BIM-based pumped storage power station building foundation dangerous point management and control system

Abstract

The invention discloses a pumped storage power station building foundation dangerous point management and control system based on BIM, and the system comprises a data layer which enables linear engineering monitoring data to be organized into a high-order tensor of pile number * time * sensor type * spatial position through four-dimensional tensor modeling, and carries out the Tucker decomposition, strain rate abrupt change points are dynamically recognized in the compression process by combining sliding window anomaly detection, original data are directly stored, and then a hierarchical storage strategy is implemented on non-abrupt change data through a data value evaluation model driven by deep reinforcement learning; the model layer is used for fusing a physical information neural network, embedding a rock-soil mechanics constitutive equation into an LSTM network to construct a mixed loss function, pre-training a general geologic model through transfer learning, and then carrying out few-sample fine tuning in combination with project data; and the application layer is used for dynamically generating a hazard source list through a BIM model and triggering instant early warning based on the compressed and reconstructed data of the data layer and the output of the model layer, and simulating a disaster situation development path in combination with a digital twinborn technology.

Description

Technical Field

[0001] The present invention relates to the technical field of dangerous point management and control of buildings or mountain rock foundations, and in particular to a BIM-based dangerous point management and control system for pumped-storage power station building foundations. Background Art

[0002] Pumped-storage power station projects include upper and lower reservoirs, underground powerhouses, and diversion channels. The Tiantai Pumped-storage Power Station in Zhejiang, currently under construction, boasts a maximum head of 724 meters. Pumped-storage power station subprojects may traverse diverse geological units, making construction disturbances prone to geological hazards. Furthermore, the dispersed nature of the work surface creates blind spots in safety oversight. During construction, various sensors, such as those for uplift pressure, seepage, and displacement, are embedded in the foundations. During construction, these sensors experience a buffering period between networking and the host computer, leading to information fragmentation, response lags, and insufficient predictive capabilities, making it difficult to meet the demands of dynamic risk management. The introduction of BIM technology, integrating multi-source data from geology, design, construction, and monitoring through the construction of a parametric 3D building model, enables the "visual location, parametric analysis, and automated early warning" of building foundation hazards. Furthermore, BIM-based construction simulations for pumped-storage power station foundations can proactively identify high-risk processes and optimize construction organization to minimize personnel exposure and potential injuries.

[0003] Existing state monitoring systems for pumped-storage power station units require the deployment of hundreds of sensors, including eddy current sensors (for measuring shaft displacement), velocity sensors (for measuring frame vibration), acceleration sensors (for measuring sub-core vibration), and pressure sensors (for measuring hydraulic system pulsation). These sensors cover parameters such as vibration, runout, pressure, temperature, and electrical quantity. Measured data shows that the sampling frequency of pressure pulsation signals can reach 1,000 times per second, with some experiments even employing high-frequency acquisition rates of 200,000 times per second to meet the needs of transient process analysis. Based on 500 sensors and 1,000 sampling times per second for a single power station, the daily data volume reaches 4.32 billion records, and the annual data volume easily exceeds terabytes. However, traditional databases rely on vertical or horizontal expansion to cope with data growth, resulting in exponentially increasing storage costs. To reduce costs, projects often use lossy compression or downsampling, but this can result in the loss of critical information. Therefore, a BIM-based management system for hazardous point control in pumped-storage power station building foundations is proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a BIM-based pumped storage power station building foundation danger point management and control system.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A BIM-based pumped storage power station building foundation hazard point management and control system, including:

[0007] Data layer: Utilizing four-dimensional tensor modeling, linear engineering monitoring data is organized into a high-order tensor of "stake number × time × sensor type × spatial location." Tucker decomposition is then combined with sliding window anomaly detection to dynamically identify strain rate mutation points during the compression process and directly store the raw data. A data value assessment model driven by deep reinforcement learning then implements a tiered storage strategy for non-mutation data.

[0008] Model layer: Integrating physical information neural networks, the geotechnical constitutive equations are embedded in the LSTM network to construct a hybrid loss function. A general geological model is pre-trained through transfer learning and then fine-tuned with project data using a small number of samples. A Bayesian adaptive inversion engine is developed to update geological parameter confidence intervals in real time.

[0009] Application layer: Based on the compressed reconstruction data of the data layer and the output of the model layer, a list of hazard sources is dynamically generated through the BIM model and immediate warnings are triggered, and digital twin technology is combined to simulate the development path of the disaster.

