Bridge sling multichannel vibration acceleration data missing filling method based on multi-environment field factor coupling

Abstract

The invention discloses a bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling, and relates to the data cleaning research aspect in the bridge structure health monitoring field, and through a series of data preprocessing technologies, a parallel deep learning network composed of a convolutional auto-encoder and a transformation neural network is trained, and the data cleaning research aspect of the bridge sling multi-channel vibration acceleration data missing filling method is realized. And the residual connection and early stop training technology is adopted to realize high-precision recovery of the abnormal and missing fragments of the vibration acceleration data of the bridge sling, and a multi-environmental factor coupling effect and a multi-channel sensor synergistic effect are introduced in the recovery process, so that the requirements of bridge health monitoring scenes in engineering are met.

Description

Technical Field

[0001] This invention belongs to the field of vibration data cleaning for bridge structural health monitoring, specifically involving a method for filling missing data of multi-channel vibration acceleration of bridge cables based on the coupling of multiple environmental field factors. Background Technology

[0002] With the rapid advancement of urbanization, long-span bridges have been developed and widely used. Ensuring the safety and reliability of these bridge structures throughout their service life is crucial, which necessitates effective and real-time monitoring of bridge structures. In recent decades, advancements in computer science and sensor technology have greatly promoted the development and application of structural health monitoring systems, providing strong technical support for maintaining structural safety. As an important branch of structural health monitoring, bridge structural health monitoring relies on sensor networks deployed on bridges to collect data and uses various computational analysis methods to assess the health status of bridges.

[0003] Structural health monitoring systems deploy various sensors at key locations on bridges to acquire structural response data. Since structural damage often induces anomalous features in the sensor response data, analyzing this data can indirectly assess the bridge's health. Therefore, diagnosing structural safety using health monitoring data essentially involves diagnosing the health status of the sensor response data itself. However, the causes of anomalous features in sensor data are not limited to structural damage; they can also stem from non-structural factors such as extreme weather, sensor equipment malfunctions, and power / communication anomalies. These factors are highly uncertain, difficult to quantify, and often severely impact the data analysis results. Therefore, before analyzing sensor response data, it is crucial to effectively identify and eliminate these anomalous features caused by non-structural factors—this is the data cleaning process in structural health monitoring.

[0004] Unstructured factors causing data anomalies are highly complex, including missing data, data trends, and data noise. Traditional data cleaning methods primarily focus on mathematical statistics, with principal component analysis, thresholding, and regression models being commonly used. While these models have their advantages in solving specific problems, they exhibit poor generalization ability in complex real-world scenarios and are computationally expensive. With the advent of the big data era, sensor response data is becoming increasingly complex, posing a significant challenge to mathematical statistics methods in handling high-dimensional, complex datasets.

[0005] To overcome the limitations of mathematical statistics methods, artificial intelligence technologies, especially deep learning methods, have been increasingly widely applied in the field of monitoring data cleaning research, and are gradually becoming the mainstream approach in this field. Some existing inventions include:

[0006] Chinese patent CN117723782A discloses a "Sensor Fault Identification and Location Method and System for Bridge Structural Health Monitoring," which preprocesses signals from a single vibration acceleration sensor, inputs them into a one-dimensional convolutional neural network for fault identification and location, and then performs post-processing using methods such as Kalman filtering. This method primarily relies on single-sensor time-series modeling, neglecting inter-node collaborative information in the sensor network and failing to consider the influence of environmental factors on the model.

[0007] Chinese patent CN111625988A discloses a “Bridge Health Management Analysis and Prediction System and Method Based on Deep Learning”, which uses a deep learning black box model to directly use sensor time-series data as input to predict stress and strain data, thereby assessing the health status of bridges. However, it does not consider the influence of environmental factors on the assessment of bridge health status.

[0008] While existing research has applied various deep learning models to the diagnosis of anomalies in sensor data and data recovery, most studies have two limitations:

[0009] On the one hand, in bridge health monitoring systems, the physical sensors in the sensor network are not independent, but have high physical and spatial correlation. Most studies use single sensor analysis, ignoring the monitoring characteristics of multi-sensor collaborative work, resulting in a lack of spatial correlation information in the analysis results.

[0010] On the other hand, environmental factors such as wind field and temperature field have obvious coupling effects, which can significantly interfere with the accuracy of diagnosis and recovery of abnormal data features, thereby affecting the accuracy of bridge health status assessment. Many studies only model and analyze vibration acceleration data or stress-strain data, failing to fully consider the influence of various environmental factors. Summary of the Invention

[0011] To address the limitations and practical engineering needs in the field of bridge health monitoring data cleaning, this invention proposes a method for filling missing data in multi-channel vibration acceleration data of bridge cables based on the coupling of multiple environmental field factors. This method can process data from multiple sensors in parallel and simultaneously, and takes into account the influence of wind speed and temperature factors in the environment, thereby achieving the identification and recovery of missing multi-channel cable vibration acceleration data under the influence of environmental factors, with high precision and efficiency.

[0012] To achieve the above-mentioned technical objectives, the present invention is based on the following solutions:

[0013] A method for filling missing data in multi-channel vibration acceleration data of bridge suspension cables based on the coupling of multiple environmental field factors includes the following steps: S1. Select multiple vibration acceleration sensors on the monitoring structure and select environmental sensors. Preprocess the vibration acceleration data and environmental data respectively, and build vibration acceleration dataset and environmental factor dataset for training and testing deep learning models in sequence.

