A dynamic monitoring and safety warning construction system for cofferdam in marine environment

Abstract

The application relates to the field of safety warning construction, in particular to a cofferdam dynamic monitoring and safety warning construction system under a marine environment, which comprises a cofferdam seepage prediction model training module, an early warning MZ criterion parameter adjustment module and a cofferdam dynamic monitoring and early warning module. The cofferdam seepage prediction model training module trains an initial cofferdam seepage prediction model through a physical penalty loss function to generate an optimized cofferdam seepage prediction model, the physical penalty loss function comprises a water head boundary constraint term and a seepage continuity constraint term, and the initial cofferdam seepage prediction model is constructed based on LSTM and an attention mechanism. The early warning MZ criterion parameter adjustment module calculates an MZ criterion threshold of a cofferdam construction early warning MZ criterion. The cofferdam dynamic monitoring and early warning module calculates and determines a safety construction early warning grade through the optimized cofferdam seepage prediction model and the MZ criterion threshold. The application improves the working condition adaptability and timeliness of the cofferdam dynamic monitoring and safety construction early warning under the marine environment.

Description

Technical Field

[0001] This invention relates to the field of safety early warning construction, and in particular to a dynamic monitoring and safety early warning construction system for cofferdams in a marine environment. Background Technology

[0002] In marine environments, cofferdams must withstand extreme loads such as typhoons, spring tides, and storm surges. However, traditional pure earth-rock or single steel sheet pile cofferdams suffer from weak seepage prevention capabilities, poor erosion resistance, and insufficient structural stability, making them prone to piping, overtopping, and instability, threatening the construction safety of drainage gates. Furthermore, the complex geological conditions of the ocean, including soft soil foundations and gravel layers, limit the adaptability of existing cofferdam designs to water depth and flow velocity. The combined effect of these multiple adverse factors makes seepage problems particularly prominent, with piping, quicksand, and even cofferdam failures caused by uncontrolled seepage becoming increasingly common, posing significant safety hazards to construction.

[0003] Therefore, seepage monitoring and early warning are core aspects of ensuring safe cofferdam construction. Traditional cofferdam seepage monitoring methods collect seepage data by deploying instruments such as pore water pressure gauges, piezometers, or water level sensors, triggering an alarm when the monitored value exceeds a preset threshold. However, because these methods use static thresholds for early warning, they cannot reflect the dynamic evolution of seepage patterns throughout the entire lifecycle of cofferdam construction, making them prone to false alarms or missed alarms. In marine environments, there are numerous driving factors for cofferdam seepage, including external water head, wave load, tidal action, soil consolidation state, and the characteristics of filling materials. Simple threshold comparison cannot integrate multi-dimensional construction status information for comprehensive judgment, resulting in low reliability of early warnings and the need for manual verification, leading to poor real-time performance of cofferdam seepage construction early warnings.

[0004] With the increasing application of machine learning technology in construction safety, data-driven intelligent prediction methods have provided new insights for cofferdam seepage monitoring. Long Short-Term Memory (LSTM) networks possess excellent time-series modeling capabilities, effectively capturing long-range temporal dependencies in seepage monitoring data, and have been preliminarily applied to seepage prediction in dams and earth-rock dams. Attention mechanisms can dynamically allocate weights to key time steps, further improving the accuracy of LSTM in representing complex seepage processes. However, purely data-driven machine learning models suffer from inconsistencies in physical mechanisms; model outputs may violate fundamental hydraulic principles such as Darcy's law of seepage and head boundary constraints, leading to a lack of interpretability and credibility of prediction results at the engineering level, and potentially resulting in significant generalization errors when data is scarce or operating conditions change abruptly.

[0005] In terms of early warning decision-making, existing cofferdam safety monitoring systems mostly rely on statistical control chart methods based on normal distribution or empirical threshold methods. These methods treat early warning parameters as static fixed quantities, making it difficult to adapt to the significant differences in seepage characteristics at different stages of cofferdam construction, and also unable to dynamically track the evolution of the deviation between the output of the seepage prediction model and the measured values. Furthermore, the traditional MZ (M estimator - confidence interval radius) criterion relies on offline statistical results for setting, and lacks dynamic adaptive adjustment parameters for marine cofferdam construction conditions, limiting its practicality and adaptability in complex marine construction environments.

[0006] In summary, how to utilize machine learning models and the MZ criterion for predictive early warning of cofferdam seepage, so as to achieve dynamic monitoring and safe construction early warning of cofferdams in marine environments with adaptability and timeliness, is a technical problem that needs to be solved. Summary of the Invention

[0007] To address this, the present invention provides a dynamic monitoring and safety early warning construction system for cofferdams in marine environments, which overcomes the problem in existing technologies that cannot utilize machine learning models and MZ criteria for predictive early warning of cofferdam seepage, thus failing to achieve the adaptability and timeliness of dynamic monitoring and safety construction early warning of cofferdams in marine environments.