[0010] The above technical solution further includes:

[0011] Furthermore, the linear engineering monitoring data is organized into a high-order tensor of "pile number × time × sensor type × spatial position" by using four-dimensional tensor modeling, and decomposed by Tucker, including the following steps:

[0012] Data collection and preprocessing: Data cleaning techniques are used to remove noise interference, standardization methods are used to eliminate the magnitude differences between monitoring data of different dimensions, and missing values ​​are filled through linear interpolation;

[0013] Four-dimensional tensor construction: The pre-processed data is organized into a four-dimensional tensor structure of "stake number × time × sensor type × spatial location" according to the physical characteristics of the linear infrastructure project;

[0014] Tucker decomposition: decompose the original tensor into core tensors and the factor matrix The product form is X≈G×1U1×2U2×3U3×4U4, and each component is solved iteratively through alternating least squares optimization, where G is the core tensor that captures multimodal associations, U1 is the pile number pattern factor matrix, U2 is the time pattern factor matrix, U3 is the sensor type pattern factor matrix, U4 is the spatial position pattern factor matrix, and R1, R2, R3, and R4 are the preset compressed dimensions.

[0015] Furthermore, the specific steps of dynamically identifying strain rate mutation points during compression and directly storing raw data in combination with sliding window anomaly detection are as follows:

[0016] Data preprocessing: Obtain time series data from the linear engineering monitoring system, set the sliding window length, apply Z-score normalization to the data within the window, and calculate the mean and standard deviation;

[0017] Sliding window scan: For each sensor's time series X i,j,k,l , apply sliding window scanning, initialize the window starting point, calculate the strain rate in the window, and the calculation formula is expressed as Δε / Δt=[ε(t)-t(tW)] / WΔt. If Δε / Δt>γ, mark the window [t, t+N-1] as the mutation point window, where γ adopts an adaptive threshold mechanism and is updated according to the historical data distribution, expressed as γ=μ+3σ. Sliding window t=t+1, repeat until the window end t+N-1=J;

[0018] Compression strategy adjustment: Perform Tucker decomposition compression on unmarked window data, and directly store the original data in the mutation point window data without compression;

[0019] Tensor reconstruction and hazard source analysis: For non-mutation point data, the approximate value is reconstructed through the core tensor G and factor matrices U1, U2, U3, and U4, and the original mutation point data is directly filled into the corresponding position of the reconstructed tensor. Based on the complete reconstructed tensor, a hazard source list is automatically generated through the BIM model.

[0020] Furthermore, the specific steps of implementing a tiered storage strategy for non-mutated data using a data value assessment model driven by deep reinforcement learning are as follows:

[0021] Data feature extraction: Extract features from each data window. The features include strain rate, data frequency, physical significance weight and time-sensitive attenuation factor. Reflecting the speed of deformation, the calculation formula is expressed as The data frequency μ is the number of data updates per unit time, the physical significance weight w is assigned according to the sensor type and spatial position, and the timeliness attenuation factor τ is the value decay of data over time. The features are combined into a vector

[0022] Deep reinforcement learning drives storage decisions: Build a deep Q network and design a reward function. The reward function includes hot storage rewards, cold storage rewards, and edge computing rewards. The hot storage reward gives high rewards to mutation point data, the cold storage reward gives low rewards to low-frequency historical data, and the edge computing reward gives medium rewards to data with high real-time requirements. The training model dynamically selects the optimal storage strategy based on data characteristics. The training process is represented by initializing the Q network parameters θ, selecting action a for each data window according to the ∈-greedy strategy, executing action a, and observing the reward R and the new state f. ', update Q network;

[0023] Storage tier division: Allocate data to different levels of storage media based on the strategy output by deep reinforcement learning.

[0024] Furthermore, the physical information fusion neural network embeds the geotechnical mechanics constitutive equation into the LSTM network to construct a hybrid loss function, including the following steps:

[0025] LSTM core construction: Use gating mechanism and memory unit to capture the time dependency of monitoring data, control the information flow and update the memory unit state through input gate, forget gate and output gate. The input gate is responsible for updating the cell state. The input gate consists of two parts: one is the sigmoid layer, which determines which information will be updated; the other is the tanh layer, which creates a new candidate value vector. The new candidate value vector is added to the cell state. Finally, the two information are multiplied to update the cell state. The forget gate is represented by f t =σ(W f ·[h t-1 ,x t ]+b f ), where W f is the weight matrix of the forget gate, b f is the bias term, σ is the sigmoid function, [h t-1 ,x t ] means h t-1 and x t Spliced ​​into a vector, the output gate determines which part of the information based on the cell state is used for output, according to the current input tx t , the state of the hidden layer at the previous moment h t-1 And the latest cell state C t , through the combined action of sigmoid function and tanh function, the output h at the current moment is determined t , output hidden state h t =o t *tanh(C t );

[0026] Physical constraints are imposed: The geotechnical constitutive equation is embedded in the LSTM hidden layer, and the residual of the physical equation is calculated by automatic differentiation. The gradient of the residual with respect to the input is expressed as The gradient is used for back propagation to force the network output to conform to physical laws;

[0027] Hybrid loss function: A hybrid loss function is constructed to fuse the data fitting error and the physical residual term, and a balance coefficient is introduced to adjust the weights of the two.

