An intelligent and accurate measurement method for bridge deflection based on artificial intelligence

Abstract

The invention discloses an intelligent and precise measurement method for the deflection of a bridge based on artificial intelligence, and relates to the technical field of bridge deflection measurement. The invention combines a Retinex algorithm with an HDR imaging technology, performs adaptive illumination correction and enhancement on images collected under different exposure conditions, effectively adjusts the brightness and contrast of images under various illumination conditions, and combines edge detection and color correction technology to significantly reduce the influence of uneven illumination on measurement. The invention adopts adaptive background modeling and a mixed Gaussian model MOG2 to achieve real-time subtraction of a dynamic background, can quickly update the background model, separate the deflection features of the bridge, extract the deflection features based on a deep learning foreground segmentation network U‑Net, filter out dynamic background noise, and ensure that the measurement results are stable and reliable. The space and time features are jointly modeled through the spatiotemporal modeling of ST‑GCN combined with LSTM, and a three-dimensional deflection field is constructed, thereby enhancing the ability to capture the dynamic behavior of the bridge structure.

Description

Technical Field

[0001] The invention relates to the technical field of bridge deflection measurement, in particular to an intelligent and precise bridge deflection measurement method based on artificial intelligence. Background Art

[0002] Currently, AI-based bridge deflection measurement methods rely on visual sensors such as cameras, millimeter-wave radars, and deep learning algorithms to measure and evaluate bridge deflection through image processing and time series prediction.

[0003] Visual sensors are easily affected by changes in ambient lighting, such as changes in light intensity between morning and evening, weather changes (such as cloudy and sunny days), and shadow interference. These problems will lead to a decrease in image quality, which in turn affects the feature extraction ability of the deep learning model. In addition, background interference is common in open environments, such as moving pedestrians, vehicles, leaves blown by the wind, etc., which will confuse the model's discrimination ability and cause unstable measurement results.

[0004] To address this problem, traditional solutions use image enhancement algorithms, such as histogram equalization or gamma correction, to preprocess images and reduce the impact of lighting changes. They also use background subtraction-based methods to build background models to identify and filter out dynamic foreground objects. However, these methods have a slow dynamic update of the background model and are easily affected by environmental changes and are difficult to adapt to complex scenes. Therefore, there is an urgent need for an intelligent and accurate measurement method for bridge deflection based on artificial intelligence to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides an intelligent and accurate measurement method for the deflection of a bridge based on artificial intelligence to solve the problems of slow dynamic update of the background model, susceptibility to environmental changes, and difficulty in adapting to complex scenes.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The present invention provides an intelligent and accurate measurement method for the deflection of a bridge based on artificial intelligence, which comprises:

[0009] Step S1, adaptive illumination correction and image preprocessing,

[0010] Deploy multiple cameras on the bridge to cover the support points, mid-span and connection points, capture the deformation video of the bridge under different environmental and load conditions, and generate image input;

[0011] Step S2, dynamic background suppression and target deflection feature extraction,

[0012] Using the image input in step S1, an adaptive background modeling method and a mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream to separate the deflection feature and the dynamic background noise;

[0013] Step S3, spatiotemporal feature extraction and three-dimensional deflection field modeling,

[0014] The extracted features in step S2 are input into the spatiotemporal graph convolutional network ST-GCN, and combined with the long short-term memory network LSTM to jointly model the temporal and spatial features. By capturing the spatiotemporal dynamic changes of deflection, the three-dimensional deflection field of the bridge under different loads and environmental conditions is established.

[0015] Integrate the image input in step S1, the deflection features in step S2 and the background noise to construct a three-dimensional deflection field, and refine the representation of the coordinated deformation of the bridge structure under multi-point and multi-time conditions;

[0016] Step S4, fusion model and parameter correction,

[0017] The attention mechanism is used to analyze the difference between the three-dimensional deflection field and the actual measurement data, and the model parameters are adaptively adjusted.

[0018] The adjusted deflection field is merged with the actual measured data.

[0019] Step S5: real-time monitoring and anomaly detection.

[0020] The adjusted three-dimensional deflection field is monitored in real time, and the unsupervised learning Autoencoder algorithm is used to detect anomalies in the monitoring data, compare the current monitoring data with historical patterns, and identify potential structural anomalies and abnormal changes in deflection.

[0021] Furthermore, in step S1, the Retinex algorithm is combined with high dynamic range HDR imaging technology to perform illumination correction on the video data, dynamically adjust the brightness and contrast of the image, and use edge detection and color correction technology to eliminate uneven illumination and shadow interference to generate image input.

[0022] Furthermore, in step S1, the Retinex algorithm is combined with HDR imaging technology to perform illumination correction:

[0023] Use the brightness information collected from multiple frames under different exposure conditions to synthesize high dynamic range images , the generation formula of HDR image is,

[0024] ,

[0025] in, represents a high dynamic range image, Indicates The pixel values ​​of the frame image at different exposures, Indicates The exposure time of the frame image, Represents the total number of image frames, represents the weight function, ,in, is the mean of all exposure times, is the parameter controlling the degree of influence;

[0026] Separate the illumination component and reflection component of the image for illumination correction.

