Island wall structure stability prediction construction method based on machine learning
By constructing a multi-source heterogeneous real-time monitoring network and a deep learning model, the stability of the island wall structure can be predicted in real time. This solves the real-time and efficiency problems of traditional methods during the construction phase, achieves accurate prediction and dynamic optimization of the island wall structure, and improves construction safety and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CCCC THIRD HARBOR ENGINEERING CO LTD
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional island wall stability analysis methods are difficult to reflect dynamic changes in real time during the construction phase, and the calculation is time-consuming, which cannot meet the needs of real-time decision-making on site. Furthermore, they fail to make full use of monitoring data to guide construction, resulting in insufficient prediction of island wall structural stability.
A multi-source heterogeneous real-time monitoring network was constructed, and a deep learning model was used to predict the stability of the island wall structure. By combining a multi-layer long short-term memory network and an attention layer, future displacement and stability were monitored and predicted in real time. A dynamic construction optimization decision-making mechanism was established to adjust construction parameters in real time.
It enables accurate prediction and dynamic optimization of island wall structures, improves construction safety and efficiency, reduces unnecessary downtime and over-reinforcement, and controls project costs.
Smart Images
Figure CN122020245A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of geotechnical engineering and artificial intelligence, specifically a construction method for predicting the stability of island wall structures based on machine learning. Background Technology
[0002] Land reclamation is a crucial means of expanding urban space and constructing major infrastructure projects such as ports and airports. In reclamation projects, the stability of island walls, as the retaining structure, directly affects the safety, schedule, and cost of the entire project. Currently, island wall structures face extremely complex engineering and environmental challenges: (1) Complex and variable geological conditions: The deep and weak seabed silt foundation has low bearing capacity and is prone to uneven settlement and lateral displacement under load, threatening the stability of the island wall.
[0003] (2) Construction load has a significant impact: large-scale backfilling, surcharge preloading, foundation reinforcement (such as crushed stone piles) and other construction activities will continuously change the stress state of the foundation and the pore water pressure field, which are the main factors that induce island wall instability.
[0004] (3) Harsh marine dynamic environment: Periodic or random hydrodynamic loads such as waves, tides and ocean currents continuously act on the island wall, which not only directly generate thrust, but also cause dynamic response of pore water pressure in the foundation soil, further weakening the soil strength.
[0005] Traditional methods for analyzing island wall stability mainly rely on limit equilibrium methods or finite element numerical simulations. While these methods play an important role in the design phase, they have significant limitations during construction: First, the calculation models are often based on simplified geological parameters and load conditions, making it difficult to reflect construction dynamics and environmental changes in real time; second, the calculation process is time-consuming, failing to meet the needs of real-time decision-making at the construction site; and finally, although there is a massive amount of on-site monitoring data, most of it is only used for post-event verification or passive early warning, failing to fully explore its predictive value to guide proactive, feedforward construction control.
[0006] Therefore, a construction method for predicting the stability of island wall structures based on machine learning is provided. Summary of the Invention
[0007] To address the aforementioned problems in existing technologies, this invention provides a machine learning-based method for predicting the stability of island wall structures during construction. This method enables accurate prediction of the future state of island wall structures and allows for dynamic and intelligent optimization of construction plans based on the prediction results, thereby improving construction safety, efficiency, and economy.
[0008] The technical solution to achieve the above objectives is: A machine learning-based method for predicting the stability of island wall structures during construction includes: Step S1: Construct a multi-source heterogeneous real-time monitoring network to comprehensively collect multi-source heterogeneous data affecting the stability of the island wall; Step S2: Establish a multi-parameter fusion dynamic prediction model. Based on historical monitoring data, predict the time series of displacement of key points of the island wall and structural stability coefficient in the future. Step S3: Establish a dynamic construction optimization decision-making mechanism, compare the output of the dynamic prediction model with the preset safety control threshold, dynamically optimize construction parameters, and form real-time dynamic decisions. In step S4, the adjusted construction behavior and the resulting actual monitoring results will be fed back to the dynamic prediction model as new data samples for training and optimization.