[0028] Furthermore, the general geological model is pre-trained by transfer learning and then fine-tuned with a small number of samples based on project data, including the following steps:

[0029] Dataset construction: A general geological model including the linear relationship between seepage velocity and hydraulic gradient is constructed using a public fault zone seepage dataset.

[0030] Pre-training objective function: learn the prior distribution of rock mass permeability k by minimizing the loss function containing only physical constraints. The loss function containing physical constraints is expressed as L pretrain =Σ(v j -k·i j ) 2 ;

[0031] Target project data preparation: 4%-6% of the target project's labeled data is required, with oversampling of key segments (such as when the excavation face advances to the fault);

[0032] Fine-tuning strategy: freeze the physical layer parameters of PINN-LSTM;

[0033] Weight update: Only LSTM recurrent kernel weights and bias updates are allowed;

[0034] Fine-tuning the loss function: Combine the labeled data of the target project to construct a hybrid loss function that includes data fitting error and optional physical residual terms, and quickly adapt to specific engineering conditions through few-shot learning.

[0035] Furthermore, the specific steps of developing the Bayesian adaptive inversion engine to update the confidence intervals of geological parameters in real time are as follows:

[0036] Parameterization and prior distribution setting: The rock mass elastic modulus is parameterized using lognormal distribution and the prior mean and standard deviation are set based on the geological survey report. The parameter uncertainty is quantified using the probability density function.

[0037] Likelihood function construction: Construct a Gaussian likelihood function to associate monitoring data with elastic modulus, and combine numerical simulation functions with measurement noise models to establish a physical connection between data and parameters;

[0038] Posterior distribution sampling: Bayesian theorem and Metropolis-Hastings algorithm are applied to perform posterior distribution sampling, and the parameter confidence interval is dynamically updated. The posterior distribution is expressed as p(E|D)∝p(D|E)·p(E), where p(E|D) is the posterior probability distribution, which represents the probability distribution of the rock elastic modulus E under the condition of the monitoring dataset D, p(D|E) is the likelihood function, which represents the probability distribution of the monitoring dataset D under the condition of the rock elastic modulus E, and p(E) is the prior probability distribution, which represents the probability distribution of the rock elastic modulus E before the monitoring dataset D.

[0039] Parameter updating and model correction: Feedback of inversion results to the physical information neural network-long short-term memory network hybrid model is achieved through an adaptive weight adjustment mechanism, strengthening the deep coupling of physical constraints and data-driven approaches. Sliding window prior updates and an adaptive MCMC algorithm are used to optimize sampling efficiency.

[0040] Effect verification: Verify through case studies and integrate the inversion parameters into the BIM model in real time.

[0041] Furthermore, the compressed reconstruction data based on the data layer and the output of the model layer dynamically generates a list of hazard sources through the BIM model and triggers an immediate warning, and combines digital twin technology to simulate the development path of the disaster, including the following steps:

[0042] Dynamic generation of hazard sources: Based on threshold comparison logic, the reconstructed monitoring data and the model output risk probability are compared with the preset thresholds. When any indicator exceeds the limit, the hazard source mark of the corresponding section in the BIM model is activated, and a list containing spatial location and risk level is generated;

[0043] Early warning triggering: Using a multi-channel push mechanism, early warning information is sent to the on-site personnel terminal through the 5G network, and the sound and light alarm device and camera steering are linked;

[0044] Digital twin disaster simulation: Using Bayesian update-driven finite element analysis, monitoring data and prior knowledge are integrated in real time to update the posterior distribution of rock mass parameters. The updated parameters are then input into the fluid dynamics equations to simulate the disaster development path with high fidelity in the digital twin.

[0045] Emergency response optimization: Through a multi-objective path planning algorithm, with escape time, path length, and risk exposure value as optimization targets, an optimal evacuation route is generated for workers to avoid dangerous areas.

