Deepwater drilling overflow early warning method based on stratum-wellbore energy evolution and multi-physical field fusion

Abstract

The application discloses a deepwater drilling overflow early warning method based on stratum-wellbore energy evolution and multi-physical field fusion, and comprises the following steps: firstly, multi-source logging data in the deepwater drilling process is acquired to form a wellbore dynamic observation sequence; secondly, a wellbore multiphase flow state representation model coupled with temperature and pressure is constructed, and a unit time energy imbalance amount input from the stratum to the wellbore is calculated based on the stratum pressure and the bottom hole pressure difference; thirdly, a multi-physical field collaborative abnormality degree and a time sequence causal consistency coefficient are constructed to generate a physical rule side risk level; meanwhile, a time sequence intelligent model with physical information constraint is adopted to generate a data driving side risk level; finally, a dual-mode redundant fusion decision is made according to the confidence of both sides to output a final overflow warning level; and the application can improve the timeliness, accuracy and interpretability of the early identification of the deepwater drilling overflow.

Description

Technical Field

[0001] This invention belongs to the field of deepwater oil and gas drilling engineering safety monitoring technology, specifically involving an early warning method for deepwater drilling overflow based on formation-wellbore energy evolution and multi-physics field fusion. Background Technology

[0002] In deepwater drilling, overflow is a major risk source threatening operational safety and causing blowouts. Due to the characteristics of the deepwater environment, such as high pressure and low temperature, long risers, narrow safety density windows, and strong temperature-pressure coupling, formation fluids, after invading the wellbore, undergo continuous gas expansion during their upward flow, accompanied by significant gas-liquid slippage, flow pattern transformation, and dynamic changes in annular pressure loss. This results in a significantly nonlinear and unsteady evolution of the wellbore flow state. Early overflow signals are extremely weak, and traditional identification methods relying on surface logging parameters such as pool volume increment, outlet flow rate difference, and riser pressure suffer from problems such as response lag, high false alarm rates, and lack of physical interpretability.

[0003] In recent years, researchers have attempted to apply deep learning models (such as recurrent neural networks and long short-term memory networks) to kick detection, aiming to achieve anomaly detection by mining the temporal features of logging data. For example, Oyedele et al., in their paper "A new methodology for kick detection during petroleum drilling using long short-term memory recurrent neural network," used LSTM-RNN to detect kicks using time-series data such as d-exponent and riser pressure. However, these methods mainly focus on learning the statistical features of logging time-series data, typically relying on the coverage of anomaly patterns by training samples. They do not couple with the real physical mechanisms downhole (such as multiphase flow evolution and formation energy transfer processes), resulting in insufficient generalization ability under complex deep-water conditions and difficulty in explaining why the model makes a certain judgment. Furthermore, existing methods generally employ a single recognition architecture. For example, Bing Lei compared RNN, LSTM, GRU, and Attention-GRU for early overflow recognition in his paper "Early Overflow Recognition and Early Warning Method Based on Attention-GRU Algorithm". In addition, patent application CN120277398A "An Early Overflow Recognition Method Based on Self-Supervised Learning" discloses an early overflow recognition scheme based on CNN-LSTM encoder and multi-branch consistency constraint Siamese network. When downhole conditions change, sensors experience short-term noise, or training data does not fully cover the complex overflow evolution patterns in deepwater drilling, the model is prone to problems such as insufficient generalization ability, false alarms, or missed alarms.

[0004] Therefore, there is an urgent need for an early warning method for deepwater drilling overflow that can integrate physical constraints and data-driven approaches and has a multi-source redundant confirmation mechanism, in order to solve the problems of slow response, high false alarm rate, lack of interpretability and weak generalization ability in existing technologies. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention aims to provide an early warning method for deepwater drilling overflow based on formation-wellbore energy evolution and multi-physics field fusion. By constructing an early warning system that combines energy imbalance as the core driving criterion, multi-physics field collaborative criteria with temporal causal chain constraints, and dual-mode redundancy fusion of physical rules and intelligent recognition, the method can achieve early, accurate, and interpretable identification of deepwater drilling overflow, reduce false alarm rate, and improve adaptability to different working conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An early warning method for deepwater drilling well blowouts based on formation-wellbore energy evolution and multiphysics field fusion includes the following steps: S1. Acquire multi-source logging data during deepwater drilling, perform time alignment, anomaly removal and standardization processing to form a dynamic observation sequence of the wellbore; S2. Based on the evolution mechanism of multiphase flow in wellbore, a temperature-pressure coupled wellbore flow state characterization model is constructed. Taking into account gas compression and expansion effects, gas-liquid slippage mechanisms, flow pattern transformation characteristics, annular frictional pressure loss variations, and the temperature-pressure coupling effect under deep-water low-temperature environments, a set of governing equations for multiphase flow in wellbore is established. The gas-liquid mixing density is obtained by solving the set of governing equations for multiphase flow in wellbore. Mixing flow velocity ,pressure Wellbore condition parameters; based on this, combined with cross-sectional gas cut. And the gas-liquid phase slip correction relationship, to determine the gas phase velocity. With liquid phase velocity Thus, the gas phase flow rate is obtained. Liquid flow rate cross-sectional gas content Gas-liquid mixture density Equivalent cyclic density (ECD), gas intrusion expansion coefficient The above parameters serve as input features for subsequent energy field analysis, multi-physics field synergistic criteria, and data-driven models. S3. Based on the pressure difference between the formation and the bottom hole, construct the energy transfer per unit time and the intrusion flow rate during the formation fluid invasion into the wellbore. A dual-path estimation method is employed, with automatic switching or weighted fusion based on data validity: S4. Multiphysics Field Collaborative Criterion System: A collaborative criterion system is constructed based on pressure field, flow field, energy field, and multiphase flow field; S5. Construct a temporal causal consistency criterion based on the overflow physical propagation mechanism; S6. Use gating and hierarchical decision-making structures to generate physical rule-based risk assessment results; S7. Construct a Physics-informed Attention-GRU time-series neural network model to extract features and identify overflow status of multi-source time series in wellbore. S8. Risk assessment results based on physical rules Risk assessment results driven by data The confidence level and consistency relationship are used to make fusion decisions.