[0014] S2, build a convolutional encoder model to process the vibration acceleration dataset described in step S1. Multiple vibration acceleration sensor data are used as multiple input channels and simultaneously input into the convolutional encoder model. After five convolutional coding layers, data features are extracted to form a high-dimensional feature vector of vibration acceleration.

[0015] S3. Construct a transform neural network encoder model to process the environmental factor dataset described in step S1. After passing through six multi-head self-attention encoding layers, obtain the high-dimensional feature vector of the environmental factors.

[0016] S4. The high-dimensional feature vector of vibration acceleration described in step S2 and the high-dimensional feature vector of environmental factors described in step S3 are concatenated to obtain a fused high-dimensional feature vector.

[0017] S5. Build a convolutional decoder model to process the fused high-dimensional feature vector described in step S4. The high-dimensional feature vector is decoded through five layers of deconvolutional decoding to reconstruct multi-channel vibration acceleration data.

[0018] S6. Combine the convolutional encoder model described in step S2 with the convolutional decoder model described in step S5 to form a convolutional autoencoder model. Synthesize the fusion model with the transform neural network encoder model described in step S3 in the form of a parallel network architecture. Select the loss function and optimizer to train the model and fine-tune the parameters.

[0019] S7 uses multiple evaluation metrics to evaluate the fusion model, verifying that the fusion model can not only sensitively identify vibration data anomalies and accurately locate them, but also accurately fill in missing vibration data.

[0020] Beneficial Effects: The method for filling missing data in multi-channel vibration acceleration data of bridge suspension cables based on the coupling of multiple environmental field factors in this invention has the following beneficial effects:

[0021] First, for different types of sensors, adopt data preprocessing methods that conform to their respective data characteristics. Through the data preprocessing process, the raw signals output by the sensors are transformed into training data that are easy for deep learning networks to learn features.

[0022] Second, because bridges are exposed to the environment for extended periods, environmental factors such as wind, rain, and temperature have a significant coupling effect on the structural health of bridges. This invention employs a parallel neural network architecture to consider environmental factors in the acceleration data cleaning process. The main network uses a convolutional autoencoder network to process acceleration data, while the auxiliary network uses a transform neural network encoder to process environmental data. The high-dimensional feature vectors extracted by both networks are fused in a feature merging layer to achieve comprehensive analysis. Since the transform neural network encoder is based on a multi-head self-attention mechanism, it can effectively capture the long-term temporal dependencies of environmental data. Test results show that this parallel network architecture enhances model performance and ensures that the influence of environmental factors is effectively considered when analyzing vibration data.

[0023] Third, the vibrations of different cables on a bridge exhibit significant correlations. As the number and scale of bridge sensor networks become increasingly sophisticated, the method of parallel input of multi-channel vibration acceleration sensor signals can simultaneously process data from multiple sensors to capture inter-sensor correlations, thereby achieving collaborative data cleaning of the entire sensor network. Compared to single-channel methods, this approach ensures the stability of data analysis through cross-sensor interaction, making it highly suitable for bridge monitoring systems with complex sensor networks.

[0024] Furthermore, as a preferred embodiment of the method of the present invention, the preprocessing of the vibration acceleration sensor data in step S1 specifically involves: subtracting the overall average value from each vibration acceleration sensor data to eliminate offset errors; then using a moving average detrending method to eliminate long-term and short-term trend anomalies; next, using a fourth-order Daubechies wavelet packet denoising method to filter noise and obtain preliminarily denoised vibration acceleration sensor data; finally, performing random slicing on the detrended and denoised vibration acceleration sensor data, and performing maximum absolute value normalization on each slice in sequence, so that the data range of each slice is within [-1, 1].

[0025] The specific preprocessing process for environmental factor data is as follows: the fifth-order Daubechies wavelet packet denoising method is selected to filter noise, and the maximum and minimum values ​​are normalized before random slicing to ensure that the overall data range is within [0, 1]. The random slicing positions of the environmental data correspond sequentially to the slicing positions of the vibration acceleration data, and the starting position of the slice is extended forward by 2.5 to 3 times the length of the bridge's natural vibration period to consider the hysteresis effect of environmental factors.

[0026] Beneficial effects: Statistical methods are used to remove easily removable anomalous features such as data offsets, trend anomalies, and simple white noise. Then, appropriate normalization methods are used to normalize the data according to the data type. After data preprocessing, the raw signals output by the sensor are transformed into training data that are easy for deep learning networks to learn features. Furthermore, the vibration acceleration data is time-aligned with environmental factor data through data alignment methods, while also considering the lag effect of environmental factors.

[0027] Furthermore, as a preferred embodiment of the method of the present invention, the environmental data in step S1 includes temperature data and wind speed data, and the vibration acceleration data, temperature data, and wind speed data are divided into training set, validation set, and test set in a ratio of 7:1:2; a random proportion of missing data (5%-15%) is introduced into the vibration acceleration data training set to improve the model's ability to recover missing data.

[0028] Beneficial effects: By dividing the dataset into appropriate proportions, sufficient data is available for training the model, while a certain amount of data is available for verifying and testing the model's training effect; introducing random missing data into the vibration acceleration data training set enhances the ability of the training data to represent missing properties, making it easier for the deep learning model to learn the missing data features, thus giving it a stronger ability and generalization in data missing completion.

