Construction optimization method and system for underwater excavation

Abstract

The invention relates to the technical field of civil engineering construction optimization, in particular to a construction optimization method and system for underwater excavation. The method comprises the following steps: constructing a dynamic three-dimensional geological probability model through multi-source sensing data and a Bayesian updating algorithm; performing environment window dynamic evaluation by utilizing fuzzy logic and time sequence analysis to obtain a dynamic environment window coefficient and an operable probability time sequence; combining the dynamic three-dimensional geological probability model, the workable probability time sequence, the equipment real-time health degree index, the three-dimensional progress deviation diagram and the dynamic risk spectrum, and generating a dynamic construction scheme adjustment vector based on a reinforcement learning thought; based on the predictive maintenance time point, the optimal environment protection path, the dynamic environment window coefficient, the workable probability time sequence and the dynamic earthwork volume matrix, a refined scheduling scheme is obtained through resource collaborative scheduling optimization; and according to the optimal environment-friendly path, the dynamic construction scheme adjustment vector and the refined scheduling scheme, construction optimization of underwater excavation is achieved.

Description

Technical Field

[0001] This invention relates to the field of civil engineering construction optimization technology, specifically to a construction optimization method and system for underwater excavation. Background Technology

[0002] Deep foundation pit underwater excavation projects generally face a triple dilemma of uncontrolled efficiency, uncontrolled costs, and uncontrolled risks due to the enclosed working environment, complex and variable geological conditions, and the difficulty of coordinating multiple systems. Existing technologies have several limitations: First, geological modeling mostly adopts a static approach, relying on preliminary survey data, and cannot capture the dynamic changes in geological conditions such as hidden karst caves and sand lenses during construction in real time. This easily leads to mismatches between equipment selection and excavation parameters, resulting in problems such as dredger jamming, over-excavation, and under-excavation, resulting in low excavation efficiency. Second, environmental assessments use fixed thresholds to determine operational feasibility, which cannot adapt to the nonlinear fluctuations of environmental factors such as water flow and waves, and is even more difficult to predict future suitable environmental windows for operation, often leading to construction interruptions due to sudden environmental changes. Third, equipment maintenance is mainly based on periodic maintenance, which suffers from over-maintenance or under-maintenance, leading to problems under high-load operation of equipment. The following issues are addressed: First, equipment malfunctions can easily lead to construction halts, increasing maintenance costs and downtime losses. Second, earthwork volume calculation and progress tracking have low accuracy and significant lag, failing to reflect construction progress deviations and their causes in real time. Third, route planning only considers distance factors, neglecting environmental constraints related to suspended solids diffusion, which can easily cause pollution of surrounding water bodies. Fourth, risk assessments are mostly qualitative or semi-quantitative analyses of single risks, failing to consider the coupling effects between risks, resulting in highly unpredictable decision-making. Fifth, adjustments to construction plans rely on manual experience, leading to slow response, low accuracy, and an inability to adapt to dynamic changes in working conditions. Sixth, resource scheduling is prone to equipment idleness, process conflicts, and disconnects from construction technical requirements, resulting in low resource utilization.

[0003] While some existing technologies address the optimization of single aspects of underwater construction, such as static geological modeling, predictive equipment maintenance, and simple path planning, they lack a comprehensive dynamic optimization system covering the entire process from geological perception and environmental assessment to equipment operation and maintenance, schedule control, risk prevention, and scheme execution. Data from each stage is isolated, hindering deep collaboration and making it difficult to overcome the complex challenges of deep foundation pit underwater excavation. Therefore, there is an urgent need for an intelligent algorithm that can integrate technologies from multiple fields to achieve dynamic optimization throughout the entire process, thereby improving construction efficiency, reducing costs, and controlling risks. Summary of the Invention

[0004] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides a construction optimization method for underwater excavation, comprising the following steps: A dynamic 3D geological probability model is constructed using multi-source sensor data and a Bayesian update algorithm. Fuzzy logic and time series analysis are used to dynamically assess the environmental window and obtain a workability probability time series. Combining the dynamic 3D geological probability model, workability probability time series, real-time equipment health index, 3D progress deviation map, and dynamic risk spectrum, a dynamic construction plan adjustment vector is generated based on reinforcement learning. Based on predictive maintenance time points, the optimal environmental path, dynamic environmental window coefficients, workability probability time series, and a dynamic earthwork matrix, a refined scheduling scheme is obtained through resource collaborative scheduling optimization. Finally, underwater excavation construction is optimized based on the optimal environmental path, the dynamic construction plan adjustment vector, and the refined scheduling scheme.

[0005] Optionally, the step of constructing a dynamic three-dimensional geological probability model using multi-source sensor data and a Bayesian update algorithm includes the following steps: Based on multi-source sensor data, an initial three-dimensional geological model is generated through interpolation calculation; a Bayesian update model is constructed, and the initial three-dimensional geological model is dynamically updated through the Bayesian update model to obtain the dynamic three-dimensional geological probability model.

[0006] Optionally, the step of using fuzzy logic and time series analysis to dynamically evaluate the environmental window and obtain the workability probability time series includes the following steps: An environmental assessment index system and a fuzzy logic assessment model are constructed. By combining the environmental assessment index system and the fuzzy logic assessment model, a time series prediction model is trained to obtain the dynamic environmental window coefficient and the workability probability time series.

[0007] Optionally, the step of combining a dynamic three-dimensional geological probability model, a workability probability time series, a real-time equipment health index, a three-dimensional progress deviation map, and a dynamic risk spectrum to generate a dynamic construction plan adjustment vector based on reinforcement learning includes the following steps: By combining the dynamic three-dimensional geological probability model, the workability probability time series, the real-time equipment health index, the three-dimensional progress deviation map, and the dynamic risk spectrum, a reinforcement learning model framework is constructed; the reinforcement learning model framework is trained and optimized, and the dynamic construction scheme adjustment vector is generated through the optimized reinforcement learning model framework.

[0008] Optionally, analyzing the real-time health index of the device includes the following steps: Construct a digital twin model of the equipment; extract equipment health features based on the digital twin model, and calculate the real-time health index of the equipment based on the equipment health features.

[0009] Optionally, extracting the dynamic risk spectrum includes the following steps: The probability distribution of risk parameters is determined based on the risk identification and indicator system; a multi-risk coupling model is constructed, and the dynamic risk spectrum is analyzed by Monte Carlo simulation based on the multi-risk coupling model.

[0010] Optionally, the refined scheduling scheme obtained through resource collaborative scheduling optimization based on predictive maintenance time points, environmentally optimal paths, dynamic environmental window coefficients, workability probability time series, and dynamic earthwork volume matrix includes the following steps: Based on the predictive maintenance time point, environmentally optimal path, dynamic environmental window coefficient, workability probability time series, and dynamic earthwork volume matrix, a scheduling optimization objective and constraints are constructed; based on the scheduling optimization objective and constraints, a mixed integer linear programming model is solved to obtain a refined scheduling scheme.

[0011] Optionally, generating the environmentally friendly optimal route includes the following steps: The dispersion of suspended matter is simulated to obtain a heat map of suspended matter concentration distribution; an environmental cost function is constructed based on the suspended matter concentration distribution heat map, and the optimal environmental path is generated by combining the A* path optimization algorithm with the environmental cost function.

[0012] Optionally, the optimization of underwater excavation construction based on the environmentally optimal path, dynamic construction plan adjustment vector, and refined scheduling plan includes the following steps: A multi-objective evaluation system is constructed; the environmentally optimal path, dynamic construction scheme adjustment vector, and refined scheduling scheme are standardized; based on the multi-objective evaluation system and the standardized processing results, underwater excavation construction optimization is achieved through a multi-objective collaborative decision-making model.

[0013] This invention achieves dynamic collaborative optimization of multiple factors such as geology, environment, equipment, schedule, and risk through multi-source data fusion and multi-algorithm collaboration. It solves problems such as static solidification, single-link optimization, multi-objective conflict, one-sided risk assessment, and delayed scheme adjustment in existing technologies, thereby improving the efficiency, economy, and safety of deep foundation pit underwater excavation construction.

[0014] Secondly, to efficiently execute the underwater excavation optimization method provided by this invention, this invention also provides an underwater excavation optimization system, including a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory stores a computer program containing program instructions. The processor is configured to invoke the program instructions to execute the underwater excavation optimization method as described in the first aspect of this invention. This underwater excavation optimization system of the present invention has a compact structure and stable performance, and can stably execute the underwater excavation optimization method provided by this invention, further enhancing the overall applicability and practical application capability of this invention. Attached Figure Description

[0015] Figure 1 A flowchart of a construction optimization method for underwater excavation provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a construction optimization system for underwater excavation provided in an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 To address the above problems, this invention provides an optimized construction method for underwater excavation, such as... Figure 1 As shown, in one embodiment, the method includes the following steps: S1. Construct a dynamic three-dimensional geological probability model using multi-source sensor data and a Bayesian update algorithm.

[0019] The underwater geological conditions of deep foundation pits are highly uncertain. Considering the dynamic changes in geological information during construction, multi-source data fusion and Bayesian dynamic updating are adopted. Based on the initial geological survey data, geological parameters are dynamically corrected through real-time sensor data to construct a three-dimensional geological probability model that is iteratively optimized with the construction process, thereby realizing a closed loop of geological information perception-update-feedback.

[0020] Specifically, the construction of a dynamic three-dimensional geological probability model using multi-source sensor data and a Bayesian update algorithm includes the following steps: S11. Based on multi-source sensor data, an initial three-dimensional geological model is generated through interpolation calculation.

[0021] In this embodiment, multi-source sensor data includes basic static data such as preliminary geological survey reports (drilling data, core sample parameters), regional geological structure maps, and hydrogeological data (groundwater level, aquifer distribution); and real-time dynamic data such as underwater geological radar detection data, sonar imaging data, slag return parameters (particle size distribution, density) during the construction of bored piles, tunnel boring machine excavation parameters (thrust, torque, tunneling speed), and pore water pressure sensor data.

