A method for eliminating water mist disturbance based on deep learning for dam surface deformation monitoring

Abstract

This invention relates to the field of dam structure safety monitoring technology, and particularly to a method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning. The method includes: deploying multiple visual targets in the dam monitoring area; acquiring image sequences of the dam surface using an image acquisition device and preprocessing them; identifying and locating the visual targets, extracting the pixel coordinates of feature points, and calculating the observation deviation of each monitoring point based on the preset spatial geometric relationship between the visual targets; constructing water mist disturbance sample data based on the observation deviation, and establishing a water mist optical path distortion error model to obtain the optical path distortion spatial error field of the monitoring area; analyzing the water mist disturbance propagation characteristics based on continuous time series error data and establishing a spatiotemporal propagation model; predicting the disturbance intensity of each monitoring point by combining the optical path distortion spatial error field and the spatiotemporal propagation model, and selecting a reliable observation subset based on the disturbance intensity to recalculate the observed displacement of the dam.

Description

Technical Field

[0001] This invention relates to the field of dam structure safety monitoring technology, and in particular to a method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning. Background Technology

[0002] Dam deformation monitoring is a crucial technical means for the safe operation of water conservancy projects. Continuous monitoring of dam structural displacement changes allows for timely understanding of the dam's structural condition and provides data support for project operation and management. With the development of computer vision technology, image-based dam deformation monitoring methods are increasingly being applied. These methods typically involve deploying visual targets on the dam surface and acquiring image sequences of the dam surface using image acquisition equipment positioned outside the monitoring area. Image processing and geometric calculations are then used to extract the pixel coordinate changes of the visual targets within the images, thereby calculating the displacement information of the dam surface. Compared to traditional contact-based monitoring methods, visual monitoring methods offer advantages such as non-contact operation, a large monitoring range, and flexible deployment, thus showing promising application prospects in the field of dam structural monitoring.

[0003] Under conditions such as dam discharge, overflow, or wave impact, a large amount of water mist is usually generated in the monitoring area. The water mist environment causes changes in the air refractive index, which causes the light propagation path to bend randomly and appears as a distortion shift in the position of the visual target pixel in the image, thus affecting the stability of the visual deformation monitoring results of the dam. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a deep learning-based method for eliminating water mist disturbance in dam surface deformation monitoring. The method aims to improve the stability of dam visual deformation monitoring results by addressing the problem of random bending of the light propagation path, which manifests as distortion shift of the visual target pixel position in the image.

[0005] This invention provides the following technical solution: a method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning, comprising the following steps:

[0006] S1. Multiple visual targets are deployed in the dam monitoring area. Image sequences of the dam surface are acquired by image acquisition equipment set outside the monitoring area, and the image sequences are preprocessed.

[0007] S2. Identify and locate visual targets in the image sequence, extract the feature point pixel coordinates of each visual target, and calculate the observation deviation of each monitoring point based on the preset spatial geometric relationship between the visual targets.

[0008] S3. Construct water mist disturbance sample data based on the observation deviation of multiple visual targets, and establish an optical path distortion error model caused by water mist disturbance in the monitoring area based on the sample data, so as to obtain the spatial error field of water mist optical path distortion in the monitoring area.

[0009] S4. Based on continuous time series optical path distortion error data, analyze the propagation characteristics of water mist disturbance in the monitoring area and establish a spatiotemporal propagation model of water mist disturbance error;

[0010] S5. Calculate the observed displacement of the dam body based on the pixel coordinate changes of the visual target in different time frames, and predict the changes in optical path distortion error of each monitoring point based on the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error, and determine the disturbance intensity corresponding to each monitoring point based on the prediction results.

[0011] S6. Based on the disturbance intensity, the monitoring area is reconstructed adaptively, and visual targets with disturbance intensity less than the set threshold are selected to form a reliable observation subset. The observation displacement is then recalculated based on the reliable observation subset.

[0012] By adopting the above technical solution, it is possible to construct water mist disturbance sample data using the observation deviations of multiple visual targets, and establish a spatial error field of water mist optical path distortion in the monitoring area. This enables the modeling and characterization of the spatial distribution of optical path distortion error in the water mist environment, thereby enabling the analysis and compensation of the distortion offset of the visual target pixel position caused by the random bending of the light propagation path. This solves the problem of the difficulty in describing and processing optical path distortion error in the water mist environment.

[0013] Preferably, in S1, the step of acquiring an image sequence of the dam surface through an image acquisition device located outside the monitoring area, and preprocessing the image sequence, includes:

[0014] The dam surface is continuously imaged using image acquisition equipment to obtain raw image data of the dam surface, and the raw image sequence is constructed according to the acquisition time sequence.

[0015] The original image sequence is processed by camera calibration parameter correction, and the pixel coordinates in the image are distorted to obtain the corrected image sequence.

[0016] The corrected image sequence is subjected to image quality normalization processing, and each frame of the image is processed to unify the resolution and image scale, resulting in a preprocessed image sequence.

[0017] Preferably, in S2, the calculation of the observation deviation of each monitoring point based on the preset spatial geometric relationship between visual targets includes:

[0018] A theoretical set of spatial coordinates for the visual targets is established based on the pre-defined spatial positional relationships of the visual targets when they are deployed on the dam surface.

[0019] Based on the imaging parameters of the image acquisition device, the theoretical spatial coordinates of the visual target are converted into the corresponding theoretical pixel coordinates;

[0020] Obtain the actual pixel coordinates of each visual target in the current image, and construct a set of observed pixel coordinates of the visual targets;

[0021] The observation deviation of each monitoring point is calculated based on the coordinate difference between the theoretical pixel coordinates and the observed pixel coordinates.

[0022] Preferably, in S3, the construction of water mist perturbation sample data based on the observation biases of multiple visual targets includes:

[0023] The observation deviation of each visual target in consecutive image frames is obtained, and the observation deviation time series is formed according to the image acquisition time sequence;

[0024] Based on the spatial relationship of each visual target in the dam monitoring area, the corresponding observation deviation time series is correlated with the spatial position of the visual target;

[0025] The observation deviation time series of each visual target are combined according to the monitoring time and spatial location to construct the observation deviation data set of the monitoring area;

[0026] Water mist disturbance sample data are generated based on the observation bias data set.