[0007] The relationship between S1 to S8 is as follows: S1 outputs standardized logging data; S2 uses the data from S1 to solve for the multiphase flow state parameters in the wellbore; S3 uses the multiphase flow parameters and bottom hole pressure data obtained from S2 to construct a formation-wellbore energy transfer model; S4 establishes a multiphysics field co-anomaly degree based on the outputs of S2 and S3; S5 calculates the temporal causality consistency coefficient based on the abnormal event sequence identified by S4; S6 combines the results of S3, S4, and S5 to generate the physical rule-based risk level and confidence level; S7 uses the temporal features formed by S1 to S3 to generate the data-driven risk level and confidence level; S8 integrates the results of S6 and S7 to output the final overflow warning level.

[0008] Step S1 multi-source logging data includes inlet flow rate, outlet flow rate, mud pit volume, riser pressure, casing pressure, drilling fluid density, and outlet drilling fluid density.

[0009] Gas phase flow rate in step S2 Liquid flow rate cross-sectional gas content Gas-liquid mixture density Equivalent cyclic density (ECD), gas intrusion expansion coefficient The formula is: Calculate the gas content of the cross section : Gas-liquid mixture density : Equivalent Cyclic Density (ECD): Invasion expansion coefficient : in, For gas phase flow rate, For liquid phase flow rate, Where A is the gas content of the cross section and A is the annular flow area. The gas phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The liquid phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The density of the gas-liquid mixture. For gas phase density, The density is the liquid phase density, and ECD is the equivalent cycle density. The pressure drop is the annular friction, where H is the well depth or the vertical depth of the calculated well section. The coefficient of gas expansion is denoted as . Let be the gas volume at time t. The gas volume is at a reference time.

[0010] The dual paths in step S3 include: (1) Path 1, flow difference method: based on mass conservation, considering the transmission delay time between the inlet flow and the outlet flow. The difference is calculated after time offset alignment of the outflow flow and then processed by low-pass filtering. (2) Path 2, pressure response inverse method: When flow data is unavailable or unreliable, the intrusion flow is inverted based on the dynamic changes of bottom hole pressure, pressure gradient information and fluid compression characteristics.

[0011] Step S3 specifically involves defining the bottom hole pressure differential: in, For formation pressure, This refers to the bottom hole pressure. Based on the principle of pressure doing work during formation fluid invasion, the instantaneous energy input from the formation to the wellbore is defined as: in, Let be the instantaneous power supplied from the formation to the wellbore at time t. Let be the intrusion flow rate of the formation fluid at time t. This represents the bottom hole pressure differential.

[0012] Considering the influence of reservoir permeability and intrusion fluid properties on energy transfer efficiency, a permeability correction factor is introduced. With fluid property correction factor The effective energy input is obtained as follows: in, Let be the effective power supplied at time t. Γ is the seepage correction factor, used to characterize the influence of reservoir permeability, differential pressure transmission efficiency, and near-wellbore flow resistance on energy input. Γ is the fluid property correction factor, used to characterize the influence of intrusive fluid density, compressibility, and gas-liquid phase differences on energy transmission efficiency.

[0013] Energy dissipation of wellbore system At least including: annular frictional dissipation Gas-liquid slip dissipation and compression power consumption ,Right now: The preferred method for dissipating frictional energy is: in, This represents the energy dissipation per unit time of the wellbore system. The power dissipated by the friction of the annular space. For gas-liquid slip dissipation power, Where is the power dissipation during compression, f is the friction coefficient, and L is the length of the calculated well section. The hydraulic diameter, The density of the gas-liquid mixture. Let be the gas-liquid mixing velocity, and Q be the wellbore mixing volumetric flow rate.

[0014] Define energy loss measurement: in, The energy loss per unit time at time t is the measure of energy loss. Let be the effective power supplied at time t. This represents the energy dissipation per unit time of the wellbore system.

[0015] Among them when When the energy input from the formation to the wellbore is greater than the energy dissipated by the system, the system is in a state of energy accumulation; when When ≤0, the system is in a relatively stable state; Further calculation of energy accumulation: in, This represents the cumulative energy change from time 0 to time t. This is the time variable for integration.

[0016] To characterize the cumulative energy changes in the formation-wellbore system; Simultaneously calculate the rate of energy change: in, Let t be the rate of energy change at time t, used to characterize the rate of abnormal energy release and the energy evolution trend of the system.

[0017] Step S4 specifically involves: First, calculating the anomaly degree of each physical field, including the riser pressure anomaly degree: Traffic anomaly level: Energy anomaly: Gas content anomaly: Weights are dynamically assigned based on the degree of abnormal contribution: Define multiphysics cooperative anomaly degree: in, , , , These are the abnormal deviations in riser pressure, flow rate, energy, and gas content, respectively. For riser pressure, This represents the average riser pressure under normal operating conditions. For riser pressure standard deviation, For export flow, For inbound traffic, The standard deviation of the flow difference The standard deviation of energy loss is the measure of energy loss. The standard deviation of gas content, Let be the dynamic weight of the anomaly deviation of the i-th physical field, and ε be a very small positive number to prevent the denominator from being zero. Let t be the multiphysics field cooperative anomaly degree.

[0018] The greater the degree of co-anomaly, the more likely the overflow will occur, indicating that multiple physics fields will have anomalous responses at the same time. This result serves as an important input for subsequent temporal causal consistency analysis and physical rule-based risk assessment.

[0019] Step S5 specifically involves establishing a standard propagation chain: Wherein, P represents a pressure anomaly, E represents an energy anomaly, Q represents a flow anomaly, α represents a gas content anomaly, and ECD represents an equivalent circulation density anomaly. The temporal causality consistency coefficient is defined by comparing the degree of matching between actual abnormal event sequences and standard propagation chains: If the event occurrence time satisfies it is considered that a physical propagation sequence is satisfied; is the number of preset sequence stages, and N is the total number of stages; When the consistency coefficient exceeds a preset threshold, it is determined that the overflow causal propagation constraint condition is satisfied; When all abnormal events appear strictly in accordance with the overflow propagation sequence, C(t)=1; when only part of the propagation sequence is satisfied, 0<C(t)<1; when the propagation sequence is not satisfied at all, C(t)=0; The causal consistency threshold is set according to the statistical result of historical wellsite data: C(t)<0.4 indicates weak causal correlation, 0.4≤C(t)<0.7 indicates partial causal correlation, and C(t)≥0.7 indicates that the overflow propagation consistency requirement is satisfied.

