invention is to provide a bridge damping ratio identification method based on monitoring data and deep learning, which can automatically identify the free decay response segments suitable for damping ratio calculation in massive monitoring data, and then calculate the damping ratio of the bridge structure using the exponential decay method.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, the bridge damping ratio identification method based on monitoring data and deep learning provided by the embodiment of the present invention includes the following steps:

[0035] Step 1: Collect bridge acceleration monitoring data, manually screen the free decay response segments with ideal decay shapes in the bridge acceleration monitoring data, and semantically annotate the time periods where they occur to construct a semantic segmentation dataset for the free decay response segments.

[0036] Step 2: Based on a deep neural network that performs image semantic segmentation in computer vision, a time series semantic segmentation model is constructed to identify free decay response segments from bridge acceleration monitoring data.

[0037] Step 3: Train, evaluate, and fine-tune the time series semantic segmentation model using the semantic segmentation dataset.

[0038] Step 4: Input the unlabeled acceleration monitoring data into the tuned semantic segmentation model to identify and locate the acceleration free decay response segment in the acceleration monitoring data;

[0039] Step 5: Correct the acceleration free decay response segment obtained by identification and positioning, pick up the peak and valley values in each vibration cycle of the free decay segment, and calculate the damping ratio using the exponential decay method.

[0040] In step 1, bridge acceleration monitoring data is collected, and free decay response segments with ideal decay shapes are manually screened from the bridge acceleration monitoring data. The time periods in which they occur are semantically annotated to construct a semantic segmentation dataset for the acceleration free decay response segments. Specifically,

[0041] Bridge acceleration monitoring data was collected. For each acceleration time series data in the bridge acceleration monitoring data, labeled time series data with the same timestamp was generated. The initial labels of all moments were set to 0. The free decay segments in the acceleration time series data were manually screened, and the initial label values of the corresponding time periods were modified to 1. The acceleration time series data were standardized so that their mean was 0 and their standard deviation was 1. The standardized data were split into training set, validation set, and test set in a ratio of 0.8:0.1:0.1.

[0042] In step 2, based on the deep neural network that performs image semantic segmentation tasks in computer vision, a time series semantic segmentation model is constructed to identify the free decay response segment from the bridge acceleration monitoring data. Specifically:

[0043] Based on the deep neural network that performs image semantic segmentation tasks in computer vision, a U-Net deep neural network model is established. The two-dimensional convolutional layer in the U-Net model is replaced with a one-dimensional convolutional layer, and the input layer length is changed. The ReLU activation function in the U-Net model is replaced with the LeakyReLU activation function. The sizes of the encoder and decoder convolutional layers in the U-Net model are modified so that the jump connections between the symmetrically positioned convolutional layers of the encoder and decoder can be realized by direct splicing. The length and number of channels of the output layer of the U-Net model are modified, and the output layer activation function is changed to Sigmoid. A time series semantic segmentation model that can identify free decay response segments from bridge acceleration monitoring data is constructed.

[0044] In step 3, the semantic segmentation model is trained, evaluated, and tuned using the semantic segmentation dataset. Specifically:

[0045] The semantic segmentation model is trained and verified using the training set and validation set. The intersection-over-union (IoU) is defined as the performance metric of the semantic segmentation model. The performance of the model is evaluated using the test set, and the convolution kernel size of the semantic segmentation model is optimized.

[0046] Step 4: Input the unlabeled acceleration monitoring data into the tuned semantic segmentation model to identify and locate the acceleration free decay response segment in the acceleration monitoring data. Specifically:

[0047] The unlabeled acceleration monitoring data is standardized and input into the tuned semantic segmentation model. The probability value of each data point in the acceleration monitoring data belonging to the free decay response segment is obtained. A judgment threshold is set, and data segments with probability values greater than the judgment threshold are identified as free decay response segments. The acceleration monitoring data input into the semantic segmentation model is denormalized, and all obtained free decay response segments are screened, and data segments with sufficient sampling length are retained.

[0048] In step 5, the acceleration free decay response segment obtained by identification and positioning is corrected, the peak and valley values in each vibration cycle of the free decay segment are picked up, and the damping ratio is calculated using the exponential decay method, specifically:

[0049] The obtained acceleration free decay response segment is subjected to trend removal and offset correction processing. The data segment with possible high-frequency noise at the tail of the free decay response segment is truncated. The acceleration amplitudes at the peaks and troughs of the remaining time periods are extracted, and their natural logarithms are calculated. The natural logarithms are fitted, and the logarithmic decay rate is calculated from the slope of the fitted straight line. The damping ratio of the bridge structure is calculated according to the exponential decay method formula.