[0046] The present invention has the following beneficial effects:

[0047] In this invention, four-dimensional tensor modeling unifies dispersed monitoring data into a four-dimensional matrix of station number × time × sensor type × spatial location. This structured representation not only preserves the temporal and spatial correlation of the data but also achieves lossless compression through Tucker decomposition, compressing terabytes of raw data to gigabytes of storage capacity, reducing storage costs. Sliding window anomaly detection performs sliding window analysis on each sensor data stream before compression to calculate the strain rate. If the strain rate exceeds a threshold, the window is immediately marked as a "micro-deformation feature segment" and compression is disabled, allowing the raw data to be stored directly, avoiding signal loss due to compression. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1This is a system block diagram of a BIM-based pumped storage power station building foundation dangerous point management and control system proposed by the present invention;

[0049] Figure 2 This is a flow chart of the method for dynamically identifying the strain rate mutation point during the compression process in the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] See also Figure 1-Figure 2 As shown, the present invention is a BIM-based pumped storage power station building foundation danger point management and control system, including:

[0052] Data layer: Utilizing four-dimensional tensor modeling, linear engineering monitoring data is organized into a high-order tensor of "stake number × time × sensor type × spatial location." Tucker decomposition is then used to achieve lossless compression to reduce storage costs. Sliding window anomaly detection (SWAD) is combined with dynamic identification of strain rate mutation points (such as micro-deformation characteristic segments exceeding 1e-5 / s) during the compression process and directly stores the raw data. A data value assessment model driven by deep reinforcement learning (DRL) implements a tiered storage strategy for non-mutation data, ensuring zero loss of critical information and efficient use of storage resources.

[0053] Model layer: Integrating physical information neural networks (PINNs), the geotechnical constitutive equations (such as the Mohr-Coulomb criterion) are embedded in the LSTM network to construct a hybrid loss function. Pre-training general geological models (such as the seepage law in fault zones) through transfer learning is then combined with project data for small-sample fine-tuning. Furthermore, a Bayesian adaptive inversion engine is developed to update geological parameter confidence intervals in real time, achieving deep coupling of physical constraints and data-driven approaches.

[0054] Application layer: Based on the compressed reconstruction data of the data layer and the output of the model layer, a list of hazard sources is dynamically generated through the BIM model and immediate warnings are triggered, and digital twin technology is combined to simulate the development path of the disaster.

[0055] In one embodiment, the linear engineering monitoring data is organized into a high-order tensor of "stake number × time × sensor type × spatial position" using four-dimensional tensor modeling and decomposed by Tucker, including the following steps:

[0056] Data collection and preprocessing: Data cleaning techniques are used to remove noise interference caused by sensor failures, standardization methods are used to eliminate the magnitude differences between different dimensional monitoring data (such as strain, displacement, and pressure), and linear interpolation is used to fill in missing values ​​caused by communication interruptions.

[0057] Four-dimensional tensor construction: The pre-processed data is organized into a four-dimensional tensor structure of "pile number × time × sensor type × spatial location" according to the physical characteristics of the linear infrastructure project. For example, a tunnel project can be represented as The dimensional design of 1,000 monitoring sections, 365-day time series, 8 types of sensors (such as strain gauges and displacement gauges) and 5 spatial monitoring points (such as vaults and side walls) realizes the structured integration of multi-source monitoring data;

[0058] Tucker decomposition: decompose the original tensor into core tensors and the factor matrix The product form is X≈G×1U1×2U2×3U3×4U4, and each component is solved iteratively through alternating least squares (ALS) optimization, where G is the core tensor that captures multimodal associations, U1 is the pile number pattern factor matrix, U2 is the time pattern factor matrix, U3 is the sensor type pattern factor matrix, U4 is the spatial position pattern factor matrix, and R1, R2, R3, and R4 are the preset compressed dimensions.

[0059] Initialization: Randomly initialize the core tensor G and factor matrices U1 to U4;

[0060] Alternating Least Squares (ALS) Optimization:

[0061] Fix U2, U3, U4, G, and optimize U1: U1←argmin||XG×1U1×2U2×3U3×4U4||_F 2 ;

[0062] Repeat optimization of U2, U3, U4, G until convergence (e.g., residual < 1e-6);

[0063] Truncated Singular Value Distortion (SVD): Perform SVD on each factor matrix and retain the first R singular values ​​to control the compression ratio.

[0064] In one embodiment, the specific steps of dynamically identifying strain rate mutation points during compression and directly storing raw data in combination with sliding window anomaly detection are as follows:

[0065] Data preprocessing: Obtain time series data from the linear engineering monitoring system, set the sliding window length, apply Z-score normalization to the data within the window, and calculate the mean and standard deviation;

[0066] Sliding window scan: For each sensor's time series X i,j,k,l (corresponding to stake number i, time j, sensor type k, spatial position l), apply sliding window scanning, initialize the window starting point, calculate the strain rate in the window, and the calculation formula is expressed as Δε / Δt=[ε(t)-t(tW)] / WΔt. If Δε / Δt>γ, mark the window [t, t+N-1] as the mutation point window, where γ adopts an adaptive threshold mechanism and is updated according to the historical data distribution, expressed as γ=μ+3σ. Sliding window t=t+1, repeat until the window end t+N-1=J;