[0027] Preferably, in S3, the step of establishing the optical path distortion error model caused by water mist disturbance within the monitoring area based on sample data includes:

[0028] Obtain water mist disturbance sample data and use the water mist disturbance sample data as model input data;

[0029] Feature extraction was performed on the water mist disturbance sample data to obtain characteristic data representing the changes in water mist disturbance in the monitoring area;

[0030] Feature data is input into a deep learning model for training, and the model parameters are updated according to the training process;

[0031] Based on the trained model parameters, a model of optical path distortion error caused by water mist disturbance in the monitoring area is constructed.

[0032] Preferably, in S4, the analysis of the propagation characteristics of water mist disturbance within the monitoring area based on continuous time-series optical path distortion error data includes:

[0033] Acquire optical path distortion error data for each visual target in consecutive image frames, and construct an optical path distortion error time series in chronological order;

[0034] Based on the spatial relationship of each visual target in the dam monitoring area, the time series of optical path distortion error is correlated with the spatial position of the corresponding visual target;

[0035] Time variation analysis was performed on the optical path distortion error time series to obtain the error variation relationship between different monitoring points;

[0036] The propagation characteristics of water mist disturbance within the monitoring area are determined based on the error variation relationship between each monitoring point.

[0037] Preferably, in S4, the establishment of the spatiotemporal propagation model for water mist disturbance error includes:

[0038] Acquire water mist disturbance propagation characteristic data and construct disturbance time series data corresponding to each monitoring point within the monitoring area;

[0039] Spatiotemporal characteristic data of water mist disturbance are constructed based on the spatial location relationship of each monitoring point and the corresponding disturbance time series data.

[0040] The spatiotemporal feature data of water mist disturbance are input into a deep learning time series model for training, and the model parameters are updated according to the training process.

[0041] A spatiotemporal propagation model of water mist disturbance error is established based on the trained model parameters.

[0042] Preferably, in S5, the prediction of the optical path distortion error change at each monitoring point based on the spatiotemporal propagation model of the water mist optical path distortion spatial error field and the water mist disturbance error includes:

[0043] Acquire spatial error field data of water mist optical path distortion at each spatial location within the monitoring area;

[0044] Obtain the spatial coordinates of each visual target's monitoring point, and extract the optical path distortion error data of the corresponding position based on the spatial coordinates;

[0045] The optical path distortion error data and the spatiotemporal propagation model of water mist disturbance error are input into the prediction calculation process to obtain the optical path distortion error prediction data of each monitoring point in the subsequent time series.

[0046] Based on the optical path distortion error prediction data, a sequence of optical path distortion error changes for each monitoring point is constructed.

[0047] Preferably, in S5, determining the disturbance intensity corresponding to each monitoring point based on the prediction results includes:

[0048] Obtain the optical path distortion error change sequence at each monitoring point;

[0049] Extract the predicted optical path distortion error value for each monitoring point at the current time;

[0050] Calculate the error amplitude corresponding to each monitoring point based on the predicted value of optical path distortion error;

[0051] The disturbance intensity corresponding to each monitoring point is determined based on the error amplitude.

[0052] Preferably, in S6, the adaptive observation area reconstruction of the monitoring area based on the disturbance intensity includes:

[0053] Obtain the disturbance intensity corresponding to each monitoring point;

[0054] The disturbance intensity at each monitoring point is compared with the preset disturbance threshold.

[0055] When the disturbance intensity of a monitoring point exceeds a preset disturbance threshold, the corresponding visual target will be marked as an abnormal monitoring point.

[0056] A reliable observation subset is constructed based on the remaining visual targets that are not marked as abnormal monitoring points, and the reconstructed monitoring area is determined based on the reliable observation subset.

[0057] The present invention has the following beneficial effects:

[0058] 1. In this invention, by utilizing the observation deviations of multiple visual targets to construct water mist disturbance sample data and establishing a spatial error field of water mist optical path distortion within the monitoring area, the spatial distribution model of optical path distortion error in the water mist environment is realized, solving the problem that the optical path is difficult to describe and quantify due to the random bending of the optical path caused by water mist turbulence during the visual monitoring of the dam body.

[0059] 2. In this invention, by analyzing the optical path distortion error data of continuous time series and establishing a spatiotemporal propagation model of water mist disturbance error, the modeling and prediction of the propagation law of water mist disturbance in the monitoring area are realized, solving the problem that traditional visual monitoring methods are difficult to utilize the spatiotemporal variation characteristics of water mist disturbance.

[0060] 3. In this invention, by adaptively reconstructing the observation area based on the disturbance intensity and selecting visual targets with disturbance intensity less than a set threshold to form a reliable observation subset, dynamic identification and elimination of monitoring points affected by water mist disturbance are realized, solving the problem of reduced reliability of dam deformation calculation caused by unstable observation data of some monitoring points in the water mist environment. Attached Figure Description

[0061] Figure 1 This is a flowchart of a deep learning-based method for monitoring dam surface deformation and eliminating water mist disturbances, as proposed in this invention. Detailed Implementation

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1:

[0064] In the first embodiment of the present invention, the present invention provides a method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning, such as... Figure 1 As shown, it includes the following steps:

[0065] S1. Multiple visual targets are deployed in the dam monitoring area. Image sequences of the dam surface are acquired by image acquisition equipment set outside the monitoring area, and the image sequences are preprocessed.

[0066] Furthermore, in S1, the acquisition of image sequences of the dam surface using an image acquisition device located outside the monitoring area, and the preprocessing of the image sequences, includes:

[0067] The dam surface is continuously imaged using image acquisition equipment to obtain raw image data of the dam surface, and the raw image sequence is constructed according to the acquisition time sequence.

[0068] The original image sequence is processed by camera calibration parameter correction, and the pixel coordinates in the image are distorted to obtain the corrected image sequence.

[0069] The corrected image sequence is subjected to image quality normalization processing, and each frame of the image is processed to unify the resolution and image scale, resulting in a preprocessed image sequence.