[0009] Preferably, in step S1, constructing a multi-source heterogeneous real-time monitoring network includes: High-precision, automated monitoring equipment was deployed at key locations along the inner and outer slopes, top and bottom of the island wall, using distributed fiber Bragg grating sensor arrays and GPS / GNSS receivers to acquire the vertical, horizontal and deep displacements of the island wall in real time. Sensors were installed at different depths and locations inside the island wall to monitor key soil mechanical parameters in real time. The excess pore water pressure and its dissipation during the soil consolidation process are monitored in real time using a vibrating wire or fiber optic grating pore water pressure gauge. By integrating construction quality control data, the source, gradation, and compaction process parameters of the backfill soil are recorded and input in real time as indirect input parameters characterizing the backfill density. GPS positioning systems are installed on construction machinery to obtain the number, location and operation trajectory of large machinery in real time. Through electronic scales and transportation records, the volume of each layer of backfill soil and the load of surcharge preloading are accurately recorded. Wave and tide gauges are installed near the wave-facing side of the island wall to monitor changes in wave height, wave period, and tide level in real time, thereby quantifying the dynamic impact of wave loads on the island wall.
[0010] Preferably, in step S1, all sensors are equipped with 5G wireless communication technology to transmit the collected data to the cloud data center in real time. The raw data from different displacement, soil properties, loads and environments are denoised, aligned and fused to form a unified, high-dimensional multivariate time series input feature vector, thus forming a structured time series database.
[0011] Preferably, in step S2, the dynamic prediction model is a deep learning model specifically designed for processing sequence data, with an encoder-decoder framework as its main architecture, and further optimized based on this framework: The encoder uses a multi-layer long short-term memory network as its core component for processing the input. Encode the multi-source feature vectors of each historical time step, extract and memorize their long-term temporal dependencies; An attention layer is introduced between the encoder and decoder to assign different weights to the input data at different times and monitoring points, enabling the model to focus on the key time periods and key areas most relevant to future instability. The decoder is also based on a multi-layer long short-term memory network, which gradually generates future context vectors based on the encoder's output and attention weights. The predicted sequence has a time step.
[0012] Preferably, in step S2, the inputs to the dynamic prediction model include: Structural response characteristics: including historical and real-time key point horizontal displacement rate, vertical settlement rate, and deep displacement increment; Internal state characteristics of soil and rock: including pore water pressure values at different depths, pore pressure coefficients, and estimated degrees of consolidation calculated based on pore pressure dissipation; Construction activity characteristics include: cumulative fill height, fill rate, and density of heavy machinery on the current work surface; Environmental disturbance characteristics: including significant wave height, tide level, and estimated wave run-up calculated from the combination of wave height and tide level; The above feature vectors are aligned with time to construct training samples, with each sample based on past time. The feature tensors with time steps are used as input: ; In the formula, Given the input feature tensor, This is the current time step; Feature vector at each time step It integrates four types of features: ; In the formula, For structural response characteristics, Characteristics of the internal state of rock and soil. Characteristics of construction activities, Characteristics of environmental disturbance; The dynamic prediction model adopts a multi-task learning paradigm and simultaneously outputs two physically meaningful and interrelated target variables: future displacement prediction and dynamic stability coefficient. in, Future displacement prediction: Directly output future displacement. The absolute displacement of each key monitoring point within each time step: ; In the formula, The output model, For the future The predicted absolute displacement value at each time point; Dynamic stability coefficient: Outputs a continuous island wall dynamic stability coefficient between 0 and 1. This coefficient is a data-driven, probabilistic comprehensive safety status indicator, which, after learning from massive historical data, predicts the system's future stability. A comprehensive assessment of the probability of maintaining a stable state over a given period: ; In the formula, For the future moment stability coefficient, The function is to standardize an interval and map it to the range (0,1). and The weight matrix and bias terms for the output layer are shown below. For the model decoder at time The state; If the coefficient is close to 1, it indicates a highly stable state; when the coefficient continues to decrease and approaches the preset alarm threshold, it indicates a significant increase in the risk of instability.
[0013] Preferably, in step S2, the training and validation of the dynamic prediction model are as follows: From the structured time-series database in the cloud data center, we divide the data into training, validation, and test sets. To ensure the predictive ability of the dynamic prediction model for extreme conditions, we use oversampling techniques to enhance the weight of unstable samples. Design a composite loss function, which consists of two weighted parts: ; In the formula, and The weights of the two parameters, This is the mean square error of the displacement prediction, used to ensure the accuracy of the displacement prediction. is the cross-entropy loss function for the stability coefficient, used to optimize the classification ability of the stability coefficient.