[0067] Compression strategy adjustment: Perform Tucker decomposition compression on unmarked window data, and directly store the original data in the mutation point window data without compression;

[0068] Tensor reconstruction and hazard source analysis: For non-mutation point data, the approximate value is reconstructed through the core tensor G and factor matrices U1, U2, U3, and U4, and the original mutation point data is directly filled into the corresponding position of the reconstructed tensor. Based on the complete reconstructed tensor (including mutation point data), a hazard source list is automatically generated through the BIM model.

[0069] In one embodiment, the specific steps of implementing a tiered storage strategy for non-mutated data using a data value assessment model driven by deep reinforcement learning (DRL) are as follows:

[0070] Data feature extraction: Extract features from each data window. The features include strain rate, data frequency, physical significance weight and time-sensitive attenuation factor. Reflecting the speed of deformation, the calculation formula is expressed as The data frequency μ is the number of data updates per unit time. The physical significance weight w is assigned according to the sensor type (e.g., displacement meter has a higher weight than thermometer) and spatial position (e.g., vault has a higher weight than side wall). The timeliness attenuation factor τ is the value decay of data over time. The features are combined into a vector

[0071] Deep reinforcement learning (DRL) drives storage decisions: A deep Q-network (DQN) is constructed and a reward function is designed. The reward function includes hot storage rewards, cold storage rewards, and edge computing rewards. The hot storage reward gives high rewards to mutation point data, the cold storage reward gives low rewards to low-frequency historical data, and the edge computing reward gives medium rewards to data with high real-time requirements (such as μ>0.5). The training model dynamically selects the optimal storage strategy (hot storage, cold storage, or edge computing) based on data characteristics. The training process is represented by initializing the Q-network parameters θ, selecting action a for each data window according to the ∈-greedy strategy, executing action a, observing the reward R and the new state f', and updating the Q-network. Critical data is stored in high-performance storage media (such as the in-memory database Redis) to ensure real-time availability, while non-critical data is stored in low-cost storage media (such as the distributed file system HDFS) to reduce long-term storage costs.

[0072] Storage level division: Data is allocated to storage media at different levels according to the strategy output by deep reinforcement learning (DRL). Hot storage stores high-frequency critical data (such as mutation point windows and real-time monitoring data), using in-memory databases (such as Redis). Cold storage stores low-frequency historical data (such as monitoring records of completed construction sections), using distributed file systems (such as HDFS). Edge computing performs local calculations on edge servers (such as on-site industrial computers) for data with high real-time requirements (such as monitoring values ​​of the current excavation surface) to reduce cloud transmission delays.

[0073] In one embodiment, the fusion of physical information neural network, embedding the geotechnical mechanics constitutive equation into the LSTM network to construct a hybrid loss function, includes the following steps:

[0074] LSTM core construction: Use gating mechanism and memory unit to capture the time dependency of monitoring data, control the information flow and update the memory unit state through input gate, forget gate and output gate. The input gate is responsible for updating the cell state. The input gate consists of two parts: one is the sigmoid layer, which determines which information will be updated; the other is the tanh layer, which creates a new candidate value vector. The new candidate value vector is added to the cell state. Finally, the two information are multiplied to update the cell state. The forget gate is represented by f t =σ(W f ·[h t-1 ,x t ]+b f ), where W f is the weight matrix of the forget gate, b f is the bias term, σ is the sigmoid function, [h t-1 ,x t ] means ht-1 and x t Spliced ​​into a vector, the output gate determines which part of the information based on the cell state is used for output, according to the current input tx t , the state of the hidden layer at the previous moment h t-1 And the latest cell state C t , through the combined action of sigmoid function and tanh function, the output h at the current moment is determined t , output hidden state h t =o t *tanh(C t );

[0075] Physical constraints are imposed: The geotechnical constitutive equations (such as the Mohr-Coulomb criterion) are embedded in the LSTM hidden layer, and the residual of the physical equation is calculated by automatic differentiation. The gradient of the residual with respect to the input is expressed as The gradient is used for back propagation to force the network output to conform to physical laws;

[0076] Hybrid loss function: A hybrid loss function is constructed to fuse the data fitting error and the physical residual term, and a balance coefficient is introduced to adjust the weights of the two.

[0077] In one embodiment, the method of pre-training a general geological model (such as the seepage law of a fault zone) through transfer learning and then fine-tuning it with a small number of samples based on project data includes the following steps:

[0078] Dataset construction: A general geological model including the linear relationship between seepage velocity and hydraulic gradient (v = k·i) was constructed using a public fault zone seepage dataset.