[0022] First, spatiotemporal alignment is performed: Spatial coordinates of each sensor are obtained using the GPS positioning module, and combined with millisecond-level timestamps, data from multiple sources such as underwater ground-penetrating radar, sonar imaging, and tunnel boring machine systems are anchored to a unified spatiotemporal coordinate system, avoiding biases in correlation analysis caused by data misalignment. To address noise in the ground-penetrating radar and sonar data caused by underwater electromagnetic wave attenuation and water flow interference, a 3-level wavelet decomposition using the db4 wavelet basis is employed. High-frequency noise components are removed using a thresholding method before signal reconstruction, resulting in clearer reflection characteristics of the geological interface. To eliminate the influence of different dimensions of data, Z-Score standardization is used to convert all data into standard data with a mean of 0 and a standard deviation of 1. For extreme outliers that may exist in the slag return parameters (particle size distribution, density) and tunneling parameters (thrust, torque) (such as a sudden increase in thrust caused by instantaneous equipment jamming), the 3σ criterion is used for identification and elimination: first, the mean μ and standard deviation σ of the parameters are calculated, and data that exceed the interval [μ-3σ, μ+3σ] are marked as outliers. Then, normal data from adjacent time points are used to complete the sequence by linear interpolation to ensure the continuity of the parameter sequence.

[0023] Based on the processed static data, an initial model reflecting macroscopic geological characteristics is constructed to provide a reliable prior basis for dynamic updates. In practice, the borehole data from previous explorations (including borehole coordinates and geological stratification information at each depth) and the physical and mechanical parameters of core samples (elastic modulus, internal friction angle, etc., measured through indoor triaxial tests) are first imported into GeoStudio software. Based on the regional geological structure map, the macroscopic trend of geological stratification is determined, and the variogram model of Kriging interpolation is set as a spherical model (to adapt to the continuity characteristics of soil layer spatial distribution), with the search radius set to 5m (to match the size of subsequent construction units). An initial three-dimensional geological model is generated through interpolation calculations. The spatial distribution range of clay, sand, and pebble layers is clearly marked in the model (presented in the form of isosurfaces). Initial values ​​of physical and mechanical parameters are assigned to each layer. For example, the elastic modulus of clay is 15 MPa, the internal friction angle is 18°, and the cohesion is 35 kPa. The elastic modulus of sand is 25 MPa, the internal friction angle is 30°, and the cohesion is 10 kPa. All parameter values ​​are calibrated with reference to the "Code for Design of Building Foundations" (GB50007-2011) to ensure the engineering applicability of the initial model.

[0024] S12. Construct a Bayesian update model, and dynamically update the initial three-dimensional geological model through the Bayesian update model to obtain the dynamic three-dimensional geological probability model.

[0025] The initial geological model parameters are used as the prior probability distribution P(θ), where θ is a high-dimensional geological parameter vector covering key parameters such as the thickness, elastic modulus, internal friction angle, and cohesion of each geological layer. The prior distribution of each parameter adopts a normal distribution, with the mean being the initial model value and the standard deviation being 15% of the initial value (reflecting the uncertainty of the exploration data). Real-time sensor data (such as tunnel boring machine thrust, torque, tunneling speed, and slag particle size distribution and density) are used as observations D. The correlation between the observation data and geological parameters is established by constructing a likelihood function P(D|θ). The core of the likelihood function construction is to train a random forest regression model: using matching samples of geological parameters and observation data from historical construction as the training set, selecting tunneling thrust, torque, and average slag particle size as input features, and geological layer thickness and elastic modulus as output targets; setting the number of decision trees in the random forest to 100, with a maximum depth of 8 layers per tree, and optimizing the model parameters through 5-fold cross-validation, the final model's goodness of fit R is determined. 2 The likelihood ratio reaches above 0.85. This regression model can be used to calculate the probability of observed data D occurring given geological parameters θ, thus completing the construction of the likelihood function P(D|θ).

[0026] Real-time iteration of the geological model is achieved based on Bayes' theorem, ensuring that the model remains synchronized with actual construction geological conditions. In the Bayesian theorem P(θ|D)=P(D|θ)P(θ) / P(D), P(D) is the marginal likelihood function, serving as a normalization constant to ensure that the posterior distribution satisfies the probability axiom. In practical calculations, it can be solved using numerical integration methods. The iteration trigger condition is the completion of a 5m×5m construction unit. This size setting considers both the continuity of equipment operation and the representativeness of real-time data.

[0027] The specific update process involves: collecting real-time sensor data within the current construction unit (such as the average thrust and torque of the tunnel boring machine within the unit, and the corresponding particle size distribution analysis results of the returned slag), substituting this data into a trained random forest model to obtain the likelihood function value P(D|θ), and combining it with the prior distribution P(θ) to calculate the posterior probability distribution of geological parameters using Bayes' theorem. For example, if the average thrust of the tunnel boring machine in a certain construction unit is 20% higher than the initial model's expected value, and the pebble content in the returned slag increases, then the thickness parameter of the pebble layer in that area in the plateau model will be adjusted through posterior calculation, and the standard deviation of its probability distribution will be increased (reflecting the uncertainty of changes in geological conditions). The updated 3D geological model will simultaneously correct the spatial distribution boundaries and parameter values ​​of each geological layer, and the posterior distribution obtained in this calculation will be directly used as the prior distribution for the next construction unit, forming a cyclical iterative mechanism of construction-observation-update.

[0028] The final output dynamic three-dimensional geological probability model contains three core components: First, the spatial coordinate information of each geological stratum, with the geological category of each grid unit (0.5m×0.5m×0.5m) marked in the form of a three-dimensional grid; second, the physical and mechanical parameters of each grid unit and their probability distribution, such as the elastic modulus of a sand layer marked in a grid unit being 28±3MPa (95% confidence interval), clearly reflecting the uncertainty range of the parameters; and third, the marking of geological anomaly areas, such as hidden caves (marked with the center coordinates and approximate volume of the caves) and special geological bodies such as sand layer lenses identified through abrupt changes in parameter distribution.

[0029] This invention transforms construction process data, such as tunnel boring machine parameters and muck return characteristics, into geological sensing data, achieving synchronization between geological exploration and construction progress. Through Bayesian updates, it enables probabilistic expression of geological parameters, accurately quantifying geological uncertainties and providing suitable input for subsequent risk assessment. By capturing real-time changes in geological conditions during construction (such as hidden karst caves and sand lens bodies), it provides precise geological support for equipment selection and excavation parameter settings, reducing problems such as equipment jamming, over-excavation, and under-excavation caused by geological misjudgments, thereby lowering construction risks and rework costs.

[0030] S2. Use fuzzy logic and time series analysis to dynamically evaluate the environmental window and obtain the workability probability time series.

[0031] Underwater construction in deep foundation pits is dynamically affected by environmental factors such as water flow velocity, wave intensity, water temperature, and water quality (including sediment concentration). By combining the uncertainty reasoning of fuzzy logic with the trend prediction capability of time series analysis, the environmental suitability is dynamically assessed, and the workability probability time series is output, providing a time basis for construction scheduling.

[0032] Specifically, the method of using fuzzy logic and time series analysis to dynamically evaluate the environmental window and obtain the dynamic environmental window coefficient and workability probability time series includes the following steps: S21. Construct an environmental assessment indicator system and a fuzzy logic assessment model.

[0033] In this embodiment, four major indicators—water flow velocity, wave height, suspended solids concentration, and visibility—are selected to construct an environmental assessment index system. Water flow velocity directly affects the positioning accuracy of equipment and excavation resistance; wave height determines the stability of the construction platform; suspended solids concentration is related to environmental compliance and equipment heat dissipation; and visibility affects underwater monitoring and operational safety.

[0034] Furthermore, the weights were determined using the analytic hierarchy process (AHP): ① Establish a hierarchical structure: the target layer is environmental suitability assessment, the criterion layer is four major indicators, and the solution layer is suitable operation, cautious operation, and prohibited operation; ② Construct a judgment matrix: Invite senior underwater construction engineers to compare the importance of indicators pairwise, such as the impact of water flow velocity on construction being 2.5 times that of visibility, to form a consistency matrix; ③ Consistency check and weight calculation: The matrix weight vector is solved by the eigenvalue method. The consistency ratio CR = 0.042 < 0.1, which meets the check requirements. In the example, the weights of water flow velocity, wave height, suspended solids concentration, and visibility are finally determined to be 0.4, 0.25, 0.2, and 0.15, thereby constructing a three-in-one indicator system of indicators, weights, and evaluation standards.

[0035] To address the nonlinear fuzzy relationship between environmental indicators and construction suitability, a fuzzy logic evaluation model is constructed in three stages: ① Determine the fuzzy input: First, clarify the quantitative range of each indicator and the correspondence between the fuzzy linguistic variables. Refer to the "Safety Standard for Waterway Engineering Construction" (JTS205-2021) to set the following: Water flow velocity (m / s) is divided into weak (0-0.5), medium (0.5-1.0), strong (1.0-1.5), and extremely strong (1.5+); wave height (m) is divided into light waves (0-0.3), small waves (0.3-0.6), medium waves (0.6-0.9), and large waves (0.9+); suspended solids concentration (mg / L) is divided into low (0-200), medium (200-500), high (500-800), and extremely high (800+); visibility (m) is divided into good (3+), fair (1-3), poor (0.5-1), and very poor (0.5-). Furthermore, a triangular membership function is used for fuzzification. Taking water flow velocity as an example, the membership function expressions are μ(x)=0 (x≤0.5), μ(x)=(x-0.5) / 0.5 (0.5<x<1.0), and μ(x)=(1.0-x) / 0.5 (x≥1.0), which accurately characterizes the degree of membership of the index value to the fuzzy language.

[0036] ② Establish a fuzzy rule base: Following the logic of multi-indicator combination-suitability output, rules are constructed by combining construction cases and expert experience to cover all indicator combination scenarios. The rule expression adopts a if-then structure. For example, if the water flow velocity is weak (0-0.5m / s), the wave height is small (0.3-0.6m), the suspended solids concentration is low (0-200mg / L), and the visibility is good (3+m), then the environmental suitability is extremely high. If the water flow velocity is strong (1.0-1.5m / s) and the wave height is medium (0.6-0.9m), then the environmental suitability is low. Each rule is marked with a confidence level (≥0.9).

[0037] ③ Fuzzy Reasoning and Defuzzification: The Mamdani reasoning method is adopted. First, the premise aggregation of fuzzy rules is achieved through minimum operation. Then, the fuzzy output set of environmental suitability (divided into 5 levels: extremely high, high, low, and extremely low) is calculated through maximum-minimum synthesis. Defuzzification adopts the centroid method, which transforms the fuzzy set into a precise environmental window coefficient C of 0-1, which directly reflects the degree of job suitability of the current environment.

[0038] S22. Combining the environmental assessment index system and the fuzzy logic assessment model, train the time series prediction model to obtain the workability probability time series.

[0039] The training of time series forecasting models aims to accurately capture the temporal variation patterns of environmental indicators, especially eliminating the interference of tidal cycles on forecast accuracy. The data base consists of hourly environmental monitoring data from the same period of the past year and real-time data from the 30 days prior to the construction period. During data preprocessing, missing values ​​are first filled in using linear interpolation, and then the STL decomposition method is used to separate the trend term, periodic term, and residual term of the indicators.