[0029] Furthermore, as a preferred embodiment of the method of the present invention, the convolutional encoder model in step S2 consists of five convolutional coding layers. Each convolutional coding layer consists of a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. The number of input channels in the first convolutional coding layer is equal to the number of vibration acceleration sensors. Data from multiple vibration acceleration sensors are processed simultaneously in a multi-channel parallel input manner. After passing through five convolutional coding layers, the dimensionality is gradually increased and the length is reduced. At the same time, the individual data features of each vibration acceleration sensor data and the synergistic relationship between each sensor are obtained, thereby forming a high-dimensional feature vector of vibration acceleration.

[0030] Beneficial effects: Convolutional neural networks excel at capturing local features of data, making them suitable for feature extraction from vibration acceleration data. Simultaneous processing of data from multiple vibration acceleration sensors via multi-channel parallel input ensures that each sensor has its own corresponding channel while simultaneously inputting multiple sensors into the same convolutional encoder. This allows the high-dimensional feature vector of vibration acceleration to fully consider the unique data characteristics and coordinating relationships of each vibration acceleration data point. Batch normalization layers accelerate training and optimize training results, and the combination with LeakyReLU activation function layers enhances the data feature extraction capability and training stability of each convolutional coding layer.

[0031] Furthermore, as a preferred embodiment of the method of the present invention, the transform neural network encoder described in step S3 uses a six-layer multi-head self-attention coding layer as its core to extract high-dimensional feature vectors of environmental factors. After the environmental factor data undergoes input rearrangement, input projection through two fully connected layers, and position encoding, it enters the six-layer multi-head self-attention coding layer. Then, after output rearrangement and output projection through two fully connected layers, it outputs high-dimensional feature vectors of environmental factors.

[0032] Beneficial effects: Transformation neural networks excel at capturing long-term features and dependencies in data, making them suitable for processing environmental factor data that changes slowly over time; the selection of six multi-head self-attention encoding layers can fully capture the long-term changes and data features of multi-scale environmental factor data, while also possessing strong generalization ability.

[0033] Furthermore, as a preferred embodiment of the method of the present invention, in step S4, the high-dimensional feature vector of vibration acceleration and the high-dimensional feature vector of environmental factors are concatenated. Specifically, the concatenation method is to superimpose and concatenate them in terms of dimensions while keeping the feature length unchanged, thereby obtaining a fused high-dimensional feature vector.

[0034] Beneficial effect: By performing feature stitching of vibration acceleration data and environmental factor data at a high-dimensional scale, the environmental factors are applied to the vibration acceleration data, thus giving the vibration acceleration data features that characterize the influence of environmental factors in subsequent model propagation.

[0035] Furthermore, as a preferred embodiment of the method of the present invention, the convolutional decoder model described in step S5 is symmetrical to the convolutional encoder model described in step S2, and consists of five deconvolutional decoding layers. Each deconvolutional decoding layer consists of a deconvolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. After fusing high-dimensional feature vectors into the deconvolutional decoding layer, as the deconvolutional decoding layer gradually reduces its dimension and increases its length, the number of output channels of the last convolutional decoding layer is equal to the number of vibration acceleration sensors, thereby realizing the reconstruction of vibration acceleration data.

[0036] Beneficial effects: The convolutional decoder uses five deconvolutional layers symmetrical to the convolutional encoder, ensuring that the vibration acceleration data decoding process is symmetrical with the encoding process, making the model complete and interpretable.

[0037] Furthermore, as a preferred embodiment of the method of the present invention, in step S6, the convolutional encoder model and the convolutional decoder model are used as the main network architecture to process vibration acceleration data; the transform neural network encoder model is used as an auxiliary neural network to process environmental factor data; the main network architecture and the auxiliary neural network run in parallel, thereby realizing vibration acceleration data reconstruction under the consideration of multiple environmental field factor coupling; at the same time, before the second convolutional encoding layer and the third convolutional encoding layer, a residual connection is set between the deconvolutional decoding layer at its symmetrical position, and the data features are prevented from being lost when the data is transmitted in the deep neural network through these two residual connections;

[0038] The MSE loss function and Adam optimizer were selected for fusion model training and parameter tuning. The model training adopted a learning rate decay strategy and an early stopping function to control the model training process. At the same time, the total weight decay regularization method was used to prevent overfitting.

[0039] Beneficial effects: The parallel neural network architecture simultaneously processes vibration acceleration data and environmental factor data, ensuring that both types of neural networks independently extract high-dimensional feature vectors, and that these high-dimensional feature vectors can be merged through a feature merging layer; two residual connections prevent the loss of data features during data transmission in the deep neural network. By transmitting the original features of the data, the residual connections balance feature extraction between shallow and deep layers, allowing the network to learn deeper abstract representations. This ensures that key information is retained throughout the entire deep neural network without loss, effectively improving the accuracy of data processing; due to the complexity of the model architecture, methods such as learning rate decay, early stopping, and weight decay regularization can be used to control the model training and convergence process, optimizing the model training effect.

[0040] Furthermore, as a preferred embodiment of the method of the present invention, in step S7, the mean square error score and the Pearson correlation coefficient score are selected as evaluation indicators to evaluate the deep learning model's ability to reconstruct each vibration acceleration data channel and its ability to clean the data of the overall sensor network, respectively. In addition, the input-output comparison image and the Fourier spectrum comparison image of the fusion model are plotted to verify that the fusion model can not only sensitively identify vibration data anomalies and accurately locate them, but also accurately fill in the missing vibration data.

