Bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning method and system, and storage medium

Abstract

The invention relates to a bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning method and system, and a storage medium, and the method comprises the steps: constructing a first training set based on historical wind speed-vibration response data with a timestamp; and meanwhile, constructing a second training set based on the historical wind speed feature vector marked with the abnormal state. And respectively training the CNN-LSTM network and the KNN classification model by using the first training set and the second training set. And inputting real-time wind speed data into the CNN-LSTM network to obtain a non-stationary wind speed prediction feature vector in a future time period. And the prediction feature vector is imported into a t-SNE-KNN dimension reduction classification algorithm, unstable abnormal wind and normal wind are distinguished, and an early warning boundary and safe and dangerous areas are established. Compared with the prior art, the method has the advantages that an accurate intelligent prediction and abnormal vibration early warning decision method can be constructed for extreme wind events such as downburst, the problems in the prior art are solved, and reliable guarantee is provided for safe operation of a large-span bridge under the extreme wind events.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge engineering, and in particular to a method, system, and storage medium for intelligently predicting non-stationary wind speed and early warning of abnormal vibration of a bridge. Background Art

[0002] In the field of bridge engineering, long-span bridges are susceptible to wind loads due to their flexibility and low damping properties. Extreme wind events such as downbursts pose a serious threat to their safety. Downbursts are intense storm phenomena in which downdrafts impact the ground and spread outward. They are characterized by high and erratic wind speeds, especially near the surface, with maximum speeds exceeding 30m / s. These can be devastating to long-span bridges, causing structural failures, pedestrian discomfort, and driving hazards.

[0003] Traditional bridge design is primarily based on boundary layer wind characteristics. However, the non-stationary characteristics and complex aeroelastic behavior of downbursts differ significantly from boundary layer winds, making it difficult for existing design methods to accurately predict and mitigate the impacts of these extreme wind events. With the intensification of climate change, extreme meteorological conditions such as microbursts are occurring more frequently in climate-sensitive areas, posing an increasingly severe threat to local infrastructure. Therefore, in-depth research on the non-stationary characteristics of downbursts and the aeroelastic behavior of bridges is of paramount practical significance for improving the safety and functionality of bridge structures.

[0004] Data-driven artificial intelligence (AI) methods, with their significant advantages in capturing nonlinear features, have shown great potential in deconstructing the short-term nonlinear response prediction of downbursts. However, existing related technologies still have many shortcomings. For example, the existing technology CN105205495A proposes a non-stationary fluctuating wind speed prediction method based on EMD-ELM. Although it improves the speed and effect of wind speed prediction to a certain extent, this method mainly focuses on the prediction accuracy and computational efficiency of wind speed, does not fully combine AI technology to mine nonlinear features, and lacks an effective early warning decision-making process, making it difficult to effectively warn of extreme wind weather in advance. When processing non-stationary wind speed data, the existing technology CN105205495A mainly relies on the combination of empirical mode decomposition (EMD) and extreme learning machine (ELM). Although it can predict non-stationary fluctuating wind speed to a certain extent, it lacks adaptability to complex wind conditions such as downbursts, especially the ability to identify and warn of abnormal wind speeds is limited.

[0005] Another example is the prior art CN104951798A, which provides an EMD-based LSSVM non-stationary pulsating wind speed prediction method, which improves the accuracy of the prediction by optimizing the model parameters through the introduction of particle swarm optimization (PSO). However, this method mainly focuses on how to improve the prediction accuracy by optimizing the model parameters, and also does not establish a complete set of early warning decision-making processes, and cannot meet the needs of early warning of extreme wind events in actual engineering. Although the prior art CN104951798A uses a combination of EMD decomposition and LSSVM model when processing non-stationary wind speed, its ability to identify and classify abnormal wind speeds is insufficient, making it difficult to effectively distinguish normal wind speeds from abnormal wind speeds, thereby achieving accurate early warning.

[0006] Furthermore, existing technologies often require extensive historical data for model training when processing non-stationary wind speed data. These models also have limited adaptability and generalization capabilities, making them incapable of handling complex, real-world wind conditions. Furthermore, existing wind speed prediction methods mostly focus on predicting single-point wind speeds, lacking the ability to analyze and provide early warnings for the overall response of bridge structures to extreme winds.

