A method for seepage prevention monitoring of a cofferdam for construction of a pile cap in a tidal area

Abstract

The present invention provides a method for anti-seepage monitoring of a cofferdam for a pile cap construction in a tidal area. This method realizes the real-time monitoring and early warning of the cofferdam state. First, the acquisition module obtains the high-definition image data of the cofferdam surface, the water pressure of the cofferdam, and the tidal data, and ensures the synchronization of timestamps; then, computer vision and deep learning algorithms are used to identify cracks, attitude deviations, and deformation conditions in the images; a deep learning model is used to combine the tidal data and the water pressure data to predict the future water pressure changes; when the detection result exceeds the preset threshold, the early warning mechanism is immediately triggered. In addition, all data is stored in the database, and the model parameters are updated regularly to improve the accuracy. The present invention integrates multi-source data, and uses computer vision technology and deep learning technology to detect the possible factors of cofferdam seepage, significantly improving the accuracy and dynamic prediction ability of cofferdam anti-seepage monitoring.

Description

Technical Field

[0001] The invention relates to the fields of water conservancy engineering and marine engineering, and specifically belongs to the technical field of cofferdam construction monitoring and anti-seepage management. Background Art

[0002] Tidal zone cofferdams are common temporary structures in water conservancy projects and marine engineering projects. Their main function is to provide a dry working environment for construction in the tidal impact area. Due to the special geographical conditions of the tidal zone, the cofferdam not only needs to withstand the external water pressure, but also needs to cope with the periodic changes in tidal height, wave height and current speed, which poses great challenges to the stability and impermeability of the cofferdam.

[0003] In practical applications, the anti-seepage performance of the cofferdam is crucial to the smooth progress of the project and the safety of construction. If the cofferdam leaks, it may cause water to invade the internal construction area, or even cause the cofferdam to collapse, seriously threatening the safety of construction workers and increasing project costs. Therefore, how to conduct real-time and accurate anti-seepage monitoring and status assessment of the cofferdam is an important technical problem that needs to be solved in the current engineering field.

[0004] Traditional cofferdam monitoring methods rely on manual inspection or simple single sensor monitoring. Manual inspection is not only time-consuming and laborious, but also has subjective judgment errors, and a single sensor cannot fully reflect the overall status of the cofferdam. With the rapid development of computer vision and deep learning technology, new solutions have been provided for cofferdam anti-seepage monitoring. Summary of the invention

[0005] The present invention proposes a method for monitoring the anti-seepage of cofferdams in tidal area foundation construction. The method can not only monitor abnormal conditions on the surface of the cofferdam through image recognition technology, but also detect pressure changes outside the cofferdam using water pressure sensors and predict future water pressure changes through artificial intelligence algorithms, thereby more comprehensively evaluating the anti-seepage status of the cofferdam.

[0006] To this end, the technical solution adopted in the present invention is as follows:

[0007] S1. Collect high-definition image data of the cofferdam surface, collect the time series of the cofferdam water pressure, and collect tidal data, including tidal height, wave height and current speed. The timestamps of tidal data and cofferdam water pressure data are synchronized;

[0008] S2. Use computer vision and deep learning algorithms to identify the collected high-definition image data and determine whether there are factors that cause leakage in the cofferdam, including crack detection, posture deviation detection, and deformation detection;

[0009] Detect the length and width of the cofferdam cracks, the rotation angle and translation of the cofferdam posture deviation, and whether the cofferdam is deformed;

[0010] S3. Build a deep learning model, and use the collected tidal data and cofferdam water pressure data to predict the cofferdam water pressure for a period of time in the future;

[0011] S4. Set the danger thresholds for crack length and width, rotation angle and translation amount. When the crack detection and attitude deviation detection results exceed the set thresholds or the deformation detection result is convex and concave, immediately trigger the warning mechanism;

[0012] Set the danger threshold of cofferdam water pressure and immediately trigger the warning mechanism;

[0013] S5. Establish a cofferdam status database, and store the high-definition image data of the cofferdam and the cofferdam status detection results in the database.