[0008] To achieve the above objectives, this invention proposes a dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, comprising: The cofferdam seepage prediction model training module is used to generate an initial cofferdam seepage prediction value by passing the first historical monitoring data of the construction status in the early stage of cofferdam construction through an initial cofferdam seepage prediction model. The initial cofferdam seepage prediction value and the historical cofferdam seepage real value are then used to train the initial cofferdam seepage prediction model based on a physical penalty loss function to generate an optimized cofferdam seepage prediction model. The physical penalty loss function includes a head boundary constraint term and a seepage continuity constraint term. Both the initial cofferdam seepage prediction model and the optimized cofferdam seepage prediction model are built based on LSTM and attention mechanism. The early warning MZ criterion parameter adjustment module is used to take the second historical monitoring data of the construction status during the stable period of cofferdam construction and use the optimized cofferdam seepage prediction model to generate the first predicted value of cofferdam seepage. Based on the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage, the MZ criterion threshold of the early warning MZ criterion for cofferdam construction is calculated. The cofferdam dynamic monitoring and early warning module is used to obtain real-time monitoring data of the construction status during the cofferdam construction early warning period. Through the optimized cofferdam seepage prediction model, a second predicted value of cofferdam seepage is generated. Based on the real-time monitoring data of the construction status, the second predicted value of cofferdam seepage, and the MZ criterion threshold, the safe construction early warning level is calculated and determined.

[0009] Furthermore, the cofferdam seepage prediction model training module includes: The prediction error loss term calculation submodule is used to calculate the smoothed L1 loss function based on the initial cofferdam seepage prediction value and the historical cofferdam seepage true value, so as to generate the prediction error loss term; The head boundary constraint term calculation submodule is used to calculate the head boundary constraint term based on the difference between the initial cofferdam seepage prediction value and the upstream tidal head value. The seepage continuity constraint term calculation submodule is used to calculate the difference between the two initial cofferdam seepage prediction values ​​and the reference hydraulic gradient at adjacent detection points, and to calculate the seepage continuity constraint term based on the ratio of the difference to the reference hydraulic gradient. The loss function calculation submodule is used to calculate the physical penalty loss function based on the weighted sum of the prediction error loss term, the head boundary constraint term, and the seepage continuity constraint term. The physical penalty loss function also includes a prediction error loss term.

[0010] Furthermore, the seepage continuity constraint term calculation submodule includes: The reference hydraulic gradient calculation unit is used to generate the seepage spatial mean by dividing the difference between the two historical cofferdam seepage values ​​at adjacent detection points by the distance between the adjacent detection points. The arithmetic mean of multiple seepage spatial means within a time window is calculated to generate the reference hydraulic gradient.

[0011] Furthermore, the loss function calculation submodule includes: The head boundary constraint weighting coefficient calculation unit is used to calculate the current head boundary constraint weighting coefficient by taking the maximum value of the head boundary constraint weighting coefficient and the current training period through the Sigmoid function. The loss function weighted calculation unit is used to calculate the physical penalty loss function by weighting and summing the head boundary constraint term and the seepage continuity constraint term based on the weighting coefficient of the current head boundary constraint term and the weighting coefficient of the set seepage continuity constraint term, and then adding the sum to the prediction error loss term.

[0012] Furthermore, the cofferdam dynamic monitoring and early warning module includes: The input layer operation submodule is used to process the real-time monitoring data of the construction status through the input layer to generate an input feature vector; The LSTM attention mechanism operation submodule is used to process the input feature vector through LSTM and the attention mechanism to generate percolation prediction features; The fully connected output layer computation submodule is used to pass the seepage prediction features through the fully connected output layer to generate the second predicted value of the cofferdam seepage. The optimized cofferdam seepage prediction model further includes an input layer and a fully connected output layer.

[0013] Furthermore, the LSTM attention mechanism operation submodule includes: The feature attention operation unit is used to perform attention weighting on key factors through the feature attention mechanism on the input feature vector to generate factor-weighted features; The LSTM encoder operation unit is used to pass the factor-weighted features through the LSTM encoder to generate time-series trend percolation features; The temporal attention operation unit is used to apply attention weights to key time sequences through the temporal attention mechanism to generate temporal weighted percolation features; The LSTM decoder operation unit is used to pass the time-series trend percolation features and time-series weighted percolation features through the LSTM decoder to generate the percolation prediction features; The LSTM includes an LSTM encoder and an LSTM decoder, and the attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0014] Furthermore, the early warning MZ criterion parameter adjustment module includes: The robust scalar calculation submodule is used to iteratively calculate the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage using the M estimator to generate a robust scalar. The predicted confidence interval radius calculation submodule is used to multiply the variance of multiple deviation values ​​by the standard normal distribution quantile to generate the predicted confidence interval radius; The MZ criterion threshold calculation submodule is used to perform a standardized calculation after weighted summation of the robustness scale and the predicted confidence interval radius to generate the MZ criterion threshold.

[0015] Furthermore, the cofferdam dynamic monitoring and early warning module also includes: The MZ criterion anomaly index calculation submodule is used to calculate the absolute value of the difference between the real-time monitoring data of the construction status and the second predicted value of the cofferdam seepage, and to calculate the ratio of the absolute value of the difference to the MZ criterion threshold to generate the MZ criterion anomaly index. The level determination submodule is used to determine the safety construction warning level based on the comparison result between the abnormal index of the MZ criterion and the warning threshold.

[0016] Furthermore, the early stage of cofferdam construction refers to the early stage of steel sheet pile cofferdam construction, and the stable stage of cofferdam construction refers to the later stage of steel sheet pile cofferdam construction and the period when no typhoon warning has been issued.

[0017] Furthermore, the cofferdam includes a sheet pile cofferdam and an earth-rock slope cofferdam.

[0018] Compared with the prior art, the beneficial effects of the present invention are that, through the dual physical constraint design of head boundary constraint term and seepage continuity constraint term, the basic laws of seepage mechanics are deeply embedded into the model training process, which completely solves the problem that pure data-driven models are easily affected by monitoring noise and construction interference. The prediction results are completely consistent with the actual seepage field change characteristics of marine cofferdams, realize the prediction of cofferdam seepage conditions using machine learning, and thus realize predictive early warning of cofferdam seepage conditions, and realize the adaptability and timeliness of dynamic monitoring and safe construction early warning of cofferdams in marine environments.