[0040] The construction process of the Time Series Forecasting Model (ARIMA) is as follows: ①Stationarity test: Perform ADF test on each environmental indicator series. If it is not stationary (P>0.05), perform differencing (d value is 1 or 2) until the series is stationary; ② Model order determination: The orders of AR(p) and MA(q) are determined by the autocorrelation function (ACF) and partial autocorrelation function (PACF). For example, the ACF of the water flow velocity sequence is truncated at order 2 and the PACF is truncated at order 3, thus determining p=3 and q=2. ③ Parameter estimation and validation: The maximum likelihood method is used to estimate the model parameters. The optimal model (with the smallest AIC value) is selected by the Akaike Information Criterion (AIC). The rolling prediction method is used for model validation. The data of the first 25 days is used as the training set and the data of the last 5 days is used as the test set to ensure that the prediction error RMSE < 0.1 (water flow velocity) and MAE < 0.05 (wave height).

[0041] For the tidal cycle, the tidal cycle is transformed into two dummy variables (sin(2πt / 12.42) and cos(2πt / 12.42), where t is the number of hours), which are then incorporated into the ARIMA model as additional input features to form the ARIMA-tidal factor composite model.

[0042] The environmental index values ​​predicted by the ARIMA-tidal factor composite model for the next 24 hours are input one by one into the constructed fuzzy logic model to calculate the environmental window coefficients at the corresponding times, forming an initial suitability sequence. Subsequently, probability correction is performed by combining the differentiated environmental thresholds of construction procedures: different procedures have different tolerances to the environment. For example, the maximum allowable water flow velocity for excavation is 1.2 m / s, for slag removal it is 0.8 m / s, and for support it is 0.6 m / s. Therefore, the environmental window coefficients need to be converted into workable probabilities for the current construction task type. If the window coefficient is ≥0.8, the workable probability is taken as 0.9-1.0; if the coefficient is 0.5-0.8, it is taken as 0.5-0.9; if the coefficient is 0.2-0.5, it is taken as 0.1-0.5; and if the coefficient is <0.2, it is taken as 0. For example, at 10:00, the predicted water flow velocity is 0.4 m / s, wave height is 0.4 m, suspended solids concentration is 180 mg / L, and visibility is 3.5 m. The fuzzy logic output window coefficient is 0.92, corresponding to a workable probability of 0.95 for the excavation process. At 14:00, due to high tide, the predicted water flow velocity is 1.3 m / s, the window coefficient drops to 0.28, and the workable probability is adjusted to 0.25.

[0043] Finally, the 95% confidence interval is calculated: Based on the normal distribution characteristics of the model prediction error (mean is 0, standard deviation is 10% of the predicted value), the confidence interval is determined by the interval estimation formula P(C-1.96σ<Cactual<C+1.96σ)=95%. For example, the probability of work being possible at 10:00 is marked as 0.95 (0.92-0.98), and at 14:00 it is marked as 0.25 (0.18-0.32), clearly reflecting the range of uncertainty of the probability and providing a dual reference of probability and risk for the formulation of construction plans.

[0044] This invention uses fuzzy logic to handle the nonlinear and fuzzy correlation between environmental indicators and construction suitability, combining environmental assessment with time series forecasting. It not only determines whether work is currently feasible but also predicts when work will be suitable in the future, enabling forward-looking planning of construction scheduling. By dynamically capturing real-time changes and future trends of environmental factors, it accurately identifies suitable environmental windows for construction, avoiding construction interruptions or safety accidents caused by sudden environmental changes, and improving the flexibility and feasibility of construction plans.

[0045] S3. Combining dynamic three-dimensional geological probability model, workability probability time series, real-time equipment health index, three-dimensional progress deviation map and dynamic risk spectrum, dynamic construction scheme adjustment vector is generated based on reinforcement learning.

[0046] Underwater excavation equipment (such as tunnel boring machines) operates in high-load and highly corrosive environments, and equipment failures can easily lead to construction stoppages. By constructing a digital twin of the equipment, a real-time mapping between the physical equipment and the virtual model can be achieved. Combined with predictive maintenance algorithms, the deterioration trend of the equipment can be identified in advance, and the real-time health status and maintenance time points can be output to ensure continuous operation of the equipment.

[0047] Specifically, analyzing the real-time health index of the device includes the following steps: S311. Construct a digital twin model of the equipment.

[0048] First, a joint modeling approach using Autodesk Revit and Unity was employed: In Revit, a 1:1 scale 3D geometric model of the equipment was constructed, with geometric errors controlled within ±0.1mm. This model fully recreated the core components of the tunnel boring machine, including its body structure, excavating device, and hydraulic system, with particular attention to detailing the dimensions and installation locations of vulnerable components such as bearings and seals. After importing into Unity, material rendering and kinematic constraints were added, enabling the virtual model to simulate the equipment's excavation and rotational movements. Second, an integrated equipment dynamics model was developed. Based on Newton's second law, a quantitative correlation was established between excavation force and geological hardness: F_excavation = K × σ_geology × S_contact, where F_excavation is the output force of the excavating device (in kN), K is the equipment power coefficient (calibrated according to the equipment model, e.g., K=0.85 for a WZ300 dredger), σ_geology is the geological hardness (a real-time parameter from the dynamic geological model, in MPa), and S_contact is the contact area between the bucket and the soil / rock (in m²). 2 This system enables real-time linkage between geological conditions and equipment stress. The sensor data interface is built using the OPCUA industrial communication protocol, supporting real-time transmission of 16 types of data, including motor current (accuracy ±0.1A), hydraulic pressure (accuracy ±0.05MPa), and vibration frequency (sampling rate 1000Hz). Simultaneously, digital sensors corresponding one-to-one with physical sensors are set in the virtual model to achieve synchronous mapping of operating parameters, location information (based on GPS+IMU fusion positioning), and operating conditions. Finally, a component-level correlation model is established: a directed graph structure is used to label the logical relationships of each component, such as the power transmission path from motor to hydraulic pump to hydraulic motor to excavator bucket, and the fault transmission path from abnormal bearing temperature to hydraulic system pressure fluctuations, providing a correlation basis for subsequent degradation location.

[0049] S312. Extract equipment health features based on the equipment digital twin model, and calculate the real-time health index of the equipment based on the equipment health features.

[0050] First, health features are extracted. For real-time operating data, in addition to calculating the mean and variance, time-domain analysis adds four new features: peak factor (peak value / effective value) and kurtosis (reflecting the steepness of the data distribution). For example, if the kurtosis of bearing vibration data is greater than 3, it indicates potential wear. Frequency-domain analysis uses Fast Fourier Transform (FFT) to convert the vibration signal to the frequency domain, extracting characteristic frequencies (e.g., characteristic frequency of motor bearing = motor speed × number of bearing balls / (60 × bearing pitch circle diameter)), spectral peaks and harmonic components, and identifying fault features such as frequency shift and the appearance of new frequencies.

[0051] By fusing time-domain and frequency-domain features to construct a device health feature vector, and then using principal component analysis (PCA) to reduce dimensionality and lower model computation costs, a SVM degradation classification model is constructed: valid samples are selected based on historical device data, and sample labels are determined by combining fault diagnosis reports and expert annotations. The SMOTE algorithm is used for sample preprocessing to address class imbalance. The radial basis function (RBF) is selected as the model kernel function, and the penalty coefficient C=10 and kernel function parameter γ=0.1 are optimized using a grid search method.

[0052] Furthermore, the degradation stage classification adopts a dual criterion of feature threshold and trend change: normal state (all feature values ​​are within the equipment's factory threshold range, with no obvious trend of change), slight degradation (1-2 features are close to the threshold, weekly change rate <5%), moderate degradation (2-3 features exceed the threshold by less than 10%, weekly change rate 5%-15%), and severe degradation (3 or more features exceed the threshold by 10%, weekly change rate >15%), ensuring that the classification results are consistent with the actual engineering situation.

[0053] By scientifically assigning weights and using quantitative scoring, we achieve intuitive quantification and tiered early warning of equipment health status. The weights are determined using the entropy weight method, based on the information entropy of the health characteristics of each component (reflecting data uncertainty). The lower the information entropy, the greater the contribution of the characteristic to the health assessment, and the higher the weight.

[0054] The specific calculation process is as follows: ① Standardize the health characteristic data of each component (positive index x'=x / x_max, negative index x'=x_min / x); ② Calculate the information entropy Hj of the j-th component; ③ Calculate the weight wj=(1-Hj) / ∑(1-Hj). In the embodiment, after correction based on engineering experience, the final weights are determined as follows: motor 0.3 (core power source, with the widest impact of failure), hydraulic system 0.28 (controls digging actions, susceptible to underwater corrosion), bearing 0.22 (high-speed rotating component, frequent wear and failure), and digging device 0.2 (directly in contact with rock and soil, subject to complex forces).

[0055] The health score conversion uses interval mapping and trend correction rules: Normal state (no abnormal features) is directly assigned 100 points; slight degradation is assigned 70-90 points based on the degree of feature exceeding the threshold (e.g., bearing temperature close to the threshold but without trend change, assigning 85 points); moderate degradation is assigned 40-69 points (e.g., hydraulic pressure fluctuation exceeding the threshold by 5%, assigning 55 points); severe degradation is assigned 0-39 points (e.g., abnormal motor current fluctuation exceeding the threshold by 15%, assigning 30 points). If a component's health feature shows an accelerated degradation trend (e.g., daily decline rate > 3%), an additional 10-20 points are deducted from the base score. The overall equipment health index is calculated using a weighted summation formula: H = ∑(wj × hj), where H is the overall health score (0-100 points), and hj is the health score of the j-th component. Warning thresholds are set based on equipment failure statistics: scores below 60 points indicate a need for attention (historical data shows that the probability of equipment failure in this range rises to 15%), triggering encrypted monitoring; scores below 40 points indicate a need for maintenance, and maintenance reminders are directly pushed to the management platform.

[0056] Furthermore, through trend modeling and lifespan prediction, accurate prediction and prioritization of maintenance plans are achieved. First, a degradation rate model is constructed: using the health index of the past 7 days as the base data, linear regression is used to fit the degradation curve. The model expression is H(t) = a × t + b, where t is time (days), H(t) is the health score at time t, a is the daily average degradation rate (slope, negative values ​​indicate degradation), and b is the initial score (intercept). To improve fitting accuracy, weighted linear regression is used, with the weight of the most recent 3 days of data set at 0.6 and the weight of the earliest 4 days of data set at 0.4, ensuring that the model focuses on the latest degradation trend and achieves a goodness-of-fit R-value. 2 The value must be ≥0.8; otherwise, a quadratic regression model should be used for correction.