[0020] Step S6 is specifically: First layer: energy imbalance gating, when and the duration does not exceed the preset time window , it is determined that the system is in a normal or to-be-observed state, and subsequent judgments are not triggered temporarily; When , or but the multi-physics field collaborative anomaly degree index exceeds the preset threshold , the process proceeds to the second layer of judgment; Second layer: multi-physics field collaborative anomaly evaluation, calculate the collaborative anomaly degree index D(t), if , it is determined as normal or weak anomaly; otherwise, proceed to the third layer; Third layer: time-series causal consistency verification: calculate the consistency coefficient , if , it is determined that overflow is in the evolution process; otherwise, it is marked as suspected disturbance, and the risk level is reduced; Fourth layer: risk level determination, define the comprehensive risk index: Wherein: wherein, is the comprehensive risk index, is the normalized energy imbalance amount, is the normalized multi-physics field collaborative anomaly degree, is the time-series causal consistency coefficient, , , are respectively the weights of the energy imbalance amount, the collaborative anomaly degree and the causal consistency coefficient.

[0021] Based on the degree of energy imbalance, the degree of anomaly in multiphysics fields, and causal consistency, a risk level determination rule is constructed; the risk level is divided into: 0 represents normal state, 1 represents Level 1 warning, 2 represents Level 2 warning, and 3 represents Level 3 warning. Step S7 is as follows: (1) Input layer features include: A time series sample is constructed using a 60-second sliding time window with a time step of Δt=1s, and each sample has a dimension of 60×11. (2) GRU layer, number of hidden GRU units in the first layer The value ranges from 64 to 256, and its function is to extract the short-period dynamic variation patterns of pressure, flow rate, and energy characteristics, and output the hidden state. Number of hidden units in the second-layer GRU The value is 32~128, and its function is to extract long-term temporal dependencies and output the hidden state. ; (3) Self-Attention layer: The self-attention mechanism is used to enhance the features of key time segments. The attention calculation formula is as follows: in, , , These are the query matrix, key matrix, and value matrix in the attention mechanism, respectively. Let T be the dimension of the key vector, T denote the matrix transpose, and Softmax be the normalized exponential function.

[0022] The focus is on strengthening the characteristics of ECD mutations, energy anomalies, pressure differential anomalies, and rapid gas content increases. The output characteristics are represented as follows: ; (4) Fully connected layer: 32 neuron nodes are set, and the ReLU function is used as the activation function to complete high-dimensional feature compression and risk mapping; (5) Output layer: Softmax classifier is used. Output risk level : Simultaneously output the model prediction confidence level (t): in, Let be the predicted probability of the i-th risk level. Let i be the unnormalized score of the output layer corresponding to the i-th risk level, and j be the risk level indices. For data-driven risk assessment results, Calculate the confidence level for model prediction.

[0023] (6) Loss function: The total loss function is defined as the sum of the cross-entropy loss and the physical constraint loss: The cross-entropy loss function is: The physical constraint loss function is: Where L is the total loss function, For cross-entropy loss, Let λ represent the physical constraint loss, and λ be the weight of the physical constraint loss. For real labels, For the predicted probability, ΔÊ(t) is the energy loss per unit time estimated by the model.

[0024] (7) Model training: The Adam optimizer was used for training with a learning rate of 10. - ³, Batch Size is 64, Epoch is 200; The training data is divided into three sets: 70% training set, 15% validation set, and 15% test set; the final output data drives the risk level. and corresponding confidence level (t).

[0025] Step S8 specifically involves defining the confidence level of the physical model. for and The weighted average; in, For the confidence level of physical rules, To normalize the weights of collaborative anomalies, The weight of the temporal causal consistency coefficient.

[0026] The fusion rules are as follows: when = When the risk level is reached, output the risk level directly. When | - When | = 1, confidence-weighted fusion is used: in, For the final risk level, For data-driven side confidence, For the confidence level of physical rules, To drive risk levels with output data; The risk level is determined by the physical rules.

[0027] When | - When | ≥2, the conflict suppression mechanism is activated, and the safety priority principle is adopted: Where max(·) represents taking the higher risk level, To drive risk levels with output data; The risk level is determined by the physical rules.

[0028] If the weighted result is lower than the maximum of the two values, it is rounded up.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention uses energy imbalance as the core driving criterion, which can detect abnormal energy injection in the formation during the micro-overflow stage (when the pool volume and flow difference have not changed significantly), thus achieving true early warning.

[0030] (2) The multi-physics field collaborative criterion and dynamic weighting mechanism of the present invention enable the system to adaptively focus on the currently dominant abnormal physical field, thereby improving the recognition sensitivity.

[0031] (3) The timing causal consistency criterion of the present invention strictly follows the physical propagation sequence of the overflow, effectively suppressing false alarms caused by sensor noise or random fluctuations.

[0032] (4) The dual-mode redundancy fusion architecture of physical rules and data-driven in this invention greatly improves the robustness of the system and avoids false alarms caused by the failure of a single module.

[0033] (5) The consistency fusion decision-making mechanism of this invention can suppress conflicts when there are conflicts between the outputs of the physical rule side and the data-driven side, thereby improving the reliability under complex working conditions. At the same time, it can output conflict feature identifiers, which facilitates the rapid location of the source of anomalies by manual intervention. Attached Figure Description

[0034] Figure 1 This is a flowchart of an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the coupling between the wellbore multiphase flow model and energy evolution in an embodiment of the present invention.

[0036] Figure 3This is a schematic diagram of the multi-physics field collaborative criterion and adaptive weighting mechanism in an embodiment of the present invention.

[0037] Figure 4 This is a diagram of the consistency fusion decision architecture according to an embodiment of the present invention. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.

[0039] Reference Figure 1 In this embodiment, S1 to S8 are executed sequentially to form a complete early warning chain of "data preprocessing - wellbore multiphase flow state characterization - energy transfer and imbalance model - multi-physics field collaborative criterion - temporal causal consistency - physical rule-based risk assessment - data-driven risk assessment - dual-mode redundancy fusion decision".