[0014] Establish a cofferdam water pressure database, store the water pressure sensor data, local tidal data and future cofferdam water pressure prediction results in the database, and regularly update the deep learning model parameters in S3 using the latest water pressure data and tidal data.

[0015] Furthermore, before identifying the high-definition image data, the image data is also preprocessed by grayscale conversion, smoothing filtering and histogram equalization. Specifically, the grayscale conversion is as follows:

[0016] The grayscale conversion is to calculate the average value of the three color channels (RGB) in the original image to obtain a grayscale value. The formula is:

[0017]

[0018] Among them, Gray(i,j) represents the pixel grayscale value of the grayscale image at the i-th row and j-th column, R(i,j) represents the R value of the original high-definition image at the i-th row and j-th column, G(i,j) represents the G value of the original high-definition image at the i-th row and j-th column, and B(i,j) represents the B value of the original high-definition image at the i-th row and j-th column;

[0019] Furthermore, the smoothing filtering uses the Gaussian function as the weight to calculate the weighted average value of the pixel and its neighboring pixels. The shape of the Gaussian function is controlled by its standard deviation σ. For a given pixel position (i,j), its smoothed grayscale value SF_Gray(i,j) is calculated by the following formula:

[0020] SF_Gray(i,j) = ∑ k,l∈W G σ (k,l)·Gray(i + k,j + l)

[0021] Among them, W is the neighborhood window centered on (i,j), the window size is k*l, and G σ (k,l) is the value of the Gaussian function, defined as:

[0022]

[0023] Further, the histogram equalization enhances the contrast of the image by adjusting the histogram distribution of the image. Its mapping function uses the cumulative distribution function, and the formula is:

[0024]

[0025] where L is the total number of gray levels, M and N are the length and width of the image respectively, and HE_Gray(i, j) is the gray value of the (i, j)-th pixel after histogram equalization.

[0026] Further, the specific steps of the crack detection are as follows.

[0027] 1) For each pixel gray value HE_Gray(i, j) in the preprocessed image HE_Gray, convolve it with G i and G j respectively to obtain the horizontal direction gradient approximation value I i and the vertical direction gradient approximation value I j . The formulas are as follows.

[0028] I i = HE_Gray(i, j) * G i

[0029] I j = HE_Gray(i, j) * G j

[0030] where * represents the two-dimensional convolution operation, and G i and G j are Sobel operators.

[0031] 2) Calculate the gradient magnitude and gradient direction of each pixel. The gradient magnitude M(i, j) represents the intensity of the edge, and the gradient direction θ(i, j) represents the direction of the edge. The formulas are as follows.

[0032]

[0033] 3) Perform non-maximum suppression on the gradient magnitude image M(i, j). Specifically, quantize θ(i, j) to one of the four main directions: horizontal, vertical, diagonal, and anti-diagonal. According to the quantized direction, find the adjacent pixel closest to the gradient direction. If the pixel (i, j) is greater than the gradient magnitude of the adjacent pixel, retain the gradient magnitude of this pixel; otherwise, set it to 0.

[0034] 4) Threshold the gradient magnitude image M(i,j), select a fixed threshold T, consider all pixels with gradient magnitude greater than T as edge pixels, and other pixels as non-edge pixels;

[0035] 5) Use dilation operation to connect the broken crack edges to form a complete crack contour. Specifically, select a rectangular structuring element, align the origin of the structuring element with each pixel in HE_Gray, and determine whether the area covered by the structuring element intersects with the edge pixels in the image. If there is an intersection, set the gray value of this pixel to the average value of all covered edge pixels; if there is no intersection, keep the original gray value of this pixel;

[0036] The dilated image is denoted as IF_Gray;

[0037] 6) Select a fixed threshold T b , convert IF_Gray to a binary image, that is, if IF_Gray(i,j)>T b, then set the pixel value of pixel (i,j) in the binary image to 255, otherwise set it to 0;

[0038] Use a skeleton extraction algorithm to reduce the lines of the binary image to single-pixel width, calculate the pixel length of the crack skeleton, and then convert the pixel length to the actual physical size according to the resolution of the image and the scale at the time of shooting to obtain the crack length;

[0039] Use the orthogonal skeleton line method to calculate the pixel width of the crack in the binary image, and then convert the pixel distance to the actual physical size according to the resolution of the image and the scale at the time of shooting to obtain the crack width.

