Moving ring multi-modal perception and digital twin AI early warning linkage method and system

Abstract

The present application relates to the technical field of digital twin modeling and multi-physical field simulation, and discloses a dynamic environment multi-modal perception and digital twin AI early warning linkage method and system, which comprises the following steps: constructing a state matrix by completing spatial interpolation on dynamic environment perception data and extracting a main mode coefficient vector; completing online update of a projection basis vector by using Givens rotation transformation when power jumps; solving a three-dimensional twin prediction vector by using a thermal load input physical constraint deduction operator on the main mode coefficient vector; constructing a drift energy functional by solving a measured and predicted residual and locating an abnormal source coordinate; solving a control increment vector according to a target functional sensitivity matrix, and generating a control message after superimposing differential lead compensation. The present application takes fluid continuity and energy conservation as a punishment boundary, suppresses state divergence of the model when the load suddenly changes to complete high-fidelity deduction, and actively offsets pipeline transmission lag to eliminate temperature field oscillation.

Description

Technical Field

[0001] This invention belongs to the field of digital twin modeling and multiphysics simulation technology, and particularly relates to a method and system for linkage between dynamic environment multimodal perception and digital twin AI early warning. Background Technology

[0002] In the monitoring of the dynamic environment in densely populated places such as data centers and communication equipment rooms, building a digital twin system that integrates multi-physics field simulation is the main way to extrapolate the state of continuous temperature field and flow field. By collecting multi-modal sensing data such as temperature, flow velocity and bus electrical parameters, and inputting these data into the extrapolation model, the limitation of discrete sensing points that can only reflect local state can be overcome, and the state distribution of the dynamic environment in continuous space can be obtained.

[0003] Beyond the hardware-level sensor point layout and physical state acquisition, existing control methods and twin-based inference algorithms also have limitations. For example, Chinese invention patent application CN119758764A discloses a data center environment management method, system, and equipment based on digital twin technology. It predicts heating and cooling capacity through regression algorithms and combines computational fluid dynamics to simulate heat load distribution, and then dynamically adjusts the environmental control strategy based on the PUE index. Although this scheme attempts to combine computational fluid dynamics for heat load prediction, it still essentially relies on traditional steady-state fluid numerical calculations and static regression mapping. When facing dynamic and unsteady scenarios such as transient changes in electrical load in the computer room, the calculation of high-order partial differential equations is computationally intensive and time-consuming, making it difficult to meet the timeliness requirements of online simulation. Moreover, its control logic only conducts steady-state optimization based on the overall PUE statistical index, without applying physical constraints of continuity and energy conservation at the surface flow level. It also lacks a dynamic compensation mechanism for the mismatch of the orthogonal projection substrate and the inherent heat transfer lag of the terminal valves and fans. When facing local flow channel obstruction or sudden load changes, it is easy to cause state drift, control overshoot and temperature field oscillation, making it difficult to achieve advanced positioning and high-precision closed-loop coordination of weak precursor anomalies.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a digital twin model that satisfies physical conservation constraints and has millisecond-level extrapolation capability when multimodal operating conditions change drastically, and how to eliminate the influence of end mechanism response lag on reverse closed-loop control while accurately extracting precursor abnormal features. Summary of the Invention

[0005] This invention aims to solve the problems in existing dynamic environment monitoring, such as the excessive time consumption of full-order physical simulation calculations, which cannot meet the requirements of real-time extrapolation; the tendency of black-box proxy models to drift under complex working conditions, violating the laws of physical conservation; and the control overshoot oscillation induced by basis mismatch and mechanism response lag during model inversion and control.

[0006] In this technical solution, a method for linking dynamic environmental multimodal perception with digital twin AI early warning includes the following steps: Step S1: Spatial interpolation is performed on the dynamic environmental sensing data containing temperature and flow rate to construct a multidimensional state matrix, and the multidimensional state matrix is ​​input into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors. Step S2: When the electrical power jump exceeds the preset threshold, extract the sensing data to construct the perturbation snapshot matrix, and use Givens rotation transformation to complete the online update of the orthogonal projection basis vector; Step S3: Input the updated principal mode coefficient vector under the orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint deduction operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; Step S4: Calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the warning threshold, locate the coordinates of the anomaly source and output the warning level. Step S5: Construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation in combination with a preset time constant, and send it to the controlled unit.