[0050] One embodiment of the present invention is a 45-week long-term monitoring of the Z24 Bridge, a well-known benchmark model in the field of bridge SHM. The monitoring specifically includes the following steps:

[0051] Step 1: Select the acceleration monitoring data of the Z24 bridge for the first four weeks to construct a semantic segmentation dataset for the free decay response segment;

[0052] For each acceleration time series data stored in one hour, generate label time series data with the same timestamp, set the initial labels of all moments to 0, and manually filter the data segments with ideal decay shapes in each acceleration time series data, such as Figure 2 As shown in the figure, the label value of the corresponding period in the initial label is modified to 1. This operation is performed on all acceleration data of the first 4 weeks. For each acceleration time series data, the original data is standardized so that its mean is 0 and its standard deviation is 1. The final labeled dataset is split into training set, validation set, and test set in a ratio of 0.8:0.1:0.1;

[0053] Step 2: If Figure 3 As shown in the figure, a time series semantic segmentation model is constructed based on the U-Net neural network model that performs image semantic segmentation tasks in deep learning methods, which can identify free decay response segments from acceleration time series.

[0054] All two-dimensional convolutional layers in the original U-Net model were modified into one-dimensional convolutional layers to adapt to the processing of acceleration time series; the model input layer length was changed to 65536 to match the length of acceleration data collected every hour, and the ReLU activation function in the original U-Net model was replaced by the LeakyReLU activation function to solve the problem of neuron "death" caused by the gradient of the negative region being 0 during the backpropagation process. A maximum pooling operation was performed at a downsampling rate of 2 between each two encoder convolution modules to shorten the length of the input sequence, and an upsampling operation was performed at a 2-fold upsampling rate between each two decoder convolution modules to output feature maps of the same length as the symmetrical position encoder convolution module. Jump connections between the symmetrical position convolution layers of the encoder and decoder were implemented by direct splicing to improve the refinement of semantic segmentation. The output layer length was set to be the same as the input layer length, with a channel number of 1. The Softmax activation function of the output layer of the original model was replaced by the Sigmoid activation function to output the probability value of each data point belonging to the free decay response segment.

[0055] Step 3: Training, evaluating, and fine-tuning the acceleration time series semantic segmentation model. Train the improved U-Net model using the split training set, and evaluate the model's performance using the test set.

[0056] The improved U-Net semantic segmentation model is trained on the training set with the set training parameters (learning rate, number of iterations, etc.). Conventional performance metrics such as accuracy are used to track the model's performance at each training session on the validation set. For acceleration time series data, the intersection of union (IoU) is defined to measure the model's semantic segmentation performance. The specific expression is as follows. The semantic segmentation performance of the model is then evaluated on the test set.

[0057]

[0058] Where n ii Indicates a data point that belongs to the free decay response segment and is correctly identified, n ij represents a data point that belongs to the free decay response segment but is not correctly identified, n ji Indicates a data point that does not belong to the free decay response segment but is identified as a free decay response segment

[0059] A grid search algorithm was used to adjust and optimize the convolution kernel size, a key hyperparameter that affects the performance of the semantic segmentation model. Ultimately, a larger convolution kernel size of 35 was selected. This completed the training, evaluation, and tuning of the optimal model.

[0060] Step 4: Extract the acceleration free decay response segment data. Input the entire acceleration response into the trained improved U-Net model to identify and locate the free decay response segment.

[0061] The remaining 41 weeks of unlabeled acceleration monitoring data of the Z24 bridge were standardized and input into the final improved U-Net model. The probability value of each data point in the acceleration time series belonging to the free decay response was obtained from the output of the Sigmoid activation function of the output layer (e.g. Figure 4 A smaller threshold (0.2 in this example) is set for the probability value. If the threshold is exceeded, the data point is judged to belong to the free decay response, and the corresponding label is 1 (e.g. Figure 4 ), extract the data segment with label 1 as the identified free decay response segment, perform denormalization on it with the previously recorded mean and standard deviation to restore the normalized data, filter all the extracted free decay response segments, delete the data segments with a sampling length of less than 400 (about the length of 10 first-order vibration cycles), and complete the extraction of the target acceleration free decay response segment;

[0062] Step 5: Figure 5 As shown in Figure 1, the damping ratio is calculated by the exponential decay method. The damping ratio of the finally selected acceleration free decay response segment is calculated using the exponential decay method;

[0063] For all free decay response segments, operations such as removing trend items and offset correction are performed to return the equilibrium position of the acceleration time series vibration to 0. The threshold is set according to the vibration amplitude and sampling length. The data segment with possible high-frequency noise at the tail of the processed free decay segment is truncated. The acceleration amplitude at each peak and trough of the retained data segment is extracted, and the natural logarithm of the extracted peak or trough is calculated. The natural logarithm of the peak or trough is fitted by linear regression, and the logarithmic decay rate is calculated from the slope of the fitting line. The calculation of the bridge structure damping ratio is completed according to the following formula. The time history scatter plot of the first-order modal damping ratio calculated using all acceleration monitoring data of the Z24 bridge is shown as follows: Figure 6 As shown, the formula is as follows:

[0064]

[0065] In the formula, δ represents the logarithmic attenuation rate, and ξ represents the damping ratio of the structure.