[0027] ,in, represents the reflection component, i.e. the image correction result, are the pixel horizontal and vertical coordinates of the light intensity, Represents the pixel value of the image after HDR fusion, Represents the illumination component, i.e., the illumination intensity,

[0028] ,in, are the horizontal and vertical coordinates of the pixel area covered by the Gaussian kernel, Respectively represent the integral width along the horizontal and vertical axes, is a two-dimensional Gaussian filter function, defined as ,in, represents the standard deviation of Gaussian filtering;

[0029] The image after illumination correction is used for edge detection and color correction to adjust the brightness and contrast of the image. The color correction formula is: ,in, represents the pixel value of the image after color correction, Represents the brightness adjustment coefficient, Represents the contrast adjustment coefficient. The Retinex algorithm is combined with HDR technology to enhance the robustness of the image under the influence of lighting changes and shadows.

[0030] Further, in step S2, based on the separation result, the deflection features are extracted using a deep learning foreground segmentation network U-Net;

[0031] The dynamic background noise in step S2 includes pedestrians and vehicles;

[0032] The deflection features extracted in step S2 include the displacement and change direction of the beam body, support points, mid-span and bridge connection points.

[0033] Furthermore, the dynamic background suppression and target deflection feature extraction method in step S2 is:

[0034] Using the image input corrected in step S1, adaptive background modeling and mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream.

[0035] Construct a multi-Gaussian distribution combination for each pixel.

[0036] ,in, Represents pixel In time The probability distribution of time, represents the number of Gaussian distributions, representing different background modes for each pixel, Indicates The weight of a Gaussian distribution, that is, the proportion of this mode, Indicates Gaussian distribution with mean parameter and the covariance matrix , Represents pixel In time The intensity value at the moment;

[0037] Each time a new frame is input, the corresponding Gaussian distribution parameters are updated according to the current pixel value, including:

[0038] Mean update,

[0039] ,in, represents the updated mean, represents the learning rate;

[0040] Covariance matrix update,

[0041] ,in, represents the updated covariance matrix, Indicates the deviation of the current pixel value from the mean;

[0042] Weight update,

[0043] ,in, represents the updated weight, , represents the learning rate adjustment factor, Indicates the matching degree. If the current pixel value meets the If there is a distribution, it is 1, otherwise it is 0;

[0044] Compare the current value of the pixel with the mean and covariance of each Gaussian distribution to determine whether the pixel belongs to the background or foreground. ,but , otherwise 0,

[0045] Indicates the judgment result. If it is 1, it is judged as the foreground, that is, the deflection feature. If it is 0, it is the background. Indicates the threshold value;

[0046] The extracted foreground is further processed to extract deflection features and filter out dynamic background noise.

[0047] ,in, represents the processed foreground image, Represents morphological processing operations, such as opening and closing, used to eliminate small noise areas and retain major deflection features.

[0048] Furthermore, in step S2, deflection feature extraction is performed based on U-Net:

[0049] U-Net consists of two parts: encoder and decoder, which are used for feature extraction and feature reconstruction respectively. The encoder gradually extracts spatial features, and the decoder gradually restores the features into high-resolution deflection feature maps.

[0050] The encoder part of U-Net performs feature extraction and inputs the foreground image , extract features based on the convolutional layer, ,in, Indicates the encoder The output feature map of the layer, Indicates the encoder The input feature map of the layer, the initial input is the foreground image , Indicates The convolution kernel matrix of the layer, Indicates The bias term of the layer, represents the ReLU function, represents the convolution operator;

[0051] The pooling layer downsamples the feature map to extract high-level features.

[0052] ,in, Indicates The feature map after layer pooling, represents the pooling operation,

[0053] The decoder uses transposed convolution to gradually reconstruct the high-resolution deflection feature map and uses the encoder's jump connection to retain the detailed features.

[0054] Use transposed convolution to upscale the feature map to its original resolution, ,in, Indicates the decoder The output feature map of the layer, Indicates the decoder The input feature map of the layer, the initial input is the deep feature output by the encoder, Indicates The transposed convolution kernel matrix of the layer, Indicates The bias term of the layer;

[0055] The skip connection is used to pass the high-resolution features of the encoder directly to the decoder, retaining the detailed information. ,in, Indicates The skip connection result of the layer, Represents the output feature map of the corresponding layer of the encoder, retaining the detailed features for reconstruction;

[0056] The final output of the decoder is converted into a segmentation probability map through an activation function to extract the deflection features.

[0057] ,in, It represents the deflection feature map finally output by U-Net, and determines whether each pixel belongs to the deflection change area. represents the convolution kernel matrix of the output layer, Represents the bias term of the output layer, , which realizes the deflection feature extraction in the foreground.

[0058] Furthermore, in step S3, the spatiotemporal feature extraction and three-dimensional deflection field modeling method is:

[0059] Using the deflection features extracted in step S2, the spatial and temporal features are jointly modeled through the spatiotemporal graph convolutional network ST-GCN combined with the long short-term memory network LSTM, and the image input of step S1, the deflection features of step S2 and the background noise are integrated to construct a three-dimensional deflection field;

[0060] According to the deflection characteristics, define the space-time diagram ,in Represents a collection of nodes. Represents a set of edges, each node represents a deflection feature point, and the edge represents the association in time and space.

[0061] ,in, represents the node set, i.e., each deflection measurement point, Indicates the number of nodes, corresponding to the total number of measurement points,

[0062] ,in, Represents a spatial edge set, connecting adjacent nodes in space, represents a set of temporal edges, connecting the states of the same node at adjacent times;

[0063] Use graph convolution to update node features and capture spatial features.

[0064] ,in, Indicates The node feature matrix of the layer, Indicates Node in layer The characteristic vector of Representation Node The set of neighbor nodes of represents the normalization coefficient, Indicates The convolution weight matrix of the layer, Indicates The bias vector of the layer, Represents the ReLU function;

[0065] Combined with the long short-term memory network LSTM, the features after graph convolution are modeled in time to capture the dynamic changes of node features over time.