[0007] Preferably, the sub-step of inputting the multidimensional state matrix into the orthogonal modal decomposition model with orthogonal projection basis vectors in step S1 includes: step S11, obtaining a historical full-condition finite element simulation snapshot dataset; step S12, solving the singular value decomposition on the snapshot dataset and extracting the set of orthogonal spatial basis vectors; step S13, according to the criterion that the cumulative energy contribution rate reaches a preset proportion threshold, extracting the top K order spatial orthogonal basis vectors arranged in the front row to construct a reduced-order projection matrix, where K is a positive integer; step S14, using the reduced-order projection matrix to linearly map the multidimensional state matrix to the low-dimensional coefficient space and output the principal modal coefficient vector.

[0008] Preferably, the sub-steps of solving the three-dimensional twin prediction vector by the physical constraint deduction operator in step S3 include: step S31, using a differentiable neural network to fit the nonlinear mapping trend of the mode evolution over time; step S32, configuring the non-steady Navier-Stokes equation discrete residual penalty term and the convective heat transfer energy balance residual penalty term in the loss function of the differentiable neural network; step S33, according to the loss function, constraining the convergence of mass continuity flux and heat flux of the physical field of the entire computer room at each discrete grid node within the prediction step.

[0009] Preferably, the sub-step of constructing the residual functional in step S4 includes: step S41, extracting the temperature measurement value and flow velocity measurement value contained in the measured state vector, and the twin predicted temperature value and twin predicted flow velocity value contained in the three-dimensional twin prediction vector; step S42, solving the residual functional according to the following formula. : ,in, For the first Temperature measurements at each measuring point For the first Twin predicted temperature values ​​for each measuring point For the first Flow velocity measurements at each measuring point For the first Twin predicted flow velocity values ​​at each measuring point The preset temperature normalization constant, The preset flow rate normalization constant, This is the temperature residual weighting coefficient. Let be the weighting coefficient for the velocity residual, and satisfy . , This represents the total number of temperature measurement points. This represents the total number of flow velocity measurement points.

[0010] Preferably, the sub-steps in step S4 for locating the coordinates of the anomaly source and outputting the warning level include: step S43, calculating the first derivative of the residual functional with respect to time; step S44, comparing the instantaneous value and the first derivative of the residual functional with the multi-level safety boundary respectively; step S45, when the instantaneous value exceeds the warning threshold and the first derivative is positive, extracting the physical coordinates corresponding to the local grid cell with the largest absolute value of the residual as the coordinates of the anomaly source.

[0011] Preferably, the sub-step of updating the orthogonal projection basis vector online using Givens rotation transformation in step S2 includes: step S21, capturing time-series sensing data and assembling it into a perturbation snapshot matrix within a continuous sliding sampling period after the rack electrical power jump; step S22, constructing a Givens plane rotation matrix sequence to eliminate off-diagonal perturbation components; step S23, applying the Givens plane rotation matrix sequence to the original orthogonal projection basis vector, refreshing the spatial orientation and constraining orthogonal normalization, and outputting the updated orthogonal projection basis vector.

[0012] Preferably, the sub-step of calculating the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit in step S5 includes: step S51, constructing a target functional with integrated energy consumption and temperature rise penalty as the target functional, using the three-dimensional twin prediction vector at the coordinates of the anomaly source to approximate the preset safety reference temperature and minimize the valve opening and fan speed increment of the controlled unit as constraints; step S52, assembling the Jacobian sensitivity matrix by traversing the first-order partial derivatives of the disturbance of the infinitesimal elements of each controlled unit parameter with respect to the target functional through an automatic differentiation algorithm.

[0013] Preferably, the sub-steps for solving the optimal control increment vector in step S5 include: step S53, extracting the adjustment difference between the measured temperature of the current controlled area and the preset safe target temperature; step S54, extracting the first partial derivative of the hot spot temperature with respect to the cold aisle air velocity and the second partial derivative of the hot spot temperature with respect to the supply air temperature using the Jacobian sensitivity matrix; step S55, projecting the adjustment difference onto the air volume channel and the temperature control channel respectively according to the preset air volume adjustment weight and the cooling temperature control weight, and calculating the air velocity adjustment increment and the temperature adjustment amount.

[0014] Preferably, the sub-step of superimposing first-order differential advance compensation on the temperature regulation amount in step S5 includes: step S56, reading the preset time constant of the cold water regulating valve; step S57, calculating the difference quotient between the temperature regulation amount calculated in the current control cycle and the temperature regulation amount calculated in the previous control cycle; step S58, using the product of the difference quotient and the time constant as the advance damping amount, superimposing it on the temperature regulation amount in the current control cycle to generate the actual control increment that offsets the transmission delay of pipeline thermal resistance, and encapsulating it into the control message; after completing the multi-physics closed-loop control, the sub-step of digital twin verification is also included: step S59, after the control message is sent and a delay period aligned with the time constant is elapsed, collecting a new round of measured data from each physical measurement point; step S510, comparing the new round of measured data with the temperature drop convergence slope of the three-dimensional twin prediction vector in the controlled area; step S511, when the temperature drop convergence slope deviation exceeds the preset tolerance range, triggering the state space covariance adaptive correction process of the digital twin model.