[0040] Furthermore, the attitude deviation detection is specifically as follows: use a CNN model to identify the HE_Gray image after data preprocessing. This model consists of 3 convolutional layers, 1 pooling layer, 1 fully connected layer, and 1 output layer in sequence. The number of neurons in the output layer is 6, representing the rotation angles around the x, y, and z axes and the translation amounts in the x, y, and z directions of the cofferdam in three-dimensional space respectively.

[0041] Furthermore, the deformation detection is specifically as follows: use a CNN model to identify the HE_Gray image after data preprocessing. The model consists of 3 convolutional layers, 1 pooling layer, 1 fully connected layer, and 1 classification layer in sequence. The number of neurons in the classification layer is 3, representing normal, convex, and concave of the cofferdam respectively.

[0042] Furthermore, the main body of the deep learning model in step S3 is a Bi-GRU structure. Bi-GRU includes a forward GRU structure and a backward GRU structure. The forward GRU structure can be expressed as

[0043] Forward update gate:

[0044]

[0045] Among them, σ is the sigmoid activation function, W z is the weight matrix of the update gate, h t-1 is the hidden state at the previous time step, x t is the tidal data at the current time step, b z is the bias term,

[0046] Forward reset gate:

[0047]

[0048] Among them, W r is the weight matrix of the reset gate, b r is the bias term,

[0049] Forward new candidate state:

[0050]

[0051] Among them, W h is the weight matrix of the new candidate state, b h is the bias term, and * represents element-wise multiplication.

[0052] Forward update hidden state:

[0053]

[0054] Among them, is the hidden state at the current time step,

[0055] The backward GRU structure is exactly the same as the forward one. The final Bi-GRU output is,

[0056]

[0057] Among them, is the hidden state of the forward GRU, is the hidden state of the backward GRU,

[0058] A fully connected layer is connected after the Bi-GRU structure.

[0059] Furthermore, the input to the deep learning model is the time series of tidal height, wave height, sea current velocity, and cofferdam water pressure within a fixed time window T s and the output is the time series of water pressure within a fixed time window T s with a lag time τ,

[0060] To find the optimal time window, different time window sizes T sPerform verification. To find the optimal lag time τ, different lag times are selected for verification.

[0061] Compare the mean squared error (MSE) of the deep learning model under different time windows and different lag times, determine the optimal time window and lag time, and input the time series of real-time tidal data and cofferdam water pressure data into the model to predict the cofferdam water pressure data τ time in the future.

[0062] Compared with the prior art, the advantages of the present invention are as follows:

[0063] 1. The present invention improves the accuracy of cofferdam seepage prevention monitoring. By combining high-definition images and water pressure sensor data, it not only monitors the changes in the external structure of the cofferdam but also can reflect the water pressure changes in the underwater part in real time, comprehensively covering the key areas of cofferdam seepage prevention monitoring. The present invention uses computer vision technology and deep learning technology to process and identify images, and can efficiently and accurately detect cofferdam cracks, deformations, and attitude deviations, ensuring early detection of seepage prevention problems.

[0064] 2. The present invention introduces a Bi-GRU structure model, utilizes the forward and backward features of time series to comprehensively capture the dynamic change laws of tides and water pressure, significantly improves the accuracy of water pressure prediction, and regularly updates the parameters of the deep learning model based on the latest data to ensure that the prediction results always maintain high efficiency and accuracy.

[0065] 3. When the detection result exceeds the set threshold, the present invention can automatically trigger an alarm and notify relevant personnel to take repair or reinforcement measures in a timely manner, reducing manual intervention and saving maintenance costs. The present invention constructs a comprehensive cofferdam state and water pressure database to provide data support for subsequent maintenance and management, facilitating trend analysis of cofferdam states and optimized maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0067] Figure 1 is the framework flowchart of the present invention;

[0068] Figure 2 is the cofferdam crack detection flowchart of the present invention;

[0069] Figure 3 is the schematic diagram of the generation of the data set for cofferdam water pressure prediction of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0070] To achieve the above object, the present invention is realized by the following technical solutions. The present invention provides a method for monitoring the anti-seepage of a cofferdam for a pile cap construction in a tidal area, in combination with Figure 1 , the method includes:

[0071] S1. Data acquisition

[0072] First, use a number of high-resolution cameras to regularly take full-range photos of the cofferdam to obtain high-definition image data of the surface of the cofferdam, and the image resolution is 1920*1080.