[0057] The remaining lifespan is calculated using a segmented formula: when the health score is between 60 and 100 (before requiring attention), the remaining lifespan T1 = (60 - H current) / |a|; when the score is between 40 and 60 (the interval requiring attention), the remaining lifespan T2 = (40 - H current) / |a| (the maintenance threshold is set to 40); if the score is < 40, the remaining lifespan is set to 0, triggering emergency maintenance. For example, if a motor currently has a health score of 52 and an average daily degradation of 0.8 points / day, then T2 = (40 - 52) / 0.8 = 15 days, predicting that maintenance is required after 15 days. The maintenance time needs to be adjusted based on the construction plan and environmental window: if the predicted maintenance time falls during an unsuitable working period (probability of operation < 0.6), the maintenance plan should be adjusted 1-2 days in advance; if this period is a critical construction node, degradation can be mitigated by reducing equipment load (e.g., reducing excavation speed by 0.1 m / s), and the maintenance time can be postponed until after the node ends. Finally, maintenance priorities are ranked: a two-dimensional matrix of component importance and remaining life is used. Core components (motors, hydraulic pumps) with a remaining life of less than 10 days are listed as the highest priority (maintenance to be arranged within 24 hours); important components (bearings) with a remaining life of less than 7 days are listed as high priority; and auxiliary components (such as cooling systems) with a remaining life of less than 5 days are listed as ordinary priority, ensuring that limited maintenance resources are focused on key risk points.

[0058] This invention integrates geological hardness from a dynamic geological model with equipment operating data to construct a correlation model of operating conditions, loads, and health. Then, it integrates maintenance time prediction functions into a digital twin to achieve integrated status monitoring, fault early warning, and maintenance planning. This enables real-time monitoring of equipment status and fault early warning, transforming the maintenance mode from periodic and passive to predictive and proactive, reducing unplanned equipment downtime, lowering maintenance costs (avoiding over-maintenance), and reducing failure losses (avoiding construction interruptions).

[0059] Deep foundation pit underwater excavation faces multiple risks, including geological risks (such as mudslides and water inrushes), equipment risks (such as machine downtime due to malfunctions), environmental risks (such as strong water flow), and environmental protection risks (such as excessive suspended solids). Furthermore, there are coupling effects between these risks (such as excessive geological hardness leading to equipment failure and project delays). By employing the probabilistic statistical capabilities of Monte Carlo simulation and combining it with real-time data to dynamically update risk parameters, we can achieve multi-risk coupling assessment and output a dynamic risk spectrum.

[0060] Specifically, extracting the dynamic risk spectrum includes the following steps: S321. Determine the probability distribution of risk parameters based on the risk identification and indicator system.

[0061] To construct a risk identification and indicator system, the following steps were taken: First, by searching literature on deep foundation pit underwater excavation projects over the past 10 years, high-frequency risk types such as geology, equipment, and environment were identified. Second, based on typical accident cases (such as a mudslide and water inrush accident in a cross-sea tunnel and an equipment jamming accident in a wharf project), accident causes and loss characteristics were extracted. Finally, a review group composed of geological survey experts, equipment operation and maintenance experts, and environmental engineers was organized to complete risk screening and classification using the Delphi method. In the embodiment, 5 primary risk categories and 12 secondary risk categories were identified, as follows: ① Geological risk (accounting for 35%, as geological abrupt changes are the core risk source): This includes 3 secondary categories: mudslide and water inrush (triggering condition: pore water pressure > 0.8 MPa and cohesion < 20 kPa), sand liquefaction (triggering condition: earthquake intensity ≥ 6 degrees or vibration frequency matching the liquefaction cycle), and hidden karst caves (triggering condition: karst cave diameter > 1 m and distance from the excavation face < 3 m); ② Equipment risk (accounting for 25%): This includes equipment downtime (triggering condition: health index < 40 points) and a sharp drop in excavation efficiency (triggering condition: actual efficiency < planned efficiency 60%). The risk assessment is categorized into three secondary indicators: ① **Risk Risk (15%):** ③ **Environmental Risk (15%):** Subcategories include strong water flow (trigger condition: flow velocity > 1.5 m / s) and extreme waves (trigger condition: wave height > 1.2 m); ④ **Environmental Protection Risk (15%):** Subcategories include suspended solids exceeding standards (trigger condition: concentration > 1.2 times the GB3838-2002 limit) and water pollution (trigger condition: detection of oil pollutants); ⑤ **Schedule Risk (10%):** Subcategories include key node delays (trigger condition: delay duration > 24 hours). Quantitative assessment standards are established for each indicator category. For example, the risk of mudslides and water inrushes is classified into extremely high (probability > 30%), high (15%-30%), medium (5%-15%), low (1%-5%), and extremely low (< 1%) levels to ensure that the risk assessment is quantifiable and comparable.

[0062] Furthermore, following the principle of prioritizing real-time data and supplementing it with historical data, we select an appropriate distribution model based on the statistical characteristics of risk types to determine the probability distribution of risk parameters.

[0063] ① Geological risk: Beta distribution is adopted (parameters α and β are calculated by drilling data and real-time geological model). For example, the distribution parameters of mudslide and water inrush risk are α=4.2 and β=12.6. The advantage of Beta distribution is that it can accurately depict the probability distribution in the interval [0,1], which is suitable for the probabilistic characteristics of geological risk. ② Equipment risk: The exponential distribution (failure rate λ is calculated based on the time between failures (MTBF)) is adopted. The exponential distribution can effectively reflect the memoryless characteristics of equipment failure. ③ Environmental risk: A normal distribution is adopted (mean μ is the real-time monitoring mean, and standard deviation σ is 1.2 times the standard deviation of the monitoring data). For example, if the water flow velocity monitoring mean is 0.8 m / s and the standard deviation is 0.2 m / s, then the distribution parameters μ=0.8 and σ=0.24. The normal distribution is suitable for the continuous fluctuation characteristics of environmental indicators. ④ Environmental risks: A binomial distribution is adopted (the success probability p is the frequency of exceeding the standard). For example, if 15 out of 100 suspended solids monitoring tests exceed the standard, then p=0.15. The binomial distribution is suitable for the binary risk state of exceeding the standard / not exceeding the standard. ⑤ Schedule risk: The Weibull distribution (shape parameter k=2.3, scale parameter η=48 hours, based on the lag data of 50 key nodes) is adopted. The Weibull distribution can flexibly characterize the different stages of schedule lag.

[0064] All distribution parameters are updated every 2 hours based on real-time data from the preceding modules. For example, geological risk parameters are iteratively corrected with the dynamic geological model, and equipment risk parameters are adjusted synchronously with the health index to ensure that the distribution model is consistent with the actual working conditions.

[0065] S322. Construct a multi-risk coupling model, and based on the multi-risk coupling model, analyze the dynamic risk spectrum through Monte Carlo simulation.

[0066] The Copula function is introduced to solve the problem of nonlinear correlation between risks. By quantifying the linkage effect of risks through coupling coefficients, it breaks through the limitations of traditional independent risk assessment.

[0067] ① Copula function selection: Select the appropriate function based on the correlation characteristics between risks. Geological risks and equipment risks are upper-tail correlated (geological changes can easily lead to equipment failures), so the GumbelCopula function is selected; environmental risks and environmental protection risks are symmetrically correlated, so the GaussianCopula function is selected; schedule risks and other risks are lower-tail correlated (multiple risks can easily lead to schedule delays), so the ClaytonCopula function is selected.

[0068] ② Coupling Coefficient Calculation: Taking geological risk (Pi) and equipment risk (Pj) as examples, the rank correlation coefficient τ = 0.62 is first calculated based on historical data (obtained through Spearman correlation analysis). Then, using the parameter correspondence θ = 1 / (1-τ) of Gumbel Copula, the coupling coefficient Cij = θ = 2.63 is calculated (the larger the coupling coefficient, the stronger the risk correlation). Finally, a 5×5 risk coupling matrix is ​​constructed. The core coupling relationships include: geology-equipment (C = 0.6), environment-environmental protection (C = 0.55), and geology-progress (C = 0.45). Weak coupling relationships (C < 0.2), such as equipment-environmental protection risk, can be simplified.

[0069] ③ Multi-risk joint probability model: Based on the Copula function, the marginal distributions are merged into a joint distribution, which satisfies: the total risk-free probability Ptotal = the product of the independent risk-free probabilities of each risk.

[0070] Furthermore, Monte Carlo simulation analysis is conducted to achieve quantitative coupling of risk probability and loss through large-sample random simulation, providing a probability-loss dual-dimensional basis for risk decision-making.

[0071] ① Simulation parameter settings: In this example, the number of simulations is set to 10,000, the random number generation adopts the MersenneTwister algorithm, and the time step of each simulation is matched with the construction unit.

[0072] ② Simulation process: First, generate Pi values ​​randomly based on the probability distribution of each risk (e.g., generate geological risk probabilities by generating random numbers from the Beta distribution), and substitute them into the coupled model to calculate P_total; second, combine the risk loss matrix to calculate the total loss value L = ∑(Pi × Li × ∑(Cij × Pj)), where Li is the baseline loss of the i-th type of risk.

[0073] ③ Statistical Analysis of Results: The simulated P total and L values ​​were statistically analyzed, and a risk probability-loss biaxial distribution curve was plotted. The horizontal axis represents the total risk probability (0-1), and the vertical axis represents the loss value. The peak value of the curve corresponds to the most likely risk scenario (e.g., P total = 0.18, L = 1.2 million yuan). Simultaneously, the Value at Risk (VaR) index was calculated. For example, at a 95% confidence level, VaR = 1.8 million yuan, meaning there is a 95% certainty that the construction risk loss will not exceed 1.8 million yuan; the extreme risk value (ES) = 2.2 million yuan, representing the average loss under extreme risk scenarios, providing a basis for risk reserve provisioning.

[0074] Furthermore, the dynamic changes of risk are presented in a two-dimensional visualization format of time and risk, enabling precise risk positioning and targeted prevention and control.

[0075] ① Risk level determination: The probability-loss two-factor matrix method is used to classify the risk level into 1-5 levels: Level 1 (extremely low risk: P≤5% and L≤500,000 yuan), Level 2 (low risk: 5%<P≤10% or 500,000 yuan<L≤1,000,000 yuan), Level 3 (medium risk: 10%<P≤20% and 1,000,000 yuan<L≤2,000,000 yuan), Level 4 (high risk: 20%<P≤30% or 2,000,000 yuan<L≤3,000,000 yuan), and Level 5 (extremely high risk: P>30% and L>3,000,000 yuan).