[0070] Specifically, multiple visual targets are first deployed in the dam monitoring area to form a stable visual observation benchmark. These targets are installed on the dam surface or in structurally stable areas near the dam, and their spatial positions are measured and recorded during the deployment phase as reference information for subsequent visual measurements. The visual targets can be high-contrast circular markers, checkerboard markers, or other artificial markers with distinct geometric features. In one possible approach, the visual targets are made of weather-resistant materials and fixedly installed on the dam structure surface to maintain a stable and identifiable state in the long-term monitoring environment. After the visual targets are deployed, continuous imaging of the dam surface is performed using image acquisition equipment located outside the monitoring area. This image acquisition equipment can be an industrial camera, a network camera, or other image acquisition device with stable imaging capabilities. Its installation position is typically set on a stable support structure outside the dam monitoring area, ensuring that the image field of view covers the entire visual target area. The image acquisition equipment periodically images the dam surface according to a preset sampling frequency and records multiple frames of images in chronological order to construct the original image sequence of the dam surface.

[0071] In the specific implementation process, after obtaining the original image sequence, camera calibration parameter correction processing is performed on the image sequence to establish the mapping relationship between image pixel coordinates and real imaging geometry; in one possible implementation, calibration data is obtained by pre-acquiring calibration board images, and the camera intrinsic parameter matrix is ​​solved using Zhang Zhengyou's camera calibration method; the camera intrinsic parameter matrix can be expressed as:

[0072] ;

[0073] in Represents the camera intrinsic parameter matrix; and These represent the focal length parameters of the image in the horizontal and vertical directions, respectively. and This represents the position of the principal point of the image in the pixel coordinate system; the geometric relationship between pixel coordinates and camera coordinates can be established through this intrinsic parameter matrix; during camera imaging, lens distortion causes pixel positions in the image to shift, therefore, distortion parameters are needed to correct the pixel coordinates; in one possible approach, a radial distortion model is used to correct the pixel coordinates, and its calculation form is:

[0074] ;

[0075] ;

[0076] in and Represents uncorrected normalized pixel coordinates; and Indicates the corrected pixel coordinates; This represents the distance from a pixel to the center of the image; , , This represents the radial distortion coefficient; the above calculation process can correct the geometric distortion in the original image and obtain the corrected image sequence.

[0077] In the specific implementation process, after distortion correction is completed, the corrected image sequence undergoes image quality normalization processing. Image quality normalization includes two processes: image resolution normalization and image scale normalization. In one possible approach, image resampling algorithms are used to normalize the size of each frame, for example, using bilinear interpolation for pixel reconstruction. The calculation process is as follows:

[0078] ;

[0079] in This represents the interpolated pixel grayscale value; , , , Indicates the grayscale value of adjacent pixels; , , , This represents the corresponding interpolation weight. Through the above image resampling process, each frame in the image sequence maintains a uniform resolution and image scale, thus obtaining a preprocessed image sequence. This image sequence serves as the data input basis for subsequent visual target recognition, observation deviation calculation, and dam deformation monitoring calculation.

[0080] S2. Identify and locate visual targets in the image sequence, extract the feature point pixel coordinates of each visual target, and calculate the observation deviation of each monitoring point based on the preset spatial geometric relationship between the visual targets.

[0081] Furthermore, in S2, the calculation of the observation deviation of each monitoring point based on the preset spatial geometric relationship between visual targets includes:

[0082] A theoretical set of spatial coordinates for the visual targets is established based on the pre-defined spatial positional relationships of the visual targets when they are deployed on the dam surface.

[0083] Based on the imaging parameters of the image acquisition device, the theoretical spatial coordinates of the visual target are converted into the corresponding theoretical pixel coordinates;

[0084] Obtain the actual pixel coordinates of each visual target in the current image, and construct a set of observed pixel coordinates of the visual targets;

[0085] The observation deviation of each monitoring point is calculated based on the coordinate difference between the theoretical pixel coordinates and the observed pixel coordinates.

[0086] Specifically, after obtaining the preprocessed image sequence, it is necessary to identify and locate the visual targets in the image to obtain the precise pixel position of each visual target in the image. Visual target identification can be achieved by combining image feature detection with a deep learning target detection model. In one possible approach, the preprocessed image is input into a convolutional neural network target detection model for processing. This model can be a target detection structure based on a convolutional neural network, such as a detection network containing a feature extraction layer, a region candidate generation layer, and a target classification regression layer. The convolutional neural network extracts image features layer by layer through multi-layer convolution operations, so that the visual target forms a stable feature response region in the feature map. The detection network outputs visual target candidate regions on the feature map and performs position regression to obtain the center position and bounding box information of the visual target in the image. After obtaining the boundary region of the visual target, the pixel coordinates of the feature points of the visual target are calculated by extracting sub-pixel-level feature points in the boundary region. In one possible approach, the centroid localization method can be used to calculate the pixel coordinates of the visual target region. The feature point position is obtained by calculating the center of the gray-level distribution of the region. The calculation form is as follows:

[0087] ;

[0088] in and Represents the pixel coordinates of visual target feature points; and Indicates the first The pixel coordinates of each pixel; This represents the grayscale value of the pixel. This indicates the number of pixels involved in the calculation; through the above calculation process, the feature point pixel position of each visual target in the image coordinate system can be obtained, thus forming a set of visual target pixel coordinates;

[0089] After extracting the pixel coordinates of the visual targets, it is necessary to establish a theoretical set of spatial coordinates for the visual targets based on the spatial geometric relationships recorded during the deployment phase on the dam surface. During deployment, the spatial position of each visual target in the local coordinate system of the dam body is obtained through measurement or design, for example, by obtaining the three-dimensional spatial coordinates of the visual targets using 3D measuring equipment or construction design data. In one possible approach, the spatial position of the visual targets can be represented as a three-dimensional coordinate vector.

[0090] ;

[0091] in Indicates the first The position vector of a visual target in the spatial coordinate system of the dam body; , , These represent the coordinate components of the visual target in the three directions of the spatial coordinate system. By uniformly recording all visual targets, a theoretical spatial coordinate set of visual targets can be constructed for subsequent visual geometry calculations.