[0040] An early warning method for deepwater drilling well blowouts based on formation-wellbore energy evolution and multiphysics field fusion includes the following steps: S1. Data Acquisition and Preprocessing: Acquire multi-source logging data during deepwater drilling, including inlet flow rate, outlet flow rate, mud pit volume, riser pressure, casing pressure, drilling fluid density, outlet drilling fluid density, etc., and perform time alignment, anomaly removal and standardization processing to form a dynamic wellbore observation sequence. S2. Characterization of multiphase flow state in wellbore: Refer to Figure 2 The formation fluid source area at the bottom of the wellbore inputs fluid and pressure energy into the wellbore. During the upward return process, the gas expands and changes the gas content, mixing density and ECD. Figure 2 The dual-path flow estimation corresponds to step S3, which is used to determine the formation fluid intrusion flow rate. . Figure 2 The term "energy input" refers to the instantaneous power supplied from the formation to the wellbore. After considering reservoir permeability and fluid properties, the effective power supply was obtained. ; Figure 2 The term "energy dissipation" in this context corresponds to the wellbore energy dissipation in this paper. The difference between the two constitutes the energy loss measurement. The above results are further passed to steps S4 to S6 for multi-physics collaborative anomaly analysis, temporal causal consistency determination, and physical rule-based risk assessment.

[0041] Based on the evolution mechanism of multiphase flow in wellbore, a temperature-pressure coupled wellbore flow state characterization model is constructed. Taking into account the gas compression and expansion effect, gas-liquid slip mechanism, flow pattern conversion characteristics, annular friction pressure loss change and temperature-pressure coupling effect in deep water low temperature environment, a set of wellbore multiphase flow control equations is established based on the mass conservation equation, momentum conservation equation and energy conservation equation. The mass conservation equation is: in, This refers to the density of the gas-liquid mixture. The area of ​​the circular flow; The mixing flow velocity; The coordinates are the well depth. This is the mass source term for formation fluid intrusion.

[0042] The momentum conservation equation is: in, For pressure; It is the acceleration due to gravity; The coefficient of friction; This refers to the density of the gas-liquid mixture. The mixing flow velocity; It is the hydraulic diameter.

[0043] The energy conservation equation is: in, Enthalpy; For heat transfer; For external heat source items; This refers to the density of the gas-liquid mixture. The mixing flow velocity; The gas-liquid mixture density was obtained by solving the multiphase flow control equations in the wellbore. Mixing flow velocity ,pressure Wellbore condition parameters; based on this, combined with cross-sectional gas cut. And the gas-liquid phase slip correction relationship, to determine the gas phase velocity. With liquid phase velocity When gas-liquid slip is ignored or an equivalent homogeneous flow approximation is used, we can take... When considering gas-liquid slip, and Due to the mixing flow velocity cross-sectional gas content The relationship between gas-liquid slip closure is used to determine this.

[0044] Thus, the gas phase flow rate is obtained. With liquid phase flow rate : in, The gas content of the cross section, For the circular flow area, The gas phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The liquid phase velocity is determined by the gas-liquid slip relationship or the equivalent homogeneous flow assumption. and It is not obtained directly and independently from the above-mentioned mixed flow control equations, but is further determined based on the mixed flow state parameters, combined with the cross-sectional gas holdup and interphase slip relationship.

[0045] Calculate the gas content of the cross section: Gas-liquid mixture density: Equivalent Cyclic Density (ECD): Invasion expansion coefficient: in, For gas phase flow rate, For liquid phase flow rate, Where A is the gas content of the cross section and A is the annular flow area. The gas phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The liquid phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The density of the gas-liquid mixture. For gas phase density, The density is the liquid phase density, and ECD is the equivalent cycle density. The pressure drop is the annular friction, where H is the well depth or the vertical depth of the calculated well section. The coefficient of gas expansion is denoted as . Let be the gas volume at time t. The gas volume is at a reference time.

[0046] The above parameters serve as input features for subsequent energy field analysis, multi-physics field synergistic criteria, and data-driven models; S3. Formation-Wellbore Energy Transfer Characterization: Based on the pressure difference between the formation and the bottom hole, the energy transfer per unit time and the intrusion flow rate during the formation fluid invasion into the wellbore are constructed. A dual-path estimation method is employed, with automatic switching or weighted fusion based on data validity: (1) Path 1 (Flow Difference Method): Based on the conservation of mass, it considers the transmission delay time between the inlet flow and the outlet flow. The difference is calculated after time offset alignment of the outflow flow and then processed by low-pass filtering. (2) Path 2 (Pressure response back-calculation method): When the flow rate data is unavailable or unreliable, the intrusion flow rate is inverted based on the dynamic changes of bottom hole pressure, pressure gradient information and fluid compression characteristics (specific implementation methods include but are not limited to pressure derivative method, material balance method, etc., which can be selected by those skilled in the art according to the site conditions). Define bottom hole pressure differential: in, For formation pressure, This refers to the bottom hole pressure. Based on the principle of pressure doing work during formation fluid invasion, the instantaneous energy input from the formation to the wellbore is defined as: in, Let be the instantaneous power supplied from the formation to the wellbore at time t. Let be the intrusion flow rate of the formation fluid at time t. This refers to the bottom pressure differential.

[0047] Considering the influence of reservoir permeability and intrusion fluid properties on energy transfer efficiency, a permeability correction factor is introduced. With fluid property correction factor The effective energy input is obtained as follows: in, Let be the effective power supplied at time t. Γ is the seepage correction factor, used to characterize the influence of reservoir permeability, differential pressure transmission efficiency, and near-wellbore flow resistance on energy input. Γ is the fluid property correction factor, used to characterize the influence of intrusive fluid density, compressibility, and gas-liquid phase differences on energy transmission efficiency.

[0048] Energy dissipation of wellbore system At least including: annular frictional dissipation Gas-liquid slip dissipation and compression power consumption ,Right now: The preferred method for dissipating frictional energy is: in, This represents the energy dissipation per unit time of the wellbore system. The power dissipated by the annular friction is For gas-liquid slip dissipation power, Where is the power dissipation during compression, f is the friction coefficient, and L is the length of the calculated well section. The hydraulic diameter, The density of the gas-liquid mixture. Let be the gas-liquid mixing velocity, and Q be the wellbore mixing volumetric flow rate.

[0049] Other specific calculation methods can be obtained using empirical formulas, mechanistic models, or numerical solutions.

[0050] Define energy loss measurement: in, The energy loss per unit time at time t is the measure of energy loss. Let be the effective power supplied at time t. This represents the energy dissipation per unit time of the wellbore system.

[0051] Among them when When the energy input from the formation to the wellbore is greater than the energy dissipated by the system, the system is in a state of energy accumulation; when When ≤0, the system is in a relatively stable state; Further calculation of energy accumulation: in, This represents the cumulative energy change from time 0 to time t. This is the time variable for integration.