[0076] ② Dynamic Risk Spectrum Construction: The horizontal axis is divided into three stages according to the construction progress: preliminary preparation (days 1-3), core excavation (days 4-15), and final cleaning (days 16-20). The vertical axis represents the risk level (levels 1-5), using color coding (green level 1, blue level 2, yellow level 3, orange level 4, red level 5) to mark the risk status of each time period. For example, on the 8th day of core excavation, the dynamic geological model monitors that the cohesion in a certain area drops to 18 kPa. Combined with the strong water flow (velocity 1.6 m / s) predicted by the environmental module, the simulation yields P total = 28% and L = 2.6 million yuan, with the risk level set at level 4, marked in orange in the spectrum.

[0077] ③ Dominant Risk Identification and Prevention Recommendations: Dominant risks are identified through risk contribution analysis (the proportion of total loss caused by a particular risk). For example, if the geological risk contribution on day 8 is 65% and the environmental risk contribution is 35%, the dominant risk is mudslide or water inrush. The corresponding prevention recommendations are to immediately reduce the excavation speed to 0.1 m / s, increase pore water pressure monitoring (sampling rate increased to 10 Hz), and pre-cast concrete grout-stopping walls. In this example, the risk spectrum is updated along with the simulation results and simultaneously pushed to the project management platform. High-risk (level 4-5) states automatically trigger warning SMS messages to the project manager and supervisor.

[0078] This invention uses real-time data as risk parameter input to achieve data-driven dynamic updates for risk assessment. By quantifying the coupling relationship between risks through the Copula function, it accurately quantifies the coupling effect and dynamic changes of multiple risks, providing targeted suggestions for construction risk prevention and control, and reducing the blindness of risk decision-making.

[0079] During underwater construction of deep foundation pits, dynamic changes in geological, environmental, and equipment conditions can cause the initial construction plan to fail. Traditional methods that rely on manual experience to adjust the plan are slow and inaccurate. Based on the trial-and-error learning and dynamic decision-making capabilities of reinforcement learning, an intelligent decision-making model is constructed with construction efficiency, cost, and risk as reward objectives. This model dynamically outputs adjustment vectors for the construction plan based on real-time working conditions, achieving adaptive optimization of the plan.

[0080] Specifically, the process of generating a dynamic construction plan adjustment vector based on reinforcement learning, by combining a dynamic three-dimensional geological probability model, a workability probability time series, a real-time equipment health index, a three-dimensional progress deviation map, and a dynamic risk spectrum, includes the following steps: S331. Combining the aforementioned dynamic three-dimensional geological probability model, workability probability time series, real-time equipment health index, three-dimensional progress deviation map, and dynamic risk spectrum, a reinforcement learning model framework is constructed.

[0081] First, the state space S is designed: a state vector is constructed, with each dimension clearly corresponding to the core parameters of the preceding steps. Specifically, it is broken down into: a geological feature sub-vector (including the geological hardness, cohesion, pore water pressure, layer thickness, and corresponding probability standard deviation of the current working area), an environmental state sub-vector (including environmental window coefficient, workability probability, real-time flow velocity, and wave height), an equipment state sub-vector (including the health index of motors / hydraulic systems / bearings and overall equipment efficiency), a progress state sub-vector (including cumulative deviation rate, daily deviation, efficiency deviation coefficient, and lag time of key nodes), and a risk state sub-vector (including the real-time probability of 5 types of primary risks and the coupling coefficient of 4 types of core risks). All features are standardized using Min-Max (mapped to the [0,1] interval) to eliminate dimensional differences. Among them, interval data such as risk probability are standardized in segments to improve the feature discrimination of extreme risks.

[0082] Then, the action space A is defined as follows: Based on the adjustable parameters of construction, eight specific actions are identified and rigid constraints are set to form a discrete-continuous hybrid action vector: a1-excavation speed adjustment (-0.2~+0.2m / s, step size 0.05m / s, upper limit not exceeding the rated speed of the equipment 0.5m / s), a2-excavation depth adjustment (-0.5~+0.5m, step size 0.1m, lower limit not lower than the design pit bottom elevation), a3-process priority adjustment (3 combinations: excavation → cleaning → support / excavation → support → cleaning / The process is as follows: a4-Slag removal → excavation → support, which must meet the logical constraints of the process; a5-Equipment selection adjustment (equipment switching, based on geological hardness matching); a6-Slag discharge parameter adjustment (slag discharge pump power ±10%, related to suspended solids diffusion control requirements); a7-Work area switching (switching between adjacent work units, which must conform to the environmentally optimal path of path planning); a8-Monitoring frequency adjustment (sensor data sampling rate ±500Hz, forced increase when risk level ≥4); a9-Load distribution adjustment (load ratio ±15% when multiple equipment work together).

[0083] Third, the reward function R is constructed using a basic reward plus dynamic adjustment mechanism. The basic formula is R = 0.4 × R_efficiency + 0.3 × R_cost + 0.3 × R_risk. The calculation logic for each item is as follows: R_efficiency = min(actual excavation efficiency / planned efficiency, 1.5). When the efficiency is ≥ 1.2 times the planned value, an additional reward of 0.2 is added (to encourage efficient operation), and when it is less than 0.8 times, a reward of 0.3 is deducted; R_cost = min(planned unit earthwork cost / actual unit earthwork cost, 1.4). When the cost is more than 10% lower than the planned value, a reward of 0.15 is added, and when it is more than 20% lower, a reward of 0.25 is deducted; R_risk = (5 - risk level) / 5. If a risk of level 4-5 occurs and is avoided in time, an additional reward of 0.3 is added. The dynamic adjustment rules for weights are as follows: In the early stage of construction (the first 30% of the construction period), the efficiency weight is increased to 0.5, and the cost weight is reduced to 0.2; during environmentally sensitive periods (such as the flood season), the environmentally related risk weight is increased to 0.4, and the efficiency weight is reduced to 0.3.

[0084] S332. Train and optimize the reinforcement learning model framework, and generate the dynamic construction scheme adjustment vector through the optimized reinforcement learning model framework.

[0085] In this embodiment, the training data is based on historical data from similar deep foundation pit underwater excavation projects, forming a state-action-reward-next state sample set. The sample time span covers different tidal cycles, geological types (clay / sand / pebble layers), and extreme environmental scenarios (typhoons, strong currents). During sample preprocessing, abnormal data (such as invalid actions caused by equipment failure) are removed, and the SMOTE algorithm is used to expand the samples for low-probability risk scenarios to ensure a balanced sample distribution.

[0086] Furthermore, the DQN model structure and parameters were set: the network was constructed using the TensorFlow 2.x framework, with 3 hidden layers (128 neurons → 64 neurons → 32 neurons), and ReLU, LeakyReLU, and ReLU activation functions respectively. The output layer was an 8-dimensional action Q-value (using a linear activation function to avoid limiting the output range). Core training parameters were set as follows: the experience replay buffer capacity was set to 100,000 sets, the batch size was 64, the initial learning rate was 0.001 (dynamically decaying using cosine annealing, minimum 0.0001), the discount factor γ = 0.9 (balancing immediate rewards and long-term benefits), and the target network update cycle was 200 steps (synchronizing the target network every 200 parameter updates).

[0087] Third, constraint integration and stability optimization: Hard construction constraints are transformed into model regularization terms. For example, equipment speed constraints are implemented by adding an L2 regularization term (regularization coefficient λ=0.01) to the loss function, which is the difference between the action value and the rated value. Environmental standards (such as the limit for suspended solids concentration) are reflected in the penalty term of the reward function (if an action causes the environmental risk level to increase by 1 level, 0.5 of the base reward is deducted). A dual DQN strategy is adopted to avoid overestimation. The action is selected by the current network, the Q-value is calculated by the target network, and gradient pruning (pruning threshold 0.5) is used to suppress gradient explosion. The model training stops when the RMSE of the Q-value prediction on the validation set is less than 0.03.

[0088] Furthermore, a full-process response mechanism is constructed, encompassing real-time data acquisition, Q-value calculation, action selection, and vector output, to generate the dynamic construction scheme adjustment vector. Specifically, firstly, real-time data acquisition: data from each preceding module is synchronously acquired via industrial Ethernet, with the acquisition frequency matching the construction unit. The data is preprocessed by edge computing nodes (outlier removal and format standardization) before being input into the reinforcement learning model framework to avoid noise from the raw data affecting decision-making. Secondly, Q-value calculation and action selection: the model calculates the Q-value (long-term expected benefit of the action) for each of the 8-dimensional actions, and employs a primary action priority + secondary action compensation strategy to select combinations: priority is given to the primary action with the largest Q-value (such as excavation speed adjustment), followed by 2-3 secondary actions with high synergy with the primary action (such as corresponding adjustments to slag discharge parameters and load allocation). Synergy is determined through the action correlation matrix (e.g., the correlation coefficient of slag discharge pump power adjustment is 0.8 when excavation speed is increased). If multiple actions have a Q-value difference < 0.02 (approximately optimal), the final combination is determined by combining the current dominant risk (e.g., prioritizing excavation depth adjustment when geological risk is high). Third, adjust vector generation: The output vector adopts a structured format of parameter encoding + priority labeling, which includes 6 core fields: action type (e.g., a1-excavation speed), adjustment range (e.g., +0.1m / s), constraint basis (e.g., equipment health score of 92 meets the speed-up condition), implementation priority (level 1-5, level 1 is the highest), expected effect (e.g., efficiency improvement of 15%), and confidence level (calculated based on model prediction error, e.g., 0.94). Complete vector example: [{Action type: a1, Adjustment range: +0.1m / s, Constraint basis: Equipment health score 92, Geological hardness 2.3MPa ≤ Rated excavation limit 5MPa, Priority: 1, Expected effect: Excavation efficiency increased by 12%, Confidence: 0.93}, {Action type: a5, Adjustment range: +8%, Constraint basis: Suspended solids diffusion simulation concentration not exceeded, Priority: 2, Expected effect: Slag discharge efficiency matches excavation speed, Confidence: 0.89}, {Action type: a6, Adjustment range: → Region B, Constraint basis: Region B environmental window coefficient 0.85 > Current region 0.62, Priority: 1, Expected effect: Operational continuity improved, Confidence: 0.91}].

[0089] In the embodiments, a closed-loop iterative mechanism of construction feedback-reward update-parameter optimization can also be established to enable the model to continuously adapt to changes in working conditions as the construction progresses.