[0007] In summary, existing technologies have obvious shortcomings in dealing with the impact of extreme wind events such as downbursts on long-span bridges. These shortcomings are mainly reflected in insufficient consideration of non-stationary wind speed characteristics, limited prediction accuracy, failure to fully integrate AI technology to explore nonlinear characteristics, and the lack of an effective early warning decision-making process. Summary of the Invention

[0008] The present invention aims to address these deficiencies and provide a method, system, and storage medium for intelligent prediction of non-steady wind speeds and early warning of abnormal vibrations for bridges. Based on multi-source monitoring data and data-driven AI methods, this invention constructs an accurate intelligent prediction and abnormal vibration early warning decision-making method for extreme wind events such as downbursts, overcomes the difficulties of existing technologies, remedies the defects, and provides reliable guarantees for the safe operation of large-span bridges under extreme wind events.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A first aspect of the present invention provides a method for intelligently predicting non-stationary wind speed and abnormal vibration of a bridge, comprising the following steps:

[0011] The first training set is constructed based on the historical wind speed-vibration response data with timestamps and the labels are the true wind speed values ​​in the future period;

[0012] Constructing a second training set based on the historical wind speed feature vectors that mark abnormal conditions;

[0013] Using the first training set to train a multi-feature sequence classification CNN-LSTM network;

[0014] Training the KNN classification model using the second training set;

[0015] Input a section of wind speed data monitored in real time into the trained CNN-LSTM network to obtain the feature vector for predicting non-stationary wind speed in the future period;

[0016] The predicted feature vector is introduced into the t-SNE-KNN dimensionality reduction classification algorithm, and the t-SNE is used to reduce the dimensionality of the data to fully distinguish between unstable abnormal wind and normal wind. The KNN classification algorithm is used to determine the sample category to be classified by finding the K most similar samples in the features based on the principle of feature similarity, and establish the warning boundary and the corresponding safe and dangerous areas;

[0017] The real-time predicted data segments after t-SNE dimensionality reduction processing are converted into data points in sequence. According to the position of the data points relative to the warning boundary, an early warning of abnormal bridge vibration is realized, and an alarm is triggered when the predicted data points are in the danger zone.

[0018] Furthermore, the construction process of the first training set includes: collecting time-stamped historical wind speed-vibration response data from the long-span bridge health monitoring system, arranging these data in chronological order, and annotating each section of historical data with the corresponding true value of the wind speed in the future period as a supervision label. After preprocessing steps such as data cleaning, normalization, and sliding window interception, an ordered data set is formed for training the CNN-LSTM network.

[0019] Furthermore, the historical wind speed-vibration response data may be in the form of:

[0020] Structured data with timestamps. Each data record contains a timestamp, wind speed measurement, and bridge vibration response parameters. Each field is separated by a comma or other delimiter.

[0021] The bridge vibration response parameters are parameters that reflect the dynamic response characteristics of the bridge under different wind speed conditions. The bridge vibration response parameters include one or more of acceleration, velocity, displacement, strain, and frequency response function.

[0022] Furthermore, the process of constructing the second training set includes:

[0023] Feature extraction is performed on historical wind speed data to generate historical wind speed feature vectors containing multiple statistical features. These feature vectors are analyzed to mark abnormal samples. The marked normal and abnormal feature vectors are organized into a data set for training the KNN classification model.

[0024] Furthermore, the training process of the multi-feature sequence classification CNN-LSTM network includes:

[0025] The first training set is imported into the network. The multi-dimensional spatial features of the data are first extracted through the convolution layer using a wide convolution kernel. After the dimensionality is reduced by the maximum pooling layer to retain the key features, the feature sequence is input into the double-layer LSTM network. Its gating mechanism is used to autonomously learn and train the temporal features. During this period, the Adam optimization algorithm is used to backpropagate the training error, and the model parameters are updated layer by layer. The iteration is continuous until the model achieves the best performance in terms of root mean square error, determination coefficient, and mean square error indicators, thereby completing the training of the network and enabling it to have the ability to predict non-stationary wind speed sequences.

[0026] Furthermore, the process of training the KNN classification model includes: based on the second training set, taking the historical wind speed feature vectors that mark abnormal states as input samples, using the feature similarity principle of the KNN algorithm to calculate the distance between each sample, determining the category label of each sample, and by finding the most similar K neighbor samples in the feature space, using voting to determine the category to which the sample belongs, and then establishing a warning boundary, dividing the safe and dangerous areas, completing the training of the KNN classification model so that it can effectively identify the characteristic patterns of normal wind and unstable abnormal wind.

[0027] Furthermore, t-SNE is used to reduce the data dimension and fully distinguish between unstable abnormal wind and normal wind. The specific process includes:

[0028] The t-SNE algorithm uses the high-dimensional feature vectors predicted by the CNN-LSTM network as input. It constructs a similarity matrix by calculating the pairwise probabilities between high-dimensional data points and optimizes the corresponding probability distribution in low-dimensional space. It uses the Gaussian kernel transform to measure the similarity between high-dimensional data points and then uses the t-distribution to calculate the similarity between low-dimensional embedding points. The dimensionality reduction process is optimized by minimizing the Kullback-Leibler divergence, mapping the high-dimensional feature vectors to two-dimensional or three-dimensional space. This allows normal wind and unstable abnormal wind to form a distinguishable clustering structure in low-dimensional space, providing a clear data distribution for subsequent classification and early warning.

[0029] Furthermore, the real-time predicted data segments after the t-SNE dimensionality reduction process are sequentially converted into data points. The specific process of implementing the early warning of abnormal bridge vibration according to the position of the data points relative to the early warning boundary includes:

[0030] The feature vectors obtained by real-time prediction are sequentially input into the trained t-SNE model. After dimensionality reduction, the data points in two-dimensional or three-dimensional space are obtained. The data points are arranged in the order of the predicted time series to form a trajectory.