[0052] To characterize the cumulative energy changes in the formation-wellbore system; Simultaneously calculate the rate of energy change: in, Let t be the rate of energy change at time t, used to characterize the rate of abnormal energy release and the energy evolution trend of the system.

[0053] S4. Multiphysics Collaborative Criterion System: (Refer to...) Figure 3 The pressure field, flow field, energy field, and multiphase flow field each contribute to the abnormal deviation. Dynamic weights Based on the adaptive adjustment of the anomaly contribution, the final output is the multi-physics collaborative anomaly degree. ;Should The temporal causality consistency coefficient with step S5 A joint risk assessment based on physical rules is conducted. A collaborative criterion system is constructed based on pressure field, flow field, energy field, and multiphase flow field. First, calculate the anomaly degree of each physical field, including the riser pressure anomaly degree: Traffic anomaly level: Energy anomaly: Gas content anomaly: Weights are dynamically assigned based on the degree of abnormal contribution: Define multiphysics cooperative anomaly degree: in, , , , These are the abnormal deviations in riser pressure, flow rate, energy, and gas content, respectively. For riser pressure, This represents the average riser pressure under normal operating conditions. For riser pressure standard deviation, For export flow, For inbound traffic, The standard deviation of the flow difference The standard deviation of energy loss is the measure of energy loss. The standard deviation of gas content, Let be the dynamic weight of the anomaly deviation of the i-th physical field, and ε be a very small positive number to prevent the denominator from being zero. Let t be the multiphysics field cooperative anomaly degree.

[0054] The greater the degree of co-anomaly, the more anomalous responses occur in multiple physics fields simultaneously, and the higher the probability of overflow. This result serves as an important input for subsequent temporal causal consistency analysis and physical rule-based risk assessment. S5. Temporal Causality Consistency Criterion: Based on the overflow physical propagation mechanism, a temporal causal consistency criterion is constructed. When formation fluid invades the wellbore, abnormal information usually follows certain physical propagation laws. The typical propagation sequence is: formation pressure imbalance → enhanced energy supply from the formation to the wellbore → abnormal energy at the bottom of the well → abnormal flow response → gas intrusion and diffusion → increased gas content → change in ECD → abnormal manifestation at the wellhead. A standard propagation chain is established based on the above mechanism: Wherein, P represents a pressure anomaly, E represents an energy anomaly, Q represents a flow anomaly, α represents a gas content anomaly, and ECD represents an equivalent circulation density anomaly. The temporal causality consistency coefficient is defined by comparing the degree of matching between actual abnormal event sequences and standard propagation chains: If the event The occurrence time satisfies it is considered that one physical propagation order is satisfied; where s is the number of preset order stages, and N is the total number of stages; when the consistency coefficient exceeds a preset threshold, it is determined that the overflow causal propagation constraint condition is satisfied; when all abnormal events appear strictly in the order of overflow propagation, C(t)=1; when only part of the propagation order is satisfied, 0<C(t)<1; when the propagation order is not satisfied at all, C(t)=0; the larger the consistency coefficient, the more the observed anomalies conform to the real overflow propagation mechanism, and the higher the credibility of overflow occurrence; the causal consistency threshold is set according to the statistical result of historical well site data: C(t)<0.4 indicates weak causal correlation, 0.4≤C(t)<0.7 indicates partial causal correlation, and C(t)≥0.7 indicates that the overflow propagation consistency requirement is satisfied; S6. Physical rule side risk assessment (gating + hierarchical decision structure): A gating and hierarchical decision structure is adopted to generate a physical rule side risk assessment result, which specifically includes: First layer: Energy imbalance gating: When and the duration does not exceed the preset time window , it is determined that the system is in a normal or to-be-observed state, and subsequent judgments are not triggered temporarily; When , or but the multi-physics field collaborative anomaly index exceeds the preset threshold , the process proceeds to the second layer of judgment; Second layer: Multi-physics field collaborative anomaly assessment: Calculate the collaborative anomaly index D(t), if , it is determined as normal or slight anomaly; otherwise, the process proceeds to the third layer; Third layer: Time-series causal consistency verification: Calculate the consistency coefficient , if , it is determined that overflow is in the evolution process; otherwise, it is marked as suspected disturbance, and the risk level is reduced; Fourth layer: Risk level determination: Define the comprehensive risk index: Where: wherein, is the comprehensive risk index, is the normalized energy imbalance, is the normalized multi-physics field collaborative anomaly, is the time-series causal consistency coefficient, , , These are the weights for energy loss measurement, cooperative anomaly degree, and causal consistency coefficient, respectively.

[0055] Based on the degree of energy imbalance, the degree of anomaly in multiple physics fields, and causal consistency, risk level determination rules are constructed; risk levels are divided as follows: 0 represents normal state, 1 represents Level 1 warning, 2 represents Level 2 warning, and 3 represents Level 3 warning; or, depending on specific working conditions, it can be divided into more or fewer levels (e.g., Level 5 or Level 6). S7. Data-driven risk assessment: A Physics-informed Attention-GRU temporal neural network model was constructed to extract features and identify overflow status of multi-source time series in wellbore. (1) Input layer features include A time series sample is constructed using a 60-second sliding time window with a time step of Δt=1s, and each sample has a dimension of 60×11. (2) GRU layer, number of hidden GRU units in the first layer The value ranges from 64 to 256, and its function is to extract the short-period dynamic variation patterns of pressure, flow rate, and energy characteristics, and output the hidden state. Number of hidden units in the second-layer GRU The value is 32~128, and its function is to extract long-term temporal dependencies and output the hidden state. ; (3) Self-Attention layer: The self-attention mechanism is used to enhance the features of key time segments. The attention calculation formula is as follows: in, , , These are the query matrix, key matrix, and value matrix in the attention mechanism, respectively. Let T be the dimension of the key vector, T denote the matrix transpose, and Softmax be the normalized exponential function.

[0056] The focus is on strengthening the characteristics of ECD mutations, energy anomalies, pressure differential anomalies, and rapid gas content increases. The output characteristics are represented as follows: ; (4) Fully connected layer: 32 neuron nodes are set, and the ReLU function is used as the activation function to complete high-dimensional feature compression and risk mapping; (5) Output layer: Softmax classifier is used. Output risk level : Simultaneously output the model prediction confidence level (t): in, Let be the predicted probability of the i-th risk level. Let i be the unnormalized score of the output layer corresponding to the i-th risk level, and j be the risk level indices. For data-driven risk assessment results, Calculate the confidence level for model prediction.