[0073] Second, install water pressure sensors on the part of the cofferdam below the water surface to monitor the change of the water pressure outside the cofferdam in real time. The water pressure sensors are horizontally distributed on the surface of the cofferdam at fixed intervals to capture the pressure information in different areas. At each time step t, collect the water pressure data of the cofferdam; in addition, obtain the local tidal data from the China Meteorological Network, including tidal height, wave height and sea current speed. The time step of the tidal data is also t. The timestamps of the tidal data and the water pressure data of the cofferdam are synchronized.

[0074] S2. Cofferdam state detection

[0075] Use computer vision and deep learning algorithms to identify the collected high-definition image data to determine whether there are factors causing leakage in the cofferdam, including cracks, attitude deviation and deformation.

[0076] S21. Preprocess the high-definition image data through three methods: grayscale conversion, smoothing filtering and histogram equalization.

[0077] Grayscale conversion is to convert a color image into a grayscale image to reduce the amount of calculation and highlight the edge information. Specifically, the average value of the three color channels (RGB) in the original image is calculated to obtain a grayscale value. The formula is:

[0078]

[0079] Among them, Gray(i,j) represents the pixel grayscale value of the grayscale image at the i-th row and j-th column, R(i,j) represents the R value of the original high-definition image at the i-th row and j-th column, G(i,j) represents the G value of the original high-definition image at the i-th row and j-th column, and B(i,j) represents the B value of the original high-definition image at the i-th row and j-th column.

[0080] Smoothing filtering is to remove the random noise in the image and improve the accuracy of subsequent analysis. Specifically, the Gaussian function is used as the weight to calculate the weighted average value of the pixel and its neighboring pixels. The shape of the Gaussian function is controlled by its standard deviation σ. For a given pixel position (i,j), its smoothed grayscale value SF_Gray(i,j) is calculated by the following formula:

[0081] SF_Gray(i,j) = ∑ k,l∈W G σ (k,l)·Gray(i + k,j + l)

[0082] where W is the neighborhood window centered at (i,j), the window size is k*l, and G σ (k,l) is the value of the Gaussian function, defined as:

[0083]

[0084] Histogram equalization enhances the contrast of an image by adjusting the histogram distribution, making features such as cracks and deformations more obvious. The mapping function of histogram equalization uses the cumulative distribution function, and the formula is:

[0085]

[0086] Here, L is the total number of gray levels, M and N are the length and width of the image respectively, and HE_Gray(i,j) is the gray value of the (i,j)-th pixel after histogram equalization.

[0087] S22. Crack Detection

[0088] 1) As Figure 2 shown, use the edge detection algorithm for crack detection. Convolve the preprocessed image with the Sobel operator to highlight the regions with significant gray value changes in the image, that is, the locations of the cracks. The specific steps are as follows:

[0089] For each pixel HE_Gray(i,j) in the image, convolve it with G i and G j respectively to obtain the horizontal direction gradient approximation value I i and the vertical direction gradient approximation value I j .

[0090] I i = HE_Gray(i,j) * G i

[0091] I j = HE_Gray(i,j) * G j

[0092] where * represents the two-dimensional convolution operation.

[0093] 2) Through I i and I j , the gradient magnitude and gradient direction of each pixel can be calculated. The gradient magnitude M(i,j) represents the strength of the edge, and the gradient direction θ(i,j) represents the direction of the edge.