[0041] Beneficial effects: By using multi-index quantitative evaluation and multi-type graph intuitive evaluation of deep learning models, the performance of the model in various aspects can be fully assessed, ensuring the accuracy and robustness of the model in real-world application scenarios. Attached Figure Description

[0042] Figure 1 This describes the sensor network layout and output signals for a typical long-span suspension bridge.

[0043] Figure 2The process for data preprocessing and dataset construction;

[0044] Figure 3 The images show a comparison of the data before and after wavelet packet denoising.

[0045] Figure 4 Comparison image before and after introducing some randomly missing segments into the vibration data training set;

[0046] Figure 5 For the convolutional autoencoder network architecture and data flow;

[0047] Figure 6 For the Transformer encoder network architecture and data flow;

[0048] Figure 7 For the overall parallel network architecture and data flow;

[0049] Figure 8 The curve shows the overall network training loss value as a function of the number of training iterations.

[0050] Figure 9 Locate the missing image segments in the input data for the model;

[0051] Figure 10 Compare the model's input and output values ​​with health data in an image;

[0052] Figure 11 The images show the comparison between the model's input and output values ​​and health data, as well as the comparison images using Fourier spectra. Detailed Implementation

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Implementation method of data cleaning as described in this invention:

[0055] The first step is to attach Figure 1 The bridge health monitoring sensor network shown acquires real-time monitoring data, including vibration acceleration data output by vibration acceleration sensors on the six suspension cables, wind speed data output by wind speed and direction sensors installed on the tower top, and temperature data output by temperature sensors deployed on the bridge deck.

[0056] These data were processed Figure 2 The data preprocessing steps shown generate a dataset for training and testing deep learning models. Specifically, the vibration sensor data preprocessing includes subtracting the overall mean, detrending using moving averages, wavelet packet denoising, random data slicing, and normalization.

[0057] Each vibration acceleration sensor data point is subtracted from the overall average value, and then a moving average detrending method is used to eliminate trend anomalies. The expression is as follows:

[0058] (1);

[0059] in This indicates the input vibration acceleration data, where M represents the size of the moving window. This represents detrended vibration acceleration data; Indicates that the input vibration acceleration data is forward. Time step sampling.

[0060] Next, wavelet packet denoising is used to process the vibration acceleration data. The expression is as follows:

[0061] (2);

[0062] in, represents the wavelet packet basis function; a fourth-order Daubechies wavelet basis is selected for the bridge cable vibration acceleration data; L represents the number of decomposition layers. Represented as the wavelet packet decomposition coefficients of the j-th frequency band component; This represents the inverse wavelet packet transform, used for signal reconstruction. The threshold based on the absolute deviation of the median is expressed by formula (3). The soft thresholding process is expressed by formula (4):

[0063] (3);

[0064] (4);

[0065] Then, the detrended and denoised vibration acceleration data is randomly sliced. Since the vibration acceleration sensor sampling frequency is 20Hz, the slice length is 1200, corresponding to 1 minute of vibration acceleration data. Maximum absolute value normalization is then performed on each slice sequentially, ensuring that the data range of each slice is within [-1, 1].

[0066] For environmental factor data, the preprocessing steps include wavelet packet denoising, normalization, and random data slicing. For wavelet packet denoising, a fifth-order Daubechies wavelet basis is selected. A comparison of the effects before and after wavelet packet denoising is shown in the attached figure. Figure 3As shown; the normalization process uses the maximum-minimum normalization method to ensure the overall data range is within [0, 1] before random slicing. The random slice positions for environmental factor data correspond sequentially to the slice positions for vibration acceleration data. Since the sampling frequency of the environmental factor sensor data is 1Hz, a slice length of 60 corresponds to 1 minute of sampling. Considering the hysteresis effect of environmental factors, according to the specifications, the starting time value of the environmental data slice is extended forward by 2.5 to 3 bridge natural vibration cycles. In this project, 15 seconds is used, resulting in a total slice length of 75 and a time scale of 75 seconds.

[0067] Vibration acceleration data and environmental factor data (including temperature and wind speed data) were divided into training, validation, and test sets in a 7:1:2 ratio. To enhance the data's expressive power and optimize the deep learning model's training performance, 10 random missing data segments were introduced into part of the vibration acceleration data training set, as shown in the attached figure. Figure 4 As shown, a training set of vibration acceleration data containing random missing slices is constructed.

[0068] The second step is to build as shown in the attached diagram. Figure 5 The convolutional encoder model shown on the left consists of five convolutional coding layers. Each layer comprises a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. The LeakyReLU activation function is chosen, with a gradient of 0.01 in the negative region to prevent the vanishing gradient problem. The first convolutional coding layer has an input channel count equal to the number of vibration accelerometers, meaning the input data has a dimension of 6 and a length of 1200, processing data from multiple vibration accelerometers simultaneously in a multi-channel parallel input manner. Through five convolutional coding layers, the dimension is gradually increased while the length is reduced, simultaneously obtaining the individual data features of each vibration accelerometer and the collaborative relationships between the sensors, thus forming a high-dimensional feature vector of vibration acceleration with a dimension of 256 and a length of 75, shortening the time scale by a factor of 16.