[0015] A dynamic environment multimodal perception and digital twin AI early warning linkage system is used to realize the dynamic environment multimodal perception and digital twin AI early warning linkage method, including: data reduction module, basis update module, physical twin inference module, anomaly diagnosis and early warning module and sensitivity control linkage module; The data reduction module is used to perform spatial interpolation on the dynamic environmental sensing data containing temperature and flow rate to construct a multi-dimensional state matrix, and input the multi-dimensional state matrix into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors; The basis update module is used to extract sensing data to construct a perturbation snapshot matrix when the electrical power jump exceeds a preset threshold, and to complete the online update of the orthogonal projection basis vector using Givens rotation transformation; The physical twin inference module is used to input the principal mode coefficient vector under the updated orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint inference operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; The anomaly diagnosis and early warning module is used to calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the early warning threshold, the coordinates of the anomaly source are located and the early warning level is output. The sensitivity control linkage module is used to construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation amount along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation amount with a preset time constant and sending it to the controlled unit.

[0016] Compared with existing technologies, the present invention provides a method and system for multimodal perception of environmental conditions and digital twin AI early warning linkage, which has the following advantages: 1. In the linkage between dynamic environment multimodal perception and digital twin AI early warning, the computational dimension of multiphysics state is reduced by orthogonal mode decomposition, and fluid and energy conservation constraints are applied to the inference process by using physical constraint inference operators. This can reduce the large amount of computational overhead brought by full-order three-dimensional physical field simulation, enabling the digital twin model to meet the millisecond-level online inference requirements. At the same time, physical conservation constraints can suppress the state divergence and physical distortion that occur in pure data-driven models when the working conditions change drastically. This allows the prediction results to maintain consistency with the actual physical process while improving the response speed, providing reliable twin prediction results for timely early warning.

[0017] 2. When the electrical power of the rack undergoes a large jump, the Givens rotation transformation is used to update the orthogonal projection basis vectors online, so that the reduced-order model can adapt to the state distribution after the load change in a timely manner. This can reduce the modal truncation error caused by the sudden load change, avoid the fixed orthogonal projection basis vectors from gradually becoming mismatched due to the deviation of the operating conditions, and enable the digital twin model to still accurately follow the actual state changes under unsteady conditions, thereby improving the stability and representation accuracy of continuous simulation.

[0018] 3. By comparing the deviation between the measured state and the three-dimensional twin prediction vector, and combining the residual functional and its changing trend for early warning judgment, the interference of normal operation fluctuations and sensor measurement noise on anomaly judgment can be reduced. For anomalies such as local air duct obstruction and gradual heat accumulation that have not yet caused obvious temperature over-limits, they can also be identified in advance based on the continuously increasing deviation between the measured state and the twin prediction state, thereby moving the anomaly warning point forward to before obvious over-temperature alarm and improving the ability to identify early anomalies in the computer room. Attached Figure Description

[0019] Figure 1 This is a flowchart of the physical field conservation constraint derivation and time delay compensation control of the present invention; Figure 2 This is a diagram of the multidimensional state reduction and sensitivity adaptive linkage system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] A method for linking dynamic environmental multimodal perception with digital twin AI early warning includes the following steps: Step S1: Spatial interpolation is performed on the dynamic environmental sensing data containing temperature and flow rate to construct a multidimensional state matrix, and the multidimensional state matrix is ​​input into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors. Step S2: When the electrical power jump exceeds the preset threshold, extract the sensing data to construct the perturbation snapshot matrix, and use Givens rotation transformation to complete the online update of the orthogonal projection basis vector; Step S3: Input the updated principal mode coefficient vector under the orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint deduction operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; Step S4: Calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the warning threshold, locate the coordinates of the anomaly source and output the warning level. Step S5: Construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation in combination with a preset time constant, and send it to the controlled unit.

[0022] Preferably, the sub-step of inputting the multidimensional state matrix into the orthogonal modal decomposition model with orthogonal projection basis vectors in step S1 includes: step S11, obtaining a historical full-condition finite element simulation snapshot dataset; step S12, solving the singular value decomposition on the snapshot dataset and extracting the set of orthogonal spatial basis vectors; step S13, according to the criterion that the cumulative energy contribution rate reaches a preset proportion threshold, extracting the top K order spatial orthogonal basis vectors arranged in the front row to construct a reduced-order projection matrix, where K is a positive integer; step S14, using the reduced-order projection matrix to linearly map the multidimensional state matrix to the low-dimensional coefficient space and output the principal modal coefficient vector.