[0014] Preferably, in step S3, a dynamic construction optimization decision-making mechanism is established: Set multi-level early warning thresholds, including: The stability coefficient threshold is set to a three-level threshold for the output dynamic stability coefficient: Safe zone (green): Stability coefficient > 0.8, indicating that the system is in good condition and construction can proceed as planned; Warning zone (yellow): 0.6 < stability coefficient ≤ 0.8, indicating an unfavorable trend in the system state, requiring attention and preparation for intervention measures; Danger zone (red): Stability coefficient ≤ 0.6, indicating a high risk of system instability, requiring immediate intervention; Displacement and velocity thresholds: At the same time, parallel three-level thresholds are set for predicted displacement and predicted displacement rate as a supplement and verification of the stability coefficient. Real-time early warning and decision support, including: When the prediction result reaches the threshold, the system automatically triggers a tiered warning: Yellow alert: Visually highlighted in the digital twin interface of the management platform, and an alert notification pushed to relevant engineers; Red Alert: In addition to sending the highest level of alarm information to project management personnel, it will also automatically send lock or restrict operation instructions to the construction machinery control system in the relevant area, forcibly suspending high-risk operations from a technical perspective. When an early warning is triggered, the system will automatically evaluate various preset control schemes, including but not limited to "stop filling", "slow down the filling rate" and "start precipitation", to assess their effects on improving the future stability coefficient and recommend the expected control effect of the optimal control strategy.
[0015] Preferably, in step S3, the dynamically optimized construction procedures and parameters include: Based on the generated control strategy, the construction plan is dynamically and precisely adjusted; Filling loading control: Based on pore water pressure prediction and stability coefficient, dynamically adjust the thickness, interval time and daily filling rate of layered filling; Reinforcement measures triggering and strengthening: Based on the deformation trend, the pre-set reinforcement plan is activated in advance in the weak area, or the construction density and depth of crushed stone piles / vibratory piles and the spacing of drainage boards are dynamically adjusted. Surcharge preloading management: Optimize the timing of preloading application and removal based on predicted post-construction settlement and degree of consolidation; Construction machinery scheduling: Based on the real-time stability map, the system intelligently plans the travel path and work area of large machinery.
[0016] Preferably, in step S4, two main dynamic prediction model update triggering modes are established, which trigger model updates based on fixed time intervals or after encountering significant changes in operating conditions: Scheduled triggering: The model update process is automatically started based on a fixed time period; Event Trigger: Automatically triggered when a significant change in on-site conditions is detected. Specific conditions include: the site has entered a completely new construction phase; a significant shift in the statistical characteristics of the input data is detected. The dynamic prediction model iteration includes the following steps: Added data quality verification: Automatically clean and label the collected real-time data to form a new high-quality dataset; Incremental training and weight fine-tuning: An incremental learning algorithm is used to mix the new dataset with some historical core datasets, and to perform supervised fine-tuning with the current model weights as the initial values.
[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention, through accurate prediction of future displacement, can identify potential instability risks in advance, transforming the traditional "post-event alarm" into "pre-event control," thus gaining valuable time to take countermeasures and improving construction safety. This invention uses machine learning models to replace some of the judgments that rely on engineers' experience, thus making construction decisions more scientific and refined. By dynamically optimizing construction parameters, it can maximize construction efficiency while ensuring safety and avoid unnecessary work stoppages or overly conservative construction. This invention, through real-time data feedback and dynamic optimization, can avoid excessive displacement, rework, or over-reinforcement caused by improper construction plans, thereby effectively controlling project costs. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a construction method for predicting the stability of island wall structures based on machine learning, according to the present invention. Figure 2 This is a schematic diagram of displacement monitoring data for key points of the island wall structure in this invention; Figure 3 This is a schematic diagram of future displacement prediction based on a multilayer long short-term memory network in this invention; Figure 4 This is a schematic diagram of the training loss function in this invention; Figure 5 This is a schematic diagram of the early warning system based on the dynamic prediction model in this invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] like Figure 1As shown, a machine learning-based method for predicting the stability of island wall structures includes: Step S1: Construct a multi-source heterogeneous real-time monitoring network to comprehensively collect multi-source heterogeneous data affecting the stability of the island wall.