[0069] The third step is to build two as shown in the attached diagram. Figure 6The transform neural network encoder model shown serves as two auxiliary network architectures, taking environmental wind speed and environmental temperature data as inputs respectively. After processing by the transform neural network encoder model, it outputs high-dimensional feature vectors of environmental wind speed and environmental temperature. Specifically, the transform neural network encoder uses a six-layer multi-head self-attention coding layer as its core to extract high-dimensional feature vectors of environmental factor data, and its expression is formula (5). After input rearrangement, input projection through two fully connected layers, and position encoding, the environmental factor data forms a 32-dimensional tensor, which enters the six-layer multi-head self-attention coding layer. Then, after output rearrangement and output projection through two fully connected layers, it outputs a 4-dimensional high-dimensional feature vector of environmental factors. The expression for the position encoding process is formula (6):

[0070] (5);

[0071] in, Indicates the i-th self-attention head; This indicates a self-attention operation; These represent the query vector, key vector, and value vector, respectively. Indicates the dimension of the key vector; The function is used to normalize the attention score; This indicates the matrix transpose.

[0072] (6);

[0073] in, Indicates the position of data in the sequence, an integer starting from 0; i represents the dimension index; Indicates the first in the sequence The encoded value at that location; This indicates the dimensionality of the input data.

[0074] The fourth step involves concatenating the high-dimensional feature vectors of vibration acceleration and environmental factors. Specifically, this concatenation is done by superimposing the features in terms of dimension while maintaining their length, resulting in a fused high-dimensional feature vector. Since the vibration acceleration feature vector has a dimension of 256 and the environmental factor feature vectors both have a dimension of 4, the fused high-dimensional feature vector has a dimension of 264. Furthermore, since both the vibration acceleration and environmental factor feature vectors have a length of 75, the fused high-dimensional feature vector also has a length of 75.

[0075] Step 5, construct as shown in the attached diagram. Figure 5The convolutional decoder model shown on the right side is symmetrical to the convolutional encoder model architecture and consists of five deconvolutional decoding layers. Each deconvolutional decoding layer consists of a deconvolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. After fusing high-dimensional feature vectors into the deconvolutional decoding layer, as the deconvolutional decoding layer gradually reduces the dimensionality and increases the length, the number of output channels of the final convolutional decoding layer is equal to the number of vibration acceleration sensors, that is, the output data has a dimension of 6 and a length of 1200, thereby realizing the reconstruction of vibration acceleration data.

[0076] Step 6: Combine the convolutional autoencoder model with the two transform neural network encoder models in a parallel network architecture to create a fused model, as shown in the attached diagram. Figure 7 As shown, a convolutional encoder and convolutional decoder model form the main network architecture for processing vibration acceleration data, while a transform neural network encoder model serves as the auxiliary neural network for processing environmental factor data. The main and auxiliary networks operate in parallel, thereby achieving vibration acceleration data reconstruction while considering the coupling of multiple environmental field factors. Simultaneously, residual connections are established between the second and third convolutional encoding layers and their symmetrically positioned deconvolutional decoding layers. These residual connections prevent the loss of data features during data transmission within the deep neural network.

[0077] The model was trained using the vibration acceleration data training set (containing partially missing data segments) and the environmental data training set from Step 1. The mean squared error loss function (MSE loss function) was chosen as the loss function, and the Adam optimizer was selected. A learning rate decay strategy was employed during model training, with an initial learning rate of 0.001 and a decay coefficient of 0.5 every 50 epochs. An early stopping function was used to control the model training process, enabling early stopping validation after 100 epochs. The number of training epochs was controlled by comparing the loss values ​​on the training and validation sets, with a maximum of 150 epochs. Weighted decay regularization with a decay coefficient of 0.001 was used to prevent overfitting. The loss function curves during model training are attached. Figure 8 As shown, it can be seen that the loss values ​​of both the training set and the validation set decrease rapidly at first and then slowly with the number of training rounds, eventually approaching 0.0027 and 0.0031 respectively, indicating that the model training effect is good.

[0078] Step 7: Test the performance of the deep learning model using multiple evaluation metrics. First, the mean squared error and Pearson correlation coefficient are used as two scoring metrics to evaluate the model's performance, expressed as formulas (7) and (8). Verification showed that the mean squared error was 0.002998, and the Pearson correlation coefficient was 0.9831, both remaining at an excellent level, indicating that the model accurately located the missing data and precisely filled it with healthy data.

[0079] (7);

[0080] Where n represents the number of data points, This represents the true value of the i-th data point. This represents the model output value for the i-th data point;

[0081] (8);

[0082] in, This represents the average value of the actual data. This represents the average value of the model's output data.

[0083] Next, the model's missing data filling capability is visually demonstrated using images. The model's input and output data are plotted on a single graph, as shown in the attached image. Figure 9 As shown, the model can accurately identify and label the missing input data. Next, the model output data and health data are plotted together on a single graph, with the missing input data segments labeled, as shown in the attached figure. Figure 10 As shown, the deep learning network model accurately fills in missing data fragments into healthy data states, and the model output value can be regarded as healthy data, indicating that the model effectively completes the data cleaning task. Furthermore, the model input data and output data are compared sequentially with healthy data, and comparison images and Fourier spectrum comparison images are plotted, as shown in the attached figure. Figure 11 As shown, missing data fragments affect data integrity and cause changes in data characteristics. However, the data cleaned by the deep learning model restores data integrity and ensures that the data characteristics are consistent with healthy data, further verifying the model's efficient data reconstruction and missing data imputation capabilities.

[0084] The tested model was applied to long-term health monitoring of actual bridges. After preprocessing the sensor data collected on-site as described in the first step, the data was input into a deep learning neural network, outputting healthy and complete vibration acceleration sensor data. This completed the data cleaning process while considering the coupling effects of multiple environmental fields, demonstrating high real-time performance. For vibration acceleration data cleaning of cables in other parts of the bridge, leveraging the model's generalization ability, a high-precision data cleaning deep learning model can be obtained simply by fine-tuning the parameters and retraining the deep learning model.