[0066] The features of each node are input into LSTM for time series modeling.

[0067] ,in, Represents LSTM at time The hidden state of the moment, Indicates at time The input features at the moment come from the output of ST-GCN. Represents the hidden state of LSTM at the previous moment, , Represents the weight matrix of input and hidden state, represents the bias vector, represents the tanh function;

[0068] The final output time feature is mapped by the output gate of LSTM.

[0069] ,in, represents the output of LSTM, represents the output weight matrix, Indicates output bias;

[0070] Integrate the image input of step S1, the deflection features and background noise of step S2, and combine the results of ST-GCN and LSTM to construct a three-dimensional deflection field.

[0071] ,in, represents the three-dimensional deflection field, that is, the deflection at each point in space and time, Represents a nonlinear mapping function that combines spatial and temporal features.

[0072] Furthermore, in step S4, the method of using the Attention mechanism to perform adaptive parameter adjustment and data fusion is as follows:

[0073] The attention mechanism is used to calculate the difference between the 3D deflection field and the actual measured data and give higher weights to important features. The input of the attention includes the 3D deflection field predicted by the model. and actual measurement data ,

[0074] Constructing a query ,key ,value matrix,

[0075] ,

[0076] ,

[0077] ,in, For actual measurement data, represents the query matrix, generated by the predicted three-dimensional deflection field, i.e., the features to be matched, represents the key matrix, generated from actual measurement data, i.e., the index of known features, represents the value matrix, generated from actual measurement data, Represents the weight matrix, which is used to construct the query, key and value matrices respectively;

[0078] Use scaled dot product Attention to calculate the match between the predicted and measured data.

[0079] ,

[0080] in, is the Attention output matrix, which represents the fusion result of the predicted data and the measured data. represents the dot product of the query and the key, i.e. the matching degree, represents the scaling factor, is the dimension of the key, Indicates that the matching degree is normalized to obtain the importance weight of each feature;

[0081] Use Attention output to adjust model parameters to make the model closer to the actual measured data and improve accuracy.

[0082] Calculate the difference loss based on the Attention output and the prediction results, and adjust the model parameters.

[0083] ,in, represents the discrepancy loss, i.e., the deviation between the model prediction and the measured data, represents the three-dimensional deflection field of the model, Represents the fused data after adjustment by the Attention mechanism;

[0084] Using the gradient descent method, the model parameters are updated according to the difference loss.

[0085] ,in, represents the current parameter vector of the model, represents the parameter vector after the model is updated, represents the learning rate, Represents the gradient of the loss function with respect to the parameters;

[0086] The adjusted deflection field is fused with the actual measurement data to generate an updated three-dimensional deflection field.

[0087] ,in,

[0088] represents the fused three-dimensional deflection field, Represents the fusion coefficient, effectively identifying and emphasizing the differences between the predictions and the actual measurements.

[0089] Furthermore, in step S5, when abnormal deflection or a tendency of bridge structure instability is detected, an alarm is automatically triggered.

[0090] Furthermore, the method of performing real-time monitoring and abnormality detection in step S5 is:

[0091] Based on the real-time monitoring of the adjusted three-dimensional deflection field, the Autoencoder unsupervised learning algorithm is used to detect anomalies in the monitoring data, compare the current monitoring data with historical patterns, and identify potential structural anomalies and abnormal changes in deflection.

[0092] The three-dimensional deflection field adjusted in step S4 Enter Autoencoder for real-time monitoring,

[0093] ,in, represents the real-time monitoring data input to the Autoencoder, i.e., the adjusted deflection value at each time and space position,

[0094] Autoencoder consists of an encoder and a decoder. The encoder is used for feature compression, and the decoder is used for feature reconstruction to detect the difference between current data and historical patterns.

[0095] Compress the input data into a low-dimensional feature space and use multi-layer nonlinear transformation to represent it.

[0096] ,in, represents the encoded feature vector, that is, the compressed deflection feature, represents the weight matrix of the encoder, represents the bias vector of the encoder, Represents the ReLU function;

[0097] Reconstruct compressed features into an approximation of the original input, detect reconstruction errors, and identify anomalies.

[0098] ,in, represents the decoder reconstructed output, which approximates the input deflection data, represents the weight matrix of the decoder, mapping the feature vector back to the original space, represents the bias vector of the decoder, Represents the activation function of the decoder, which is used to output the reconstruction result;

[0099] Compare the input data with the reconstruction result and calculate the reconstruction error as an anomaly detection indicator.

[0100] ,in, represents the reconstruction error, i.e., the difference between the current deflection field and the historical pattern, represents the actual monitored deflection data, Represents the deflection data reconstructed by Autoencoder;

[0101] Set thresholds to determine reconstruction errors, identify potential anomalies and automatically trigger alarms.

[0102] The beneficial effects of the present invention are:

[0103] The present invention combines the Retinex algorithm with the HDR imaging technology to perform adaptive illumination correction and enhancement on images collected under different exposure conditions, effectively adjust the image brightness and contrast under various illumination conditions, and significantly reduce the impact of uneven illumination on measurement by combining edge detection and color correction technology.

[0104] The present invention adopts adaptive background modeling and mixed Gaussian model MOG2 to achieve real-time subtraction of dynamic background, can quickly update the background model, separate the deflection characteristics of the bridge, extract the deflection characteristics based on the deep learning foreground segmentation network U-Net, filter out dynamic background noise, and ensure that the measurement results are stable and reliable.