[0079] Pre-training objective function: learn the prior distribution of rock mass permeability k by minimizing the loss function containing only physical constraints. The loss function containing physical constraints is expressed as L pretrain =Σ(v j -k·i j ) 2 ;

[0080] Target project data preparation: 4%-6% of the target project's labeled data is required (e.g., a three-month monitoring record of a soft rock tunnel, including vault settlement S and surrounding rock stress σ). Oversampling is performed on key segments (e.g., when the excavation face advances to a fault) to improve event coverage.

[0081] Fine-tuning strategy: Freeze the physical layer parameters of PINN-LSTM to prevent overfitting;

[0082] Weight update: Only LSTM recurrent kernel weights and bias updates are allowed;

[0083] Fine-tuning the loss function: Combine the labeled data of the target project to construct a hybrid loss function that includes data fitting error and optional physical residual terms, and quickly adapt to specific engineering conditions through few-shot learning.

[0084] In one embodiment, the specific steps of developing a Bayesian adaptive inversion engine to update geological parameter confidence intervals in real time are as follows:

[0085] Parameterization and prior distribution setting: The rock mass elastic modulus is parameterized using lognormal distribution and the prior mean and standard deviation are set based on the geological survey report. The parameter uncertainty is quantified using the probability density function.

[0086] Likelihood function construction: Construct a Gaussian likelihood function to associate monitoring data with elastic modulus, and combine numerical simulation functions with measurement noise models to establish a physical connection between data and parameters;

[0087] Posterior distribution sampling: Bayesian theorem and Metropolis-Hastings algorithm are applied to perform posterior distribution sampling, and the parameter confidence interval is dynamically updated. The posterior distribution is expressed as p(E|D)∝p(D|E)·p(E), where p(E|D) is the posterior probability distribution, which represents the probability distribution of the rock elastic modulus E under the condition of the monitoring dataset D, p(D|E) is the likelihood function, which represents the probability distribution of the monitoring dataset D under the condition of the rock elastic modulus E, and p(E) is the prior probability distribution, which represents the probability distribution of the rock elastic modulus E before the monitoring dataset D.

[0088] Parameter updating and model correction: The inversion results are fed back to the physical information neural network-long short-term memory network (PINN-LSTM) hybrid model through an adaptive weight adjustment mechanism, strengthening the deep coupling of physical constraints and data-driven. Sliding window prior updates and adaptive MCMC algorithms are used to optimize sampling efficiency and ensure real-time performance.

[0089] Effect verification: Through case studies, the effects of improved prediction accuracy, accelerated convergence speed, and quantitative decision support are verified, and the inversion parameters are integrated into the BIM model in real time.

[0090] In one embodiment, the compressed reconstruction data of the data layer and the output of the model layer are used to dynamically generate a list of hazard sources through the BIM model and trigger an immediate warning, and the digital twin technology is used to simulate the development path of the disaster, including the following steps:

[0091] Dynamic generation of hazard sources: Based on threshold comparison logic, the reconstructed monitoring data and the model output risk probability are compared with the preset thresholds. When any indicator exceeds the limit, the hazard source mark of the corresponding section in the BIM model is activated, and a list containing spatial location and risk level is generated;

[0092] Early warning triggering: Using a multi-channel push mechanism, early warning information is sent to the on-site personnel terminal through the 5G network, and the sound and light alarm device and camera steering are linked;

[0093] Digital twin disaster simulation: Using Bayesian update-driven finite element analysis, monitoring data and prior knowledge are integrated in real time to update the posterior distribution of rock mass parameters. The updated parameters are then input into the fluid dynamics equations to simulate the disaster development path with high fidelity in the digital twin.

[0094] Emergency response optimization: Through a multi-objective path planning algorithm, with escape time (t), path length (d), and risk exposure value (r) as optimization targets (min(w1·t+w2·d+w3·r)), an optimal evacuation route is generated for workers to avoid dangerous areas.

[0095] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A BIM-based pumped storage power station building foundation danger point management and control system, characterized by: include: Data layer: Utilizing four-dimensional tensor modeling, linear engineering monitoring data is organized into a high-order tensor of "stake number × time × sensor type × spatial location." Tucker decomposition is then combined with sliding window anomaly detection to dynamically identify strain rate mutation points during the compression process and directly store the raw data. A data value assessment model driven by deep reinforcement learning then implements a tiered storage strategy for non-mutation data. Model layer: Integrating physical information neural networks, the geotechnical constitutive equations are embedded in the LSTM network to construct a hybrid loss function. A general geological model is pre-trained through transfer learning and then fine-tuned with project data using a small number of samples. A Bayesian adaptive inversion engine is developed to update geological parameter confidence intervals in real time. Application layer: Based on the compressed reconstruction data of the data layer and the output of the model layer, a list of hazard sources is dynamically generated through the BIM model and immediate warnings are triggered, and digital twin technology is combined to simulate the development path of the disaster.