[0085] Example

[0086] Actual engineering case study of Ma'anshan Yangtze River Bridge:

[0087] The first step is to attach Figure 1The health monitoring sensor network of the Ma'anshan Yangtze River Bridge shown acquires real-time monitoring data, including vibration acceleration data output by vibration acceleration sensors on the six suspension cables, wind speed data output by wind speed and direction sensors installed on the tower top, and temperature data output by temperature sensors deployed on the bridge deck.

[0088] These data were processed Figure 2 The data preprocessing steps shown generate a dataset for training and testing deep learning models. Specifically, the vibration sensor data preprocessing includes subtracting the overall mean, detrending using moving averages, wavelet packet denoising, random data slicing, and normalization.

[0089] Each vibration acceleration sensor data point is subtracted from the overall average value, and then a moving average detrending method is used to eliminate trend anomalies. The expression is as follows:

[0090] (1);

[0091] in This indicates the input vibration acceleration data, where M represents the size of the moving window. This represents detrended vibration acceleration data;

[0092] Next, wavelet packet denoising is used to process the vibration acceleration data. The expression is as follows:

[0093] (2);

[0094] in, represents the wavelet packet basis function; a fourth-order Daubechies wavelet basis is selected for the bridge cable vibration acceleration data; L represents the number of decomposition layers. Represented as the wavelet packet decomposition coefficients of the j-th frequency band component; This represents the inverse wavelet packet transform, used for signal reconstruction. The threshold based on the absolute deviation of the median is expressed by formula (3). The soft thresholding process is expressed by formula (4):

[0095] (3);

[0096] (4);

[0097] Then, the detrended and denoised vibration acceleration data is randomly sliced. Since the vibration acceleration sensor sampling frequency is 20Hz, the slice length is 1200, corresponding to 1 minute of vibration acceleration data. Maximum absolute value normalization is then performed on each slice sequentially, ensuring that the data range of each slice is within [-1, 1].

[0098] For environmental factor data, the preprocessing steps include wavelet packet denoising, normalization, and random data slicing. For wavelet packet denoising, a fifth-order Daubechies wavelet basis is selected. A comparison of the effects before and after wavelet packet denoising is shown in the attached figure. Figure 3 As shown; the normalization process uses the maximum-minimum normalization method to ensure the overall data range is within [0, 1] before random slicing. The random slice positions for environmental factor data correspond sequentially to the slice positions for vibration acceleration data. Since the sampling frequency of the environmental factor sensor data is 1Hz, a slice length of 60 corresponds to 1 minute of sampling. Considering the hysteresis effect of environmental factors, according to the specifications, the starting time value of the environmental data slice is extended forward by 2.5 to 3 bridge natural vibration cycles. In this project, 15 seconds is used, resulting in a total slice length of 75 and a time scale of 75 seconds.

[0099] Vibration acceleration data and environmental factor data (including temperature and wind speed data) were divided into training, validation, and test sets in a 7:1:2 ratio. To enhance the data's expressive power and optimize the deep learning model's training performance, 10 random missing data segments were introduced into part of the vibration acceleration data training set, as shown in the attached figure. Figure 4 As shown, a training set of vibration acceleration data containing random missing slices is constructed.

[0100] The second step is to build as shown in the attached diagram. Figure 5 The convolutional encoder model shown on the left consists of five convolutional coding layers. Each layer comprises a convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. The LeakyReLU activation function is chosen, with a gradient of 0.01 in the negative region to prevent the vanishing gradient problem. The first convolutional coding layer has an input channel count equal to the number of vibration accelerometers, meaning the input data has a dimension of 6 and a length of 1200, processing data from multiple vibration accelerometers simultaneously in a multi-channel parallel input manner. Through five convolutional coding layers, the dimension is gradually increased while the length is reduced, simultaneously obtaining the individual data features of each vibration accelerometer and the collaborative relationships between the sensors, thus forming a high-dimensional feature vector of vibration acceleration with a dimension of 256 and a length of 75, shortening the time scale by a factor of 16.

[0101] The third step is to build two as shown in the attached diagram. Figure 6The transform neural network encoder model shown serves as two auxiliary network architectures, taking environmental wind speed and environmental temperature data as inputs respectively. After processing by the transform neural network encoder model, it outputs high-dimensional feature vectors of environmental wind speed and environmental temperature. Specifically, the transform neural network encoder uses a six-layer multi-head self-attention coding layer as its core to extract high-dimensional feature vectors of environmental factor data, and its expression is formula (5). After input rearrangement, input projection through two fully connected layers, and position encoding, the environmental factor data forms a 32-dimensional tensor, which enters the six-layer multi-head self-attention coding layer. Then, after output rearrangement and output projection through two fully connected layers, it outputs a 4-dimensional high-dimensional feature vector of environmental factors. The expression for the position encoding process is formula (6):

[0102] (5);

[0103] in, Indicates the i-th self-attention head; This indicates a self-attention operation; These represent the query vector, key vector, and value vector, respectively. Indicates the dimension of the key vector; The function is used to normalize the attention score; This indicates the matrix transpose.

[0104] (6);

[0105] in, Indicates the position of data in the sequence, an integer starting from 0; i represents the dimension index; Indicates the first in the sequence The encoded value at that location; This indicates the dimensionality of the input data.