[0019] In particular, this invention uses dual-attention adaptive weighting to accurately pinpoint core patterns and anomaly precursors, weakening the influence of irrelevant features such as construction interference and instrument noise. This solves the feature redundancy problem in multi-factor coupled scenarios, allowing the model to focus on core seepage patterns. Through an LSTM encoding and decoding architecture, it can accurately capture long-term seepage evolution patterns covering multiple complete semi-diurnal tidal cycles, deeply explore the dynamic response relationship between the periodic alternation of ocean tides and the cofferdam seepage field, and fully preserve the long-term development trend and periodic fluctuation characteristics of seepage. This ensures prediction accuracy under stable operating conditions and effectively captures seepage precursor signals under abnormal operating conditions, thereby enabling predictive early warning of cofferdam seepage using machine learning models. This achieves adaptability and timeliness of dynamic monitoring and safe construction early warning of cofferdams in marine environments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, as described in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the optimized cofferdam seepage prediction model for the dynamic monitoring and safety early warning construction system of cofferdams in a marine environment, as described in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the calculation process of the MZ criterion threshold for a dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, as described in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0024] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] like Figures 1 to 4 As shown, the present invention provides a dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, which overcomes the problem in the prior art that it is impossible to use machine learning models and MZ criteria to predict and warn of cofferdam seepage, and thus cannot achieve the adaptability and timeliness of dynamic monitoring and safety construction early warning of cofferdams in a marine environment.

[0026] like Figure 1 and 2 As shown in the figure, this embodiment proposes a dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, including: The cofferdam seepage prediction model training module is used to generate an initial cofferdam seepage prediction value by passing the first historical monitoring data of the construction status in the early stage of cofferdam construction through an initial cofferdam seepage prediction model. The initial cofferdam seepage prediction value and the historical cofferdam seepage real value are then used to train the initial cofferdam seepage prediction model based on a physical penalty loss function to generate an optimized cofferdam seepage prediction model. The physical penalty loss function includes a head boundary constraint term and a seepage continuity constraint term. Both the initial cofferdam seepage prediction model and the optimized cofferdam seepage prediction model are built based on LSTM and attention mechanism. The early warning MZ criterion parameter adjustment module is used to take the second historical monitoring data of the construction status during the stable period of cofferdam construction and use the optimized cofferdam seepage prediction model to generate the first predicted value of cofferdam seepage. Based on the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage, the MZ criterion threshold of the early warning MZ criterion for cofferdam construction is calculated. The cofferdam dynamic monitoring and early warning module is used to obtain real-time monitoring data of the construction status during the cofferdam construction early warning period. Through the optimized cofferdam seepage prediction model, a second predicted value of cofferdam seepage is generated. Based on the real-time monitoring data of the construction status, the second predicted value of cofferdam seepage, and the MZ criterion threshold, the safe construction early warning level is calculated and determined.

[0027] Understandably, the LSTM (Long Short-Term Memory) network is responsible for predicting future states, while the cofferdam construction early warning MZ (M-estimate - confidence interval radius) criterion is responsible for judging whether the current deviation is abnormal. The two complement each other, upgrading the passive alarm of cofferdam seepage to active prediction. The cofferdam construction early warning MZ criterion calculates the deviation between the historical measured cofferdam seepage value and the first predicted value of cofferdam seepage from the LSTM, and applies the MZ criterion based on the M-estimater to the residual sequence. This criterion can effectively resist the interference of outliers in the data, accurately identify the real structural anomalies, and thus generate a dynamic early warning threshold.

[0028] Understandably, during the early warning period of cofferdam construction, before the arrival of storm surge, the optimized cofferdam seepage prediction model based on LSTM and attention mechanisms may have already foreseen the accelerated upward trend of seepage pressure. At this time, if a slight increase in seepage pressure is detected, it may be interpreted by the model as an expected trend, thus the MZ score remains low, the warning level remains normal, and false alarms that may occur with traditional fixed thresholds are avoided. However, when the seepage pressure data rises sharply, far exceeding the reasonable range that the model can explain based on all environmental factors, an extremely high MZ score is calculated, and it quickly exceeds the MZ criterion threshold, triggering a red anomaly warning level.

[0029] Therefore, by deeply integrating the forward-looking prediction capability of the LSTM model with the robust detection capability of the MZ criterion, an intelligent seepage early warning system for cofferdams can be realized, which can detect risks in advance, accurately identify anomalies, and dynamically adapt to changes. This provides a guarantee for the safe operation of cofferdams in the construction of marine seawall drainage gates in complex and harsh environments, avoids the risk of disasters such as cofferdam overtopping and piping caused by extreme weather, and can effectively ensure the construction safety of the main drainage gate project.

[0030] Furthermore, the cofferdam seepage prediction model training module includes: The prediction error loss term calculation submodule is used to calculate the smoothed L1 loss function based on the initial cofferdam seepage prediction value and the historical cofferdam seepage true value, so as to generate the prediction error loss term; The head boundary constraint term calculation submodule is used to calculate the head boundary constraint term based on the difference between the initial cofferdam seepage prediction value and the upstream tidal head value. The seepage continuity constraint term calculation submodule is used to calculate the difference between the two initial cofferdam seepage prediction values ​​and the reference hydraulic gradient at adjacent detection points, and to calculate the seepage continuity constraint term based on the ratio of the difference to the reference hydraulic gradient. The loss function calculation submodule is used to calculate the physical penalty loss function based on the weighted sum of the prediction error loss term, the head boundary constraint term, and the seepage continuity constraint term. The physical penalty loss function also includes a prediction error loss term.