[0021] In this embodiment, a multi-source heterogeneous real-time monitoring network is constructed, including: Key locations along the inner and outer slopes of the island wall, at the top and bottom of the slope, such as... Figure 2 As shown, high-precision, automated monitoring equipment is deployed, using a distributed fiber Bragg grating sensor array and a GPS / GNSS receiver to acquire the vertical, horizontal, and deep displacements of the island wall in real time. Sensors were installed at different depths and locations inside the island wall to monitor key soil mechanical parameters in real time. The excess pore water pressure and its dissipation during the soil consolidation process are monitored in real time using a vibrating wire or fiber optic grating pore water pressure gauge. By integrating construction quality control data, the source, gradation, and compaction process parameters of the backfill soil are recorded and input in real time as indirect input parameters characterizing the backfill density. GPS positioning systems are installed on construction machinery to obtain the quantity, location and operation trajectory of large machinery (such as vibratory rollers, stone pile drivers and tampers) in real time. The volume of each layer of backfill soil and the load of surcharge preloading are accurately recorded through electronic scales and transportation records. Wave and tide gauges are installed near the wave-facing side of the island wall to monitor changes in wave height, wave period, and tide level in real time, thereby quantifying the dynamic impact of wave loads on the island wall.
[0022] In this embodiment, all sensors are equipped with 5G wireless communication technology to transmit the collected data to the cloud data center in real time. The raw data from different displacement, soil properties, loads and environments are denoised, aligned and fused to form a unified, high-dimensional multivariate time series input feature vector, thus forming a structured time series database.
[0023] Step S2: Establish a multi-parameter fusion dynamic prediction model. Based on historical monitoring data, predict the time series of displacement of key points of the island wall and structural stability coefficient in the future.
[0024] In this embodiment, the dynamic prediction model uses a deep learning model specifically designed for processing sequence data, with an encoder-decoder framework as its main architecture, and is further optimized based on this framework: The encoder uses a multi-layer long short-term memory network as its core component for processing the input. Encode the multi-source feature vectors of each historical time step, extract and memorize their long-term temporal dependencies; An attention layer is introduced between the encoder and decoder to assign different weights to the input data at different times and monitoring points, enabling the model to focus on the key time periods and key areas most relevant to future instability. The decoder is also based on a multi-layer long short-term memory network, which gradually generates future context vectors based on the encoder's output and attention weights. The predicted sequence has a time step.
[0025] In this embodiment, the inputs to the dynamic prediction model include: Structural response characteristics: including historical and real-time key point horizontal displacement rate, vertical settlement rate, and deep displacement increment; Internal state characteristics of soil and rock: including pore water pressure values at different depths, pore pressure coefficients, and estimated degrees of consolidation calculated based on pore pressure dissipation; Construction activity characteristics include: cumulative fill height, fill rate, and density of heavy machinery on the current work surface; Environmental disturbance characteristics: including significant wave height, tide level, and estimated wave run-up calculated from the combination of wave height and tide level; The above feature vectors are aligned with time to construct training samples, with each sample based on past time. The feature tensors with time steps are used as input: ; In the formula, Given the input feature tensor, This is the current time step; Feature vector at each time step It integrates four types of features: ; In the formula, For structural response characteristics, Characteristics of the internal state of rock and soil. Characteristics of construction activities, Characteristics of environmental disturbance; The dynamic prediction model adopts a multi-task learning paradigm and simultaneously outputs two physically meaningful and interrelated target variables: future displacement prediction and dynamic stability coefficient. in, Future displacement prediction: Directly output future displacement. Within each time step, the absolute displacement of each key monitoring point, such as Figure 3 As shown: ; In the formula, The output model, For the future The predicted absolute displacement value at each time point; Dynamic stability coefficient: Outputs a continuous island wall dynamic stability coefficient between 0 and 1. This coefficient is a data-driven, probabilistic comprehensive safety status indicator. By learning from massive historical data (including stable operating conditions and near-instability or already unstable operating conditions), it predicts the system's future stability. A comprehensive assessment of the probability of maintaining a stable state over a given period: ; In the formula, For the future moment stability coefficient, The function is to standardize an interval and map it to the range (0,1). and The weight matrix and bias terms for the output layer are shown below. For the model decoder at time The state; If the coefficient is close to 1, it indicates a highly stable state; when the coefficient continues to decrease and approaches the preset alarm threshold (such as 0.6), it indicates a significant increase in the risk of instability.