[0090] First, data collection of results: After the vector adjustment is executed, feedback data is collected in two cycles: short-term and long-term. Short-term data collection includes equipment operating parameters (such as actual excavation speed and energy consumption) and changes in progress deviation; long-term data collection includes changes in risk level and environmental monitoring data (suspended solids concentration), forming an execution results dataset.

[0091] Secondly, the reward value is updated and calculated: Based on the feedback data, the actual reward value Ractual is recalculated and compared with the model's predicted reward value Rpredicted. If |Ractual - Rpredicted| > 0.2 (the deviation is too large), the sample relabeling mechanism is triggered, and the data set is marked as a high-value sample and stored in the experience replay buffer for priority training. Example of reward value calculation: After the execution of a certain adjustment vector, the actual excavation efficiency is 1.12 times the planned efficiency (Refficiency = 1.12), the unit cost is 0.92 times the planned cost (Rcost = 1.09), and the risk level decreases from level 3 to level 2.

[0092] Third, model parameter iteration: when the reward bias of three consecutive feedbacks is greater than 0.15, or the reward bias of a single high-value sample is greater than 0.3, an instant parameter update is triggered.

[0093] This invention integrates dynamic data from multiple sources, including geology, environment, and equipment, into a state space for reinforcement learning. It then constructs a multi-objective reward function that takes into account efficiency, cost, and risk, adapting to the multi-constraint requirements of underwater construction. This enables real-time and intelligent adjustment of construction plans, rapid response to changes in working conditions, avoids the lag and subjectivity of manual decision-making, and improves the adaptability and optimization effect of construction plans.

[0094] S4. Based on predictive maintenance time points, environmentally optimal paths, dynamic environmental window coefficients, workability probability time series, and dynamic earthwork volume matrix, a refined scheduling scheme is obtained through resource collaborative scheduling optimization.

[0095] Underwater excavation generates a large amount of suspended matter (silt, rock debris). If the spread is too large, it can easily cause environmental problems such as pollution of surrounding water bodies and ecological damage. This invention first predicts the spread range of suspended matter through fluid dynamics simulation, and then combines the path optimization capability of the A* algorithm to plan the optimal path that balances excavation efficiency and environmental compliance, thereby achieving pre-control of environmental risks.

[0096] Specifically, generating the environmentally optimal path includes the following steps: S411. Simulate the diffusion of suspended matter to obtain a heat map of suspended matter concentration distribution.

[0097] Based on fluid dynamics theory, a high-fidelity three-dimensional flow field model was constructed using FLUENT software to accurately predict the diffusion process of suspended matter, providing a quantitative basis for environmental protection pathway planning. Specifically: ① Mesh generation: A hybrid approach of structured and unstructured meshes was adopted. The core construction area used a tetrahedral unstructured mesh, while the outer area used a hexahedral structured mesh, with the total number of meshes controlled within 2 million; ② Turbulence model selection: The RNG k-ε model (suitable for underwater high Reynolds number flow fields) was selected, with model constants C1=1.44, C2=1.92, and Cμ=0.09. The boundary layer was treated using the wall function method; ③ Boundary and initial condition settings: The inlet boundary was set as a velocity inlet (using three-dimensional flow field data), the outlet boundary was set as a free outflow, and the sensitive water body boundary was set as a concentration constraint boundary. The initial condition set the excavation point as a non-point source pollution source; ④ Simulation and result verification: The time step was set to 0.5s, the simulation duration was 24h (covering one tidal cycle), and suspended matter concentration cloud maps and diffusion range data were output hourly. The final output includes heat maps of suspended solids concentration distribution under different excavation paths, vector boundaries of areas exceeding standards, and time thresholds for diffusion to sensitive water bodies, providing basic data for subsequent cost calculations.

[0098] S412. Construct an environmental cost function based on the suspended solids concentration distribution heatmap, and generate the optimal environmental path by combining the A* path optimization algorithm with the environmental cost function.

[0099] Based on diffusion simulation results, a two-dimensional environmental cost function is constructed to quantify environmental risks. The calculation of the core variables in the cost function requires combining engineering standards and monitoring data: ① Determination of S-exceedance (area of ​​the exceeding limit): Using the suspended solids concentration limit of the corresponding water body functional zone in GB3838-2002 as a benchmark, the vector boundary of the area exceeding the limit is extracted from the concentration cloud map using ArcGIS, and its projected area is calculated. Simultaneously, a time correction coefficient k is introduced (k=1.2 when the duration of exceeding the limit t>4h, k=1.0 when t≤4h). Finally, S-exceedance = vector area × k, ensuring that the cost matches the sustained impact of pollution; ② Measurement of D-sensitivity (shortest distance to sensitive water bodies): Based on the legally protected boundary of sensitive water bodies, the shortest distance from each point on the centerline of the excavation path to the protected boundary is calculated using the Euclidean distance formula. If the path crosses the protected boundary, then D_sensitivity = 0, triggering the highest environmental protection cost; ③ Dynamic adjustment of weighting coefficients α and β: around drinking water sources, α = 0.7, β = 0.3 (prioritizing control of areas exceeding standards), aquaculture areas, α = 0.5, β = 0.5 (balancing the impact of exceeding standards with distance risk), and ordinary landscape water bodies, α = 0.3, β = 0.7 (prioritizing ensuring distance safety); At the same time, an environmental penalty coefficient λ is set, when S exceeds the standard by more than 1000m. 2If D sensitivity < 50m, λ = 1.5, and the cost function is modified to C environmental protection = λ × (α × S exceeding the standard + β × D sensitivity), thus strengthening the cost constraints in high-risk scenarios.

[0100] Furthermore, a dual-objective optimization model is constructed by integrating environmental protection costs and construction efficiency, and the path is accurately selected through heuristic search using the A* algorithm. ① Objective function and cost refinement: In the total cost formula, Cefficiency needs to be further quantified as a comprehensive value of time cost and energy consumption cost, that is, Cefficiency = (L / V) × CEnergy Consumption + L × CTransportation, where L is the path length, V is the equipment operating speed, CEnergy Consumption is the energy consumption cost per unit time of the equipment, and CTransportation is the cost of slag transportation per unit distance, ensuring that efficiency cost is consistent with the actual project; at the same time, a weight balance coefficient η is set (η=0.6, prioritizing environmental protection), and the total cost is finally optimized to CTotal = η × CEnvironmental Protection + (1-η) × CEfficiency. ② Path node and constraint refinement: The 1m×1m grid node division needs to be aligned with the dynamic geological model grid. Each node is associated with three core attributes: geological hardness σ, environmental cost coefficient, and operational accessibility (whether there are underwater obstacles). The node exclusion rule is clearly defined as nodes with σ> equipment excavation limit or with rigid obstacles (such as underwater pile foundations). After exclusion, connectivity analysis is used to ensure that the remaining nodes form a continuous network to avoid path breakage. ③ Design of core parameters for the A* algorithm: The heuristic function adopts a weighted combination of Manhattan distance and environmental cost estimation, i.e., h(n) = ω × (Δx + Δy) + (1 - ω) × C_environmental_cost_estimation(n), where Δx and Δy are the differences in the horizontal and vertical coordinates from node n to the endpoint, and C_environmental_cost_estimation(n) is the estimated environmental cost from node n to the endpoint, ω = 0.4 (balancing distance and environmental prediction); the cost function g(n) = g(parent(n)) + C_node(n), where C_node(n) is the real-time total cost (environmental protection + efficiency) of the current node; during the search process, a priority queue is used to store the nodes to be explored, and the node with the smallest h(n) + g(n) is expanded first. At the same time, a path deduplication mechanism is introduced to avoid repeatedly exploring the same area of ​​nodes. ④ Optimization Result Verification and Output: The optimal path obtained through the search needs to pass two verifications: first, the geological hardness of all nodes on the path is ≤ the equipment excavation limit; second, the total C corresponding to the path is ≥20% lower than the initial path. After the verification is passed, the path coordinate sequence (accurate to 0.1m, such as (10.2,25.3), (11.2,25.3)...), node operation sequence (marking the operation priority in the order from the starting point to the end point), and recommended operation parameters for each node (such as reducing the operation speed to 0.15m / s at nodes with high environmental protection costs) are output. At the same time, a three-dimensional visualization model of the path and a detailed cost composition are generated to provide accurate guidance for construction execution.

[0101] This invention deeply integrates fluid dynamics simulation with the A* algorithm, transforming the suspended matter diffusion range into a path cost parameter, thereby constructing a dual-objective cost function that balances environmental protection and efficiency. This adapts to the complex requirements of underwater construction, bringing environmental constraints forward to the path planning stage, effectively controlling the suspended matter diffusion range and avoiding environmental penalties; at the same time, it takes into account excavation efficiency, achieving synergistic optimization of environmental protection and efficiency.

[0102] Calculating earthwork volume and tracking progress for underwater excavation of deep foundation pits are crucial for cost control and project management. However, the underwater environment results in low accuracy and significant delays in traditional measurement methods. This paper proposes a method that compares BIM design models with real-time point cloud scanning to dynamically calculate the deviation between the actual excavated earthwork volume and the design value, generating a 3D progress deviation map to achieve visualized and precise progress tracking.

[0103] Specifically, the first step is 3D grid division: Based on the BIM boundary model of the excavation area, the 3D grid method is used to divide it into standardized grid units. For grid units that cross geological layers, the height ratio of each layer within the grid is automatically determined using the 3D geological interface data of the dynamic geological model (e.g., 50% of a grid is located in the clay layer and 50% in the sand layer), preparing for subsequent weighted calculations.

[0104] Then, the earthwork volume of the grid is calculated as follows: First, the design elevation Hdesign corresponding to the center point of each grid is extracted through the API interface of the BIM model; then, the actual average elevation Hactual of the point cloud within the grid is calculated using the neighborhood weighted average method. For all valid point cloud elevation values ​​within the grid, weights are assigned according to their distance from the grid center (the closer the distance, the greater the weight), and Hactual is calculated using a weighted average method with an accuracy controlled within ±2cm. The earthwork volume of a single grid is calculated according to the principle of layered accumulation: If the grid is entirely located within a single geological layer, the earthwork volume V = grid bottom area (0.25㎡) × |Hactual - Hdesign| × the unit weight γ of the soil in that layer; if it spans multiple layers, the calculation is performed by splitting the volume according to the proportion of each layer and then summing the results. The formula is V = ∑(0.25 × h_i × γ_i), where h_i is the height of the i-th layer within the grid, and γ_i is the unit weight of the corresponding layer.