[0094]

[0095] 3) Perform non-maximum suppression on the gradient magnitude image M(i, j). Specifically, quantize θ(i, j) to one of four main directions: horizontal, vertical, diagonal, and anti-diagonal. This is achieved by taking the angle differences between θ(i, j) and the angles of the four basic directions, and selecting the direction corresponding to the smallest difference. The angles of the four basic directions are: horizontal direction: 0° or 180°; vertical direction: 90° or 270°; main diagonal direction: 45° or 225°; secondary diagonal direction: 135° or 315°.

[0096] Based on the quantized direction, the adjacent pixels closest to the gradient direction can be found. Specifically,

[0097] When the direction is horizontal (0° or 180°), compare (i, j) with its two adjacent pixels to the left and right, (i + 1, j) and (i - 1, j).

[0098] When the direction is vertical (90° or 270°), compare (i, j) with its two adjacent pixels above and below, (i, j + 1) and (i, j - 1).

[0099] When the direction is diagonal (45° or 225°), compare (i, j) with its two adjacent pixels in the upper left and lower right, (i + 1, j + 1) and (i - 1, j - 1).

[0100] When the direction is anti-diagonal (135° or 315°), compare (i, j) with its two adjacent pixels in the upper right and lower left, (i + 1, j - 1) and (i - 1, j + 1).

[0101] If the pixel (i, j) is greater than the gradient magnitudes of the adjacent pixels, retain the gradient magnitude of this pixel; otherwise, set it to 0.

[0102] 4) Perform thresholding on the gradient magnitude image M(i, j). Select a fixed threshold T, and consider all pixels with gradient magnitudes greater than T as edge pixels, and other pixels as non-edge pixels.

[0103] Use dilation operation to connect the broken crack edges to form a complete crack contour. Specifically, select a rectangular structuring element, align the origin of the structuring element with each pixel in HE_Gray, and determine whether the area covered by the structuring element intersects with the edge pixels in the image. If there is an intersection, set the gray value of this pixel to the average value of all covered edge pixels; if there is no intersection, keep the original gray value of this pixel. The dilated image is denoted as IF_Gray.

[0104] 5) Select a fixed threshold Tb , convert IF_Gray to a binary image, that is, pixels greater than T b are set to 255, and vice versa to 0.

[0105] Calculate the crack length: Use the skeleton extraction algorithm to reduce the lines of the binary image to single-pixel width, calculate the pixel length of the crack skeleton, and then convert the pixel length to the actual physical size according to the resolution of the image and the scale at the time of shooting to obtain the crack length.

[0106] Calculate the crack width: Use the orthogonal skeleton line method to calculate the pixel width of the crack in the image, and then convert the pixel distance to the actual physical size according to the resolution of the image and the scale at the time of shooting to obtain the crack width.

[0107] S23. Attitude deviation detection

[0108] Use the pre-trained CNN model to identify the pre-processed HE_Gray image. The model consists of 3 convolutional layers (the number of neurons is 32, 64, and 128 respectively, and the convolutional kernels are all 3*3), 1 pooling layer, 1 fully connected layer, and 1 output layer in sequence. The number of neurons in the output layer is 6, representing the rotation angles (around the x, y, and z axes) and translation amounts (in the x, y, and z directions) in three-dimensional space respectively.

[0109] S24. Deformation detection

[0110] Use the pre-trained CNN model to identify the pre-processed HE_Gray image. The model consists of 3 convolutional layers (the number of neurons is 32, 64, and 128 respectively, and the convolutional kernels are all 3*3), 1 pooling layer, 1 fully connected layer, and 1 classification layer in sequence. The number of neurons in the classification layer is 3, representing normal, convex, and concave respectively.

[0111] S3. Cofferdam water pressure prediction

[0112] S31. Build a deep learning model. The main body of the model is a Bi-GRU structure. By considering the forward information and backward information in the time series simultaneously, Bi-GRU can capture the temporal characteristics of the data more comprehensively. The forward GRU structure can be expressed as:

[0113] Forward update gate:

[0114]

[0115] where σ is the sigmoid activation function, W z is the weight matrix of the update gate, h t-1 is the hidden state of the previous time step, x t is the tidal data of the current time step, and b z is the bias term.

[0116] Forward reset gate:

[0117]

[0118] where \(W\) r is the weight matrix of the reset gate, and \(b\) r is the bias term.