[0105] The present invention combines ST-GCN with the spatiotemporal modeling of LSTM to jointly model spatial and temporal features, construct a three-dimensional deflection field, and enhance the ability to capture the dynamic behavior of the bridge structure.

[0106] The present invention introduces the Attention mechanism, analyzes the difference between the three-dimensional deflection field and the actual measurement data, adaptively adjusts the model parameters, and adjusts and corrects the model in real time through a data-driven method, which solves the shortcomings of traditional methods in parameter dependence and computational complexity and achieves more efficient model fusion.

[0107] The present invention uses the Autoencoder unsupervised learning algorithm to analyze the adjusted three-dimensional deflection field in real time, performs anomaly detection, compares the current monitoring data according to the historical pattern, and automatically triggers an alarm when abnormal deflection or structural instability trend is detected, thereby effectively reducing the misjudgment rate and improving the reliability and response speed in long-term monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0109] Figure 1 The present invention is a schematic flow chart of the intelligent and precise measurement method for bridge deflection based on artificial intelligence. DETAILED DESCRIPTION

[0110] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0111] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0112] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0113] Example 1, reference Figure 1 This embodiment provides an intelligent and accurate measurement method for the deflection of a bridge based on artificial intelligence, comprising the following steps:

[0114] Step S1, adaptive illumination correction and image preprocessing,

[0115] Deploy multiple cameras on the bridge to cover the support points, mid-span and connection points, capture the deformation video of the bridge under different environmental and load conditions, and generate image input;

[0116] In step S1, the Retinex algorithm is combined with high dynamic range (HDR) imaging technology to perform illumination correction on the video data, dynamically adjust the brightness and contrast of the image, and use edge detection and color correction technology to eliminate uneven illumination and shadow interference to generate image input;

[0117] In step S1, the Retinex algorithm is combined with HDR imaging technology to perform illumination correction:

[0118] Use the brightness information collected from multiple frames under different exposure conditions to synthesize high dynamic range images , the generation formula of HDR image is,

[0119] ,

[0120] in, represents a high dynamic range image, Indicates The pixel values ​​of the frame image at different exposures, Indicates The exposure time of the frame image, Represents the total number of image frames, represents the weight function, ,in, is the mean of all exposure times, is the parameter controlling the degree of influence;

[0121] Separate the illumination component and reflection component of the image for illumination correction.

[0122] ,in, represents the reflection component, i.e. the image correction result, are the pixel horizontal and vertical coordinates of the light intensity, Represents the pixel value of the image after HDR fusion, Represents the illumination component, i.e., the illumination intensity,

[0123] ,in, are the horizontal and vertical coordinates of the pixel area covered by the Gaussian kernel, Respectively represent the integral width along the horizontal and vertical axes, is a two-dimensional Gaussian filter function, defined as ,in, represents the standard deviation of Gaussian filtering;

[0124] The image after illumination correction is used for edge detection and color correction to adjust the brightness and contrast of the image. The color correction formula is: ,in, represents the pixel value of the image after color correction, Represents the brightness adjustment coefficient, Represents the contrast adjustment coefficient. The Retinex algorithm is combined with HDR technology to enhance the robustness of the image under the influence of lighting changes and shadows.

[0125] Specifically, adaptive illumination correction and image preprocessing technology is used, combining the Retinex algorithm with high dynamic range (HDR) imaging technology to perform real-time illumination correction on images of bridges under different lighting conditions, effectively eliminating uneven illumination and shadow interference, and enhancing the robustness of images under lighting changes and complex environments.

[0126] In addition, most existing bridge deflection measurement methods use single-point or two-dimensional plane feature extraction and prediction, lacking a three-dimensional dynamic scene understanding of the overall deflection changes of the bridge. The model is insufficient in the multi-point coordinated deformation of large-span bridges, and time series feature extraction is limited to short-term predictions. There is a lack of effective modeling of long spans and multi-point deflection coordinated changes, making it difficult to accurately capture the overall behavior of the structure in long-term monitoring. To address this problem, traditional solutions mostly use finite element analysis (FEA) combined with sensor data to analyze the dynamic response of the bridge through physical modeling. However, these methods require detailed bridge structure parameters, are complex and time-consuming to calculate, and are difficult to apply in real time.

[0127] It can be seen that traditional image enhancement algorithms cannot make adaptive adjustments based on rapid changes in actual lighting conditions. The correction effect is limited and easily fails under extreme lighting conditions. They are less effective in dealing with sudden interferences, such as vehicles passing by suddenly. The background is not updated in time, resulting in a high misjudgment rate, making it difficult to maintain reliability in long-term monitoring.

[0128] In step S2 of the present invention, dynamic background suppression and target deflection feature extraction are performed.

[0129] Using the image input in step S1, an adaptive background modeling method and a mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream to separate the deflection feature and the dynamic background noise;

[0130] In step S2, based on the separation result, the deflection features are extracted using the deep learning foreground segmentation network U-Net;

[0131] The dynamic background noise in step S2 includes pedestrians and vehicles;

[0132] The deflection features extracted in step S2 include the displacement and change direction of the beam body, support point, mid-span and bridge connection point;

[0133] The dynamic background suppression and target deflection feature extraction method in step S2 is:

[0134] Using the image input corrected in step S1, adaptive background modeling and mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream.

[0135] Construct a multi-Gaussian distribution combination for each pixel.