[0031] The warning boundary is established in advance in the low-dimensional space through the KNN classification algorithm, and the space is divided into safe areas and dangerous areas;

[0032] The data points obtained by real-time conversion are compared with the warning boundaries. If the data points are located in the danger zone, it is determined that the bridge may be at risk of abnormal vibration and an alarm is immediately triggered. If the data points are located in the safe zone, the bridge is determined to be in normal condition and no warning is needed.

[0033] A second aspect of the present invention provides a bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning system, comprising:

[0034] Model training module: Based on the time-stamped historical wind speed-vibration response data, labeled with the true wind speed value for the future period, a first training set is constructed. Based on the historical wind speed feature vectors that mark abnormal conditions, a second training set is constructed. The first training set is used to train the multi-feature sequence classification CNN-LSTM network, and the second training set is used to train the KNN classification model.

[0035] Data prediction module: This module inputs a segment of real-time wind speed data into a trained CNN-LSTM network to obtain a feature vector for predicting non-stationary wind speeds for future periods. This feature vector is then fed into the t-SNE-KNN dimensionality reduction classification algorithm. t-SNE is used to reduce the dimensionality of the data to fully distinguish between non-stationary abnormal winds and normal winds. The KNN classification algorithm, based on the principle of feature similarity, determines the sample category to be classified by finding the K most similar samples in the features, thereby establishing warning boundaries and corresponding safe and dangerous areas.

[0036] Early warning decision module: It is used to convert the real-time predicted data segments after t-SNE dimensionality reduction processing into data points in sequence, and realize early warning of abnormal bridge vibration based on the position of the data points relative to the warning boundary. When the predicted data points are in the danger zone, an alarm is triggered.

[0037] A third aspect of the present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, is used to execute the above-mentioned method for intelligent prediction of non-stationary wind speed and abnormal vibration warning of bridges.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The method of the present invention can provide early warning of extreme wind weather, such as downbursts, before they occur, allowing relevant technical personnel to take safety precautions in advance to prevent pedestrian discomfort, driving hazards, and structural failures during the construction or operation of large-span bridges. Compared with traditional iterative methods, which rapidly accumulate errors over time, physics-based methods hinder short-term wind speed predictions due to their computationally intensive processes, and time series methods are unsatisfactory in strong nonlinear predictions; data-driven artificial intelligence (AI) methods are more efficient and accurate in capturing nonlinear characteristics, which is more effective in deconstructing the short-term nonlinear response predictions of downbursts. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the flow of the method for intelligent prediction of non-steady wind speed and abnormal vibration early warning of bridges according to the present invention;

[0041] Figure 2 This is the structure diagram of the CNN-LSTM network used in the present invention for multi-feature sequence prediction;

[0042] Figure 3 This is the empirical mode decomposition result diagram of downburst wind speed data;

[0043] Figure 4 Comparison chart of non-stationary wind speed prediction training effects under different network parameters;

[0044] Figure 5 This is a graph showing the training loss change for a 64-unit LSTM network.

[0045] Figure 6 This is a comparison chart of the predicted and measured values ​​of downburst wind speed;

[0046] Figure 7 This is a classification visualization diagram of t-SNE dimensionality reduction data under different perplexity parameters;

[0047] Figure 8 This is a decision diagram for real-time wind speed prediction and abnormal vibration warning. DETAILED DESCRIPTION

[0048] Overall, the present invention processes the non-stationary wind speed data detected by the long-span bridge health monitoring system by combining the multi-feature sequence classification CNN-LSTM network and the t-SNE-KNN dimensionality reduction classification method for anomaly identification, thereby achieving efficient prediction of wind speed for a subsequent period of time. This method can not only capture the nonlinear and non-stationary characteristics in the wind speed data and improve the prediction accuracy, but also effectively reduce the feature space dimension, reduce computing resource consumption, and improve model training and prediction efficiency. In addition, by using the t-SNE-KNN method to reduce the dimensionality of the predicted data and classify it, it is possible to accurately distinguish between normal wind and abnormal wind, establish warning boundaries and corresponding safe and dangerous areas, thereby achieving real-time warning of abnormal vibrations of the bridge and providing reliable protection for the safe operation of the bridge.

[0049] Compared to the methods of reference documents CN105205495A and CN104951798A, the present invention not only improves prediction accuracy and computational efficiency, but also places special emphasis on the ability to identify and warn of abnormal wind speeds. By introducing a data-driven AI approach, the present invention can more effectively capture the complex nonlinear characteristics of non-stationary wind speeds, and by optimizing the model structure and parameters, it improves the adaptability and generalization ability of the model, enabling it to maintain high prediction and warning performance under different wind conditions. These improvements make the present invention more advantageous in practical applications and can better meet the safety monitoring and warning needs of long-span bridges under extreme wind events.