[0106] The fourth step involves concatenating the high-dimensional feature vectors of vibration acceleration and environmental factors. Specifically, this concatenation is done by superimposing the features in terms of dimension while maintaining their length, resulting in a fused high-dimensional feature vector. Since the vibration acceleration feature vector has a dimension of 256 and the environmental factor feature vectors both have a dimension of 4, the fused high-dimensional feature vector has a dimension of 264. Furthermore, since both the vibration acceleration and environmental factor feature vectors have a length of 75, the fused high-dimensional feature vector also has a length of 75.

[0107] Step 5, construct as shown in the attached diagram. Figure 5The convolutional decoder model shown on the right side is symmetrical to the convolutional encoder model architecture and consists of five deconvolutional decoding layers. Each deconvolutional decoding layer consists of a deconvolutional layer, a batch normalization layer, and a LeakyReLU activation function layer. After fusing high-dimensional feature vectors into the deconvolutional decoding layer, as the deconvolutional decoding layer gradually reduces the dimensionality and increases the length, the number of output channels of the final convolutional decoding layer is equal to the number of vibration acceleration sensors, that is, the output data has a dimension of 6 and a length of 1200, thereby realizing the reconstruction of vibration acceleration data.

[0108] Step 6: Combine the convolutional autoencoder model with the two transform neural network encoder models in a parallel network architecture to create a fused model, as shown in the attached diagram. Figure 7 As shown, a convolutional encoder and convolutional decoder model form the main network architecture for processing vibration acceleration data, while a transform neural network encoder model serves as the auxiliary neural network for processing environmental factor data. The main and auxiliary networks operate in parallel, thereby achieving vibration acceleration data reconstruction while considering the coupling of multiple environmental field factors. Simultaneously, residual connections are established between the second and third convolutional encoding layers and their symmetrically positioned deconvolutional decoding layers. These residual connections prevent the loss of data features during data transmission within the deep neural network.

[0109] The model was trained using the vibration acceleration data training set (containing partially missing data segments) and the environmental data training set from Step 1. The mean squared error loss function (MSE loss function) was chosen as the loss function, and the Adam optimizer was selected. A learning rate decay strategy was employed during model training, with an initial learning rate of 0.001 and a decay coefficient of 0.5 every 50 epochs. An early stopping function was used to control the model training process, enabling early stopping validation after 100 epochs. The number of training epochs was controlled by comparing the loss values ​​on the training and validation sets, with a maximum of 150 epochs. Weighted decay regularization with a decay coefficient of 0.001 was used to prevent overfitting. The loss function curves during model training are attached. Figure 8 As shown, it can be seen that the loss values ​​of both the training set and the validation set decrease rapidly at first and then slowly with the number of training rounds, eventually approaching 0.0027 and 0.0031 respectively, indicating that the model training effect is good.

[0110] Step 7: Test the performance of the deep learning model using multiple evaluation metrics. First, the mean squared error and Pearson correlation coefficient are used as two scoring metrics to evaluate the model's performance, expressed as formulas (7) and (8). Verification showed that the mean squared error was 0.002998, and the Pearson correlation coefficient was 0.9831, both remaining at an excellent level, indicating that the model accurately located the missing data and precisely filled it with healthy data.

[0111] (7);

[0112] Where n represents the number of data points, This represents the true value of the i-th data point. This represents the model output value for the i-th data point;

[0113] (8);

[0114] in, This represents the average value of the actual data. This represents the average value of the model's output data.

[0115] Next, the model's missing data filling capability is visually demonstrated using images. The model's input and output data are plotted on a single graph, as shown in the attached image. Figure 9 As shown, the model can accurately identify and label the missing input data. Next, the model output data and health data are plotted together on a single graph, with the missing input data segments labeled, as shown in the attached figure. Figure 10 As shown, the deep learning network model accurately fills in missing data fragments into healthy data states, and the model output value can be regarded as healthy data, indicating that the model effectively completes the data cleaning task. Furthermore, the model input data and output data are compared sequentially with healthy data, and comparison images and Fourier spectrum comparison images are plotted, as shown in the attached figure. Figure 11 As shown, missing data fragments affect data integrity and cause changes in data characteristics. However, the data cleaned by the deep learning model restores data integrity and ensures that the data characteristics are consistent with healthy data, further verifying the model's efficient data reconstruction and missing data imputation capabilities.

[0116] The tested model was applied to the long-term health monitoring of the Ma'anshan Yangtze River Bridge. After the data preprocessing shown in the first step, the field-collected sensor data was input into a deep learning neural network, which outputs healthy and complete vibration acceleration sensor data. This completed the data cleaning work while considering the coupling effects of multiple environmental fields, demonstrating high real-time performance. For the vibration acceleration data cleaning of cables in other parts of the bridge, leveraging the model's generalization ability, a high-precision data cleaning deep learning model can be obtained simply by fine-tuning the parameters and retraining the deep learning model.