[0026] In this embodiment, the training and validation of the dynamic prediction model are as follows: From the structured time-series database in the cloud data center, we divide the data into training, validation, and test sets. To ensure the predictive ability of the dynamic prediction model for extreme conditions, we use oversampling techniques to enhance the weight of unstable samples. Design a composite loss function, which consists of two weighted parts: ; In the formula, and The weights of the two parameters, This is the mean square error of the displacement prediction, used to ensure the accuracy of the displacement prediction. The cross-entropy loss function is used to optimize the classification ability of the stability coefficient, such as... Figure 4 As shown.
[0027] Step S3: Establish a dynamic construction optimization decision-making mechanism. Compare the output of the dynamic prediction model with preset safety control thresholds, dynamically optimize construction parameters, and form real-time dynamic decisions. Early warnings based on the dynamic prediction model are as follows: Figure 5 As shown.
[0028] In this embodiment, a dynamic construction optimization decision-making mechanism is established: Set multi-level early warning thresholds, including: The stability coefficient threshold is set to a three-level threshold for the output dynamic stability coefficient: Safe zone (green): Stability coefficient > 0.8, indicating that the system is in good condition and construction can proceed as planned; Warning zone (yellow): 0.6 < stability coefficient ≤ 0.8, indicating an unfavorable trend in the system state, requiring attention and preparation for intervention measures; Danger zone (red): Stability coefficient ≤ 0.6, indicating a high risk of system instability, requiring immediate intervention; Displacement and velocity thresholds: At the same time, parallel three-level thresholds are set for predicted displacement and predicted displacement rate as a supplement and verification of the stability coefficient. Real-time early warning and decision support, including: When the prediction result reaches the threshold, the system automatically triggers a tiered warning: Yellow alert: Visually highlighted in the digital twin interface of the management platform, and an alert notification pushed to relevant engineers; Red Alert: In addition to sending the highest level of alarm information to project management personnel, it will also automatically send lock or restrict operation instructions to the construction machinery control system in the relevant area, forcibly suspending high-risk operations from a technical perspective. When an early warning is triggered, multiple preset control schemes will be automatically evaluated, including but not limited to "stop filling", "slow down the filling rate" and "start precipitation", to improve the future stability coefficient and recommend the expected control effect of the optimal control strategy (such as "the stability coefficient is expected to rise to above 0.8 24 hours after the scheme is implemented").
[0029] In this embodiment, the dynamically optimized construction procedures and parameters include: Based on the generated control strategy, the construction plan is dynamically and precisely adjusted; Filling loading control: Based on pore water pressure prediction and stability coefficient, dynamically adjust the thickness, interval time and daily filling rate of layered filling; Reinforcement measures triggering and strengthening: Based on the deformation trend, the pre-set reinforcement plan is activated in advance in the weak area, or the construction density and depth of crushed stone piles / vibratory piles and the spacing of drainage boards are dynamically adjusted. Surcharge preloading management: Optimize the timing of preloading application and removal, and maximize load benefits and minimize construction period based on predicted post-construction settlement and consolidation. Construction machinery scheduling: Based on the real-time stability map, the system intelligently plans the travel path and working area of large machinery to avoid load concentration in areas with low stability coefficients.
[0030] In step S4, the adjusted construction behavior and the resulting actual monitoring results will be fed back to the dynamic prediction model as new data samples for training and optimization.
[0031] In this embodiment, two main dynamic prediction model update triggering modes are established, which trigger model updates based on fixed time intervals or after encountering significant changes in operating conditions: Scheduled triggering: The model update process is automatically started based on a fixed time period; Event Trigger: Automatically triggered when a significant change in on-site conditions is detected. Specific conditions include: the site has entered a completely new construction phase; a significant shift in the statistical characteristics of the input data is detected. The dynamic prediction model iteration includes the following steps: Added data quality verification: Automatically clean and label the collected real-time data to form a new high-quality dataset; Incremental training and weight fine-tuning: An incremental learning algorithm is used to mix the new dataset with some historical core datasets. The current model weights are used as the initial values for supervised fine-tuning, so that the model can absorb new data while retaining its original generalization ability.