[0119] Forward new candidate state:

[0120]

[0121] where \(W\) h is the weight matrix of the new candidate state, \(b\) h is the bias term, and \(*\) represents element-wise multiplication.

[0122] Forward update hidden state:

[0123]

[0124] where is the hidden state at the current time step.

[0125] The backward GRU structure is exactly the same as the forward one. Finally, the output of the bidirectional GRU is

[0126]

[0127] where is the hidden state of the forward GRU, is the hidden state of the backward GRU.

[0128] A fully connected layer is connected after the Bi-GRU structure to integrate the learned time series features.

[0129] S32. The input to the deep learning model is the time series of tidal height, wave height, sea current velocity, and cofferdam water pressure with a fixed time window \(T\) s , and the output is the time series of water pressure with a fixed time window \(T\) s for a lag time \(\tau\). This is to take into account the lag of the influence of tidal data on the cofferdam water pressure and the regularity of the water pressure itself. Using the tidal data and cofferdam water pressure data at the current time step \(t\), the cofferdam water pressure at the lag time \(\tau\) is predicted, as Figure 3 shown.

[0130] To find the optimal time window, different time window sizes \(T\) s are selected for verification, \(T\) s={12t, 14t, 16t, 18t, 20t, 22t}. To find the optimal lag time τ, different lag times are selected for verification, τ = {2t, 4t, 6t, 8t, 10t, 12t}.

[0131] Using different time windows and lag times, the collected tidal data and cofferdam water pressure data are used to generate different data sets, and they are divided into training sets and validation sets according to the ratio of 8:2 respectively.

[0132] S33. Compare the mean squared error (MSE) of the validation sets of different data sets to determine the optimal time window and lag time. And input the time series of real-time tidal data and cofferdam water pressure data into this model to predict the cofferdam water pressure data after τ time in the future.

[0133] S4. Early warning notification

[0134] Set the danger thresholds for crack length and width, rotation angle and translation amount. When the results of crack detection and attitude deviation detection exceed the set thresholds or the result of deformation detection is bulge and depression, the early warning mechanism is immediately triggered to notify the relevant personnel to carry out anti-seepage repair work on the cofferdam. The danger threshold for crack length is set to 1 / 3 of the total length of the cofferdam, the danger threshold for crack width is 0.5 mm, the danger threshold for rotation angle is 5°, and the danger threshold for translation amount is 1 / 1000 of the height of the cofferdam.

[0135] Set the danger threshold for cofferdam water pressure. When the predicted water pressure exceeds 70% of the design limit of the cofferdam, the early warning mechanism is immediately triggered to notify the relevant personnel to carry out anti-seepage reinforcement work on the cofferdam.

[0136] S5. Data storage and management

[0137] Establish a cofferdam status database, and store the high-definition image data of the cofferdam and the results of cofferdam status detection in the database for viewing the historical status of the cofferdam in subsequent cofferdam management.

[0138] Establish a cofferdam water pressure database, and store the water pressure sensor data, local tidal data and future cofferdam water pressure prediction results in the database. Regularly update the parameters of the deep learning model in S3 with the latest water pressure data and tidal data to ensure the accuracy of cofferdam water pressure prediction.

[0139] The present invention improves the accuracy of cofferdam anti-seepage monitoring. By combining high-definition images and water pressure sensor data, it not only monitors the changes in the external structure of the cofferdam, but also can reflect the changes in water pressure in the underwater part in real time, comprehensively covering the key areas of cofferdam anti-seepage monitoring. Using computer vision technology and deep learning technology to process and identify images can efficiently and accurately detect cofferdam cracks, deformations and attitude deviations, ensuring early detection of anti-seepage problems.

[0140] The present invention introduces a Bi-GRU structural model, which utilizes the forward and backward features of time series to comprehensively capture the dynamic change laws of tides and water pressure, significantly improving the accuracy of water pressure prediction. Based on the latest data, the parameters of the deep learning model are updated regularly to ensure that the prediction results always maintain high efficiency and accuracy.