[0031] Specifically, the calculation process of the prediction error loss term is the calculation process of the smooth L1 loss function. When the absolute value of the difference between the initial predicted seepage value and the historical actual seepage value of the cofferdam is less than 1, the loss value is half of the square of the difference; when the absolute value is greater than or equal to 1, the loss value is the absolute value of the difference minus 0.5. Therefore, the smooth L1 loss function ensures that the model prediction results are numerically consistent with the observed data.

[0032] Specifically, the calculation process of the seepage continuity constraint term is as follows: First, the initial cofferdam seepage prediction value needs to be converted into a predicted head value expressed in terms of elevation using the buried elevation of the measuring point. Based on the measured data of the upstream tide gauge at the corresponding time, the highest water level currently acting on the outside of the cofferdam is determined. Considering the brief lag effect of seepage transmission in the soil or steel sheet pile gaps, the engineering allows the head height to be slightly higher than the theoretical static head momentarily. Therefore, the upper limit of the head height is usually set as the real-time head height plus a small allowable lag margin of 0.1 meters. The predicted head values ​​of multiple measuring points are compared with the upper limit of the head height one by one. If the predicted head value is less than or equal to the allowable upper limit of the head height, it means that the predicted value of the point is within the range allowed by physical laws. No penalty is imposed on the point in this calculation, and the penalty contribution is recorded as zero. Otherwise, it means that the model has predicted an abnormally high pressure that violates the law of conservation of energy. At this time, the difference between the predicted head of the point and the allowable upper limit is calculated, and this difference is squared. This squared value is the physical penalty that the point needs to bear at this moment. The individual penalties of all detection points are added together and then divided by the total number of monitoring points involved in the calculation to obtain the average penalty value of the batch of samples. This average value is the seepage continuity constraint term.

[0033] Therefore, by using the seepage continuity constraint term, when the model attempts to predict a physically impossible error such as the pressure head being higher than the tidal level, a reverse constraint is applied to the model through the loss function, forcing the model to adjust its internal parameters. This ensures that the prediction curve of the finally trained model still has a certain degree of hydraulic rationality when encountering extreme high tide levels.

[0034] Furthermore, the seepage continuity constraint term calculation submodule includes: The reference hydraulic gradient calculation unit is used to generate the seepage spatial mean by dividing the difference between the two historical cofferdam seepage values ​​at adjacent detection points by the distance between the adjacent detection points. The arithmetic mean of multiple seepage spatial means within a time window is calculated to generate the reference hydraulic gradient.

[0035] Specifically, the calculation process of the reference hydraulic gradient is as follows: For any two adjacent monitoring points, obtain the actual seepage pressure monitoring values ​​of these two points at the same historical moment, convert these two pressure values ​​into water head height expressed in terms of elevation, calculate the difference in water head between the two points, and then divide it by the actual average seepage space value of these two points in space. This average seepage space value represents the steepness of the water flow between the two points inside the cofferdam at that specific moment, and is called the instantaneous gradient of the seepage space. Considering that ocean tides have periodic fluctuations and seepage response has lag, the noise of the gradient value at a single instant is relatively large and cannot be directly used as the benchmark for physical constraints. Therefore, a time window including a complete tidal day is defined to fully include the normal hydraulic gradient fluctuation range under different states such as high tide, low tide, and slack tide. The arithmetic mean of all the seepage space values ​​calculated within this time window is taken, and the average value is the reference hydraulic gradient. This reference hydraulic gradient represents the macroscopic average slope that should normally be maintained between two adjacent points inside the cofferdam under the current tidal environment and construction conditions.

[0036] Specifically, the calculation process of the seepage continuity constraint term is as follows: Calculate the difference between the two initial cofferdam seepage prediction values ​​and the reference hydraulic gradient at adjacent monitoring points. Divide the difference by the actual distance between the two points to obtain the predicted instantaneous gradient constructed from the model prediction values. Calculate the difference between the predicted instantaneous gradient and the reference hydraulic gradient. Iterate through all preset combinations of adjacent monitoring points in the model, accumulate the difference for each pair of combinations, and finally divide by the total number of monitoring point combinations to obtain the seepage continuity constraint term for this batch of training samples. Therefore, the reference hydraulic gradient, which slides with the time window, does not require the model-predicted seepage field to be completely horizontal like still water. It teaches the model to mimic the real dynamic hydraulic gradient inside the cofferdam, affected by tidal fluctuations, thereby preventing abrupt changes in model values ​​while allowing reasonable seepage fluctuations.

[0037] Furthermore, the loss function calculation submodule includes: The head boundary constraint weighting coefficient calculation unit is used to calculate the current head boundary constraint weighting coefficient by taking the maximum value of the head boundary constraint weighting coefficient and the current training period through the Sigmoid function. The loss function weighted calculation unit is used to calculate the physical penalty loss function by weighting and summing the head boundary constraint term and the seepage continuity constraint term based on the weighting coefficient of the current head boundary constraint term and the weighting coefficient of the set seepage continuity constraint term, and then adding the sum to the prediction error loss term.