[0023] Preferably, the sub-steps of solving the three-dimensional twin prediction vector by the physical constraint deduction operator in step S3 include: step S31, using a differentiable neural network to fit the nonlinear mapping trend of the mode evolution over time; step S32, configuring the non-steady Navier-Stokes equation discrete residual penalty term and the convective heat transfer energy balance residual penalty term in the loss function of the differentiable neural network; step S33, according to the loss function, constraining the convergence of mass continuity flux and heat flux of the physical field of the entire computer room at each discrete grid node within the prediction step.

[0024] Preferably, the sub-step of constructing the residual functional in step S4 includes: step S41, extracting the temperature measurement value and flow velocity measurement value contained in the measured state vector, and the twin predicted temperature value and twin predicted flow velocity value contained in the three-dimensional twin prediction vector; step S42, solving the residual functional according to the following formula. : ,in, For the first Temperature measurements at each measuring point For the first Twin predicted temperature values ​​for each measuring point For the first Flow velocity measurements at each measuring point For the first Twin predicted velocity values ​​for each measuring point The preset temperature normalization constant, The preset flow rate normalization constant, This is the temperature residual weighting coefficient. Let be the weighting coefficient for the velocity residual, and satisfy . , This represents the total number of temperature measurement points. This represents the total number of flow velocity measurement points.

[0025] Preferably, the sub-steps in step S4 for locating the coordinates of the anomaly source and outputting the warning level include: step S43, calculating the first derivative of the residual functional with respect to time; step S44, comparing the instantaneous value and the first derivative of the residual functional with the multi-level safety boundary respectively; step S45, when the instantaneous value exceeds the warning threshold and the first derivative is positive, extracting the physical coordinates corresponding to the local grid cell with the largest absolute value of the residual as the coordinates of the anomaly source.

[0026] Preferably, the sub-step of updating the orthogonal projection basis vector online using Givens rotation transformation in step S2 includes: step S21, capturing time-series sensing data and assembling it into a perturbation snapshot matrix within a continuous sliding sampling period after the rack electrical power jump; step S22, constructing a Givens plane rotation matrix sequence to eliminate off-diagonal perturbation components; step S23, applying the Givens plane rotation matrix sequence to the original orthogonal projection basis vector, refreshing the spatial orientation and constraining orthogonal normalization, and outputting the updated orthogonal projection basis vector.

[0027] Preferably, the sub-step of calculating the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit in step S5 includes: step S51, constructing a target functional with integrated energy consumption and temperature rise penalty as the target functional, using the three-dimensional twin prediction vector at the coordinates of the anomaly source to approximate the preset safety reference temperature and minimize the valve opening and fan speed increment of the controlled unit as constraints; step S52, assembling the Jacobian sensitivity matrix by traversing the first-order partial derivatives of the disturbance of the infinitesimal elements of each controlled unit parameter with respect to the target functional through an automatic differentiation algorithm.

[0028] Preferably, the sub-steps for solving the optimal control increment vector in step S5 include: step S53, extracting the adjustment difference between the measured temperature of the current controlled area and the preset safe target temperature; step S54, extracting the first partial derivative of the hot spot temperature with respect to the cold aisle air velocity and the second partial derivative of the hot spot temperature with respect to the supply air temperature using the Jacobian sensitivity matrix; step S55, projecting the adjustment difference onto the air volume channel and the temperature control channel respectively according to the preset air volume adjustment weight and the cooling temperature control weight, and calculating the air velocity adjustment increment and the temperature adjustment amount.

[0029] Preferably, the sub-step of superimposing first-order differential advance compensation on the temperature regulation amount in step S5 includes: step S56, reading the preset time constant of the cold water regulating valve; step S57, calculating the difference quotient between the temperature regulation amount calculated in the current control cycle and the temperature regulation amount calculated in the previous control cycle; step S58, using the product of the difference quotient and the time constant as the advance damping amount, superimposing it on the temperature regulation amount in the current control cycle to generate the actual control increment that offsets the transmission delay of pipeline thermal resistance, and encapsulating it into the control message; after completing the multi-physics closed-loop control, the sub-step of digital twin verification is also included: step S59, after the control message is sent and a delay period aligned with the time constant is elapsed, collecting a new round of measured data from each physical measurement point; step S510, comparing the new round of measured data with the temperature drop convergence slope of the three-dimensional twin prediction vector in the controlled area; step S511, when the temperature drop convergence slope deviation exceeds the preset tolerance range, triggering the state space covariance adaptive correction process of the digital twin model.