[0092] After obtaining the theoretical spatial coordinate set of the visual target, it is necessary to convert the spatial coordinates into theoretical pixel coordinates by combining the imaging parameters of the image acquisition device; this process belongs to the projection calculation process from spatial coordinates to image coordinates; in one possible way, a pinhole imaging model can be used to establish the mapping relationship between spatial coordinates and pixel coordinates; the calculation form of spatial point projection onto the image plane is as follows:

[0093] ;

[0094] in and Represents the theoretical pixel coordinates of the visual target in the image; Represents the camera intrinsic parameter matrix; This represents the rotation matrix in the camera's extrinsic parameters; Represents the translation vector; This represents the three-dimensional position of the visual target in the spatial coordinate system; the theoretical pixel position of the visual target on the image plane can be calculated using this projection relationship.

[0095] After obtaining the theoretical pixel coordinates of the visual target, they are matched with the actual pixel coordinates obtained from image detection to construct the set of observed pixel coordinates of the visual target; for the th A visual target can be represented by its theoretical pixel coordinates as follows: The detected actual pixel coordinates are represented as After completing pixel coordinate matching, the observation bias is obtained by calculating the difference between the theoretical pixel coordinates and the observed pixel coordinates; the observation bias can be expressed as:

[0096] ;

[0097] in and Indicates the first The pixel deviation of each monitoring point in the horizontal and vertical directions; by performing the above calculation on all visual targets, the set of observation deviation data of each monitoring point in the monitoring area can be obtained; the observation deviation reflects the difference between the visual measurement results and the theoretical geometric relationship. This data is used as the basic input data in the subsequent construction of water mist disturbance samples and optical path distortion error modeling, thus forming a complete data processing flow.

[0098] S3. Construct water mist disturbance sample data based on the observation deviation of multiple visual targets, and establish an optical path distortion error model caused by water mist disturbance in the monitoring area based on the sample data, so as to obtain the spatial error field of water mist optical path distortion in the monitoring area.

[0099] Furthermore, in S3, the construction of water mist disturbance sample data based on the observation biases of multiple visual targets includes:

[0100] The observation deviation of each visual target in consecutive image frames is obtained, and the observation deviation time series is formed according to the image acquisition time sequence;

[0101] Based on the spatial relationship of each visual target in the dam monitoring area, the corresponding observation deviation time series is correlated with the spatial position of the visual target;

[0102] The observation deviation time series of each visual target are combined according to the monitoring time and spatial location to construct the observation deviation data set of the monitoring area;

[0103] Water mist disturbance sample data are generated based on the observation bias data set.

[0104] Furthermore, in S3, the step of establishing an optical path distortion error model based on sample data caused by water mist disturbance within the monitoring area includes:

[0105] Obtain water mist disturbance sample data and use the water mist disturbance sample data as model input data;

[0106] Feature extraction was performed on the water mist disturbance sample data to obtain characteristic data representing the changes in water mist disturbance in the monitoring area;

[0107] Feature data is input into a deep learning model for training, and the model parameters are updated according to the training process;

[0108] Based on the trained model parameters, a model of optical path distortion error caused by water mist disturbance in the monitoring area is constructed.

[0109] Specifically, after obtaining the observation bias data corresponding to each visual target, it is necessary to construct water mist disturbance sample data based on the observation bias information over continuous time. Since the visual targets can obtain corresponding pixel observation biases in consecutive image frames, the observation biases of each visual target in consecutive image frames are first collected and arranged according to the image acquisition time sequence to form an observation bias time series. For the ... The visual target, in the first The pixel observation bias obtained in the frame image can be expressed as ,in Indicates the first The visual target in the first Horizontal pixel deviation in a frame image Indicates the first The visual target in the first Vertical pixel deviation in a frame image; by recording consecutive time frames, a time series of observation deviations of the visual target can be formed:

[0110] ;

[0111] in Indicates the first A visual target in time The corresponding observation bias vector; by performing the above processing on all visual targets in the monitoring area, multiple observation bias time series can be generated, providing basic data for the construction of subsequent perturbation samples;

[0112] After generating the observation deviation time series, it is necessary to correlate the observation deviation time series with the spatial location of the visual target in the dam monitoring area. Since the spatial coordinates of the visual target have been determined during the deployment phase, each observation deviation time series can be matched with the spatial coordinates of the corresponding visual target, thus forming a data structure with spatial attributes. In one possible implementation, the spatial location of the visual target can be represented as two-dimensional planar coordinates. ,in Indicates the first The lateral position of a visual target in the coordinate system of the dam surface. This indicates the longitudinal position of the visual target in the coordinate system on the dam surface. By combining the spatial coordinates with the time series deviation, a data record with both spatial and temporal attributes can be obtained. In this data structure, each data element simultaneously contains the observation deviation between the spatial position of the visual target and the corresponding time, thereby reflecting the changes in the influence of water mist disturbance on different spatial positions within the monitoring area at different times.

[0113] After completing the spatial correlation, the observation deviation time series of each visual target are organized uniformly according to the monitoring time and spatial location, thereby constructing an observation deviation data set for the monitoring area; in one possible implementation, this data set can be represented as a three-dimensional data structure:

[0114] ;

[0115] in Indicates spatial location And the time is The data structure can simultaneously describe the deviation changes of different spatial locations within the monitoring area when disturbed at different times. The observation deviation data set is then organized and standardized to generate water mist disturbance sample data. The water mist disturbance sample data contains spatial location, time series, and observation deviation information, which can be used as input data for subsequent optical path distortion error modeling.

[0116] After obtaining water mist disturbance sample data, it is necessary to establish an optical path distortion error model caused by water mist disturbance in the monitoring area based on the sample data. First, the water mist disturbance sample data is used as the model input data, and feature extraction processing is performed on the sample data. The feature extraction process is used to extract data features that reflect the changing patterns of water mist disturbance from the original observation deviation data. In one possible implementation, feature extraction can be performed using a convolutional neural network. The convolutional neural network performs local feature extraction on the input data through convolution operations, and its convolution calculation process can be expressed as:

[0117] ;

[0118] in Indicates the first Each convolutional kernel is located at... The resulting characteristic response; Indicates the first Each convolutional kernel is located at... The weight parameters; This represents the numerical value of the input feature map at the corresponding position; and This represents the kernel size; spatial feature information can be extracted from the original sample data through the above convolution operation; after feature extraction, the obtained feature data is input into the deep learning model for training; during the training process, the deep learning model continuously adjusts the model parameters through the backpropagation algorithm, enabling the model to learn the mapping relationship between observation bias data and water mist disturbance; the backpropagation algorithm iteratively updates the model parameters by calculating the loss function, and its parameter update process can be expressed as:

[0119] ;

[0120] in Indicates the first Model parameters at the next iteration; Indicates the learning rate; This represents the gradient of the loss function with respect to the parameters; The loss function represents the model training process; stable model parameters can be obtained through continuous training; finally, an optical path distortion error model caused by water mist disturbance in the monitoring area is constructed based on the trained model parameters; this model is used to describe the optical path distortion variation law caused by water mist disturbance at different spatial locations in the monitoring area at different times, thereby forming a spatial error field of water mist optical path distortion, providing basic data for subsequent water mist disturbance propagation analysis and dam deformation monitoring calculation.