Claims

1. A bridge sling multi-channel vibration acceleration data missing filling method based on multi-environmental field factor coupling, characterized by, The method comprises the following steps: S1, selecting a plurality of vibration acceleration sensors on a monitoring structure, and selecting an environmental sensor, preprocessing vibration acceleration data and environmental data respectively, and sequentially building a vibration acceleration data set and an environmental factor data set for training and testing a deep learning model; S2, building a convolutional encoder model for processing the vibration acceleration data set described in step S1, inputting the data of the plurality of vibration acceleration sensors as a plurality of input channels into the convolutional encoder model at the same time, extracting data features through five convolutional encoding layers to form a vibration acceleration high-dimensional feature vector; S3, building a transformer neural network encoder model for processing the environmental factor data set described in step S1, obtaining an environmental factor high-dimensional feature vector through six multi-head self-attention encoding layers; S4, splicing the vibration acceleration high-dimensional feature vector described in step S2 and the environmental factor high-dimensional feature vector described in step S3 to obtain a fusion high-dimensional feature vector; S5, building a convolutional decoder model for processing the fusion high-dimensional feature vector described in step S4, decoding the high-dimensional feature vector through five inverse convolutional decoding layers to realize multi-channel vibration acceleration data reconstruction; S6, combining the convolutional encoder model described in step S2 and the convolutional decoder model described in step S5 into a convolutional autoencoder model, synthesizing the fusion model with the transformer neural network encoder model described in step S3 in the form of a parallel network architecture, and selecting a loss function and an optimizer to train the model and optimize the parameters; S7, using a plurality of evaluation indexes to evaluate the fusion model, verifying that the fusion model can not only sensitively identify vibration data anomalies and accurately locate them, but also accurately fill in missing vibration data.

2. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The vibration acceleration sensor data preprocessing in step S1 is as follows: each vibration acceleration sensor data is subtracted by the overall average value to eliminate offset errors; then a moving average detrending method is used to eliminate long-term and short-term trend anomalies; next, a fourth-order Daubechies wavelet packet denoising method is used to filter noise to obtain preliminary denoised vibration acceleration sensor data; finally, the detrended and denoised vibration acceleration sensor data is randomly sliced, and each slice is sequentially normalized by the maximum absolute value, so that the data range of each slice is within [-1, 1]; The environmental factor data preprocessing process is as follows: a fifth-order Daubechies wavelet packet denoising method is used to filter noise, and maximum and minimum value normalization is performed before random slicing to make the overall data range within [0, 1]; the random slicing positions of the environmental data correspond to the vibration acceleration data slicing positions in turn, and the slicing start position is extended forward by 2.5 to 3 times the bridge natural vibration period length to consider the lagging effect of environmental factors.

3. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The environmental data in step S1 includes temperature data and wind speed data, and the vibration acceleration data, temperature data and wind speed data are divided into training set, validation set and test set in a ratio of 7:1:2; 5%-15% of random proportion data is introduced into the vibration acceleration data training set to improve the ability of the model to recover missing data.

4. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The convolutional encoder model in step S2 is composed of five convolutional encoding layers, each of which is composed of a convolutional layer, a batch normalization layer and a LeakyReLU activation function layer; the input channel number of the first convolutional encoding layer is equal to the number of vibration acceleration sensors, and the data of multiple vibration acceleration sensors is processed in a multi-channel parallel input manner; after passing through the five convolutional encoding layers, the dimension is gradually increased and the length is gradually reduced, and the data characteristics of each vibration acceleration sensor and the synergistic relationship between the sensors are obtained, thereby forming a vibration acceleration high-dimensional feature vector.

5. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The transformation neural network encoder in step S3 is composed of six multi-head self-attention encoding layers, which extract the environment factor high-dimensional feature vector; after the input rearrangement, two fully connected layer input projections and position encoding, the environment factor data enters the six multi-head self-attention encoding layers, and then the environment factor high-dimensional feature vector is output after the output rearrangement and two fully connected layer output projections.

6. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, In step S4, the vibration acceleration high-dimensional feature vector and the environment factor high-dimensional feature vector are spliced, and the specific splicing method is to stack and splice in the dimension while keeping the feature length unchanged, thereby obtaining a fusion high-dimensional feature vector.

7. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The convolutional decoder model in step S5 is symmetrical to the convolutional encoder model in step S2, and is composed of five deconvolutional decoding layers, each of which is composed of a deconvolutional layer, a batch normalization layer and a LeakyReLU activation function layer; after the fusion high-dimensional feature vector is input into the deconvolutional decoding layer, the dimension is gradually reduced and the length is gradually increased, and the output channel number of the last convolutional decoding layer is equal to the number of vibration acceleration sensors, thereby realizing the reconstruction of the vibration acceleration data.

8. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, In step S6, the convolutional encoder model and the convolutional decoder model are combined into a convolutional autoencoder model, which is used as the main network architecture to process the vibration acceleration data; the transformation neural network encoder model is used as an auxiliary neural network to process the environment factor data; The main network architecture and the auxiliary neural network run in parallel, thereby realizing the reconstruction of the vibration acceleration data considering the coupling of multiple environmental factors; at the same time, residual connections are set between the deconvolutional decoding layers symmetrically located before the second and third convolutional encoding layers, thereby preventing the loss of data features during the transmission of the deep neural network; The MSE loss function and the Adam optimizer are selected for model training and parameter optimization; the learning rate decay strategy is adopted for model training, and the early stopping function is used to control the model training process; At the same time, the total weight decay regularization method is adopted to prevent overfitting.

9. The bridge sling multi-channel vibration acceleration data missing filling method based on multi-environment field factor coupling according to claim 1, characterized in that, The mean square error score and the Pearson correlation coefficient score are selected as evaluation indexes in step S7, and the reconstruction ability of the deep learning model for each vibration acceleration data channel and the data cleaning ability of the overall sensor network are evaluated respectively; in addition, the input and output contrast images of the fusion model and the Fourier spectrum contrast images are drawn to verify that the fusion model can sensitively identify vibration data anomalies and accurately locate them, and can accurately fill in the missing vibration data.

Images