[0030] A dynamic environment multimodal perception and digital twin AI early warning linkage system is used to realize the dynamic environment multimodal perception and digital twin AI early warning linkage method, including: data reduction module, basis update module, physical twin inference module, anomaly diagnosis and early warning module and sensitivity control linkage module; The data reduction module is used to perform spatial interpolation on the dynamic environmental sensing data containing temperature and flow rate to construct a multi-dimensional state matrix, and input the multi-dimensional state matrix into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors; The basis update module is used to extract sensing data to construct a perturbation snapshot matrix when the electrical power jump exceeds a preset threshold, and to complete the online update of the orthogonal projection basis vector using Givens rotation transformation; The physical twin inference module is used to input the principal mode coefficient vector under the updated orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint inference operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; The anomaly diagnosis and early warning module is used to calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the early warning threshold, the coordinates of the anomaly source are located and the early warning level is output. The sensitivity control linkage module is used to construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation amount along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation amount with a preset time constant and sending it to the controlled unit.

[0031] Example 1: Under the operating conditions of the modular high-density data center hot aisle closed power environment monitoring system, the rated total electrical input power of a single cabinet is set to... The rack heat conversion coefficient is set to The cross-sectional area of ​​the cold aisle air supply outlet is set as The physical properties of the air medium are set to air density. Compared with specific heat capacity at constant pressure Multi-source sensor points synchronously collect underlying power environment data during the current sampling period, including the average temperature collected on the air intake side of the cabinet. The average temperature collected on the air outlet side is 100°C. The collected airflow velocity value is Real-time electrical input power is Spatial interpolation is performed on the environmental sensing data containing temperature and flow velocity according to the spatial location of the measurement points, and the state variables within the same sampling period are arranged according to their spatial location to construct a multi-dimensional state matrix. In the single-rack projection calculation of this embodiment, the current state column corresponding to the target rack in the multi-dimensional state matrix is ​​taken to construct the current state space vector. Before constructing the multidimensional state matrix, the temperature, airflow velocity, and electrical power collected by multi-source sensors have different physical dimensions and numerical magnitudes. To avoid the introduction of calculation biases due to the dominance of a large number of physical quantities in singular value decomposition and spatial orthogonal projection, the system performs dimensionless standardization on each physical quantity. The inlet and outlet air temperatures are divided by a reference temperature of 20°C, the supply air velocity is divided by a reference velocity of 1 m / s, and the electrical input power is divided by a rated power of 10 kW, so that each dimension of physical quantity is transformed into a dimensionless scalar distributed between 0 and 2. The dimensionless standardization process ensures that when calculating the state deviation vector and solving the orthogonal spatial eigenvector, the energy contribution of each physical field participates in singular value decomposition and inner product mapping under a unified mathematical benchmark, and the system calls the pre-stored steady-state reference working point vector. Calculate the state deviation vector to obtain The multidimensional state matrix is ​​input into an orthogonal mode decomposition model constructed by singular value decomposition based on a historical full-condition multiphysics finite element numerical simulation snapshot dataset. In this embodiment, orthogonal projection basis vectors are extracted from the current state column projection. The first-order reduced-order mode coefficients are obtained by solving the orthogonal inner product operation. This linearly maps high-dimensional physical states to low-dimensional modal coordinate systems.

[0032] After obtaining the principal mode coefficient vector, it is input into the physical constraint deduction operator along with the data center heat load vector of the current time slot. In the single-rack calculation of this embodiment, the actual heat load power released by the rack is used as the heat load component corresponding to the target rack in the data center heat load vector. The physical constraint deduction operator uses a differentiable neural network to fit the nonlinear mapping trend of the mode evolution over time. The loss function is configured with a non-steady Navier-Stokes equation discrete residual penalty term and a convective heat transfer energy balance residual penalty term. Forward multiphysics deduction is carried out based on the mass and heat balance physical model. At each discrete grid node, the mass continuity residual is formed by the velocity field divergence and the mass continuity flux is constrained. The energy balance residual is formed by the temperature field and airflow parameters and the heat flux convergence is constrained. The cold aisle supply air volume flow rate is calculated. Then calculate the airflow mass flow rate. The airflow heat capacity flow rate is obtained. The actual heat load power released by the cabinet is determined by the electrical input power and the heat conversion coefficient. The theoretical convection core temperature rise of the cabinet is calculated as follows: The modal space hotspot temperature coupling correction coefficient is a calibration parameter characterizing the influence of the first-order principal mode dynamic fluctuation on the temperature rise of the local turbulent micro-region on the air outlet side of the cabinet. Its unit of measurement is °C. This coefficient is determined during the initial system calibration phase by comparing the linear regression slope of the maximum local hotspot temperature at the air outlet under different load disturbances in the full-order finite element simulation. Its calibration value is 0.2 °C, used to compensate for the high-order perturbation energy missed at the local geometric boundary due to the low-dimensional principal mode truncation. A coupling correction coefficient is configured for the hotspot temperature in the modal space. Under the given conditions, the physical constraint deduction operator solves for the three-dimensional twin prediction vector. In this embodiment, the core hotspot temperature component on the air outlet side of the cabinet is used for calculation, and its theoretical steady-state prediction temperature is... This constrains the convergence of mass continuity flux and heat flux across all grid nodes, reducing the solution time of higher-order partial differential equations and preventing computational state divergence.