2. A BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 1, characterized in that: The four-dimensional tensor modeling is used to organize the linear engineering monitoring data into a high-order tensor of "pile number × time × sensor type × spatial position" and decompose it through Tucker, including the following steps: Data collection and preprocessing: Data cleaning techniques are used to remove noise interference, standardization methods are used to eliminate the magnitude differences between monitoring data of different dimensions, and missing values ​​are filled through linear interpolation; Four-dimensional tensor construction: The preprocessed data is organized into a four-dimensional tensor structure of "stake number × time × sensor type × spatial location" according to the physical characteristics of the linear infrastructure project; Tucker decomposition: decompose the original tensor into core tensors and the factor matrix The product form is X≈G×1U1×2U2×3U3×4U4, and each component is solved iteratively through alternating least squares optimization, where G is the core tensor that captures multimodal associations, U1 is the pile number pattern factor matrix, U2 is the time pattern factor matrix, U3 is the sensor type pattern factor matrix, U4 is the spatial position pattern factor matrix, and R1, R2, R3, and R4 are the preset compressed dimensions.

3. A BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 2, characterized in that: The specific steps of combining sliding window anomaly detection to dynamically identify strain rate mutation points during compression and directly store raw data are as follows: Data preprocessing: Obtain time series data from the linear engineering monitoring system, set the sliding window length, apply Z-score normalization to the data within the window, and calculate the mean and standard deviation; Sliding window scan: For each sensor's time series X i,j,k,l , apply sliding window scanning, initialize the window starting point, calculate the strain rate in the window, and the calculation formula is expressed as Δε / Δt=[ε(t)-t(tW)] / WΔt. If Δε / Δt>γ, mark the window [t, t+N-1] as the mutation point window, where γ adopts an adaptive threshold mechanism and is updated according to the historical data distribution, expressed as γ=μ+3σ. Sliding window t=t+1, repeat until the window end t+N-1=J; Compression strategy adjustment: Perform Tucker decomposition compression on unmarked window data, and directly store the original data in the mutation point window data without compression; Tensor reconstruction and hazard source analysis: For non-mutation point data, the approximate value is reconstructed through the core tensor G and factor matrices U1, U2, U3, and U4, and the original mutation point data is directly filled into the corresponding position of the reconstructed tensor. Based on the complete reconstructed tensor, a hazard source list is automatically generated through the BIM model.

4. A BIM-based pumped storage power station building foundation danger point management and control system according to claim 1, characterized in that: The specific steps of implementing a tiered storage strategy for non-mutated data using a data value assessment model driven by deep reinforcement learning are as follows: Data feature extraction: Extract features from each data window. The features include strain rate, data frequency, physical significance weight and time-sensitive attenuation factor. Reflecting the speed of deformation, the calculation formula is expressed as The data frequency μ is the number of data updates per unit time, the physical significance weight w is assigned according to the sensor type and spatial position, and the timeliness attenuation factor τ is the value decay of data over time. The features are combined into a vector Deep reinforcement learning drives storage decisions: Build a deep Q network and design a reward function. The reward function includes hot storage rewards, cold storage rewards, and edge computing rewards. The hot storage reward gives high rewards to mutation point data, the cold storage reward gives low rewards to low-frequency historical data, and the edge computing reward gives medium rewards to data with high real-time requirements. The training model dynamically selects the optimal storage strategy based on data characteristics. The training process is represented by initializing the Q network parameters θ, selecting action a for each data window according to the ∈-greedy strategy, executing action a, observing the reward R and the new state f', and updating the Q network. Storage tier division: Allocate data to different levels of storage media based on the strategy output by deep reinforcement learning.