[0136] ,in, Represents pixel In time The probability distribution of time, represents the number of Gaussian distributions, representing different background modes for each pixel, Indicates The weight of a Gaussian distribution, that is, the proportion of this mode, Indicates Gaussian distribution with mean parameter and The covariance matrix, Represents pixel In time The intensity value at the moment;

[0137] Each time a new frame is input, the corresponding Gaussian distribution parameters are updated according to the current pixel value, including:

[0138] Mean update,

[0139] ,in, represents the updated mean, represents the learning rate;

[0140] Covariance matrix update,

[0141] ,in, represents the updated covariance matrix, Indicates the deviation of the current pixel value from the mean;

[0142] Weight update,

[0143] ,in, represents the updated weight, , represents the learning rate adjustment factor, Indicates the matching degree. If the current pixel value meets the If there is a distribution, it is 1, otherwise it is 0;

[0144] Compare the current value of the pixel with the mean and covariance of each Gaussian distribution to determine whether the pixel belongs to the background or foreground. ,but , otherwise 0,

[0145] Indicates the judgment result. If it is 1, it is judged as the foreground, that is, the deflection feature. If it is 0, it is the background. Indicates the threshold value;

[0146] The extracted foreground is further processed to extract deflection features and filter out dynamic background noise.

[0147] ,in, represents the processed foreground image, Represents morphological processing operations, such as opening and closing, used to eliminate small noise areas and retain major deflection features.

[0148] In step S2, deflection feature extraction is performed based on U-Net:

[0149] U-Net consists of two parts: encoder and decoder, which are used for feature extraction and feature reconstruction respectively. The encoder gradually extracts spatial features, and the decoder gradually restores the features into high-resolution deflection feature maps.

[0150] The encoder part of U-Net performs feature extraction and inputs the foreground image , extract features based on the convolutional layer, ,in, Indicates the encoder The output feature map of the layer, Indicates the encoder The input feature map of the layer, the initial input is the foreground image , Indicates The convolution kernel matrix of the layer, Indicates The bias term of the layer, represents the ReLU function, represents the convolution operator;

[0151] The pooling layer downsamples the feature map to extract high-level features.

[0152] ,in, Indicates The feature map after layer pooling, Represents a pooling operation;

[0153] The decoder uses transposed convolution to gradually reconstruct the high-resolution deflection feature map and uses the encoder's jump connection to retain the detailed features.

[0154] Use transposed convolution to upscale the feature map to its original resolution, ,in, Indicates the decoder The output feature map of the layer, Indicates the decoder The input feature map of the layer, the initial input is the deep feature output by the encoder, Indicates The transposed convolution kernel matrix of the layer, Indicates The bias term of the layer;

[0155] The skip connection is used to pass the high-resolution features of the encoder directly to the decoder, retaining the detailed information. ,in, Indicates The skip connection result of the layer, Represents the output feature map of the corresponding layer of the encoder, retaining the detailed features for reconstruction;

[0156] The final output of the decoder is converted into a segmentation probability map through an activation function to extract the deflection features.

[0157] ,in, It represents the deflection feature map finally output by U-Net, and determines whether each pixel belongs to the deflection change area. represents the convolution kernel matrix of the output layer, represents the bias term of the output layer,, which realizes the deflection feature extraction in the foreground;

[0158] Specifically, through dynamic background suppression and target deflection feature extraction, adaptive background modeling and mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream, effectively separating the deflection features and dynamic background noise, pedestrians, and vehicles. The real-time background update mechanism significantly reduces the impact of environmental dynamic interference on the measurement accuracy, ensuring the system's accurate recognition of target features. At the same time, combined with the U-Net deep learning network, the accuracy of deflection feature extraction is further enhanced.

[0159] Step S3, spatiotemporal feature extraction and three-dimensional deflection field modeling,

[0160] The extracted features in step S2 are input into the spatiotemporal graph convolutional network ST-GCN, and combined with the long short-term memory network LSTM to jointly model the temporal and spatial features. By capturing the spatiotemporal dynamic changes of deflection, the three-dimensional deflection field of the bridge under different loads and environmental conditions is established.

[0161] Integrate the image input in step S1, the deflection features in step S2 and the background noise to construct a three-dimensional deflection field, and refine the representation of the coordinated deformation of the bridge structure under multi-point and multi-time conditions;

[0162] In step S3, the spatiotemporal feature extraction and three-dimensional deflection field modeling method is:

[0163] Using the deflection features extracted in step S2, the spatial and temporal features are jointly modeled through the spatiotemporal graph convolutional network ST-GCN combined with the long short-term memory network LSTM, and the image input of step S1, the deflection features of step S2 and the background noise are integrated to construct a three-dimensional deflection field;

[0164] According to the deflection characteristics, define the space-time diagram ,in Represents a collection of nodes. Represents a set of edges, each node represents a deflection feature point, and the edge represents the association in time and space.

[0165] ,in, represents the node set, i.e., each deflection measurement point, Indicates the number of nodes, corresponding to the total number of measurement points,

[0166] ,in, Represents a spatial edge set, connecting adjacent nodes in space, represents a set of temporal edges, connecting the states of the same node at adjacent times;

[0167] Use graph convolution to update node features and capture spatial features.

[0168] ,in, Indicates The node feature matrix of the layer, Indicates Node in layer The characteristic vector of Representation Node The set of neighbor nodes of represents the normalization coefficient, Indicates The convolution weight matrix of the layer, Indicates The bias vector of the layer, Represents the ReLU function;

[0169] Combined with the long short-term memory network LSTM, the features after graph convolution are modeled in time to capture the dynamic changes of node features over time.