[0038] Specifically, the calculation process for the weighting coefficient of the head boundary constraint term is as follows: The maximum value of the weighting coefficient of the head boundary constraint term is preset to 0.1, which indicates that the head constraint is a hard boundary, and the consequences of violation are serious, requiring stronger constraints. The current weighting coefficient of the head boundary constraint term is calculated using the following formula in the form of the Sigmoid function: ; in, This represents the weighting coefficient of the current head boundary constraint term. represents the weighting coefficient of the preset head boundary constraint term, and k represents the slope factor of the Sigmoid function, which determines how quickly the physical constraint weights transition from near zero to near their maximum value during training. 4.4 is divided by 20% of the total number of training rounds to obtain k. This calculation process is based on the mathematical properties of the Sigmoid function. Indicates the current training cycle. The training period representing the turning point is preferably 30% of the total training rounds. This achieves a weighting strategy that is relatively flat in the early stages and gradually increases in the later stages. This avoids underfitting caused by forcibly applying seepage continuity constraints before the model has learned the basic tidal patterns, and thus ensures that the weighting coefficient of the head boundary constraint term remains stable at 0.95 times in approximately the last 20% of the training rounds. The above measures are taken to ensure that the final model is fully constrained by physical laws.

[0039] Specifically, the calculation process of the physical penalty loss function is as follows: the weighting coefficients of the weighted summation of the prediction error loss term and the seepage continuity constraint term are 1.0 and 0.01, respectively. The weighting coefficient of 1.0 is used to ensure that the prediction error loss term contributes as the main loss term, and the weighting coefficient of 0.01 is used to avoid the model being unable to learn the abnormal signals of local seepage points if the hydraulic gradient is forcibly constrained to be continuous.

[0040] like Figure 3 As shown, the cofferdam dynamic monitoring and early warning module further includes: The input layer operation submodule is used to process the real-time monitoring data of the construction status through the input layer to generate an input feature vector; The LSTM attention mechanism operation submodule is used to process the input feature vector through LSTM and the attention mechanism to generate percolation prediction features; The fully connected output layer computation submodule is used to pass the seepage prediction features through the fully connected output layer to generate the second predicted value of the cofferdam seepage. The optimized cofferdam seepage prediction model further includes an input layer and a fully connected output layer.

[0041] Specifically, the real-time monitoring data of the construction status includes the tide level, wave height, wind speed, rainfall intensity, cofferdam sheet pile penetration depth, and cofferdam earth-rock filling elevation recorded by the construction management system, as well as the seepage pressure, displacement, and phreatic line elevation of the cofferdam at previous times. After normalizing these variables in the input layer, a standardized input feature vector is constructed. Therefore, the input feature vector provides the LSTM with complete working condition context information, enabling the model to capture the temporal causal chain of rapid tide rise and lagged seepage pressure response. The fully connected output layer is used to map the seepage prediction features to a second predicted value of the cofferdam seepage.

[0042] Specifically, the structure of the initial cofferdam seepage prediction model is completely consistent with that of the optimized cofferdam seepage prediction model. The only difference is whether the model is in the training phase or the deployment and implementation phase.

[0043] like Figure 3 As shown, the LSTM attention mechanism operation submodule further includes: The feature attention operation unit is used to perform attention weighting on key factors through the feature attention mechanism on the input feature vector to generate factor-weighted features; The LSTM encoder operation unit is used to pass the factor-weighted features through the LSTM encoder to generate time-series trend percolation features; The temporal attention operation unit is used to apply attention weights to key time sequences through the temporal attention mechanism to generate temporal weighted percolation features; The LSTM decoder operation unit is used to pass the time-series trend percolation features and time-series weighted percolation features through the LSTM decoder to generate the percolation prediction features; The LSTM includes an LSTM encoder and an LSTM decoder, and the attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

[0044] Specifically, the feature attention mechanism calculates the original score of each feature using a learnable parameter matrix and the ReLU activation function. This score is then transformed into a weight distribution using the Softmax function. The weight distribution is then multiplied element-wise with the input feature vector, achieving dynamic scaling of each factor element in the input feature vector. Therefore, the feature attention mechanism avoids the drawback of traditional LSTM treating all input variables equally, significantly improving the predictive adaptability to complex conditions. For example, during typhoons, the model automatically assigns higher attention weights to tide levels and wave heights, while downplaying the influence of the cofferdam sheet pile penetration depth and the cofferdam earth-rock filling elevation on the construction progress.

[0045] Specifically, the LSTM encoder controls the flow of information through internal input gates, forget gates, and output gates, memorizes long-term dependencies, and learns long-term trend changes caused by the slow rise of the phreatic line and soil consolidation during the construction of a cofferdam.

[0046] Specifically, the temporal attention mechanism uses the dot product similarity score between the current temporal trend seepage feature (which serves as the hidden state) output by the LSTM encoder and the temporal trend seepage features of all historical time steps of the encoder. This dot product similarity score is then processed by a Softmax function to generate temporal attention weights. These weights are then summed over the temporal trend seepage features to obtain the temporally weighted seepage features. Therefore, during the alternating construction of sheet piles and earth-rock cofferdams, the temporal attention mechanism enables the model to recall response characteristics under similar past perturbations, significantly improving the reliability of predictions during the construction transition period and effectively mitigating the concept drift problem.

[0047] Specifically, the LSTM decoder's computation process is as follows: The temporal trend seepage features and temporal weighted seepage features are mapped to the decoder's hidden state dimension space via a fully connected network. These are then concatenated along the feature dimension and input into the LSTM decoder. The LSTM decoder uses its input gate, forget gate, candidate memory unit, memory update unit, and output gate to generate seepage prediction features. Therefore, when the LSTM decoder predicts seepage 6 hours after a typhoon's landfall, the seepage prediction features generate a high similarity score with the hidden state of the encoder during the previous typhoon landfall. This ensures that the temporal weighted seepage features primarily contain the feature information of that typhoon response process. This adaptive historical retrieval mechanism allows the LSTM decoder to proactively review and learn from seepage evolution patterns under similar historical conditions when the construction system slightly transitions from sheet piles to earth-rock cofferdams, rather than mechanically extrapolating.