[0050] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0051] Example 1

[0052] Step 1: Take the current segment of wind speed data from the long-span bridge health monitoring system and input it into the multi-feature sequence classification CNN-LSTM network.

[0053] Step 2: The input time history is processed by a method combining empirical mode decomposition and sliding window technique, which can produce five intrinsic mode functions (IMFs) representing different frequency components and a residual trend term.

[0054] Step 3: The data segments are fed into the CNN convolutional layer, where wider convolution kernels with more channels are used to adaptively extract features.

[0055] Step 4: The extracted features are passed through the maximum pooling layer of CNN to take the maximum value of each sliding window, which effectively reduces the feature space dimension while retaining the key discriminant features of the original signal.

[0056] Step 5: Input the reduced dimension feature data as features into the two-layer LSTM network to promote autonomous training and learning of sequence features.

[0057] Step 6: Backpropagate the training error through the Adam algorithm and update the model parameters layer by layer.

[0058] Step 7: Use addition operation to achieve feature fusion, thereby completing the prediction task of non-stationary time course.

[0059] Step 8: Import the predicted data into the t-SNE-KNN dimensionality reduction classification algorithm and use t-SNE to fully distinguish between unstable abnormal wind and normal wind. It is better than revealing the nonlinear structure of data and is suitable for high-dimensional data and complex pattern recognition.

[0060] Step 9: Use the KNN classification algorithm and the principle of feature similarity to determine the sample category to be classified by finding the K most similar samples in the features to establish the warning boundary and the corresponding safe and dangerous areas.

[0061] Step 10: The data segments for real-time prediction are sequentially converted into data points through t-SNE, and the alarm trigger is determined by its position relative to the warning boundary.

[0062] in:

[0063] The model training process in this embodiment includes:

[0064] The first training set is constructed based on the historical wind speed-vibration response data with timestamps and the labels are the true wind speed values ​​in the future period;

[0065] Constructing a second training set based on the historical wind speed feature vectors that mark abnormal conditions;

[0066] Using the first training set to train a multi-feature sequence classification CNN-LSTM network;

[0067] The KNN classification model is trained using the second training set.

[0068] In specific implementation, the construction process of the first training set includes: collecting time-stamped historical wind speed-vibration response data from the long-span bridge health monitoring system, arranging these data in chronological order, and annotating each section of historical data with the corresponding true value of the wind speed in the future period as a supervision label. After preprocessing steps such as data cleaning, normalization, and sliding window interception, an ordered data set is formed for training the CNN-LSTM network.

[0069] In specific implementation, the historical wind speed-vibration response data may be in the form of:

[0070] Structured data with timestamps. Each data record contains a timestamp, wind speed measurement, and bridge vibration response parameters. Each field is separated by a comma or other delimiter.

[0071] The bridge vibration response parameters are parameters that reflect the dynamic response characteristics of the bridge under different wind speed conditions. The bridge vibration response parameters include one or more of acceleration, velocity, displacement, strain, and frequency response function.

[0072] In specific implementation, the process of constructing the second training set includes:

[0073] Feature extraction is performed on historical wind speed data to generate historical wind speed feature vectors containing multiple statistical features. These feature vectors are analyzed to mark abnormal samples. The marked normal and abnormal feature vectors are organized into a data set for training the KNN classification model.

[0074] In specific implementation, the training process of the multi-feature sequence classification CNN-LSTM network includes:

[0075] The first training set is imported into the network. The multi-dimensional spatial features of the data are first extracted through the convolution layer using a wide convolution kernel. After the dimensionality is reduced by the maximum pooling layer to retain the key features, the feature sequence is input into the double-layer LSTM network. Its gating mechanism is used to autonomously learn and train the temporal features. During this period, the Adam optimization algorithm is used to backpropagate the training error, and the model parameters are updated layer by layer. The iteration is continuous until the model achieves the best performance in terms of root mean square error, determination coefficient, and mean square error indicators, thereby completing the training of the network and enabling it to have the ability to predict non-stationary wind speed sequences.

[0076] In specific implementation, the process of training the KNN classification model includes: based on the second training set, taking the historical wind speed feature vectors that mark abnormal states as input samples, using the feature similarity principle of the KNN algorithm to calculate the distance between each sample, determining the category label of each sample, and by finding the most similar K neighbor samples in the feature space, using voting to determine the category to which the sample belongs, and then establishing a warning boundary, dividing the safe and dangerous areas, completing the training of the KNN classification model so that it can effectively identify the characteristic patterns of normal wind and unstable abnormal wind.

[0077] In step 2, empirical mode decomposition demonstrates excellent ability to adaptively process nonlinear and nonstationary signals by decomposing them into a set of IMFs and residuals of non-predefined basis functions. Each IMF must exhibit an almost equal number of extreme values ​​and zero crossings (difference ≤ 1), and its maximum and minimum envelopes have zero-mean symmetry to ensure effective time-frequency.

[0078] The original time sequence can be obtained by It is calculated that the eigenmode function can be obtained by Calculated.