[0105] Finally, dynamic earthwork volume matrix generation: construct a two-dimensional matrix of earthwork volume data of all grids according to their spatial coordinates (X,Y,Z) (rows represent X coordinates, columns represent Y coordinates, and matrix values ​​are the cumulative earthwork volume of the corresponding grid), and mark the geological stratification information and calculation accuracy level of each grid (Level A: error <2%, Level B: error 2%-3%), to provide refined data support for schedule analysis and cost accounting.

[0106] Furthermore, the schedule deviation is quantitatively calculated: based on the dynamic earthwork volume matrix and the construction schedule plan, two types of core deviation indicators are generated: cumulative deviation = cumulative actual earthwork excavation volume - cumulative planned earthwork volume, with positive values ​​indicating that the schedule is ahead of schedule and negative values ​​indicating that it is behind schedule; daily deviation = actual earthwork excavation volume on the day - planned earthwork volume on the day, which is used to monitor short-term construction efficiency fluctuations.

[0107] This invention integrates soil density data from geological stratification into earthwork volume calculation, and also correlates progress deviations with geological conditions to achieve real-time and accurate earthwork volume calculation. This solves the problems of invisible and inaccurate underwater excavation progress, and provides a direct basis for cost accounting (earthwork transportation costs) and construction plan adjustments.

[0108] Deep foundation pit underwater excavation involves multiple pieces of equipment, multiple construction areas and multiple processes. By adopting the precise constraint optimization capability of mixed integer linear programming (MILP), combined with constraints such as equipment capacity, maintenance plan, and environmental window, we can achieve refined collaborative scheduling of resources and improve resource utilization efficiency.

[0109] Specifically, the refined scheduling scheme obtained through resource collaborative scheduling optimization based on predictive maintenance time points, environmentally optimal paths, dynamic environmental window coefficients, workability probability time series, and dynamic earthwork volume matrix includes the following steps: S421. Based on the predictive maintenance time point, environmentally optimal path, dynamic environmental window coefficient, workability probability time series, and dynamic earthwork volume matrix, construct the scheduling optimization objective and constraints.

[0110] The objective function is constructed using a dual-objective weighted fusion and dynamic correction model. Objective 1 is to minimize the total resource cost, with the cost composition broken down into four quantifiable items: equipment energy consumption cost, labor cost, equipment depreciation cost, and maintenance reserve. Objective 2 is to minimize the total project duration, with key node project durations as rigid anchor points. A project duration penalty coefficient is introduced, increasing the total cost by 3% for each day of delay and awarding a 1.5% bonus to the total cost for each day of early completion.

[0111] The weights were determined using expert scoring and project type adjustments: scheduling experts were invited to score the importance of cost and schedule (1-10 points). In the example, the basic weights were 0.55 for cost and 0.45 for schedule. For government livelihood projects (focusing on schedule assurance), the weights were adjusted to 0.6 for cost and 0.4 for schedule. For commercial development projects (focusing on cost control), the weights were adjusted to 0.7 for cost and 0.3 for schedule. The final objective function expression was: MinZ=0.6×(∑equipment energy consumption cost+∑labor cost+∑depreciation cost+∑maintenance reserve fund+schedule penalty / reward)+0.4×(total schedule / planned schedule×100).

[0112] Refined Constraints: Equipment maintenance constraints (hard constraints) distinguish between predictive and emergency maintenance. Predictive maintenance time points are accurate to the hour; emergency maintenance is triggered when equipment health score is <30, immediately interrupting work and updating the scheduling plan. Time constraints (hard constraints) are associated with the workability probability time series. Only time periods with a workability probability ≥0.6 are included in the effective work time, and the continuous operation time of the same equipment is ≤8 hours (to avoid fatigue work), with a daily rest interval ≥2 hours. Task logic constraints (hard constraints) clearly define the sequence and duration of procedures. For example, after the excavation procedure is completed, a 1-hour interval is required for slag removal, achieved by introducing a 0-1 variable zjk (the completion status of the procedure in the j-th time period of the k-th region). For example, if the excavation procedure z1jk=1, then the slag removal procedure z2jk can only be 1 in time periods j+1 and later. Spatial constraints are also included: equipment transportation paths must avoid the core protection zone of the environmentally optimal path, with a path overlap rate ≤10%, incorporated into the objective function through a distance penalty term.

[0113] S422. Solve the mixed-integer linear programming model based on the scheduling optimization objective and constraints to obtain a refined scheduling scheme.

[0114] To establish a mixed-integer linear programming model, the decision variable system is first constructed as follows: a two-dimensional variable matrix consisting of core variables and auxiliary variables is built, covering all dimensions of equipment, time, region, and process. Core variables: xijk represents the effective working time of the i-th equipment (i=1,2,...,n, where n is the total number of equipment) in the j-th time period (j=1,2,...,m, where m is the total number of time periods, with each hour constituting one time period) to the k-th work area (k=1,2,...,p, where p is the number of areas), expressed in hours. xijk ≥ 0 and is a continuous variable. yijk is a 0-1 variable; yijk = 1 indicates that the i-th equipment is assigned to the k-th area in the j-th time period, and yijk = 0 indicates the opposite, used to correlate equipment allocation with working time. Auxiliary variables: zijkq are 0-1 variables, representing the execution status of the q-th process (q=1 excavation, 2 slag removal, 3 support) of the i-th equipment in the k-th region during the j-th time period, used to reinforce the process logic; tijk is the transportation time of the i-th equipment from the current region to the k-th region, data obtained from real-time calculations of GIS route planning (e.g., if the transportation time from region A to B is 20 minutes, then tijk=0.33). Secondly, the objective function is linearized, transforming it into a standard linear expression to clarify the calculation logic of each cost item. Thirdly, the constraints are linearized: all constraints are transformed into standard linear inequalities to ensure the model is solvable.

[0115] Furthermore, the mixed-integer linear programming model is optimized by dimensionality reduction, redundant constraints are removed, and similar variables are merged. The Gurobi 10.0 solver is used, and the solution strategy adopts a coarse-to-fine approach: in the initial stage, a heuristic algorithm is used to quickly generate feasible solutions as the starting point for the precise solution; in the precise solution stage, the branch and bound method is used to prioritize branching on key variables to shorten the solution time.

[0116] Furthermore, a refined scheduling plan is generated: a structured output consisting of a master table and several sub-tables, directly connecting to the construction execution layer. The core master table is the "Comprehensive Equipment Scheduling Plan," which includes fields such as equipment number, type, work area (including coordinate range), time slot arrangement (accurate to the hour), work duration, process type, transport vessel matching number, and estimated volume. The sub-tables contain three types of specialized plans: equipment start-up, shutdown, and relocation plans; transport equipment loading plans; and manpower scheduling plans.

[0117] This invention integrates dynamic constraints such as predictive maintenance time points and dynamic environmental windows into the model to achieve dynamic adjustment of the scheduling scheme. It combines the environmentally optimal path and dynamic construction adjustment vector to determine the model parameters, so that the scheduling scheme not only meets resource optimization but also adapts to construction technical requirements. It realizes collaborative scheduling optimization of multiple equipment, multiple regions, and multiple processes, reduces equipment idleness and process conflicts, improves resource utilization, and reduces construction costs.

[0118] S5. Based on the optimal environmental protection path, dynamic construction plan adjustment vector, and refined scheduling plan, the underwater excavation construction is optimized.

[0119] Underwater excavation construction optimization is a complex decision-making problem involving multiple conflicting objectives such as efficiency, cost, safety, and environmental protection (e.g., improving excavation efficiency may increase environmental risks). This paper adopts a multi-objective collaborative decision-making algorithm, integrates the optimization results of previous steps, and outputs a final construction optimization scheme that takes into account multiple objectives through dynamic allocation of objective weights and comprehensive evaluation of the scheme.

[0120] Specifically, the optimization of underwater excavation construction based on the environmentally optimal path, dynamic construction plan adjustment vector, and refined scheduling plan includes the following steps: S51. Construct a multi-objective evaluation system.

[0121] The indicator system is further refined into eight secondary indicators under four primary objectives, with specific definitions and scoring criteria as follows: Efficiency targets: Excavation efficiency (actual earthwork volume / hour ÷ planned earthwork volume / hour × 100, full marks 100, ≥120% gets 100 points, <80% gets below 50 points), progress achievement rate (cumulative actual progress ÷ cumulative planned progress × 100, full marks 100, ≥100% gets 100 points, <90% gets below 60 points); Cost targets: Unit earthwork cost (actual cost ÷ planned cost × 100, full marks 100, ≤90% gets 100 points, >110% gets below 50 points), resource waste rate (idle equipment time ÷ total operation time × 100, full marks 100, ≤5% gets 100 points, >15% gets below 40 points); Safety targets: risks, etc. Level compliance rate (duration of low-risk and below working conditions ÷ total construction time × 100, full score 100 points, ≥95% gets 100 points, <85% gets less than 50 points); Safety hazard rectification rate (number of rectified hazards ÷ total number of hazards × 100, full score 100 points, 100% gets 100 points, <90% gets less than 60 points); Environmental protection targets: Suspended solids control compliance rate (duration of concentration not exceeding the standard ÷ total operation time × 100, full score 100 points, ≥98% gets 100 points, <90% gets less than 40 points); Sensitive water body protection compliance rate (duration of not breaching the protection boundary ÷ operation time in sensitive areas × 100, full score 100 points, 100% gets 100 points, <95% gets less than 50 points).

[0122] Furthermore, the weight determination process is as follows: a combination of objective weighting using the entropy weighting method and subjective correction using the analytic hierarchy process (AHP) is adopted. First, the real-time monitoring data of various indicators over the past 30 days (such as excavation efficiency fluctuation data and cost deviation data) is processed using the entropy weighting method to calculate the objective weights (the smaller the entropy value, the higher the weight; for example, the cost indicator has an entropy value of 0.32 and an objective weight of 0.31). Then, a decision-making team (including project managers, supervising engineers, and environmental experts) is organized to construct a judgment matrix using the AHP. The weights are then adjusted based on the project type (government livelihood project / commercial development project) to finally determine the initial weights: efficiency 0.25, cost 0.3, safety 0.25, and environmental protection 0.2.

[0123] Finally, a dynamic adjustment mechanism is implemented, setting three types of trigger conditions to automatically adjust weights: environmental protection trigger (sudden environmental inspection, water quality warning for sensitive water bodies): environmental protection weight increases to 0.35, cost weight decreases to 0.2; schedule trigger (critical node delay > 12 hours): efficiency weight increases to 0.35, environmental protection weight decreases to 0.15; safety trigger (risk level rises to level 4 or above): safety weight increases to 0.4, efficiency weight decreases to 0.15.

[0124] S52. Standardize the environmentally optimal path, dynamic construction plan adjustment vector, and refined scheduling plan.