[0032] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A construction method for predicting the stability of island wall structures based on machine learning, characterized in that, include: Step S1: Construct a multi-source heterogeneous real-time monitoring network to comprehensively collect multi-source heterogeneous data affecting the stability of the island wall; Step S2: Establish a multi-parameter fusion dynamic prediction model. Based on historical monitoring data, predict the time series of displacement of key points of the island wall and structural stability coefficient in the future. Step S3: Establish a dynamic construction optimization decision-making mechanism, compare the output of the dynamic prediction model with the preset safety control threshold, dynamically optimize construction parameters, and form real-time dynamic decisions. In step S4, the adjusted construction behavior and the resulting actual monitoring results will be fed back to the dynamic prediction model as new data samples for training and optimization.
2. The construction method for predicting the stability of island wall structures based on machine learning according to claim 1, characterized in that, In step S1, constructing a multi-source heterogeneous real-time monitoring network includes: High-precision, automated monitoring equipment was deployed at key locations along the inner and outer slopes, top and bottom of the island wall, using distributed fiber Bragg grating sensor arrays and GPS / GNSS receivers to acquire the vertical, horizontal and deep displacements of the island wall in real time. Sensors were installed at different depths and locations inside the island wall to monitor key soil mechanical parameters in real time. The excess pore water pressure and its dissipation during the soil consolidation process are monitored in real time using a vibrating wire or fiber optic grating pore water pressure gauge. By integrating construction quality control data, the source, gradation, and compaction process parameters of the backfill soil are recorded and input in real time as indirect input parameters characterizing the backfill density. GPS positioning systems are installed on construction machinery to obtain the number, location and operation trajectory of large machinery in real time. Through electronic scales and transportation records, the volume of each layer of backfill soil and the load of surcharge preloading are accurately recorded. Wave and tide gauges are installed near the wave-facing side of the island wall to monitor changes in wave height, wave period, and tide level in real time, thereby quantifying the dynamic impact of wave loads on the island wall.
3. The construction method for predicting the stability of island wall structures based on machine learning according to claim 2, characterized in that, In step S1, all sensors are equipped with 5G wireless communication technology to transmit the collected data to the cloud data center in real time. The raw data from different displacement, soil properties, loads and environments are denoised, aligned and fused to form a unified, high-dimensional multivariate time series input feature vector, thus forming a structured time series database.
4. The construction method for predicting the stability of island wall structures based on machine learning according to claim 3, characterized in that, In step S2, the dynamic prediction model uses a deep learning model specifically designed for processing sequence data, with an encoder-decoder framework as its main architecture, and is further optimized based on this framework. The encoder uses a multi-layer long short-term memory network as its core component for processing the input. Encode the multi-source feature vectors of each historical time step, extract and memorize their long-term temporal dependencies; An attention layer is introduced between the encoder and decoder to assign different weights to the input data at different times and monitoring points, enabling the model to focus on the key time periods and key areas most relevant to future instability. The decoder is also based on a multi-layer long short-term memory network, which gradually generates future context vectors based on the encoder's output and attention weights. The predicted sequence has a time step.
5. The construction method for predicting the stability of island wall structures based on machine learning according to claim 4, characterized in that, In step S2, the inputs to the dynamic prediction model include: Structural response characteristics: including historical and real-time key point horizontal displacement rate, vertical settlement rate, and deep displacement increment; Internal state characteristics of soil and rock: including pore water pressure values at different depths, pore pressure coefficients, and estimated degrees of consolidation calculated based on pore pressure dissipation; Construction activity characteristics include: cumulative fill height, fill rate, and density of heavy machinery on the current work surface; Environmental disturbance characteristics: including significant wave height, tide level, and estimated wave run-up calculated from the combination of wave height and tide level; The above feature vectors are aligned with time to construct training samples, with each sample based on past time. The feature tensors with time steps are used as input: ； In the formula, Given the input feature tensor, This is the current time step; Feature vector at each time step It integrates four types of features: ； In the formula, For structural response characteristics, Characteristics of the internal state of rock and soil. Characteristics of construction activities, Characteristics of environmental disturbance; The dynamic prediction model adopts a multi-task learning paradigm and simultaneously outputs two physically meaningful and interrelated target variables: future displacement prediction and dynamic stability coefficient. in, Future displacement prediction: Directly output future displacement. The absolute displacement of each key monitoring point within each time step: ； In the formula, The output model, For the future The predicted absolute displacement value at each time point; Dynamic stability coefficient: Outputs a continuous island wall dynamic stability coefficient between 0 and 1. This coefficient is a data-driven, probabilistic comprehensive safety status indicator, which, after learning from massive historical data, predicts the system's future stability. A comprehensive assessment of the probability of maintaining a stable state over a given period: ； In the formula, For the future moment stability coefficient, The function is to standardize an interval and map it to the range (0,1). and The weight matrix and bias terms for the output layer are shown below. For the model decoder at time The state; If the coefficient is close to 1, it indicates a highly stable state; when the coefficient continues to decrease and approaches the preset alarm threshold, it indicates a significant increase in the risk of instability.