[0079] In step 5, the LSTM network employs a gating mechanism involving an input gate, a forget gate, and an output gate. The forget gate is activated using a sigmoid function, which uses the previous hidden state and the current input to determine irrelevant information to discard from the memory cell. The input gate uses the same input and selectively updates the memory cell by combining the sigmoid-controlled correlations with the candidate values ​​of the Tanh transform. The output gate filters the updated memory cell through the Tanh function to produce the current hidden state, which serves as the output for the subsequent step.

[0080] In step 6, the Adam algorithm uses the root mean square error (RMSE), the coefficient of determination (R 2 ) and mean squared error (MSE) to evaluate whether the model can accurately reflect the training and prediction results. By meeting these criteria, the model can achieve the best performance because the error function stabilizes at the lowest level, which theoretically produces the maximum prediction accuracy.

[0081] In step eight, the t-SNE method replaces the Gaussian distribution with the t distribution in low-dimensional similarity modeling, provides a larger stretch for the distance between dissimilar data and implements tighter clustering for similar data; implements adaptive gradient descent with momentum in cost function optimization to escape local minima; introduces perplexity-guided bandwidth calibration to essentially suppress outlier interference while maintaining topological fidelity.

[0082] In step nine, the KNN algorithm uses the cross entropy function to evaluate the classification accuracy because the cross entropy function is better than the MSE in aligning probability distributions, especially for classification tasks, which has logarithmic sensitivity to prediction confidence.

[0083] Application Example 1

[0084] Step 1: Take the current segment of wind speed data from the long-span bridge health monitoring system and input it into the multi-feature sequence classification CNN-LSTM network.

[0085] Step 2: The input time history is processed using a method that combines empirical mode decomposition and sliding window techniques. Empirical mode decomposition demonstrates excellent adaptive processing capabilities for nonlinear and nonstationary signals by decomposing them into a set of IMFs and residuals of non-predefined basis functions. IMFs are calculated using equations (1) and (2). Each IMF must exhibit an almost equal number of extreme values ​​and zero crossings (with a difference of ≤ 1), and its maximum and minimum envelopes must have zero-mean symmetry to ensure effective time-frequency distribution.

[0086]

[0087] yi(t) represents the original signal value of the i-th measurement point or the i-th signal channel at time t.

[0088] represents the summation of m intrinsic mode functions (IMFs).

[0089] Each ai,j(t) is an IMF component of the original signal yi(t), representing the characteristics of the signal at different frequencies or time scales.

[0090] ri,m(t) represents the residual term, which is the remaining part of the original signal yi(t) after decomposing it into m IMF components.

[0091] IMFj(t) is the final target to be solved, which represents the average value of the j-th IMF component.

[0092] N is the total number of signals.

[0093] The IMF and residuals extracted from the training samples are as follows Figure 3 As shown in Figure 2, empirical mode decomposition (EMD) was performed on the wind speed time history. The five IMFs, labeled IMF1 to IMF5, exhibited temporally nonstationary statistical characteristics. However, features such as IMF1 and IMF2 were found to be oscillations that represent random fluctuations in normal wind speed. In contrast, IMFs 3 and 4 exhibited similar trends and peaks to the training samples, implying a greater weighting of the network when predicting downbursts.

[0094] Step 3: The data segments are fed into the CNN convolutional layer, where wider convolution kernels with more channels are used to adaptively extract features.

[0095] Step 4: The extracted features are passed through the maximum pooling layer of CNN to take the maximum value of each sliding window, which effectively reduces the feature space dimension while retaining the key discriminant features of the original signal.

[0096] Step 5: The reduced dimension feature data is input as features into the two-layer LSTM network. LSTM uses time backpropagation for parameter update and is paired with the Adam optimizer (a hybrid algorithm that integrates momentum-driven gradient descent and root mean square propagation), as shown in Figure 2 This combination ensures adaptive learning rate adjustment, computational efficiency, and robust convergence, making it the first choice for LSTM optimization. The outputs it, ft, and ot of the input gate, forget gate, and output gate at time step t are given by the following equations:

[0097] it=sigmoid(Wi[ht-1,yt]+bi)(3)

[0098] ft=sigmoid(Wf[ht-1,yt]+bf)(4)

[0099] ot=sigmoid(Wo[ht-1,yt]+bo)(5)

[0100] Among them, Wi, Wf and Wo represent the corresponding weight matrices respectively.

[0101] The weight matrix Wi of the input gate is used to weight the previous hidden state ht-1 and the current input yt, indicating the sensitivity of the input gate to the input data and the hidden state at the previous moment.

[0102] The logic of Wf and Wo is similar to that of Wi, so they will not be introduced here.

[0103] bi, bf and bo are corresponding bias terms.

[0104] The hidden state value ht is the output of LSTM at time step t and can be calculated by the following formula:

[0105]

[0106] ht=ot·tanh(ct)(8)

[0107] in, Represents the candidate storage unit value at time step t, which is used to update information; ct represents the pixel updated at time step t.