[0125] First, the indicators of the plan are mapped: the optimal environmental protection path is mapped to the compliance rate of suspended solids control and the compliance rate of sensitive water body protection; the dynamic construction adjustment vector is mapped to the excavation efficiency and the unit earthwork cost; the refined scheduling plan is mapped to the resource waste rate (equipment idle time statistics) and the progress compliance rate (the progress deviation correction value after scheduling); and the safety indicators are directly taken as the risk level compliance rate and the hidden danger rectification rate.

[0126] Secondly, the indicators were classified and standardized. The eight secondary indicators were divided into positive indicators (the larger the value, the better, such as excavation efficiency and rectification rate) and negative indicators (the smaller the value, the better, such as resource waste rate and unit cost). The corresponding methods were used for standardization. Positive indicators: x'=(x-x_min) / (x_max-x_min) (x is the actual value, x_max is the historical best value, and x_min is the historical worst value); Negative indicators: x'=(x_max-x) / (x_max-x_min).

[0127] The processed index values ​​are subjected to a consistency check, requiring all index values ​​to fall within the [0,1] interval. The final output is a standardized index matrix, as shown in the example below: [Excavation efficiency: 0.75, Schedule achievement rate: 0.8, Unit cost: 0.5, Resource waste rate: 0.9, Risk achievement rate: 0.85, Hidden danger rectification rate: 1.0, Suspended solids achievement rate: 0.92, Sensitive protection achievement rate: 0.95], providing a unified input for subsequent decision-making calculations.

[0128] S53. Based on the multi-objective evaluation system and standardized processing results, underwater excavation construction optimization is achieved through a multi-objective collaborative decision-making model.

[0129] A fusion algorithm combining TOPSIS objective ranking and fuzzy comprehensive evaluation subjective calibration is adopted to achieve a balance between the accuracy and practicality of decision-making. First, the TOPSIS algorithm calculates and constructs a weighted standardized matrix. The standardized index matrix is ​​multiplied by the dynamic weight vector to obtain the weighted values. Then, the ideal solutions are determined: the positive ideal solution S+ is the maximum weighted value of each index, and the negative ideal solution S- is the minimum weighted value of each index. Next, the Euclidean distance is calculated: the distance D+ between the solution and the positive ideal solution, and the distance D- between the solution and the negative ideal solution. Finally, the proximity score C = D- / (D++D-) is calculated; the closer the proximity score is to 1, the better the solution.

[0130] Furthermore, through fuzzy comprehensive evaluation calibration, an expert review group was formed, and a three-level fuzzy evaluation system of indicators, factors, and schemes was constructed. The evaluation factor set (8 secondary indicators) and the comment set (Excellent / Good / Medium / Poor, corresponding to quantitative values ​​of 1 / 0.8 / 0.5 / 0.3) were determined. Experts scored the evaluations on a 10-point scale, which were then converted into a fuzzy relation matrix (e.g., expert 1 scored 8 points for excavation efficiency, corresponding to Good, with a fuzzy value of 0.8). Combining the indicator weights calculated using the entropy weight method, a comprehensive fuzzy evaluation result was obtained through fuzzy synthesis operations. Next, a fusion correction is performed: the two results are fused using a linear weighting method, and the corrected comprehensive score is 0.6×C×100+0.4×fuzzy evaluation score×100 (TOPSIS objective result weight 0.6, expert subjective evaluation weight 0.4).

[0131] Furthermore, based on the principles of optimal score and controllable risk, a structured and executable solution is output. The alternative solutions (such as those generated based on different weights) are ranked in descending order of their comprehensive scores, with the highest-scoring solution becoming the candidate solution. Simultaneously, the constraint satisfaction of the candidate solutions is verified: equipment scheduling must comply with predictive maintenance times, construction parameters must match geological hardness, and environmental control must meet concentration limits. If any constraint is violated, the next solution is selected in order of priority.

[0132] Structured output of the plan: adopting a general outline + volume format, with the general outline clearly defining the project objectives (e.g., excavation to be completed in 30 days, unit cost ≤ 280 yuan / m). 3 The risk level is ≤3), and the core optimization direction is (such as prioritizing environmental risk control through path optimization and simultaneously improving equipment coordination efficiency). The volume contains four core contents: equipment scheduling plan (including equipment number, type, work area coordinates, daily work time period, and maintenance window), construction parameter settings, risk prevention and control measures, and environmental control requirements.

[0133] In this embodiment, four types of core data—construction execution data, cost data, safety data, and environmental protection data—can be collected synchronously through an industrial internet platform. This data is then used to dynamically recalculate the comprehensive score, generating trend charts of indicator changes and score fluctuation analysis tables. Different responses are triggered by setting score thresholds: ≥90 points (Excellent): Maintain the current plan and record the optimal parameters for future reference; 80 points ≤ Score < 90 points (Pass): Only minor adjustments to construction parameters (such as adjusting the power of the slag discharge pump) are made, without triggering module re-optimization; Score < 80 points (Fail): Immediately trigger re-optimization of preceding modules, recalculate the scheduling plan based on the latest equipment status, regenerate the adjustment vector based on the new operating conditions, and recalculate the environmental protection path based on the new water flow data, achieving seamless integration of execution-feedback-optimization.

[0134] Please see Figure 2In an embodiment, to efficiently execute the underwater excavation construction optimization method provided by the present invention, the present invention also provides an underwater excavation construction optimization system, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory contains program instructions for the steps of the underwater excavation construction optimization method. The underwater excavation construction optimization system of the present invention has a compact structure and stable performance, and can stably execute the underwater excavation construction optimization method of the present invention, further enhancing the overall applicability and practical application capability of the present invention.

[0135] In summary, this invention achieves dynamic collaborative optimization of multiple factors such as geology, environment, equipment, schedule, and risk through multi-source data fusion and multi-algorithm collaboration. It solves problems such as static rigidity, single-link optimization, multi-objective conflict, one-sided risk assessment, and delayed scheme adjustment in existing technologies, thereby improving the efficiency, economy, and safety of deep foundation pit underwater excavation construction.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A construction optimization method for underwater excavation, characterized in that, Includes the following steps: A dynamic three-dimensional geological probability model is constructed using multi-source sensor data and a Bayesian update algorithm. Fuzzy logic and time series analysis are used to dynamically evaluate the environmental window and obtain the time series of workability probability. By combining a dynamic three-dimensional geological probability model, a workability probability time series, a real-time equipment health index, a three-dimensional progress deviation map, and a dynamic risk spectrum, a dynamic construction scheme adjustment vector is generated based on reinforcement learning. Based on predictive maintenance time points, environmentally optimal paths, dynamic environmental window coefficients, workability probability time series, and dynamic earthwork volume matrix, a refined scheduling scheme is obtained through resource collaborative scheduling optimization. By using the environmentally optimal path, dynamic construction plan adjustment vector, and refined scheduling plan, underwater excavation construction is optimized.

2. The construction optimization method for underwater excavation according to claim 1, characterized in that, The construction of a dynamic three-dimensional geological probability model using multi-source sensor data and a Bayesian update algorithm includes the following steps: An initial three-dimensional geological model is generated based on multi-source sensor data through interpolation calculation. A Bayesian update model is constructed, and the initial three-dimensional geological model is dynamically updated through the Bayesian update model to obtain the dynamic three-dimensional geological probability model.

3. The construction optimization method for underwater excavation according to claim 1, characterized in that, The method of using fuzzy logic and time series analysis to dynamically evaluate the environmental window and obtain the workability probability time series includes the following steps: Construct an environmental assessment indicator system and a fuzzy logic assessment model; By combining the environmental assessment index system and the fuzzy logic assessment model, a time series prediction model is trained to obtain the workability probability time series.

4. The construction optimization method for underwater excavation according to claim 1, characterized in that, The process of combining a dynamic three-dimensional geological probability model, a workability probability time series, a real-time equipment health index, a three-dimensional progress deviation map, and a dynamic risk spectrum to generate a dynamic construction plan adjustment vector based on reinforcement learning includes the following steps: By combining the dynamic three-dimensional geological probability model, the workability probability time series, the real-time equipment health index, the three-dimensional progress deviation map, and the dynamic risk spectrum, a reinforcement learning model framework is constructed. The reinforcement learning model framework is trained and optimized, and the dynamic construction scheme adjustment vector is generated through the optimized reinforcement learning model framework.

5. The construction optimization method for underwater excavation according to claim 4, characterized in that, Analyzing the real-time health index of the device includes the following steps: Build digital twin models of equipment; Based on the digital twin model of the device, the device health features are extracted, and the real-time health index of the device is calculated according to the device health features.

6. The construction optimization method for underwater excavation according to claim 4, characterized in that, Extracting the dynamic risk spectrum includes the following steps: Determine the probability distribution of risk parameters based on risk identification and indicator system; A multi-risk coupling model is constructed, and the dynamic risk spectrum is analyzed by Monte Carlo simulation based on the multi-risk coupling model.

7. The construction optimization method for underwater excavation according to claim 1, characterized in that, The refined scheduling scheme, based on predictive maintenance time points, environmentally optimal paths, dynamic environmental window coefficients, workability probability time series, and dynamic earthwork volume matrix, is obtained through resource collaborative scheduling optimization, including the following steps: Based on the predictive maintenance time point, environmentally optimal path, dynamic environmental window coefficient, workability probability time series, and dynamic earthwork volume matrix, a scheduling optimization objective and constraints are constructed. A refined scheduling scheme is obtained by solving a mixed-integer linear programming model based on scheduling optimization objectives and constraints.

8. The construction optimization method for underwater excavation according to claim 7, characterized in that, Generating the optimal environmentally friendly route includes the following steps: Simulate the diffusion of suspended matter to obtain a heat map of suspended matter concentration distribution; An environmental cost function is constructed based on the suspended solids concentration distribution heatmap. The optimal environmental path is then generated by combining the A* path optimization algorithm with the environmental cost function.

9. The construction optimization method for underwater excavation according to claim 1, characterized in that, The optimization of underwater excavation construction based on the environmentally optimal path, dynamic construction plan adjustment vector, and refined scheduling plan includes the following steps: Construct a multi-objective evaluation system; Standardize the optimal environmental protection path, dynamic construction plan adjustment vector, and refined scheduling plan; Based on the aforementioned multi-objective evaluation system and standardized processing results, underwater excavation construction optimization is achieved through a multi-objective collaborative decision-making model.

10. A construction optimization system for underwater excavation, characterized in that, The construction optimization system for underwater excavation includes: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory includes program instructions for executing the construction optimization method for underwater excavation as described in any one of claims 1-9.

Images