[0048] Specifically, the fully connected output layer is used to map the high-dimensional seepage prediction features to the target physical quantity space through a linear transformation to generate a second predicted value of the cofferdam seepage.

[0049] like Figure 4 As shown, the early warning MZ criterion parameter adjustment module further includes: The robust scalar calculation submodule is used to iteratively calculate the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage using the M estimator to generate a robust scalar. The predicted confidence interval radius calculation submodule is used to multiply the variance of multiple deviation values ​​by the standard normal distribution quantile to generate the predicted confidence interval radius; The MZ criterion threshold calculation submodule is used to perform a standardized calculation after weighted summation of the robustness scale and the predicted confidence interval radius to generate the MZ criterion threshold.

[0050] Specifically, the robustness calculation process is as follows: For each time point, calculate the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage, forming a deviation value sequence. Calculate the standardized median of the deviation value sequence. Calculate the absolute value of the difference between each deviation value and this median. Then, take the median of these absolute values. Finally, multiply this median by a constant factor of 1.4826 to obtain the estimator. The constant factor ensures that the estimator maintains consistency with the standard deviation when the data follows a normal distribution, thus initializing the M-estimator. Divide each deviation value by the estimated value to perform standardization. The standardization process involves calculating the ratio of the adjustment constant to the standardized value to obtain the weight value. The adjustment constant is 1.345, meaning that the greater the deviation, the smaller the weight value. The weights of all data points are collected, and the average of all weights is calculated to obtain the updated estimate after this iteration. The difference between the estimated value and the updated estimate is then checked against a preset threshold of 0.001. If not, the updated estimate is used as the new estimated value for iterative calculation using the M-estimator until the difference between the estimated value and the updated estimate is less than the preset threshold of 0.001. The updated estimate is then used as the robust metric. Therefore, the robust metric is a robust statistic that excludes a few abnormal disturbances and purely reflects the fluctuation range of the model's normal prediction accuracy. It represents a key component of the normal fluctuation benchmark and provides a stable and reliable benchmark for subsequent real-time early warning.

[0051] Specifically, the standard normal distribution quantile represents the tolerance for false alarm risk, determining the probability that the predicted interval will cover the true seepage value. This probability is called the confidence level; at a confidence level of 95%, the standard normal distribution quantile is 1.96.

[0052] Specifically, the MZ criterion threshold is calculated as follows: the robustness scale is multiplied by a safety factor of 1.0, and then added to the radius of the prediction confidence interval to obtain the MZ criterion threshold. This safety factor of 1.0 is the same as the warning trigger line for a yellow alert, thus realizing the mapping relationship between statistical distribution characteristics and warning level classification.

[0053] Furthermore, the cofferdam dynamic monitoring and early warning module also includes: The MZ criterion anomaly index calculation submodule is used to calculate the absolute value of the difference between the real-time monitoring data of the construction status and the second predicted value of the cofferdam seepage, and to calculate the ratio of the absolute value of the difference to the MZ criterion threshold to generate the MZ criterion anomaly index. The level determination submodule is used to determine the safety construction warning level based on the comparison result between the abnormal index of the MZ criterion and the warning threshold.

[0054] Specifically, the absolute value of the difference between the real-time monitoring data of the construction status and the second predicted value of the cofferdam seepage can reflect the degree to which the current actual seepage status deviates from the normal expectation. The ratio of the absolute value of the difference to the threshold of the MZ criterion is used as the MZ criterion anomaly index, which is a dimensionless anomaly score.

[0055] Specifically, the safety construction early warning levels include: a normal warning level when the MZ criterion anomaly index is less than 1.0, meaning no seepage warning is issued and the seepage status fluctuates within the expected range; a yellow warning level when the MZ criterion anomaly index is less than 1.5 and greater than or equal to 1.0, indicating a trend deviation exceeding the normal noise level, suggesting an increase in the frequency of on-site manual inspections; an orange warning level when the MZ criterion anomaly index is less than 2.0 and greater than or equal to 1.5, indicating a significant deviation and the possibility of structural abnormalities or measurement malfunctions, requiring immediate on-site verification; and a red warning level when the MZ criterion anomaly index is greater than or equal to 2.0, indicating a very high probability of leakage, requiring immediate activation of the emergency plan.

[0056] Specifically, eight sets of piezometers (four upstream and four downstream), two ultrasonic tide gauges, one water level gauge inside the weir, three water measuring weirs (for seepage monitoring), and four settlement / horizontal displacement measuring points were deployed along the cofferdam axis. The sampling frequency was once per hour to construct a dataset. The dataset consisted of 1008 time-series samples from the first historical monitoring data during the early construction phase, divided into training and validation subsets in a 7:2 ratio according to time sequence. The threshold calibration set consisted of 960 time-series samples from the second historical monitoring data during the stable construction period. The test set consisted of 1008 time-series samples from real-time monitoring data during the construction early warning period, including three spring tides, two typhoon peripheral impact periods, and six controlled simulated abnormal working conditions. All data were Min-Max normalized to the [0,1] interval. Compared with ordinary LSTM, LSTM, and the optimized cofferdam seepage prediction model described in this embodiment without physical constraints, the goodness of fit R0 was higher. 2 The values ​​are 0.886, 0.924, and 0.968 respectively, and their mean absolute errors (MAE) are 0.427, 0.315, and 0.142 respectively. It can be seen that the goodness of fit of the optimized cofferdam seepage prediction model described in this embodiment is improved and the mean absolute error is reduced. By using the physical penalty loss function, the proportion of invalid prediction samples that violate the laws of seepage mechanics is reduced, and the problem of false fitting of pure data-driven models is completely solved.