[0108] Step 6: Backpropagate the training error using the Adam algorithm, gradually updating the model parameters layer by layer. Methods such as root mean square error (RMSE), coefficient of determination (R²), and mean square error (MSE) can be used to assess whether the model accurately reflects training and prediction results. Meeting these criteria leads to optimal model performance, as the error function stabilizes at a minimum level, theoretically resulting in maximum prediction accuracy.

[0109] Formulas (9), (10), and (11) show how to calculate the loss for a dataset with n data size.

[0110]

[0111] Among them, yi and are the actual data and prediction of the model at the i-th data point, respectively. is the average value of nyi.

[0112] Step 7: Use addition operation to achieve feature fusion, thereby completing the prediction task of non-stationary time course. Figure 4The following figure shows the wind speed training results for different numbers of iterations and LSTM sizes. The left figure shows the effect of the number of iterations for an LSTM size of 64, while the right figure shows the effect of the LSTM size for 100 iterations. The results show that the network's accuracy improves with increasing numbers of iterations and LSTM size. Computation time increases exponentially with LSTM size. For this downburst wind speed prediction, an LSTM network with a size of 64, trained for 100 iterations, achieved good fitting results with optimal computational efficiency. Figure 5 The training loss curves for the first 200 iterations of an LSTM network of size 64 are shown according to formulas (9), (10), and (11). It can be observed that the values ​​of MSE and RMSE decrease rapidly with the increase in iterations, and eventually stabilize and approach zero after about 100 iterations. At the same time, R increases rapidly and approaches 1. The rapid convergence of the loss value indicates effective training performance.

[0113] Based on the developed multi-feature sequence prediction CNN-LSTM network, wind speed prediction is performed on the instance starting from 17.5 hours. The results are as follows: Figure 6 The prediction results (red line) follow the overall trend and local fluctuations (black line) of the actual measurements very closely, which means that the dynamic and non-stationary characteristics of wind speed are effectively captured by the network. In addition, the behavior of the prediction accuracy during the downburst is further verified by power spectral density analysis, as shown in Figure 1. Figure 6 As shown in the figure on the right, the predicted power spectral density closely matches the measured data in all frequency bands, demonstrating high accuracy.

[0114] Step 8: Import the predicted data into the t-SNE-KNN dimensionality reduction classification algorithm. t-SNE is a nonlinear dimensionality reduction technique derived from stochastic neighbor embedding (SNE) that is specifically used to project data into 2-3 dimensions for visualization. It constructs pairwise probabilities, where similar objects exhibit a higher likelihood of co-selection, while the probability of dissimilar pairs of objects decreases. In order to minimize the Kullback-Leibler divergence, the dimensionality reduction process requires optimizing the corresponding probability distribution in the low-dimensional space. The pairwise similarity measure between high-dimensional data points xi and xj is obtained by the Gaussian kernel transformation of their Euclidean distance, as calculated in Equation (12).

[0115]

[0116] where σi represents the Gaussian variance centered at the data point xi.

[0117] Correspondingly, the low-dimensional similarity between embeddings yi and yj can be calculated as follows:

[0118]

[0119] t-SNE is used to fully distinguish non-stationary abnormal wind from normal wind. It is superior to revealing the nonlinear structure of data and is suitable for high-dimensional data and complex pattern recognition.

[0120] Step 9: Use the KNN classification algorithm and the principle of feature similarity to determine the sample category to be classified by finding the K most similar samples in the features to establish the warning boundary and the corresponding safe and dangerous areas. After dimensionality reduction through t-SNE, the data is simplified to two dimensions, and then the KNN algorithm is used for binary classification training. As described in the algorithm, an important parameter that affects its performance is perplexity, which represents the number of similar adjacent points around a given point. t-SNE is relatively robust to the adjustment of perplexity. The perplexity value is set to start from 10 and increments from 10 to 60. The visualization results of data classification under different perplexity values ​​are shown below. Figure 7 As shown in Figure 2, it reveals that when the perplexity is 10, the clustering effect of the dangerous wind speed data is poor due to the significant distribution of clusters. When the perplexity exceeds 30, the data is overfitted. When the perplexity is 20, the optimal dimensionality reduction and classification effect is achieved.

[0121] Step 10: The real-time prediction data segments are sequentially converted into data points through t-SNE, and the alarm trigger is determined by its position relative to the warning boundary. The real-time prediction data segments predicted by the CNN-LSTM network are sequentially converted into data points through the t-SNE-KNN method, and the results are as follows: Figure 8 As shown in Figure 2, the prediction data segment for the warning decision has a period of 0.1 hours, and the calculation time for each decision is approximately 14.25 seconds. The warning decision is determined by its position (black dot) relative to the warning boundary (red line). It can be seen that the prediction result (black) is in the danger zone during the downburst, while the prediction result (black) falls into the safe zone immediately after the downburst. Therefore, the warning system and warning decision model are verified with high accuracy.