6. The construction method for predicting the stability of island wall structures based on machine learning according to claim 5, characterized in that, In step S2, the training and validation of the dynamic prediction model are performed. From the structured time-series database in the cloud data center, we divide the data into training, validation, and test sets. To ensure the predictive ability of the dynamic prediction model for extreme conditions, we use oversampling techniques to enhance the weight of unstable samples. Design a composite loss function, which consists of two weighted parts. composition: ； In the formula, and The weights of the two parameters, This is the mean square error of the displacement prediction, used to ensure the accuracy of the displacement prediction. is the cross-entropy loss function for the stability coefficient, used to optimize the classification ability of the stability coefficient.
7. The construction method for predicting the stability of island wall structures based on machine learning according to claim 6, characterized in that, In step S3, a dynamic construction optimization decision-making mechanism is established: Set multi-level early warning thresholds, including: The stability coefficient threshold is set to a three-level threshold for the output dynamic stability coefficient: Safe zone (green): Stability coefficient > 0.8, indicating that the system is in good condition and construction can proceed as planned; Warning zone (yellow): 0.6 < stability coefficient ≤ 0.8, indicating an unfavorable trend in the system state, requiring attention and preparation for intervention measures; Danger zone (red): Stability coefficient ≤ 0.6, indicating a high risk of system instability, requiring immediate intervention; Displacement and velocity thresholds: At the same time, parallel three-level thresholds are set for predicted displacement and predicted displacement rate as a supplement and verification of the stability coefficient. Real-time early warning and decision support, including: When the prediction result reaches the threshold, the system automatically triggers a tiered warning: Yellow alert: Visually highlighted in the digital twin interface of the management platform, and an alert notification pushed to relevant engineers; Red Alert: In addition to sending the highest level of alarm information to project management personnel, it will also automatically send lock or restrict operation instructions to the construction machinery control system in the relevant area, forcibly suspending high-risk operations from a technical perspective. When an early warning is triggered, the system will automatically evaluate various preset control schemes, including but not limited to "stop filling", "slow down the filling rate" and "initiate precipitation", to assess their effects on improving the future stability coefficient and recommend the expected control effect of the optimal control strategy.
8. The construction method for predicting the stability of island wall structures based on machine learning according to claim 7, characterized in that, In step S3, the dynamically optimized construction procedures and parameters include: Based on the generated control strategy, the construction plan is dynamically and precisely adjusted; Filling loading control: Based on pore water pressure prediction and stability coefficient, dynamically adjust the thickness, interval time and daily filling rate of layered filling; Reinforcement measures triggering and strengthening: Based on the deformation trend, the pre-set reinforcement plan is activated in advance in the weak area, or the construction density and depth of crushed stone piles / vibratory piles and the spacing of drainage boards are dynamically adjusted. Surcharge preloading management: Optimize the timing of preloading application and removal based on predicted post-construction settlement and degree of consolidation; Construction machinery scheduling: Based on the real-time stability map, the system intelligently plans the travel path and work area of large machinery.
9. The construction method for predicting the stability of island wall structures based on machine learning according to claim 8, characterized in that, In step S4, two main dynamic prediction model update triggering modes are established, which trigger model updates based on fixed time intervals or after encountering significant changes in operating conditions: Scheduled triggering: The model update process is automatically started based on a fixed time period; Event Trigger: Automatically triggered when a significant change in on-site conditions is detected. Specific conditions include: the site has entered a completely new construction phase; a significant shift in the statistical characteristics of the input data is detected. The dynamic prediction model iteration includes the following steps: Added data quality verification: Automatically clean and label the collected real-time data to form a new high-quality dataset; Incremental training and weight fine-tuning: An incremental learning algorithm is used to mix the new dataset with some historical core datasets, and to perform supervised fine-tuning with the current model weights as the initial values.