[0033] During the current monitoring period, the actual hotspot temperature measured by the infrared temperature measurement array deployed at the air outlet of the server rack was: The flow velocity sensor measured the local airflow velocity as The temperature and velocity values ​​of the corresponding measuring points are extracted from the measured state vector and the three-dimensional twin prediction vector, and the residual tensor is calculated to obtain the temperature state residual. and airflow velocity state residual In the calculation of the velocity state residual, since the theoretically predicted velocity of the cabinet air supply measurement point obtained by the digital twin model based on the fluid continuity conservation under the baseline operating condition without local flow channel anomalies completely coincides with the set air supply velocity of 2 m / s at the cold aisle inlet under the steady-state baseline, the theoretically predicted value substituted at the local measurement point is the baseline velocity of that measurement point under the calibration condition. When the actual measured local velocity is obstructed and decays to 1.8 m / s, the difference between the measured value and the twin theoretical prediction value accurately characterizes the local flow resistance disturbance at that point, and the temperature normalization reference constant is set. Flow velocity normalization reference constant Configure temperature residual weighting coefficient Weighting coefficient of flow velocity residual And satisfy Constructing residual functionals as follows: ,in, and These are the measured temperature value and the twin-predicted temperature value, respectively. and These are the measured airflow velocity value and the supply airflow velocity value, respectively. This is the temperature normalization reference constant. The reference constant for normalized flow velocity, and These are the weighting coefficients for temperature residuals and flow velocity residuals, respectively; substituting them into the numerical calculation yields... The temperature residual weighting coefficient is set to 0.7, and the flow velocity residual weighting coefficient is set to 0.3. This is based on measurements from a computer room thermodynamic sensitivity experiment. Temperature, as a core safety assessment indicator, has a higher early warning weight, while flow velocity is used to detect obstructed precursor characteristics in advance. The first-level precursor warning threshold is set to 0.05, which is a statistical safety boundary determined by the sum of the mean of the residual functional calculated from 500 sets of baseline data continuously collected under normal steady-state fault-free conditions and three times the standard deviation. Values ​​below 0.05 are considered measurement noise and small steady-state disturbances, while values ​​above 0.05 are considered deterministic anomalies. The system compares the instantaneous value and first derivative of the residual functional with the multi-level safety boundaries. Each level of safety boundary corresponds to the instantaneous value boundary and the first derivative boundary of the residual functional, respectively, and outputs the corresponding warning level based on the comparison results. This embodiment explains the first-level boundary; the first-level yellow precursor warning threshold is... The first derivative is used as the boundary of its growth trend with zero, due to the current residual functional calculation results. Furthermore, the first derivative of the residual functional with respect to time is positive, and the system outputs a first-level early warning. At the same time, the physical coordinates corresponding to the local grid cell with the largest absolute value of the residual are extracted as the coordinates of the anomaly source. In the specific implementation of locating the coordinates of the anomaly source, the digital twin system subtracts the theoretical prediction values ​​of the 10,000 grid nodes generated by forward extrapolation from the measured physical field of the whole domain reconstructed by multi-point spatial interpolation at each grid node to generate a 10,000-dimensional spatial residual distribution matrix. The system traverses the residual distribution matrix and extracts the index number of the discrete grid cell with the largest weighted comprehensive absolute value of the temperature residual and velocity residual. Based on the coordinates of the three-dimensional geometric center point of the grid cell in the preset spatial topological coordinate system, the coordinates of the anomaly source in the computer room where the heat accumulation or air duct blockage is located are directly output.