[0057] (6) Loss function: The total loss function is defined as the sum of the cross-entropy loss and the physical constraint loss: The cross-entropy loss function is: The physical constraint loss function is: Where L is the total loss function, For cross-entropy loss, Let λ be the physical constraint loss, and λ be the weight of the physical constraint loss. For real labels, For the predicted probability, ΔÊ(t) is the energy loss per unit time estimated by the model.

[0058] By introducing energy conservation constraints, the physical interpretability and generalization ability of the model are improved; (7) Model training: The Adam optimizer was used for training with a learning rate of 10. - ³, Batch Size is 64, Epoch is 200; The training data is divided into three sets: 70% training set, 15% validation set, and 15% test set; the final output data drives the risk level. and corresponding confidence level (t); S8, Dual-Mode Redundancy Fusion Decision: Refer to Figure 4 Risk assessment results based on physical rules Data-driven risk assessment results A fusion decision is made; if the two are consistent, the output is given directly; if they differ by one level, a confidence-weighted fusion is used; if they differ by two levels or more, a conflict suppression mechanism is activated and a safety-first principle is adopted, thereby achieving complementary verification and conflict suppression between the physical rule side and the data-driven side. Define the confidence level of the physical model. for and The weighted average; in, For the confidence level of physical rules, To normalize the weights of collaborative anomalies, The weight of the temporal causal consistency coefficient.

[0059] The fusion rules are as follows: when = When the risk level is reached, output the risk level directly. When | - When | = 1, confidence-weighted fusion is used: in, For the final risk level, For data-driven side confidence, For the confidence level of physical rules, To drive risk levels with output data; The risk level is determined by the physical rules, and the weighted result is mapped to an integer risk level using rounding or a safe rounding method.

[0060] When | - When | ≥2, the conflict suppression mechanism is activated, and the safety priority principle is adopted: Where max(·) represents taking the higher risk level, To drive risk levels with output data; The risk level is determined by the physical rules.

[0061] If the weighted result is lower than the maximum of the two values, it is rounded up. S1 outputs standardized logging data; S2 uses the S1 data to solve the wellbore multiphase flow control equations to obtain gas cut, mixing density, ECD change rate, and gas intrusion expansion coefficient; S3 uses the S2 results and bottom hole pressure data to construct a formation-wellbore energy transfer model to obtain energy loss, cumulative energy, and energy change rate; S4 establishes a multiphysics field collaborative anomaly degree based on the outputs of S2 and S3; S5 calculates the temporal causal consistency coefficient based on the abnormal event sequence identified in S4; S6 combines the results of S4 and S5 to generate the physical rule-based risk level and confidence level; S7 uses the temporal features formed from S1 to S3 to generate a data-driven risk level and confidence level; S8 integrates the results of S6 and S7 to output the final overflow warning result, thus forming a complete early warning chain of "formation energy release → wellbore multiphase flow evolution → multiphysics field anomaly response → causal consistency verification → intelligent identification → dual-mode fusion decision-making".

[0062] Example 1: Taking a deep water well as an example, oil-based drilling fluid is used. Real-time logging data is collected: inlet flow rate. (L / s), Outbound Flow (L / s), mud pit volume (m³), riser pressure (MPa), casing pressure (MPa), drilling fluid density (g / cm³), outlet drilling fluid density (g / cm³), etc., with a sampling interval of 1 second.

[0063] Step 1: Data preprocessing: Remove outliers, linearly imput missing values, use Savitzky-Golay filter for noise reduction and smoothing, and perform Min-Max normalization; Step 2: Multiphase flow model calculation: Using wellhead pressure, temperature, and flow rate as boundary conditions, the wellbore is discretized using the finite difference method, and the gas-liquid two-phase flow control equations are solved to obtain the gas cutoff of the cross section. ECD change rate , gas intrusion expansion coefficient wait; Step 3: Energy Model Calculation: Estimate Intrusion Flow Rate Based on Pool Volume Change Rate Bottom hole pressure differential ; Calculate instantaneous power supply For gas intrusion, a seepage correction factor is introduced. Gas risk factors (Specific values ​​are calibrated based on well conditions) to obtain the effective power supply. Accumulated and rate of change of energy ; Step 4: Multi-physics collaborative criterion: Set characteristic thresholds for each field (dynamically calculated based on historical normal operating condition statistics). Calculate the anomaly deviation degree. The dynamic weight formula is: in Determined dynamically based on historical data.

[0064] Further calculation of multiphysics cooperative anomaly degree Simultaneously, abnormal response sequences are detected within a fixed time window, and their consistency with theoretical sequences [pressure, energy, flow, multiphase flow] is calculated, along with the consistency coefficient. Based on whether the collaborative function value and consistency coefficient exceed a preset threshold, a collaborative early warning or collaborative alarm will be output. Step 5: Temporal Intelligent Recognition Model: Construct a two-layer GRU network (preferred, but not limited), input multi-dimensional features, and output overflow probability and risk level after passing through a self-attention layer. At the same time, output the confidence level. Adaptive weights are generated before the input layer based on the degree of abnormal deviation. The input features are then weighted after being smoothed by an exponential moving average. Step 6: Fusion Decision: Assume that there is a first-order difference between the model output and the collaborative criterion output at a certain moment. Use confidence-weighted fusion to obtain the final risk level and display the current high-weight physical fields (such as energy fields and multiphase flow fields).

[0065] Sensor noise can cause a difference of two or more levels between the model output and the physical rule output (e.g., severe overflow in the model output while the collaborative criterion output is normal). A conflict suppression mechanism is triggered: the confidence and consistency scores of the two modules are compared. If one is significantly higher than the other, the result from the higher confidence level is taken; if the confidence levels are close, a review mechanism is triggered, outputting a conflict flag (e.g., "Data model conflicts with physical rules, sensor check recommended"), without triggering an alarm or handling it as a lower-risk scenario, effectively avoiding false alarms.

[0066] The calibration data for this embodiment is shown in Table 1.