[0121] S4. Based on continuous time series optical path distortion error data, analyze the propagation characteristics of water mist disturbance in the monitoring area and establish a spatiotemporal propagation model of water mist disturbance error;

[0122] Furthermore, in S4, the analysis of the propagation characteristics of water mist disturbance within the monitoring area based on continuous time-series optical path distortion error data includes:

[0123] Acquire optical path distortion error data for each visual target in consecutive image frames, and construct an optical path distortion error time series in chronological order;

[0124] Based on the spatial relationship of each visual target in the dam monitoring area, the time series of optical path distortion error is correlated with the spatial position of the corresponding visual target;

[0125] Time variation analysis was performed on the optical path distortion error time series to obtain the error variation relationship between different monitoring points;

[0126] The propagation characteristics of water mist disturbance within the monitoring area are determined based on the error variation relationship between each monitoring point.

[0127] Furthermore, in S4, the establishment of the spatiotemporal propagation model for water mist disturbance error includes:

[0128] Acquire water mist disturbance propagation characteristic data and construct disturbance time series data corresponding to each monitoring point within the monitoring area;

[0129] Spatiotemporal characteristic data of water mist disturbance are constructed based on the spatial location relationship of each monitoring point and the corresponding disturbance time series data.

[0130] The spatiotemporal feature data of water mist disturbance are input into a deep learning time series model for training, and the model parameters are updated according to the training process.

[0131] A spatiotemporal propagation model of water mist disturbance error is established based on the trained model parameters.

[0132] Specifically, after obtaining the optical path distortion error data corresponding to each visual target within the monitoring area, the error data of each visual target in consecutive image frames are processed to form an optical path distortion error time series reflecting the change of disturbance over time; for the th A visual target, in time The corresponding optical path distortion error is denoted as ,in This indicates that the visual target is at time [time]. The corresponding error values; by arranging the error data at each moment according to the image acquisition time sequence, an error time series corresponding to the visual target can be constructed, thereby reflecting the change process of disturbance in the monitoring area in the time dimension; then the error time series is correlated with the spatial position of the visual target in the dam monitoring area; the visual target already has definite spatial coordinates during the deployment stage, so the first... The position of each visual target is represented as follows: ,in This represents the horizontal coordinate in the plane coordinate system of the monitoring area. Representing the vertical axis; by using the error time series Spatial coordinates By matching these data, a disturbance data record containing information on spatial location and temporal changes can be generated, thereby constructing an error spatiotemporal data structure for the monitoring area.

[0133] After forming the spatiotemporal data structure of the error, time variation analysis is performed on the error time series between different monitoring points to identify the propagation relationship of water mist disturbance in the monitoring area; in one possible implementation, the direction of disturbance propagation is determined by calculating the time-delay correlation between the error time series of different monitoring points; for the th The monitoring point and the first With multiple monitoring points, the correlation between the two under different time delay conditions can be calculated, and the calculation method is as follows:

[0134] ;

[0135] in Indicates monitoring point With monitoring points Time delay Relevant values ​​under the given conditions; Indicates the amount of time delay; Indicates the length of the time series; Indicates monitoring point In time The error value; Indicates monitoring point In time The error value; by analyzing the changes in the correlation values ​​between different monitoring points, the propagation relationship of error changes at different spatial locations can be determined, thereby obtaining the propagation characteristics of water mist disturbance in the monitoring area; this propagation characteristic reflects the spatial propagation direction of the disturbance and the propagation time delay relationship;

[0136] After obtaining the propagation characteristics of water mist disturbance, it is necessary to construct the time-series data of the disturbance for model training; for each monitoring point in the monitoring area, continuous data before the current time are selected. Error data from each sampling time point form a time window, thereby constructing perturbation time series data; Each monitoring point at time The corresponding disturbance time series data can be represented as:

[0137] ;

[0138] in Indicates monitoring point In time The corresponding disturbance time series data; Indicates the length of the time window; by mapping the disturbance time series data to the spatial location of the corresponding monitoring points. By combining these elements, spatiotemporal characteristic data of water mist disturbance can be generated, which simultaneously contains both temporal and spatial information.

[0139] After obtaining the spatiotemporal characteristic data of water mist disturbance, this data is input into a deep learning time series model for training, thereby establishing a spatiotemporal propagation model of water mist disturbance error. The deep learning time series model can be a long short-term memory network, a gated recurrent neural network, or other neural network structures capable of processing time series data. During the training process, the spatiotemporal characteristic data of water mist disturbance is used as the model input, and the model parameters are iteratively updated through the backpropagation algorithm, enabling the model to learn the propagation law of error in time and space. After the model training is completed, a spatiotemporal propagation model that can describe the propagation relationship of water mist disturbance error within the monitoring area is obtained. This model is used to characterize the propagation relationship of error changes over time between different monitoring points, thus providing a basic model for subsequent optical path distortion error prediction.

[0140] S5. Calculate the observed displacement of the dam body based on the pixel coordinate changes of the visual target in different time frames, and predict the changes in optical path distortion error of each monitoring point based on the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error, and determine the disturbance intensity corresponding to each monitoring point based on the prediction results.

[0141] Furthermore, in S5, the prediction of the optical path distortion error change at each monitoring point based on the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error includes:

[0142] Acquire spatial error field data of water mist optical path distortion at each spatial location within the monitoring area;

[0143] Obtain the spatial coordinates of each visual target's monitoring point, and extract the optical path distortion error data of the corresponding position based on the spatial coordinates;

[0144] The optical path distortion error data and the spatiotemporal propagation model of water mist disturbance error are input into the prediction calculation process to obtain the optical path distortion error prediction data of each monitoring point in the subsequent time series.