[0122] Example 2

[0123] See also Figure 1 In this embodiment, a bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning system is provided, including:

[0124] Model training module: Based on the time-stamped historical wind speed-vibration response data, labeled with the true wind speed value for the future period, a first training set is constructed. Based on the historical wind speed feature vectors that mark abnormal conditions, a second training set is constructed. The first training set is used to train the multi-feature sequence classification CNN-LSTM network, and the second training set is used to train the KNN classification model.

[0125] Data prediction module: This module inputs a segment of real-time wind speed data into a trained CNN-LSTM network to obtain a feature vector for predicting non-stationary wind speeds for future periods. This feature vector is then fed into the t-SNE-KNN dimensionality reduction classification algorithm. t-SNE is used to reduce the dimensionality of the data to fully distinguish between non-stationary abnormal winds and normal winds. The KNN classification algorithm, based on the principle of feature similarity, determines the sample category to be classified by finding the K most similar samples in the features, thereby establishing warning boundaries and corresponding safe and dangerous areas.

[0126] Early warning decision module: It is used to convert the real-time predicted data segments after t-SNE dimensionality reduction processing into data points in sequence, and realize early warning of abnormal bridge vibration based on the position of the data points relative to the warning boundary. When the predicted data points are in the danger zone, an alarm is triggered.

[0127] This embodiment does not introduce modules that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other modules in this embodiment.

[0128] In addition, those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the embodiment provided by the present invention, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0129] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0130] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing module, each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional units.

[0131] Example 3

[0132] This embodiment provides a storage medium containing computer-executable instructions. When executed by a computer processor, the storage medium containing computer-executable instructions is used to perform the above-mentioned method for intelligent prediction of non-stationary wind speed and abnormal vibration warning of bridges. When executed by a computer processor, the storage medium is used to perform the above-mentioned method for intelligent prediction of non-stationary wind speed and abnormal vibration warning of bridges. Specifically, the storage medium can be a USB flash drive, an SD card, a hard drive, a solid-state drive, an optical disk, etc., on which a computer program is stored. When executed by the processor, the program can automatically perform the following operations:

[0133] Data acquisition and preprocessing: The program first obtains time-stamped historical wind speed-vibration response data from the bridge health monitoring system and preprocesses this data, including removing outliers, filtering noise, and normalizing the data to ensure data accuracy and consistency.

[0134] Training set construction: Based on the processed data, the program constructs two training sets. The first training set is used to train the CNN-LSTM network and contains historical wind speed-vibration response data and its corresponding true wind speed labels for future time periods. The second training set is used to train the KNN classification model and contains historical wind speed feature vectors that mark abnormal conditions.

[0135] Model Training: Using the first training set, the program trains a multi-feature sequence classification CNN-LSTM network, optimizing prediction performance by adjusting network parameters. Simultaneously, the first training set is used to train a KNN classification model to establish warning boundaries and safe and dangerous areas.

[0136] Real-time prediction and warning: The program monitors bridge wind speed data in real time, feeds the data into a trained CNN-LSTM network, and predicts the non-stationary wind speed feature vector for future time periods. The predicted feature vector is then processed using the t-SNE-KNN dimensionality reduction and classification algorithm. t-SNE dimensionality reduction and KNN classification determine the category of the predicted data point, and its position relative to the warning boundary provides an effective warning of abnormal bridge vibration.

[0137] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning method, characterized in that: The following steps are involved: The first training set is constructed based on the historical wind speed-vibration response data with timestamps and the labels are the true wind speed values ​​in the future period; Constructing a second training set based on the historical wind speed feature vectors that mark abnormal conditions; Using the first training set to train a multi-feature sequence classification CNN-LSTM network; Training the KNN classification model using the second training set; Input a section of wind speed data monitored in real time into the trained CNN-LSTM network to obtain the feature vector for predicting non-stationary wind speed in the future period; The predicted feature vector is introduced into the t-SNE-KNN dimensionality reduction classification algorithm, and the t-SNE is used to reduce the dimensionality of the data to fully distinguish between unstable abnormal wind and normal wind. The KNN classification algorithm is used to determine the sample category to be classified by finding the K most similar samples in the features based on the principle of feature similarity, and establish the warning boundary and the corresponding safe and dangerous areas; The real-time predicted data segments after t-SNE dimensionality reduction processing are converted into data points in sequence. According to the position of the data points relative to the warning boundary, an early warning of abnormal bridge vibration is realized, and an alarm is triggered when the predicted data points are in the danger zone.

2. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The construction process of the first training set includes: collecting time-stamped historical wind speed-vibration response data from the long-span bridge health monitoring system, arranging this data in chronological order, and annotating each section of historical data with the corresponding true wind speed value for the future period as a supervision label. After preprocessing steps such as data cleaning, normalization, and sliding window interception, an ordered data set is formed for training the CNN-LSTM network.

3. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The historical wind speed-vibration response data includes: Structured data with timestamps. Each data record contains a timestamp, wind speed measurement, and bridge vibration response parameters. Each field is separated by a comma or other delimiter. The bridge vibration response parameters are parameters that reflect the dynamic response characteristics of the bridge under different wind speed conditions. The bridge vibration response parameters include one or more of acceleration, velocity, displacement, strain, and frequency response function.

4. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The construction process of the second training set includes: Feature extraction is performed on historical wind speed data to generate historical wind speed feature vectors containing multiple statistical features. These feature vectors are analyzed to mark abnormal samples. The marked normal and abnormal feature vectors are organized into a data set for training the KNN classification model.

5. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The training process of the multi-feature sequence classification CNN-LSTM network includes: The first training set is imported into the network. The multi-dimensional spatial features of the data are first extracted through the convolution layer using a wide convolution kernel. After the dimensionality is reduced by the maximum pooling layer to retain the key features, the feature sequence is input into the double-layer LSTM network. Its gating mechanism is used to autonomously learn and train the temporal features. During this period, the Adam optimization algorithm is used to backpropagate the training error, and the model parameters are updated layer by layer. The iteration is continuous until the model achieves the best performance in terms of root mean square error, determination coefficient, and mean square error indicators, thereby completing the training of the network and enabling it to have the ability to predict non-stationary wind speed sequences.

6. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The process of training the KNN classification model includes: based on the second training set, taking the historical wind speed feature vectors that mark abnormal states as input samples, using the feature similarity principle of the KNN algorithm to calculate the distance between each sample, determining the category label of each sample, and by finding the K most similar neighbor samples in the feature space, using voting to determine the category to which the sample belongs, and then establishing the warning boundary, dividing the safe and dangerous areas, completing the training of the KNN classification model so that it can effectively identify the characteristic patterns of normal wind and unstable abnormal wind.

7. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The specific process of using t-SNE to reduce the data dimension and fully distinguish between unstable abnormal wind and normal wind includes: The t-SNE algorithm uses the high-dimensional feature vectors predicted by the CNN-LSTM network as input. It constructs a similarity matrix by calculating the pairwise probabilities between high-dimensional data points and optimizes the corresponding probability distribution in low-dimensional space. It uses the Gaussian kernel transform to measure the similarity between high-dimensional data points and then uses the t-distribution to calculate the similarity between low-dimensional embedding points. The dimensionality reduction process is optimized by minimizing the Kullback-Leibler divergence, mapping the high-dimensional feature vectors to two-dimensional or three-dimensional space. This allows normal wind and unstable abnormal wind to form a distinguishable clustering structure in low-dimensional space, providing a clear data distribution for subsequent classification and early warning.

8. The method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to claim 1 is characterized in that: The real-time predicted data segments after the t-SNE dimensionality reduction process are sequentially converted into data points, and the specific process of implementing an early warning of abnormal bridge vibration according to the position of the data points relative to the early warning boundary includes: The feature vectors obtained by real-time prediction are sequentially input into the trained t-SNE model. After dimensionality reduction, the data points in two-dimensional or three-dimensional space are obtained. The data points are arranged in the order of the predicted time series to form a trajectory. The warning boundary is established in advance in the low-dimensional space through the KNN classification algorithm, and the space is divided into safe areas and dangerous areas; The data points obtained by real-time conversion are compared with the warning boundaries. If the data points are located in the danger zone, it is determined that the bridge may be at risk of abnormal vibration and an alarm is immediately triggered. If the data points are located in the safe zone, the bridge is determined to be in normal condition and no warning is needed.

9. A bridge non-stationary wind speed intelligent prediction and abnormal vibration early warning system, characterized in that: include: Model training module: Based on the time-stamped historical wind speed-vibration response data, labeled with the true wind speed value for the future period, a first training set is constructed. Based on the historical wind speed feature vectors that mark abnormal conditions, a second training set is constructed. The first training set is used to train the multi-feature sequence classification CNN-LSTM network, and the second training set is used to train the KNN classification model. Data prediction module: This module inputs a segment of real-time wind speed data into a trained CNN-LSTM network to obtain a feature vector for predicting non-stationary wind speeds for future periods. This feature vector is then fed into the t-SNE-KNN dimensionality reduction classification algorithm. t-SNE is used to reduce the dimensionality of the data to fully distinguish between non-stationary abnormal winds and normal winds. The KNN classification algorithm, based on the principle of feature similarity, determines the sample category to be classified by finding the K most similar samples in the features, thereby establishing warning boundaries and corresponding safe and dangerous areas. Early warning decision module: It is used to convert the real-time predicted data segments after t-SNE dimensionality reduction processing into data points in sequence, and realize early warning of abnormal bridge vibration based on the position of the data points relative to the warning boundary. When the predicted data points are in the danger zone, an alarm is triggered.

10. A storage medium containing computer-executable instructions, characterized in that: The storage medium of the computer executable instructions is used to execute the method for intelligent prediction of non-stationary wind speed and abnormal vibration early warning of a bridge according to any one of claims 1 to 8 when executed by a computer processor.

Images