[0067] Table 1 As shown in Table 1, in the early stages of minor intrusion, the flow difference has not yet reached a significant alarm level, but , and A coordinated increase has already occurred, and this method can output a Level 1 warning; as the gas content increases and the ECD decreases, the risk level gradually increases. This result is consistent with... Figure 2 The energy input-multiphase flow evolution process is shown. Figure 3 The multiphysics field co-operation criterion shown and Figure 4The fusion decision logic shown is consistent, which reflects the advantages of the method of the present invention in early identification, collaborative confirmation and conflict suppression.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method for early warning of deepwater drilling well blowouts based on formation-wellbore energy evolution and multiphysics field fusion, characterized in that, comprising the following steps: S1. Acquire multi-source logging data during deepwater drilling, and perform time alignment, anomaly elimination and standardization processing to form a wellbore dynamic observation sequence; S2. Based on the evolution mechanism of multiphase flow in wellbore, a temperature-pressure coupled wellbore flow state characterization model is constructed. Taking into account gas compression and expansion effects, gas-liquid slippage mechanisms, flow pattern transformation characteristics, annular frictional pressure loss variations, and the temperature-pressure coupling effect under deep-water low-temperature environments, a set of governing equations for multiphase flow in wellbore is established. The gas-liquid mixing density is obtained by solving the set of governing equations for multiphase flow in wellbore. Mixing flow velocity ,pressure Wellbore condition parameters; based on this, combined with cross-sectional gas cut. And the gas-liquid phase slip correction relationship, to determine the gas phase velocity. With liquid phase velocity ; thereby obtaining gas phase flow rate Liquid flow rate cross-sectional gas content Gas-liquid mixture density Equivalent cyclic density (ECD), gas intrusion expansion coefficient The above parameters serve as input features for subsequent energy field analysis, multi-physics field synergistic criteria, and data-driven models. S3. Based on the pressure difference between the formation and the bottom hole, construct the energy transfer per unit time and the intrusion flow rate during the formation fluid invasion into the wellbore. A dual-path estimation method is employed, with automatic switching or weighted fusion based on data validity: S4. Construct a collaborative criterion system based on the pressure field, flow field, energy field and multiphase flow field; S5. Construct a time-series causal consistency criterion based on the physical propagation mechanism of overflow; S6. Adopt a gating and hierarchical decision structure to generate a physical rule side risk assessment result; S7. Construct a Physics-informed Attention-GRU time-series neural network model to perform feature extraction and overflow state identification on the multi-source time series of the wellbore; S8. Risk assessment results based on physical rules Data-driven risk assessment results The confidence level and consistency relationship are used to make fusion decisions.

2. The method according to claim 1, characterized in that, The multi-source logging data in step S1 includes inlet flow rate, outlet flow rate, mud pit volume, standpipe pressure, casing pressure, drilling fluid density and outlet drilling fluid density.

3. The method according to claim 1, characterized in that, Gas phase flow rate in step S2 Liquid flow rate cross-sectional gas content Gas-liquid mixture density Equivalent cyclic density (ECD), gas intrusion expansion coefficient The formula is: Calculate the gas content of the cross section : Gas-liquid mixture density : Equivalent Circulating Density ECD: Invasion expansion coefficient : in, For gas phase flow rate, For liquid phase flow rate, Where A is the gas content of the cross section and A is the annular flow area. The gas phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The liquid phase velocity is determined by the gas-liquid slip relation or the assumption of equivalent homogeneous flow. The density of the gas-liquid mixture. For gas phase density, The density is the liquid phase density, and ECD is the equivalent cycle density. The pressure drop is the annular friction, where H is the well depth or the vertical depth of the calculated well section. The coefficient of gas expansion is denoted as . Let be the gas volume at time t. The gas volume is at a reference time.

4. The method according to claim 1, characterized in that, The dual paths in step S3 include: (1) Path 1, flow difference method: based on mass conservation, considering the transmission delay time between the inlet flow and the outlet flow. The difference is calculated after time offset alignment of the outflow flow and then processed by low-pass filtering. (2) The second path, pressure response inverse deduction method: when the flow data is unavailable or unreliable, invert the intrusion flow according to the dynamic change of bottomhole pressure, pressure gradient information and fluid compression characteristics.

5. The method according to claim 1, characterized in that, Step S3 is specifically: defining the bottomhole pressure difference: in, For formation pressure, This refers to the bottom hole pressure. According to the principle of pressure work during formation fluid intrusion, the instantaneous energy supply power input from the formation to the wellbore is defined as: in, Let be the instantaneous power supplied from the formation to the wellbore at time t. Let be the intrusion flow rate of the formation fluid at time t. This refers to the bottom hole pressure differential. Considering the influence of reservoir permeability and intrusion fluid properties on energy transfer efficiency, a permeability correction factor is introduced. With fluid property correction factor The effective energy input is obtained as follows: in, Let be the effective power supplied at time t. Γ is the seepage correction factor, used to characterize the influence of reservoir permeability, differential pressure transmission efficiency and near-wellbore flow resistance on energy input; Γ is the fluid property correction factor, used to characterize the influence of intrusive fluid density, compressibility and gas-liquid phase differences on energy transmission efficiency. Energy dissipation of wellbore system At least including: annular frictional dissipation Gas-liquid slip dissipation and compression power consumption ,Right now: Wherein, the frictional dissipation energy is preferably adopted as: in, This represents the energy dissipation per unit time of the wellbore system. The power dissipated by the annular friction is For gas-liquid slip dissipation power, Where is the power dissipation during compression, f is the friction coefficient, and L is the length of the calculated well section. The hydraulic diameter, The density of the gas-liquid mixture. Where is the gas-liquid mixing velocity, and Q is the wellbore mixing volume flow rate; Defining the energy imbalance amount: in, The energy loss per unit time at time t is the measure of energy loss. Let be the effective power supplied at time t. This refers to the energy dissipation per unit time of the wellbore system. Among them when When the energy input from the formation to the wellbore is greater than the energy dissipated by the system, the system is in a state of energy accumulation; when When ≤0, the system is in a relatively stable state; Further calculating the energy accumulation amount: in, This represents the cumulative energy change from time 0 to time t. For integration time; to characterize the cumulative energy change of the formation-wellbore system; Simultaneously calculating the energy change rate: in, Let t be the rate of energy change at time t, used to characterize the rate of abnormal energy release and the energy evolution trend of the system.