[0057] Specifically, comparing the false alarm rates of the optimized cofferdam seepage prediction model with the 3σ criterion for early warning, the static MZ criterion for early warning, and the cofferdam construction early warning MZ criterion described in this embodiment, the false alarm rates are 11.41%, 5.85%, and 1.19%, respectively, and the average early warning lead time (h) is 4.3, 7.6, and 10.4, respectively. It is evident that the cofferdam construction early warning MZ criterion completely solves the problems of false alarms and missed alarms related to the static early warning threshold, identifies hidden hazards such as local failure of the seepage prevention system and internal soil scouring in advance, and reserves sufficient window period for emergency response.

[0058] Furthermore, the early stage of cofferdam construction refers to the early stage of steel sheet pile cofferdam construction, and the stable stage of cofferdam construction refers to the later stage of steel sheet pile cofferdam construction and the period when no typhoon warning has been issued.

[0059] Furthermore, the cofferdam includes a sheet pile cofferdam and an earth-rock slope cofferdam.

[0060] Designed for tropical typhoon zones and tidal environments, the design incorporates innovative features such as the depth of steel sheet piles, the slope of the earth-rock cofferdam, and the protective structure of bagged soil. Combined with a dynamic seepage monitoring and early warning mechanism, it achieves a balance between typhoon resistance, erosion prevention, and durability. The construction process is suitable for construction in extreme environments.

[0061] It is understood that this embodiment is preferably implemented in tropical typhoon areas and tidal environments, to provide a dynamic monitoring and early warning mechanism for seepage in cofferdam structures with steel sheet pile penetration depth, so as to achieve typhoon resistance and ensure the construction safety of its construction process in extreme environments.

[0062] The early stage of cofferdam construction is the model training period, which is preferably the first 70% of the steel sheet pile cofferdam construction stage, that is, the relatively stable period after the steel sheet piles are driven into place and the first support is installed.

[0063] The stable construction period of the cofferdam is the MZ parameter calibration period, preferably the later stage of the steel sheet pile cofferdam construction and the period before the typhoon warning is issued. This stage includes at least 1 to 2 complete tidal cycles, the period of severe water level fluctuation, and the period excluding the direct impact of the typhoon. At this time, the cofferdam's support system is complete, the deformation tends to be stable, and it is in quasi-static equilibrium. There is no severe construction disturbance, so the normal prediction deviation distribution of the model under ideal working conditions can be obtained, and the MZ criterion threshold can be calculated to accurately represent the robust scale of normal fluctuations.

[0064] The early warning period for cofferdam construction is a real-time monitoring period, preferably the entire remaining construction period after the calibration period ends. This period can cover the following high-risk conditions: high-intensity filling period of earth-rock cofferdam, transition period between steel sheet pile and earth-rock cofferdam, typhoon season, and period of increased water level due to astronomical tides and storm surges.

[0065] Preferably, when entering the stable period of cofferdam construction, the data in this period can be used to perform the last few rounds of fine-tuning on the optimized cofferdam seepage prediction model through the physical penalty loss function, so that the model can adapt to the new stiffness state, and then the model can be fixed to calculate the MZ criterion threshold of the cofferdam construction early warning MZ criterion.

[0066] In this embodiment, the basic laws of seepage mechanics are deeply embedded into the model training process through the dual physical constraint design of head boundary constraint term and seepage continuity constraint term. This completely solves the problem that pure data-driven models are easily affected by monitoring noise and construction interference. The prediction results are completely consistent with the actual seepage field change characteristics of marine cofferdams. This enables the prediction of cofferdam seepage conditions using machine learning, thereby achieving predictive early warning of cofferdam seepage conditions and realizing the adaptability and timeliness of dynamic monitoring and safe construction early warning of cofferdams in marine environments. By employing dual-attention adaptive weighting, the core patterns and anomaly precursors are accurately identified, while the influence of irrelevant features such as construction interference and instrument noise is weakened. This solves the feature redundancy problem in multi-factor coupled scenarios, allowing the model to focus on the core seepage patterns. Through an LSTM encoding and decoding architecture, the model can accurately capture long-term seepage evolution patterns covering multiple complete semi-diurnal tidal cycles. It deeply explores the dynamic response relationship between the periodic alternation of ocean tides and the seepage field of the cofferdam, fully preserving the long-term development trend and periodic fluctuation characteristics of seepage. This ensures prediction accuracy under stable operating conditions and effectively captures seepage precursor signals under abnormal operating conditions. As a result, it enables predictive early warning of cofferdam seepage using machine learning models, achieving adaptability and timeliness of dynamic monitoring and safe construction early warning of cofferdams in marine environments.