[0066] The present invention provides a bridge damping ratio identification method based on monitoring data and deep learning. The method first collects bridge acceleration monitoring data, manually screens free decay response segments with ideal decay shapes in the bridge acceleration monitoring data, and semantically annotates the time periods in which they occur. A semantic segmentation dataset is constructed for the acceleration free decay response segments. Based on a deep neural network that performs image semantic segmentation tasks in computer vision, a time series semantic segmentation model is constructed that can identify free decay response segments from bridge acceleration monitoring data. The semantic segmentation model is trained, evaluated, and tuned using the semantic segmentation dataset, and the unannotated acceleration In the semantic segmentation model after the monitoring data input tuning, the acceleration free decay response segment in the acceleration monitoring data is identified and located, the acceleration free decay response segment obtained by identification and location is corrected, the peak and valley values in each vibration cycle of the free decay segment are picked up, and the damping ratio is calculated by the exponential decay method. This method solves the problem that human power cannot efficiently identify and extract the acceleration free decay response segment under massive monitoring data, so that the exponential decay method can be applied to bridge monitoring data, and realizes a more reliable calculation of the damping ratio compared with other methods, which can lay a reliable data foundation for further understanding the long-term change law of the damping ratio of bridge structures.

[0067] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A bridge damping ratio identification method based on monitoring data and deep learning, characterized in that: The steps include: Step 1: Collect bridge acceleration monitoring data, manually screen the free decay response segments with ideal decay shapes in the bridge acceleration monitoring data, and semantically annotate the time periods where they occur to construct a semantic segmentation dataset for the free decay response segments. Step 2: Based on a deep neural network that performs image semantic segmentation in computer vision, a time series semantic segmentation model is constructed to identify free decay response segments from bridge acceleration monitoring data. Step 3: Train, evaluate, and fine-tune the semantic segmentation model using the semantic segmentation dataset. Step 4: Input the unlabeled acceleration monitoring data into the tuned semantic segmentation model to identify and locate the acceleration free decay response segments in the acceleration monitoring data. Specifically, the unlabeled acceleration monitoring data is normalized and then input into the tuned semantic segmentation model to obtain the probability value of each data point in the acceleration monitoring data belonging to the free decay response segment; a judgment threshold is set to identify data segments with probability values greater than the judgment threshold as free decay response segments; the acceleration monitoring data input into the semantic segmentation model is denormalized; and all obtained free decay response segments are screened to retain data segments with sufficient sampling length. Step 5: Correct the acceleration free decay response segment obtained by identification and positioning, pick up the peak and valley values in each vibration cycle of the free decay segment, and calculate the damping ratio using the exponential decay method.

2. The bridge damping ratio identification method based on monitoring data and deep learning according to claim 1 is characterized in that: In step 1, bridge acceleration monitoring data is collected, and free decay response segments with ideal decay shapes are manually screened from the bridge acceleration monitoring data. The time periods in which they occur are semantically annotated to construct a semantic segmentation dataset for the acceleration free decay response segments. Specifically, Bridge acceleration monitoring data was collected. For each acceleration time series data in the bridge acceleration monitoring data, label time series data with the same timestamp was generated. The initial labels of all moments were set to 0. The free decay segments in the acceleration time series data were manually screened, and the label values of the initial labels of the time periods corresponding to the free decay segments were modified to 1. The acceleration time series data were standardized so that their mean was 0 and their standard deviation was 1. The standardized data were split into training set, validation set, and test set in a ratio of 0.8:0.1:0.

1.

3. The bridge damping ratio identification method based on monitoring data and deep learning according to claim 1 is characterized in that: Based on a deep neural network that performs image semantic segmentation tasks in computer vision, a time series semantic segmentation model is constructed to identify free decay response segments from bridge acceleration monitoring data. Specifically: Improve the U-Net deep neural network model for performing image semantic segmentation tasks in computer vision. Replace the two-dimensional convolutional layer in the U-Net model with a one-dimensional convolutional layer, and change the input layer length. Replace the ReLU activation function in the U-Net model with the LeakyReLU activation function. Modify the sizes of the encoder and decoder convolutional layers in the U-Net model so that the jump connections between the symmetrically positioned convolutional layers of the encoder and decoder can be implemented by direct splicing. Modify the length and number of channels of the output layer of the U-Net model, and change the output layer activation function to Sigmoid.

4. The bridge damping ratio identification method based on monitoring data and deep learning according to claim 1 is characterized in that: The semantic segmentation model is trained, evaluated, and tuned using the semantic segmentation dataset. Specifically: The semantic segmentation model is trained and verified using the training set and validation set. The intersection-over-union (IoU) is defined as the performance metric of the semantic segmentation model, and the performance of the model is evaluated using the test set. The convolution kernel size of the semantic segmentation model is optimized.

5. The bridge damping ratio identification method based on monitoring data and deep learning according to claim 1 is characterized in that: Correct the acceleration free decay response segment obtained by identification and positioning, pick up the peak and valley values in each vibration cycle of the free decay segment, and calculate the damping ratio using the exponential decay method, specifically: The obtained acceleration free decay response segment is subjected to trend removal and offset correction. The data segment with possible high-frequency noise at the tail of the free decay response segment is cut off, and the acceleration amplitudes at the peaks and troughs of the remaining time periods are extracted, and their natural logarithms are calculated. Fit each natural logarithm, calculate the logarithmic decay rate from the slope of the fitting line, and calculate the damping ratio of the bridge structure according to the exponential decay method formula.

Images