6. The method according to claim 1, characterized in that, Step S4 is specifically: firstly calculating the anomaly degree of each physical field, wherein the standpipe pressure anomaly degree: Flow anomaly degree: Energy anomaly degree: Gas holdup anomaly degree: Dynamically assigning weights according to the anomaly contribution degree: Defining the multi-physics field collaborative anomaly degree: in, , , , These are the abnormal deviations in riser pressure, flow rate, energy, and gas content, respectively. For riser pressure, This represents the average riser pressure under normal operating conditions. For riser pressure standard deviation, For export flow, For inbound traffic, The standard deviation of the flow difference The standard deviation of energy loss is the measure of energy loss. The standard deviation of gas content, Let be the dynamic weight of the anomaly deviation of the i-th physical field, and ε be a very small positive number to prevent the denominator from being zero. The multi-physics field cooperative anomaly degree at time t; A larger collaborative anomaly degree indicates that abnormal responses occur simultaneously in multiple physical fields, and the probability of overflow occurrence is higher. This result serves as an important input for subsequent time-series causal consistency analysis and physical rule side risk assessment.

7. The method according to claim 1, characterized in that, Step S5 is specifically, establishing a standard propagation chain: Wherein, P represents a pressure anomaly event, E represents an energy anomaly event, Q represents a flow anomaly event, α represents a gas holdup anomaly event, and ECD represents an equivalent circulating density anomaly event; By comparing the matching degree between the actual anomaly event sequence and the standard propagation chain, the time-series causal consistency coefficient is defined: If the event The occurrence time satisfies Then it is considered to satisfy the physical propagation sequence; To satisfy the preset number of sequential stages, N is the total number of stages; When the consistency coefficient exceeds a preset threshold, it is determined that the overflow causal propagation constraint condition is satisfied; When all anomaly events appear strictly in the order of overflow propagation, C(t)=1; when only part of the propagation order is satisfied, 0<C(t)<1; when the propagation order is not satisfied at all, C(t)=0; Set the causal consistency threshold according to the statistical results of historical well site data: C(t)<0.4 indicates weak causal correlation, 0.4≤C(t)<0.7 indicates partial causal correlation, and C(t)≥0.7 indicates that the overflow propagation consistency requirement is satisfied.

8. The method according to claim 1, characterized in that, Step S6 is specifically: First layer: Energy imbalance gating, when And the duration did not exceed the preset time window. When the system is in a normal or unobservable state, no further judgment is triggered. when ,or However, the multiphysics field cooperative anomaly index Exceeding the preset threshold Then, proceed to the second level of judgment; The second layer: Multiphysics collaborative anomaly evaluation, calculating the collaborative anomaly index D(t), if If so, it is judged as normal or slightly abnormal; otherwise, proceed to the third layer. Third layer: Temporal causality consistency check: Calculate the consistency coefficient ,like If so, it is determined to be in the process of overflow evolution; Otherwise, it is marked as a suspected disturbance, and the risk level is reduced; Fourth layer: risk level determination, defining a comprehensive risk index: Wherein: in, As a comprehensive risk index, To normalize energy loss measurement, To normalize the multiphysics cooperative anomaly degree, The coefficient of temporal causality consistency. , , These are the weights of energy loss measurement, coordination anomaly degree, and causal consistency coefficient, respectively. Based on the degree of energy imbalance, the degree of anomaly in multiphysics fields, and causal consistency, a risk level determination rule is constructed; the risk level is divided into: 0 represents normal state, 1 represents Level 1 warning, 2 represents Level 2 warning, and 3 represents Level 3 warning. .

9. The method according to claim 1, characterized in that, Step S7 is specifically: (1) The input layer features include: A 60-second sliding time window is used to construct time-series samples, with a time step Δt=1s, and the dimension of each sample is 60×11; (2) GRU layer, number of hidden GRU units in the first layer The value ranges from 64 to 256, and its function is to extract the short-period dynamic variation patterns of pressure, flow rate, and energy characteristics, and output the hidden state. Number of hidden units in the second-layer GRU The value is 32~128, and its function is to extract long-term temporal dependencies and output hidden states. ; (3) Self-Attention layer: The self-attention mechanism is used to enhance the features of key time segments. The attention calculation formula is as follows: in, , , These are the query matrix, key matrix, and value matrix in the attention mechanism, respectively. The dimension of the key vector is T, where T represents the matrix transpose, and Softmax is the normalized exponential function. The focus is on strengthening the characteristics of ECD mutations, energy anomalies, pressure differential anomalies, and rapid gas content increases. The output characteristics are represented as follows: ; (4) Fully connected layer: 32 neuron nodes are set, and the ReLU function is used as the activation function to complete high-dimensional feature compression and risk mapping; (5) Output layer: Softmax classifier is used. Output risk level : Simultaneously output the model prediction confidence level (t): in, Let be the predicted probability of the i-th risk level. Let i be the unnormalized score of the output layer corresponding to the i-th risk level, and j be the risk level indices. For data-driven risk assessment results, Predict confidence levels for the model; (6) Loss function: The total loss function is defined as the sum of the cross-entropy loss and the physical constraint loss: The cross-entropy loss function is: The physical constraint loss function is: Where L is the total loss function, For cross-entropy loss, Let λ represent the physical constraint loss, and λ be the weight of the physical constraint loss. For real labels, For the predicted probability, ΔÊ(t) is the energy loss per unit time estimated by the model; (7) Model training: The Adam optimizer was used for training with a learning rate of 10. - ³, Batch Size is 64, Epoch is 200; The training data is divided into three sets: 70% training set, 15% validation set, and 15% test set; the final output data drives the risk level. and corresponding confidence level (t).

10. The method according to claim 1, characterized in that, Step S8 specifically involves defining the confidence level of the physical model. for and The weighted average; in, For the confidence level of physical rules, To normalize the weights of collaborative anomalies, The weight of the temporal causal consistency coefficient; The fusion rules are as follows: when = When the risk level is reached, output the risk level directly. When | - When | = 1, confidence-weighted fusion is used: in, For the final risk level, For data-driven side confidence, For the confidence level of physical rules, To drive risk levels based on output data; For the risk level on the physical rule side, the weighted result is mapped to an integer risk level using rounding or safe rounding. When | - When | ≥2, the conflict suppression mechanism is activated, and the safety priority principle is adopted: Where max(·) represents taking the higher risk level, To drive risk levels based on output data; Risk level based on physical rules; If the weighted result is lower than the maximum of the two values, it is rounded up.

Images