[0067] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0068] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic monitoring and safety early warning construction system for cofferdams in a marine environment, characterized in that, include: The cofferdam seepage prediction model training module is used to generate an initial cofferdam seepage prediction value by passing the first historical monitoring data of the construction status in the early stage of cofferdam construction through an initial cofferdam seepage prediction model. The initial cofferdam seepage prediction value and the historical cofferdam seepage real value are then used to train the initial cofferdam seepage prediction model based on a physical penalty loss function to generate an optimized cofferdam seepage prediction model. The physical penalty loss function includes a head boundary constraint term and a seepage continuity constraint term. Both the initial cofferdam seepage prediction model and the optimized cofferdam seepage prediction model are built based on LSTM and attention mechanism. The early warning MZ criterion parameter adjustment module is used to take the second historical monitoring data of the construction status during the stable period of cofferdam construction and use the optimized cofferdam seepage prediction model to generate the first predicted value of cofferdam seepage. Based on the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage, the MZ criterion threshold of the early warning MZ criterion for cofferdam construction is calculated. The cofferdam dynamic monitoring and early warning module is used to obtain real-time monitoring data of the construction status during the cofferdam construction early warning period. Through the optimized cofferdam seepage prediction model, a second predicted value of cofferdam seepage is generated. Based on the real-time monitoring data of the construction status, the second predicted value of cofferdam seepage, and the MZ criterion threshold, the safe construction early warning level is calculated and determined.

2. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 1, characterized in that, The cofferdam seepage prediction model training module includes: The prediction error loss term calculation submodule is used to calculate the smoothed L1 loss function based on the initial cofferdam seepage prediction value and the historical cofferdam seepage true value, so as to generate the prediction error loss term; The head boundary constraint term calculation submodule is used to calculate the head boundary constraint term based on the difference between the initial cofferdam seepage prediction value and the upstream tidal head value. The seepage continuity constraint term calculation submodule is used to calculate the difference between the two initial cofferdam seepage prediction values ​​and the reference hydraulic gradient at adjacent detection points, and to calculate the seepage continuity constraint term based on the ratio of the difference to the reference hydraulic gradient. The loss function calculation submodule is used to calculate the physical penalty loss function based on the weighted sum of the prediction error loss term, the head boundary constraint term, and the seepage continuity constraint term. The physical penalty loss function also includes a prediction error loss term.

3. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 2, characterized in that, The seepage continuity constraint term calculation submodule includes: The reference hydraulic gradient calculation unit is used to generate the seepage spatial mean by dividing the difference between the two historical cofferdam seepage values ​​at adjacent detection points by the distance between the adjacent detection points. The arithmetic mean of multiple seepage spatial means within a time window is calculated to generate the reference hydraulic gradient.

4. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 2, characterized in that, The loss function calculation submodule includes: The head boundary constraint weighting coefficient calculation unit is used to calculate the current head boundary constraint weighting coefficient by taking the maximum value of the head boundary constraint weighting coefficient and the current training period through the Sigmoid function. The loss function weighted calculation unit is used to calculate the physical penalty loss function by weighting and summing the head boundary constraint term and the seepage continuity constraint term based on the weighting coefficient of the current head boundary constraint term and the weighting coefficient of the set seepage continuity constraint term, and then adding the sum to the prediction error loss term.

5. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 1, characterized in that, The cofferdam dynamic monitoring and early warning module includes: The input layer operation submodule is used to process the real-time monitoring data of the construction status through the input layer to generate an input feature vector; The LSTM attention mechanism operation submodule is used to process the input feature vector through LSTM and the attention mechanism to generate percolation prediction features; The fully connected output layer computation submodule is used to pass the seepage prediction features through the fully connected output layer to generate the second predicted value of the cofferdam seepage. The optimized cofferdam seepage prediction model further includes an input layer and a fully connected output layer.

6. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 5, characterized in that, The LSTM attention mechanism operation submodule includes: The feature attention operation unit is used to perform attention weighting on key factors through the feature attention mechanism on the input feature vector to generate factor-weighted features; The LSTM encoder operation unit is used to pass the factor-weighted features through the LSTM encoder to generate time-series trend percolation features; The temporal attention operation unit is used to apply attention weights to key time sequences through the temporal attention mechanism to generate temporal weighted percolation features; The LSTM decoder operation unit is used to pass the time-series trend percolation features and time-series weighted percolation features through the LSTM decoder to generate the percolation prediction features; The LSTM includes an LSTM encoder and an LSTM decoder, and the attention mechanism includes a feature attention mechanism and a temporal attention mechanism.

7. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 1, characterized in that, The early warning MZ criterion parameter adjustment module includes: The robust scalar calculation submodule is used to iteratively calculate the deviation between the first predicted value of cofferdam seepage and the historical measured value of cofferdam seepage using the M estimator to generate a robust scalar. The predicted confidence interval radius calculation submodule is used to multiply the variance of multiple deviation values ​​by the standard normal distribution quantile to generate the predicted confidence interval radius; The MZ criterion threshold calculation submodule is used to perform a standardized calculation after weighted summation of the robustness scale and the predicted confidence interval radius to generate the MZ criterion threshold.

8. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to claim 1, characterized in that, The cofferdam dynamic monitoring and early warning module also includes: The MZ criterion anomaly index calculation submodule is used to calculate the absolute value of the difference between the real-time monitoring data of the construction status and the second predicted value of the cofferdam seepage, and to calculate the ratio of the absolute value of the difference to the MZ criterion threshold to generate the MZ criterion anomaly index. The level determination submodule is used to determine the safety construction warning level based on the comparison result between the abnormal index of the MZ criterion and the warning threshold.

9. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to any one of claims 1 to 8, characterized in that, The early stage of cofferdam construction refers to the early phase of steel sheet pile cofferdam construction, while the stable phase of cofferdam construction refers to the later stage of steel sheet pile cofferdam construction and the period during which no typhoon warning has been issued.

10. The dynamic monitoring and safety early warning construction system for cofferdams in a marine environment according to any one of claims 1 to 8, characterized in that, The cofferdams include sheet pile cofferdams and earth-rock slope cofferdams.

Images