5. The BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 1 is characterized in that: The fusion of physical information neural network, embedding the geotechnical constitutive equation into the LSTM network to construct a hybrid loss function, includes the following steps: LSTM core construction: Use gating mechanism and memory unit to capture the time dependency of monitoring data, control the information flow and update the memory unit state through input gate, forget gate and output gate. The input gate is responsible for updating the cell state. The input gate consists of two parts: one is the sigmoid layer, which determines which information will be updated; the other is the tanh layer, which creates a new candidate value vector. The new candidate value vector is added to the cell state. Finally, the two information are multiplied to update the cell state. The forget gate is represented by f t =σ(W f ·[h t-1 ,x t ]+b f ), where W f It is forgotten The gate weight matrix, b f is the bias term, σ is the sigmoid function, [h t-1 ,x t ] means h t-1 and x t Spliced ​​into a vector, the output gate determines which part of the information based on the cell state is used for output, according to the current input tx t , the state of the hidden layer at the previous moment h t-1 And the latest cell state C t , through the combined action of sigmoid function and tanh function, the output h at the current moment is determined t , output hidden state h t =o t *tanh(C t ); Physical constraints are imposed: The geotechnical constitutive equation is embedded in the LSTM hidden layer, and the residual of the physical equation is calculated by automatic differentiation. The gradient of the residual with respect to the input is expressed as The gradient is used for back propagation to force the network output to conform to physical laws; Hybrid loss function: A hybrid loss function is constructed to fuse the data fitting error and the physical residual term, and a balance coefficient is introduced to adjust the weights of the two.

6. A BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 5, characterized in that: The process of pre-training the general geological model through transfer learning and then fine-tuning it with a small number of samples based on the project data includes the following steps: Dataset construction: A general geological model including the linear relationship between seepage velocity and hydraulic gradient is constructed using a public fault zone seepage dataset. Pre-training objective function: learn the prior distribution of rock mass permeability k by minimizing the loss function containing only physical constraints. The loss function containing physical constraints is expressed as L pretrain =Σ(v j -k·i j ) 2 ; Target project data preparation: 4%-6% of the target project's labeled data is required, with oversampling of key segments (such as when the excavation face advances to the fault); Fine-tuning strategy: freeze the physical layer parameters of PINN-LSTM; Weight update: Only LSTM recurrent kernel weights and bias updates are allowed; Fine-tuning the loss function: Combine the labeled data of the target project to construct a hybrid loss function that includes data fitting error and optional physical residual terms, and quickly adapt to specific engineering conditions through few-shot learning.

7. A BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 6, characterized in that: The specific steps of developing a Bayesian adaptive inversion engine to update geological parameter confidence intervals in real time are as follows: Parameterization and prior distribution setting: The rock mass elastic modulus is parameterized using lognormal distribution and the prior mean and standard deviation are set based on the geological survey report. The parameter uncertainty is quantified using the probability density function. Likelihood function construction: Construct a Gaussian likelihood function to associate monitoring data with elastic modulus, and combine numerical simulation functions with measurement noise models to establish a physical connection between data and parameters; Posterior distribution sampling: Bayesian theorem and Metropolis-Hastings algorithm are applied to perform posterior distribution sampling, and the parameter confidence interval is dynamically updated. The posterior distribution is expressed as p(E|D)∝p(D|E)·p(E), where p(E|D) is the posterior probability distribution, which represents the probability distribution of the rock elastic modulus E under the condition of the monitoring dataset D, p(D|E) is the likelihood function, which represents the probability distribution of the monitoring dataset D under the condition of the rock elastic modulus E, and p(E) is the prior probability distribution, which represents the probability distribution of the rock elastic modulus E before the monitoring dataset D. Parameter updating and model correction: Feedback of inversion results to the physical information neural network-long short-term memory network hybrid model is achieved through an adaptive weight adjustment mechanism, strengthening the deep coupling of physical constraints and data-driven approaches. Sliding window prior updates and an adaptive MCMC algorithm are used to optimize sampling efficiency. Effect verification: Verify through case studies and integrate the inversion parameters into the BIM model in real time.

8. The BIM-based pumped storage power station building foundation dangerous point management and control system according to claim 1 is characterized in that: The compressed reconstruction data based on the data layer and the output of the model layer dynamically generates a list of hazard sources through the BIM model and triggers an immediate warning, and combines digital twin technology to simulate the development path of the disaster, including the following steps: Dynamic generation of hazard sources: Based on threshold comparison logic, the reconstructed monitoring data and the model output risk probability are compared with the preset thresholds. When any indicator exceeds the limit, the hazard source mark of the corresponding section in the BIM model is activated, and a list containing spatial location and risk level is generated; Early warning triggering: Using a multi-channel push mechanism, early warning information is sent to the on-site personnel terminal through the 5G network, and the sound and light alarm device and camera steering are linked; Digital twin disaster simulation: Using Bayesian update-driven finite element analysis, monitoring data and prior knowledge are integrated in real time to update the posterior distribution of rock mass parameters. The updated parameters are then input into the fluid dynamics equations to simulate the disaster development path with high fidelity in the digital twin. Emergency response optimization: Through a multi-objective path planning algorithm, with escape time, path length, and risk exposure value as optimization targets, an optimal evacuation route is generated for workers to avoid dangerous areas.

Images