[0034] In response to a precursor warning trigger, and with the goal of eliminating heat accumulation at the coordinates of the anomaly source and reducing system energy consumption within the digital twin space, a comprehensive energy consumption and temperature rise penalty objective functional is constructed as the objective functional. Set the target hot spot temperature for safe rack operation as follows: This temperature is used as the preset safety reference temperature in the objective functional to calculate the difference in adjustment requirements between the current measured temperature and the target value. The objective functional approximates the preset safety reference temperature in the hotspot prediction vector of the three-dimensional twin prediction vector at the anomaly source coordinates as a temperature rise penalty term, and uses the weighted penalty term of the valve opening increment and fan speed increment of the controlled unit as an energy consumption term. When calculating the Jacobian sensitivity matrix of the objective functional with respect to the parameters of the controlled unit online, the speed of the terminal variable frequency fan and the opening of the chilled water regulating valve are used as parameters of the controlled unit. The first-order partial derivatives of the objective functional when the parameters of each controlled unit are subjected to infinitesimal disturbances are calculated by an automatic differential algorithm, and the Jacobian sensitivity matrix is ​​assembled according to the calculation branches of the temperature rise term and the energy consumption term. Among them, the supply air velocity and supply air temperature are used as intermediate parameters of the controlled unit action on the hotspot temperature. The calculation branch corresponding to the temperature rise term retains the first-order partial derivative of the hotspot temperature with respect to the two intermediate parameters. The first partial derivative of the hotspot temperature with respect to the supply air velocity of the cold aisle is extracted by decoupling from the Jacobian sensitivity matrix. And the second partial derivative of the hot spot temperature with respect to the supply air temperature Based on the preset airflow adjustment weight allocation ratio Weighting ratio with cooling temperature regulation The theoretical adjustment increment of the air supply velocity was calculated along the negative gradient direction. Theoretical adjustment increment of supply air temperature Among them, the theoretical adjustment increment of the supply air temperature is used as the temperature adjustment quantity, and together with the theoretical adjustment increment of the supply air velocity, they form the optimal control increment vector; when the initial supply air velocity is And the initial air supply temperature is Based on the benchmark, the inherent physical inertia hysteresis time constant of the cold water regulating valve is considered. In the discrete linkage control cycle The next extraction of the previous cycle's air supply temperature control increment Compared with the current cycle increment Applying a first-order differential lead compensation operator yields the actual control increment. To prevent excessive transient impact commands from being generated in the early stages of a step disturbance due to the large time constant of the first-order differential lead calculation (45s), the system is equipped with a first-order low-pass filter with a time constant of 2s and a dynamic anti-saturation limiting unit at the back end of the differential lead operator. When the calculated transient supply air temperature adjustment increment exceeds the maximum allowable temperature range of 15℃ or exceeds the rated physical stroke of the chilled water regulating valve, the limiting unit automatically cuts off the actual issued command and clamps it within the safe output boundary. At the same time, the low-pass filter smoothly eliminates high-frequency numerical glitches introduced by the discrete differential calculation, and while compensating for pipeline thermal hysteresis, it prevents the actuator from oversaturating and generating mechanical shock, adjusting the target supply air velocity to... Modulate the current transient air supply set temperature command to The data is distributed to the terminal inverter air conditioners and chilled water regulating valves. Simultaneously, when a step change in electrical power exceeding 30% is detected, a perturbation snapshot matrix is ​​extracted, and the orthogonal projection basis vectors are updated online using Givens rotation transformation. This offsets pipeline transmission lag and suppresses modal truncation error, driving the continuous three-dimensional temperature field of the computer room to converge smoothly to the safe operating reference range.

[0035] Example 2: This example constructs a dynamic environment physical test platform including a rack array and a closed hot aisle partition. It is used to investigate the simulation and linkage control process of the dynamic environment multimodal sensing and digital twin AI early warning linkage method under conditions of measurement noise, local airflow obstruction, and sudden changes in electrical load. The test platform is equipped with a temperature sensor array with a temperature measurement accuracy of 0.05℃, a hot-wire anemometer with a measurement range of 0.10 to 10.00 m / s and a measurement accuracy of 0.02 m / s, and a power bus analyzer with a sampling response period better than 10 ms; the data sampling period is set to 5 seconds, and... This cycle serves as the global clock for routine physical field inspections and historical database storage across all 10,000 grid nodes. The bus power analyzer and core airflow sensing points are also equipped with a 10ms hardware interrupt acquisition thread. When physical obstruction occurs in a local airflow path, the high-frequency interrupt thread captures airflow velocity fluctuations and power fluctuations within tens of milliseconds, triggering a rapid forward inference of the principal mode coefficient vector in an event-driven manner. Gaussian white noise with a signal-to-noise ratio of 20dB and 50Hz power frequency harmonic interference are superimposed in the data acquisition link. Five test units were set up for the experiment. The present invention employs a complete linkage architecture of physical conservation constraint order reduction derivation, online update of orthogonal projection basis vectors by Givens rotation, and first-order differential advance compensation. Control group 1 adopts a pure data-driven neural network derivation model that removes the boundary of the Navier-Stokes equation and the energy balance residual penalty. Control group 2 adopts a fixed orthogonal projection basis model that removes the online update mechanism of Givens rotation. Control group 3 adopts a conventional sensitivity negative gradient feedback adjustment model that removes the first-order differential advance compensation operator. Control group 4 adopts a conventional dynamic environment monitoring system based on the static over-limit threshold setting of discrete physical measurement points.