[0170] The features of each node are input into LSTM for time series modeling.

[0171] ,in, Represents LSTM at time The hidden state of the moment, Indicates at time The input features at the moment come from the output of ST-GCN. Represents the hidden state of LSTM at the previous moment, , represents the weight matrix of input and hidden state, represents the bias vector, represents the tanh function;

[0172] The final output time feature is mapped by the output gate of LSTM.

[0173] ,in, represents the output of LSTM, represents the output weight matrix, Indicates output bias;

[0174] Integrate the image input of step S1, the deflection features and background noise of step S2, and combine the results of ST-GCN and LSTM to construct a three-dimensional deflection field.

[0175] ,in, represents the three-dimensional deflection field, that is, the deflection at each point in space and time, Represents a nonlinear mapping function that combines spatial and temporal features.

[0176] Specifically, in the spatiotemporal feature extraction and three-dimensional deflection field modeling, the spatiotemporal graph convolutional network ST-GCN and the long short-term memory network LSTM are combined to capture the dynamic characteristics of the bridge in different time and space. ST-GCN is responsible for modeling the node characteristics and their associations in the space, while LSTM performs time series modeling on the spatial characteristics, effectively integrating the spatiotemporal information of the bridge deflection, establishing a high-precision three-dimensional deflection field, and refining the collaborative deformation behavior of the bridge structure under multi-point and multi-period conditions, so as to more comprehensively reflect the dynamic response of the bridge under different loads and environmental changes.

[0177] Step S4, fusion model and parameter correction,

[0178] The attention mechanism is used to analyze the difference between the three-dimensional deflection field and the actual measurement data, and the model parameters are adaptively adjusted.

[0179] The adjusted deflection field is merged with the actual measured data.

[0180] In step S4, the method of using the Attention mechanism for adaptive parameter adjustment and data fusion is as follows:

[0181] The attention mechanism is used to calculate the difference between the 3D deflection field and the actual measured data and give higher weights to important features. The input of the attention includes the 3D deflection field predicted by the model. and actual measurement data ,

[0182] Constructing a query ,key ,value matrix,

[0183] ,

[0184] ,

[0185] ,in, For actual measurement data, represents the query matrix, generated by the predicted three-dimensional deflection field, i.e., the features to be matched, represents the key matrix, generated from actual measurement data, i.e., the index of known features, represents the value matrix, generated from actual measurement data, Represents the weight matrix, which is used to construct the query, key and value matrices respectively;

[0186] Use scaled dot product Attention to calculate the match between the predicted and measured data.

[0187] ,

[0188] in, is the Attention output matrix, which represents the fusion result of the predicted data and the measured data. represents the dot product of the query and the key, i.e. the matching degree, represents the scaling factor, is the dimension of the key, Indicates that the matching degree is normalized to obtain the importance weight of each feature;

[0189] Use Attention output to adjust model parameters to make the model closer to the actual measured data and improve accuracy.

[0190] Calculate the difference loss based on the Attention output and the prediction results, and adjust the model parameters.

[0191] ,in, represents the discrepancy loss, i.e., the deviation between the model prediction and the measured data, represents the three-dimensional deflection field of the model, Represents the fused data after adjustment by the Attention mechanism;

[0192] Using the gradient descent method, the model parameters are updated according to the difference loss.

[0193] ,in, represents the current parameter vector of the model, represents the updated parameter vector of the model, represents the learning rate, Represents the gradient of the loss function with respect to the parameters;

[0194] The adjusted deflection field is fused with the actual measurement data to generate an updated three-dimensional deflection field.

[0195] ,in,

[0196] represents the fused three-dimensional deflection field, Represents the fusion coefficient, effectively identifying and emphasizing the differences between the predictions and the actual measurements.

[0197] Specifically, the Attention mechanism is introduced to adaptively adjust the model parameters by comparing the differences between the three-dimensional deflection field and the actual measurement data. The Attention mechanism effectively identifies and predicts the differences between the measurements, enabling the model to more accurately reflect the actual structural state.

[0198] Step S5: real-time monitoring and anomaly detection.

[0199] Real-time monitoring of the adjusted three-dimensional deflection field is performed, and anomaly detection of the monitoring data is performed using the unsupervised learning Autoencoder algorithm, comparing the current monitoring data with historical patterns to identify potential structural anomalies and abnormal changes in deflection;

[0200] In step S5, when abnormal deflection or a tendency of bridge structure instability is detected, an alarm is automatically triggered.

[0201] The method of performing real-time monitoring and abnormality detection in step S5 is:

[0202] Based on the real-time monitoring of the adjusted three-dimensional deflection field, the Autoencoder unsupervised learning algorithm is used to detect anomalies in the monitoring data, compare the current monitoring data with historical patterns, and identify potential structural anomalies and abnormal changes in deflection.

[0203] The three-dimensional deflection field adjusted in step S4 Enter Autoencoder for real-time monitoring,

[0204] ,in, represents the real-time monitoring data input to the Autoencoder, i.e., the adjusted deflection value at each time and space position,

[0205] Autoencoder consists of an encoder and a decoder. The encoder is used for feature compression, and the decoder is used for feature reconstruction to detect the difference between current data and historical patterns.

[0206] Compress the input data into a low-dimensional feature space and use multi-layer nonlinear transformation to represent it.