[0145] Based on the optical path distortion error prediction data, a sequence of optical path distortion error changes for each monitoring point is constructed.

[0146] Furthermore, in S5, determining the disturbance intensity corresponding to each monitoring point based on the prediction results includes:

[0147] Obtain the optical path distortion error change sequence at each monitoring point;

[0148] Extract the predicted optical path distortion error value for each monitoring point at the current time;

[0149] Calculate the error amplitude corresponding to each monitoring point based on the predicted value of optical path distortion error;

[0150] The disturbance intensity corresponding to each monitoring point is determined based on the error amplitude.

[0151] Specifically, after constructing the spatial error field of water mist optical path distortion and establishing the spatiotemporal propagation model of water mist disturbance error, it is necessary to jointly calculate the observed displacement of the dam structure and the disturbance state within the monitoring area. First, the observed displacement of the dam is calculated based on the pixel coordinate changes of the visual target in different time frames. The pixel coordinates of the visual target in consecutive image frames can reflect the change in the position of the dam surface structure over time; therefore, the corresponding pixel displacement can be obtained by comparing the pixel coordinates of the same visual target in different time frames. For the first... A visual target, in time and time The corresponding pixel coordinates are denoted as follows: and Then the pixel displacement of the visual target in the image coordinate system can be expressed as:

[0152] ;

[0153] ;

[0154] in This represents the pixel displacement of the visual target in the horizontal direction. This represents the pixel displacement of the visual target in the vertical direction. By combining the camera imaging geometry, the pixel displacement can be converted into the observed displacement of the corresponding monitoring point on the dam surface, thereby obtaining the dam observation displacement data of each monitoring point at different times.

[0155] After completing the displacement calculation of the dam body, it is necessary to predict the changes in optical path distortion error corresponding to each monitoring point in the monitoring area. First, obtain the spatial error field data of water mist optical path distortion at each spatial location within the monitoring area. This error field describes the degree of optical path distortion at different spatial locations within the monitoring area when disturbed by water mist at the current moment. Then, obtain the spatial coordinates of each monitoring point where the visual target is located, and extract the corresponding optical path distortion error data from the error field based on the spatial coordinates. For example, for the first... A visual target, the spatial location of which can be represented as: Extract the error value at the corresponding position from the spatial error field of water mist optical path distortion. The error value reflects the optical path disturbance state experienced by the monitoring point at the current moment. After obtaining the current error data, the optical path distortion error data and the spatiotemporal propagation model of water mist disturbance error are input into the prediction calculation process. The spatiotemporal propagation model infers the error changes in subsequent time series based on the current error state and historical error change trends, thereby obtaining the optical path distortion error prediction data of each monitoring point at future moments. By arranging the error data of consecutive prediction moments, a sequence of optical path distortion error changes corresponding to each monitoring point can be constructed. This sequence reflects the possible disturbance changes experienced by the monitoring point in the subsequent time range.

[0156] After obtaining the optical path distortion error change sequence at each monitoring point, it is necessary to calculate the disturbance intensity corresponding to each monitoring point based on the prediction results. First, the predicted optical path distortion error value corresponding to each monitoring point at the current time is extracted, and this predicted value is used as the basic data for calculating the disturbance intensity. When the error is represented in two-dimensional component form, the error vector can be represented as... ,in Indicates the first Predicted optical path distortion error values ​​in the horizontal direction for each monitoring point This represents the predicted value of optical path distortion error in the vertical direction. To obtain the degree of perturbation in a single scalar form, the error magnitude can be obtained by calculating the magnitude of the error vector, which is calculated as follows:

[0157] ;

[0158] in Indicates the first The amplitude of optical path distortion error at each monitoring point and These represent the predicted error values ​​in two directions, respectively. By calculating the error amplitude, the degree to which the monitoring point is affected by water mist disturbance at the current moment can be quantified. Subsequently, the disturbance intensity corresponding to each monitoring point is determined based on the error amplitude, thereby providing a basis for the selection of reliable observation subsets of the subsequent monitoring area and the reconstruction process of the observation area. The calculation results, together with the dam body observation displacement data, participate in the subsequent monitoring data processing flow to achieve stable acquisition of dam body deformation monitoring data under water mist disturbance conditions.

[0159] S6. Based on the disturbance intensity, the monitoring area is reconstructed adaptively, and visual targets with disturbance intensity less than the set threshold are selected to form a reliable observation subset. The observation displacement is then recalculated based on the reliable observation subset.

[0160] Furthermore, in S6, the adaptive observation area reconstruction of the monitoring area based on the disturbance intensity includes:

[0161] Obtain the disturbance intensity corresponding to each monitoring point;

[0162] The disturbance intensity at each monitoring point is compared with the preset disturbance threshold.

[0163] When the disturbance intensity of a monitoring point exceeds a preset disturbance threshold, the corresponding visual target will be marked as an abnormal monitoring point.

[0164] A reliable observation subset is constructed based on the remaining visual targets that are not marked as abnormal monitoring points, and the reconstructed monitoring area is determined based on the reliable observation subset.

[0165] Specifically, after obtaining the disturbance intensity corresponding to each monitoring point, the monitoring area is adaptively reconstructed. The disturbance intensity originates from the aforementioned optical path distortion error prediction results; therefore, each monitoring point corresponds to an intensity value that reflects the degree of water mist disturbance at the current moment. To distinguish between effective monitoring points that can participate in displacement calculation and abnormal monitoring points that are significantly affected by water mist disturbance, the disturbance intensity of each monitoring point is first compared with a preset disturbance threshold. The preset disturbance threshold can be determined based on the monitoring system calibration results, historical monitoring data statistics, or rules set during the operation phase. In one possible approach, the preset disturbance threshold is a fixed threshold; in another possible approach, the preset disturbance threshold can be adaptively updated based on the disturbance intensity distribution of all monitoring points within the current monitoring period. Through this comparison process, the monitoring points corresponding to each visual target within the monitoring area can be effectively classified.