[0141] When the detection result exceeds the set threshold, the present invention can automatically trigger an alarm and notify relevant personnel to take repair or reinforcement measures in a timely manner, reducing manual intervention and saving maintenance costs. The present invention constructs a comprehensive cofferdam status and water pressure database to provide data support for subsequent maintenance and management, facilitating trend analysis of the cofferdam status and optimizing maintenance decisions.

[0142] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

Claims

1. A method for monitoring the anti-seepage of cofferdams for platform construction in tidal areas, characterized in that: The method includes: S1. Collect high-definition image data of the cofferdam surface, collect the time series of the cofferdam water pressure, and collect tidal data, including tidal height, wave height and current speed. The timestamps of tidal data and cofferdam water pressure data are synchronized; S2. Use computer vision and deep learning algorithms to identify the collected high-definition image data and determine whether there are factors that cause leakage in the cofferdam, including crack detection, posture deviation detection, and deformation detection; Detect the length and width of the cofferdam cracks, the rotation angle and translation of the cofferdam posture deviation, and whether the cofferdam is deformed; S3, build a deep learning model, use the collected tidal data and cofferdam water pressure data to predict the cofferdam water pressure in the future; The input to the deep learning model is a fixed time window T s The output is a fixed time window T with a lag time τ. s The water pressure time series, To find the best time window, select different time window sizes T s To verify, in order to find the best lag time τ, different lag times are selected for verification. Compare the mean square error (MSE) of the deep learning model under different time windows and different lag times, determine the optimal time window and lag time, and input the time series of real-time tidal data and cofferdam water pressure data into the model to predict the cofferdam water pressure data after τ time in the future; S4. Setting dangerous thresholds for crack length and width, rotation angle and translation, and immediately triggering the warning mechanism when the crack detection and posture deviation detection results exceed the set thresholds or the deformation detection results are convex and concave; Set the water pressure danger threshold of the cofferdam to immediately trigger the early warning mechanism; S5. Establish a cofferdam status database, and store the high-definition image data of the cofferdam and the cofferdam status detection results in the database. A cofferdam water pressure database is established to store water pressure sensor data, local tidal data, and future cofferdam water pressure prediction results in the database, and the deep learning model parameters in S3 are regularly updated with the latest water pressure data and tidal data.

2. A method for monitoring the anti-seepage of cofferdams for platform construction in tidal areas according to claim 1, characterized in that: Before recognizing high-definition image data, the image data is preprocessed by grayscale, smoothing filtering and histogram equalization. The grayscale is as follows: Grayscale is to average the three color channel (RGB) values ​​in the original image to get a grayscale value. The formula is: Among them, Gray(i,j) represents the gray value of the pixel in the i-th row and j-th column of the grayscale image, R(i,j) represents the R value in the i-th row and j-th column of the original high-definition image, G(i,j) represents the G value in the i-th row and j-th column of the original high-definition image, and B(i,j) represents the B value in the i-th row and j-th column of the original high-definition image.

3. The anti-seepage monitoring method for cofferdam construction in tidal area according to claim 2, characterized in that: The smoothing filter uses a Gaussian function as a weight to calculate the weighted average of a pixel and its neighboring pixels. The shape of the Gaussian function is controlled by its standard deviation σ. For a given pixel position (i, j), the gray value SF_Gray(i, j) after smoothing filtering is calculated by the following formula: SF_Gray(i,j)=∑ k,l∈W G σ (k,l)·Gray(i+k,j+l), Among them, W is the neighborhood window centered at (i, j), the window size is k*l, G σ (k,l) is the value of the Gaussian function, defined as:

4. The anti-seepage monitoring method for cofferdam construction in tidal area according to claim 3, characterized in that: The histogram equalization is to enhance the contrast of the image by adjusting the histogram distribution of the image. The mapping function uses the cumulative distribution function, and the formula is: Where L is the total number of gray levels, M and N are the length and width of the image respectively, and HE_Gray(i,j) is the gray value of the (i,j)th pixel after histogram equalization.