[0207] ,in, represents the encoded feature vector, that is, the compressed deflection feature, represents the weight matrix of the encoder, represents the bias vector of the encoder, Represents the ReLU function;

[0208] Reconstruct compressed features into an approximation of the original input, detect reconstruction errors, and identify anomalies.

[0209] ,in, represents the decoder reconstructed output, which approximates the input deflection data, represents the weight matrix of the decoder, mapping the feature vector back to the original space, represents the bias vector of the decoder, Represents the activation function of the decoder, which is used to output the reconstruction result;

[0210] Compare the input data with the reconstruction result and calculate the reconstruction error as an anomaly detection indicator.

[0211] ,in, represents the reconstruction error, i.e., the difference between the current deflection field and the historical pattern, represents the actual monitored deflection data, Represents the deflection data reconstructed by Autoencoder;

[0212] Set thresholds to determine reconstruction errors, identify potential anomalies and automatically trigger alarms;

[0213] Specifically, through the real-time monitoring and anomaly detection module, the Autoencoder unsupervised learning algorithm is used to perform real-time analysis on the adjusted three-dimensional deflection field to automatically identify potential abnormal deflections and structural instability trends.

[0214] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent and accurate measurement method for bridge deflection based on artificial intelligence, characterized in that: include, Step S1, adaptive illumination correction and image preprocessing, deploying multiple cameras on the bridge to cover the support points, mid-span and connection points, capturing bridge deformation video and generating image input; In step S1, the Retinex algorithm is combined with the high dynamic range HDR imaging technology to perform illumination correction on the video data to generate image input. The steps of performing illumination correction are as follows: Use the brightness information collected from multiple frames under different exposure conditions to synthesize high dynamic range images , the generation formula of HDR image is, ,in, represents a high dynamic range image, Indicates The pixel values ​​of the frame image at different exposures, Indicates The exposure time of the frame image, Represents the total number of image frames, represents the weight function, ,in, is the mean of all exposure times, is the parameter controlling the degree of influence; Separate the illumination component and reflection component of the image for illumination correction. ,in, represents the reflection component, i.e. the image correction result, are the pixel horizontal and vertical coordinates of the light intensity, Represents the pixel value of the image after HDR fusion, Represents the illumination component, i.e., the illumination intensity, ,in, are the horizontal and vertical coordinates of the pixel area covered by the Gaussian kernel, Respectively represent the integral width along the horizontal and vertical axes, is a two-dimensional Gaussian filter function, defined as ,in, represents the standard deviation of Gaussian filtering; The image after illumination correction is used for edge detection and color correction to adjust the brightness and contrast of the image. The color correction formula is: ,in, represents the pixel value of the image after color correction, Represents the brightness adjustment coefficient, Indicates the contrast adjustment coefficient; Step S2, dynamic background suppression and target deflection feature extraction, using the image input in step S1, using the adaptive background modeling method and the mixed Gaussian model MOG2, real-time background subtraction is performed on the video stream to separate the deflection feature and the dynamic background noise; In step S2, based on the separation results, the deflection features are extracted using the deep learning foreground segmentation network U-Net. The dynamic background noise in step S2 includes pedestrians and vehicles. The deflection features extracted in step S2 include the displacement and change direction of the beam body, support point, mid-span and bridge connection point; The dynamic background suppression and target deflection feature extraction method in step S2 is: Using the image input corrected in step S1, adaptive background modeling and mixed Gaussian model MOG2 are used to perform real-time background subtraction on the video stream. Construct a multi-Gaussian distribution combination for each pixel. ,in, Represents pixel In time The probability distribution of time, represents the number of Gaussian distributions, representing different background modes for each pixel, Indicates The weight of a Gaussian distribution, that is, the proportion of this mode, Indicates Gaussian distribution with mean parameter and the covariance matrix , Represents pixel In time The intensity value at the moment; Each time a new frame is input, the corresponding Gaussian distribution parameters are updated according to the current pixel value, including: Mean update, ,in, represents the updated mean, represents the learning rate; Covariance matrix update, ,in, represents the updated covariance matrix, Indicates the deviation of the current pixel value from the mean; Weight update, ,in, represents the updated weight, , represents the learning rate adjustment factor, Indicates the matching degree. If the current pixel value meets the If there is a distribution, it is 1, otherwise it is 0; Compare the current value of the pixel with the mean and covariance of each Gaussian distribution to determine whether the pixel belongs to the background or foreground. ,but , otherwise 0, Indicates the judgment result. If it is 1, it is judged as the foreground, that is, the deflection feature. If it is 0, it is the background. Indicates the threshold value; The extracted foreground is further processed to extract deflection features and filter out dynamic background noise. ,in, represents the processed foreground image, Represents morphological processing operations; In step S2, deflection feature extraction is performed based on U-Net: The encoder part of U-Net performs feature extraction and inputs the foreground image , extract features based on the convolutional layer, ,in, Indicates the encoder The output feature map of the layer, Indicates the encoder The input feature map of the layer, the initial input is the foreground image , Indicates The convolution kernel matrix of the layer, Indicates The bias term of the layer, represents the ReLU function, represents the convolution operator; The pooling layer downsamples the feature map to extract high-level features. ,in, Indicates The feature map after layer pooling, represents the pooling operation, Use transposed convolution to upscale the feature map to its original resolution, ,in, Indicates the decoder The output feature map of the layer, Indicates the decoder The input feature map of the layer, the initial input is the deep feature output by the encoder, Indicates The transposed convolution kernel matrix of the layer, used for upsampling, Indicates The bias term of the