[0166] When the disturbance intensity of a monitoring point exceeds a preset disturbance threshold, the visual target corresponding to that monitoring point is marked as an abnormal monitoring point. An abnormal monitoring point indicates that the water mist disturbance at that location at the current moment has exceeded the preset discrimination range, and therefore it will no longer directly participate in the recalculation of the observed displacement at the current moment. When the disturbance intensity of a monitoring point is not greater than the preset disturbance threshold, the visual target corresponding to that monitoring point is retained as a valid observation point. By performing the same discrimination process on all monitoring points, a set of visual targets that have not been marked as abnormal monitoring points can be selected from the initial set of visual targets. This set constitutes a reliable observation subset. The visual targets in the reliable observation subset all meet the preset disturbance constraint conditions at the current moment, so their corresponding pixel displacement data can be used as input data for subsequent recalculation of observed displacement.

[0167] After forming a reliable observation subset, the reconstructed monitoring area is determined based on the reliable observation subset. The reconstructed monitoring area is not a physical adjustment of the original monitoring area, but rather a redefinition of the observation area range for calculation based on the spatial distribution of the current effective visual targets. In one possible approach, the set of spatial location points corresponding to all visual targets in the reliable observation subset is used as the discrete sampling points of the current effective monitoring area. In another possible approach, the outline of the reconstructed monitoring area at the current moment can be determined based on the spatial boundary relationships of each visual target in the reliable observation subset. Since the distribution of water mist disturbance has dynamic characteristics at different times, the reliable observation subset obtained through the above screening process and the corresponding reconstructed monitoring area also change dynamically with time, thus giving the monitoring area reconstruction process an adaptive characteristic.

[0168] After determining the reconstructed monitoring area, the observed displacement is recalculated based on the reliable observation subset; for the first reliable observation subset... A visual target, whose pixel coordinates at the reference time are: The pixel coordinates at the current moment are: The pixel displacement vector corresponding to the visual target can then be expressed as:

[0169] ;

[0170] in, Indicates the first The pixel displacement vector corresponding to each visual target; and These represent the horizontal and vertical pixel coordinates at the reference time, respectively. and These represent the horizontal and vertical pixel coordinates at the current moment, respectively. After obtaining the pixel displacement vectors of all visual targets in the reliable observation subset, the observed displacements can be comprehensively calculated based on the spatial distribution relationship of each visual target. In one possible approach, the displacement vectors of all valid visual targets in the reliable observation subset are fused using a mean calculation method, and the calculation form is as follows:

[0171] ;

[0172] in, This represents the comprehensive observation displacement vector corresponding to the reconstructed monitoring area; This indicates the number of visual targets in the reliable observation subset; Indicates the first The pixel displacement vector corresponding to each visual target; when it is necessary to combine the target spatial position for weighted calculation, a weighted average method can also be used to fuse the displacement; by recalculating the observed displacement based on a reliable subset of observations, the observed displacement result at the current moment is established on the basis of effective monitoring points that meet the disturbance discrimination conditions, thereby completing the adaptive observation area reconstruction and displacement recalculation process of the monitoring area.

[0173] Example 2:

[0174] Under flood discharge or overflow operation conditions, a large amount of high-speed water impacts the dam surface, forming a continuously spreading water mist environment. Water mist particles create a turbulent disturbance field with spatially uneven distribution in the air. Under these conditions, the air refractive index fluctuates with the water mist concentration, causing random bending of light during propagation. This results in unstable shifts in the pixel positions of visual targets in the image. Traditional dam visual monitoring methods typically assume relatively stable light propagation. When water mist disturbance exists in the monitoring area, the pixel positions of visual targets will exhibit additional deviations, causing fluctuations in the calculated dam displacement results and affecting the stability of the dam visual deformation monitoring results. To address these issues, this invention provides a deep learning-based method for eliminating water mist disturbance in dam surface deformation monitoring, the structure of which is as follows... Figure 1 As shown. The specific implementation process of this method is as follows:

[0175] Multiple visual targets were deployed in the dam monitoring area, and image acquisition equipment was installed outside the monitoring area to continuously image the dam surface. The image acquisition equipment periodically took pictures of the dam monitoring area and obtained continuous image sequences. At the same time, the acquired image sequences were preprocessed, including camera calibration parameter correction and image scale normalization, to obtain image data of uniform scale.

[0176] Visual targets in the image sequence are identified and located, and the feature point pixel coordinates of each visual target are extracted. The observation deviation of each monitoring point is calculated in combination with the spatial geometric relationship determined during the deployment of the visual targets on the dam surface. By sorting out the observation deviations in continuous time frames, the observation deviation data of each visual target at different times can be obtained.

[0177] After obtaining the observation deviation data, water mist disturbance sample data is constructed based on the observation deviations of multiple visual targets. Based on the sample data, an optical path distortion error model caused by water mist disturbance in the monitoring area is established, thereby obtaining the spatial error field of water mist optical path distortion in the monitoring area. This error field is used to describe the distribution of optical path distortion error that may occur at different spatial locations in the monitoring area under water mist disturbance conditions.

[0178] After obtaining the spatial error field of optical path distortion, the propagation characteristics of water mist disturbance in the monitoring area are analyzed based on continuous time series optical path distortion error data, and a spatiotemporal propagation model of water mist disturbance error is established. This model is used to describe the propagation law of water mist disturbance in the monitoring area with time and space variations.

[0179] The observed displacement of the dam body is calculated based on the pixel coordinate changes of the visual target in different time frames. Combined with the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error, the changes of optical path distortion error at each monitoring point are predicted, and the disturbance intensity corresponding to each monitoring point is calculated based on the prediction results.

[0180] Finally, the monitoring area is reconstructed adaptively based on the disturbance intensity. Visual targets with disturbance intensity greater than a set threshold are marked, and visual targets with disturbance intensity less than a set threshold are selected to form a reliable observation subset. The dam body observation displacement is recalculated based on the reliable observation subset, thus completing the dam body surface deformation monitoring process under water mist disturbance conditions.