5. The anti-seepage monitoring method for cofferdams in tidal area cap construction according to claim 4 is characterized in that: The specific steps of crack detection are: 1) For each pixel grayscale value HE_Gray(i,j) in the preprocessed image HE_Gray, respectively i and G j Perform convolution to obtain the horizontal gradient approximation I i and the vertical gradient approximation I j , the formula is as follows, I i =HE_Gray(i,j)*G i , I j =HE_Gray(i,j)*G j , Where * represents a two-dimensional convolution operation, G i and G j is the Sobel operator; 2) Calculate the gradient magnitude and gradient direction of each pixel. The gradient magnitude M(i, j) represents the strength of the edge, and the gradient direction θ(i, j) represents the direction of the edge. The formula is as follows: 3) Perform non-maximum suppression on the gradient magnitude image M(i,j), specifically quantizing θ(i,j) to one of the four main directions, namely horizontal, vertical, diagonal, and anti-diagonal. According to the quantized direction, the closest neighboring pixel in the gradient direction can be found. If the gradient magnitude of the pixel (i,j) is greater than the gradient magnitude of the neighboring pixel, the gradient magnitude of the pixel is retained, otherwise it is set to 0; 4) Perform threshold processing on the gradient magnitude image M(i, j), select a fixed threshold T, and regard all pixels with gradient magnitude greater than T as edge pixels, and other pixels as non-edge pixels; 5) Use the dilation operation to connect the broken crack edges to form a complete crack outline. Specifically, select a rectangular structure element, align the origin of the structure element with each pixel in HE_Gray, and determine whether the area covered by the structure element has an intersection with the edge pixels in the image. If there is an intersection, the grayscale value of the pixel is set to the average of all covered edge pixels; if there is no intersection, the original grayscale value of the pixel is maintained; The expanded image is represented as IF_Gray; 6) Select a fixed threshold T b , convert IF_Gray to a binary image, that is, if IF_Gray(i,j)>T b, Then the pixel value of pixel (i, j) in the binary image is set to 255, otherwise it is set to 0; Use the skeleton extraction algorithm to reduce the lines of the binary image to a single pixel width, calculate the pixel length of the crack skeleton, and then convert the pixel length into the actual physical size according to the image resolution and the scale at the time of shooting to obtain the crack length; The orthogonal skeleton method is used to calculate the pixel width of the crack in the binary image. Then, according to the resolution of the image and the scale at the time of shooting, the pixel distance is converted into the actual physical size to obtain the crack width.

6. The anti-seepage monitoring method for cofferdam construction in tidal area according to claim 4, characterized in that: The posture deviation detection is specifically to use a CNN model to identify the preprocessed HE_Gray image. The model is composed of 3 convolutional layers, 1 pooling layer, 1 fully connected layer and 1 output layer in sequence. The number of neurons in the output layer is 6, which respectively represent the rotation angles of the cofferdam around the x, y and z axes in three-dimensional space and the translation amounts in the x, y and z directions.

7. The method for monitoring the anti-seepage of cofferdams for platform construction in tidal areas according to claim 4, characterized in that: The deformation detection is specifically to use a CNN model to identify the preprocessed image HE_Gray, wherein the model is sequentially composed of 3 convolutional layers, 1 pooling layer, 1 fully connected layer and 1 classification layer. The number of neurons in the classification layer is 3, representing normal, convex and concave cofferdams respectively.

8. The anti-seepage monitoring method for cofferdams in tidal area cap construction according to claim 1, characterized in that: The main body of the deep learning model in step S3 is a Bi-GRU structure, which includes a forward GRU structure and a backward GRU structure. The forward GRU structure can be expressed as: Forward update gate: Where σ is the sigmoid activation function, W z is the weight matrix of the update gate, h t-1 is the hidden state of the previous time step, x t is the tide data of the current time step, b z is the bias term, Forward reset gate: Among them, W r is the weight matrix of the reset gate, b r is the bias term, Forward to new candidate state: Among them, W h is the weight matrix of the new candidate state, b h is the bias term, * indicates element multiplication; forward update hidden state: in, is the hidden state at the current time step, The backward GRU structure is exactly the same as the forward one, and the final Bi-GRU output is, in, is the hidden state of the forward GRU, It is the hidden state of the reverse GRU, and there is a fully connected layer after the Bi-GRU structure.

Images