layer; Make a jump connection. ,in, Indicates The skip connection result of the layer, Represents the output feature map of the corresponding layer of the encoder; The final output of the decoder is converted into a segmentation probability map through an activation function to extract the deflection features. ,in, It represents the deflection feature map finally output by U-Net, and determines whether each pixel belongs to the deflection change area. represents the convolution kernel matrix of the output layer, Represents the bias term of the output layer; Step S3, spatiotemporal feature extraction and three-dimensional deflection field modeling, input the extracted features in step S2 into the spatiotemporal graph convolutional network ST-GCN, combine the long short-term memory network LSTM to perform joint modeling of temporal and spatial features, integrate the image input in step S1, the deflection features in step S2 and the background noise, and construct a three-dimensional deflection field; In step S3, the spatiotemporal feature extraction and three-dimensional deflection field modeling method is: According to the deflection characteristics, define the space-time diagram ,in Represents a collection of nodes. Represents a set of edges, each node represents a deflection feature point, and the edge represents the association in time and space. ,in, represents the node set, i.e., each deflection measurement point, Indicates the number of nodes, corresponding to the total number of measurement points, ,in, Represents a spatial edge set, connecting adjacent nodes in space, represents a set of temporal edges, connecting the states of the same node at adjacent times; Use graph convolution to update node features and capture spatial features. ,in, Indicates The node feature matrix of the layer, Indicates Node in layer The characteristic vector of Representation Node The set of neighbor nodes of represents the normalization coefficient, Indicates The convolution weight matrix of the layer, Indicates The bias vector of the layer, represents the ReLU function, The features of each node are input into LSTM for time series modeling. ,in, Represents LSTM at time The hidden state of the moment, Indicates at time The input features at the moment come from the output of ST-GCN. Represents the hidden state of LSTM at the previous moment, represents the weight matrix of input and hidden state, represents the bias vector, represents the tanh function; The final output time feature is mapped by the output gate of LSTM. ,in, represents the output of LSTM, represents the output weight matrix, Indicates output bias; Integrate the image input of step S1, the deflection features and background noise of step S2, and combine the results of ST-GCN and LSTM to construct a three-dimensional deflection field. ,in, represents the three-dimensional deflection field, that is, the deflection at each point in space and time, Represents a nonlinear mapping function that combines spatial and temporal features; Step S4, fusion model and parameter correction, using the attention mechanism to analyze the difference between the three-dimensional deflection field and the actual measurement data, adaptively adjust the model parameters, and fuse the adjusted deflection field with the actual measurement data; In step S4, the method of using the Attention mechanism for adaptive parameter adjustment and data fusion is: Constructing a query ,key ,value matrix, ， ， ,in, For actual measurement data, represents the query matrix, generated by the predicted three-dimensional deflection field, i.e., the features to be matched, represents the key matrix, generated from actual measurement data, i.e., the index of known features, represents the value matrix, generated from actual measurement data, Represents the weight matrix, which is used to construct the query, key and value matrices respectively; Use scaled dot product Attention to calculate the match between the predicted and measured data. ， in, is the Attention output matrix, which represents the fusion result of the predicted data and the measured data. represents the dot product of the query and the key, i.e. the matching degree, represents the scaling factor, is the dimension of the key, Indicates that the matching degree is normalized to obtain the importance weight of each feature. Calculate the difference loss based on the Attention output and the prediction results, and adjust the model parameters. ,in, represents the discrepancy loss, i.e., the deviation between the model prediction and the measured data, represents the three-dimensional deflection field of the model, Represents the fused data after adjustment by the Attention mechanism; Using the gradient descent method, the model parameters are updated according to the difference loss. ,in, represents the current parameter vector of the model, represents the updated parameter vector of the model, represents the learning rate, Represents the gradient of the loss function with respect to the parameters; The adjusted deflection field is fused with the actual measurement data to generate an updated three-dimensional deflection field. ,in, represents the fused three-dimensional deflection field, represents the fusion coefficient; Step S5, real-time monitoring and anomaly detection, real-time monitoring of the adjusted three-dimensional deflection field, using an unsupervised learning Autoencoder algorithm to detect anomalies in the monitoring data, comparing the current monitoring data with historical patterns, and identifying potential structural anomalies and abnormal changes in deflection; In step S5, when abnormal deflection or instability trend of the bridge structure is detected, an alarm is automatically triggered; The method of performing real-time monitoring and abnormality detection in step S5 is: The three-dimensional deflection field adjusted in step S4 Enter Autoencoder for real-time monitoring, ,in, It represents the real-time monitoring data input into the Autoencoder, i.e., the adjusted deflection value at each spatial and temporal position; Compress the input data into a low-dimensional feature space and use multi-layer nonlinear transformation to represent it. ,in, represents the encoded feature vector, that is, the compressed deflection feature, represents the weight matrix of the encoder, represents the bias vector of the encoder, Represents the ReLU function; Reconstruct compressed features into an approximation of the original input, detect reconstruction errors, and identify anomalies. ,in, represents the decoder reconstructed output, which approximates the input deflection data, represents the weight matrix of the decoder, mapping the feature vector back to the original space, represents the bias vector of the decoder, Represents the activation function of the decoder, which is used to output the reconstruction result; Compare the input data with the reconstruction result and calculate the reconstruction error as an anomaly detection indicator. ,in, represents the reconstruction error, i.e., the difference between the current deflection field and the historical pattern, represents the actual monitored deflection data, Represents the deflection data reconstructed by Autoencoder; Set thresholds to determine reconstruction errors, identify potential anomalies and automatically trigger alarms.

Images