[0181] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based dam surface deformation monitoring water mist disturbance elimination method, characterized in that, Includes the following steps: S1. Multiple visual targets are deployed in the dam monitoring area. Image sequences of the dam surface are acquired by image acquisition equipment set outside the monitoring area, and the image sequences are preprocessed. S2. Identify and locate visual targets in the image sequence, extract the feature point pixel coordinates of each visual target, and calculate the observation deviation of each monitoring point based on the preset spatial geometric relationship between the visual targets. S3. Construct water mist disturbance sample data based on the observation deviation of multiple visual targets, and establish an optical path distortion error model caused by water mist disturbance in the monitoring area based on the sample data, so as to obtain the spatial error field of water mist optical path distortion in the monitoring area. S4. Based on continuous time series optical path distortion error data, analyze the propagation characteristics of water mist disturbance in the monitoring area and establish a spatiotemporal propagation model of water mist disturbance error; S5. Calculate the observed displacement of the dam body based on the pixel coordinate changes of the visual target in different time frames, and predict the changes in optical path distortion error of each monitoring point based on the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error, and determine the disturbance intensity corresponding to each monitoring point based on the prediction results. S6. Based on the disturbance intensity, the monitoring area is reconstructed adaptively, and visual targets with disturbance intensity less than the set threshold are selected to form a reliable observation subset. The observation displacement is then recalculated based on the reliable observation subset.

2. The method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S1, the acquisition of image sequences of the dam surface using image acquisition devices located outside the monitoring area, and the preprocessing of the image sequences, includes: The dam surface is continuously imaged using image acquisition equipment to obtain raw image data of the dam surface, and the raw image sequence is constructed according to the acquisition time sequence. The original image sequence is processed by camera calibration parameter correction, and the pixel coordinates in the image are distorted to obtain the corrected image sequence. The corrected image sequence is subjected to image quality normalization processing, and each frame of the image is processed to unify the resolution and image scale, resulting in a preprocessed image sequence.

3. The method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S2, the calculation of the observation deviation of each monitoring point based on the preset spatial geometric relationship between visual targets includes: A theoretical set of spatial coordinates for the visual targets is established based on the pre-defined spatial positional relationships of the visual targets when they are deployed on the dam surface. Based on the imaging parameters of the image acquisition device, the theoretical spatial coordinates of the visual target are converted into the corresponding theoretical pixel coordinates; Obtain the actual pixel coordinates of each visual target in the current image, and construct a set of observed pixel coordinates of the visual targets; The observation deviation of each monitoring point is calculated based on the coordinate difference between the theoretical pixel coordinates and the observed pixel coordinates.

4. The method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S3, the construction of water mist disturbance sample data based on the observation bias of multiple visual targets includes: The observation deviation of each visual target in consecutive image frames is obtained, and the observation deviation time series is formed according to the image acquisition time sequence; Based on the spatial relationship of each visual target in the dam monitoring area, the corresponding observation deviation time series is correlated with the spatial position of the visual target; The observation deviation time series of each visual target are combined according to the monitoring time and spatial location to construct the observation deviation data set of the monitoring area; Water mist disturbance sample data are generated based on the observation bias data set.

5. The method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S3, the establishment of the optical path distortion error model caused by water mist disturbance in the monitoring area based on sample data includes: Obtain water mist disturbance sample data and use the water mist disturbance sample data as model input data; Feature extraction was performed on the water mist disturbance sample data to obtain characteristic data representing the changes in water mist disturbance in the monitoring area; Feature data is input into a deep learning model for training, and the model parameters are updated according to the training process; Based on the trained model parameters, a model of optical path distortion error caused by water mist disturbance in the monitoring area is constructed.

6. The method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S4, the analysis of the propagation characteristics of water mist disturbance within the monitoring area based on continuous time-series optical path distortion error data includes: Acquire optical path distortion error data for each visual target in consecutive image frames, and construct an optical path distortion error time series in chronological order; Based on the spatial relationship of each visual target in the dam monitoring area, the time series of optical path distortion error is correlated with the spatial position of the corresponding visual target; Time variation analysis was performed on the optical path distortion error time series to obtain the error variation relationship between different monitoring points; The propagation characteristics of water mist disturbance within the monitoring area are determined based on the error variation relationship between each monitoring point.

7. The method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S4, the establishment of the spatiotemporal propagation model for water mist disturbance error includes: Acquire water mist disturbance propagation characteristic data and construct disturbance time series data corresponding to each monitoring point within the monitoring area; Spatiotemporal characteristic data of water mist disturbance are constructed based on the spatial location relationship of each monitoring point and the corresponding disturbance time series data. The spatiotemporal feature data of water mist disturbance are input into a deep learning time series model for training, and the model parameters are updated according to the training process. A spatiotemporal propagation model of water mist disturbance error is established based on the trained model parameters.

8. The method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S5, the prediction of the optical path distortion error change at each monitoring point based on the spatial error field of water mist optical path distortion and the spatiotemporal propagation model of water mist disturbance error includes: Acquire spatial error field data of water mist optical path distortion at each spatial location within the monitoring area; Obtain the spatial coordinates of each visual target's monitoring point, and extract the optical path distortion error data of the corresponding position based on the spatial coordinates; The optical path distortion error data and the spatiotemporal propagation model of water mist disturbance error are input into the prediction calculation process to obtain the optical path distortion error prediction data of each monitoring point in the subsequent time series. Based on the optical path distortion error prediction data, a sequence of optical path distortion error changes for each monitoring point is constructed.

9. The method for eliminating water mist disturbance during dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S5, determining the disturbance intensity corresponding to each monitoring point based on the prediction results includes: Obtain the optical path distortion error change sequence at each monitoring point; Extract the predicted optical path distortion error value for each monitoring point at the current time; Calculate the error amplitude corresponding to each monitoring point based on the predicted value of optical path distortion error; The disturbance intensity corresponding to each monitoring point is determined based on the error amplitude.

10. The method for eliminating water mist disturbance in dam surface deformation monitoring based on deep learning according to claim 1, characterized in that, In S6, the adaptive observation area reconstruction based on the disturbance intensity includes: Obtain the disturbance intensity corresponding to each monitoring point; The disturbance intensity at each monitoring point is compared with the preset disturbance threshold. When the disturbance intensity of a monitoring point exceeds a preset disturbance threshold, the corresponding visual target will be marked as an abnormal monitoring point. A reliable observation subset is constructed based on the remaining visual targets that are not marked as abnormal monitoring points, and the reconstructed monitoring area is determined based on the reliable observation subset.

Images