[0036] At the start of the test, the rated total electrical input power of a single cabinet was set to 12kW, the initial air supply velocity in the cold aisle was 2m / s, and the initial air supply temperature was 20℃. When the system was running for 60 seconds, a local physical grid with a 20% opening was applied to the air intake duct at the bottom of the target cabinet to create an airflow short-circuit anomaly. When the system was running for 180 seconds, the electrical load power of the cabinet was increased from 12kW to 16.5kW in 8 seconds, with a load jump of 37.5%. The test continuously recorded the original multimodal sensor signals, principal mode coefficient vector, twin predicted temperature, residual functional value, anomaly warning response delay time, linkage control command adjustment amount, and steady-state convergence time and temperature overshoot of the temperature field in the controlled area before and after the occurrence of the anomaly and the sudden change in load.

[0037] During the steady-state baseline operation phase, the load power of the sample group of this invention was 12kW, the measured temperature was 40.72℃, the twin prediction temperature corresponding to the three-dimensional twin prediction vector was 40.7℃, the residual functional value was 0.00042, the system had no alarms, the steady-state convergence time was 0s, and the temperature overshoot was 0℃. Under the same load, the measured temperature of the control group was 40.68℃, the twin prediction temperature was 40.85℃, the residual functional value was 0.0031, and there were also no alarms, the steady-state convergence time was 0s, and the temperature overshoot was 0℃.

[0038] After applying localized airflow obstruction, the sample group of this invention still maintained a load power of 12kW, with the measured temperature rising to 43.5℃ and the twin predicted temperature being 40.7℃. Subtracting the corresponding components of the measured state vector and the three-dimensional twin predicted vector yielded a temperature residual of 2.8℃ and a wind speed residual of -0.2m / s, from which the residual functional was derived. The value is 0.06688, which exceeds the warning threshold. With a boundary value of 0.05, the system completes the early warning judgment and locates the coordinates of the anomaly source within 1.2 seconds. After entering the reverse linkage control, it follows the time constant. The temperature regulation was superimposed with first-order differential lead compensation. The actual transient control increment was -10.8℃. The controlled area converged to the target temperature of 35℃ within 35s, with a temperature overshoot of 0.28℃. Under the same local duct obstruction condition, the load power of control group 1 was 12kW, the measured temperature was 43.52℃, the twin predicted temperature rose to 42.15℃, the residual functional value was 0.0215, no anomaly was detected, the steady-state convergence time was 115s, and the temperature overshoot was 2.45℃. The load power of control group 3 was also 12kW. The measured temperature was 43.48℃, the twin-predicted temperature was 40.7℃, the residual functional value was 0.0668, the warning delay was 1.2s, the steady-state convergence time was 92s and the temperature overshoot was 4.82℃ without first-order differential lead compensation. The load power of the control group 4 was 12kW, the measured temperature was 43.5℃. Since twin prediction was not performed in this group, there was no twin-predicted temperature and the residual functional could not be calculated. Its warning delay was 48.5s, the steady-state convergence time was 160s and the temperature overshoot was 5.3℃.

[0039] After the load step change was triggered at 180s, the load power of the sample group of this invention increased to 16.5kW, the measured temperature was 48.2℃, and the twin predicted temperature was 47.75℃, with a prediction deviation of 0.45℃. The residual functional function value was 0.0382, the warning delay was 2s, the steady-state convergence time was 42s, and the temperature overshoot was 0.35℃. Under the same load step conditions, the control group II used a fixed orthogonal projection basis, the measured temperature was 48.15℃, and the twin predicted temperature was 44.33℃, with a prediction deviation of 3.82℃ from the difference between the two. The residual functional function value was 0.1625, the warning delay was extended to 18.5s, the steady-state convergence time was 138s, and the corresponding temperature overshoot was 3.82℃. The sample group of this invention constructed a perturbation snapshot matrix after the power jump and used Givens rotation transformation to update the orthogonal projection basis vector online. The modal truncation prediction error was controlled within 0.45℃. The measured temperature of the control group 1 under a load of 16.5kW was 48.25℃, the twin predicted temperature was 52.8℃, the residual functional value was 0.231, a false alarm occurred, the steady-state convergence time was 185s, and the temperature overshoot was 6.15℃.

[0040] After completing the closed-loop control, the sample group of this invention maintained a load power of 16.5kW, the measured temperature of the controlled area was 34.92℃, the twin predicted temperature was 35℃, the residual functional value dropped to 0.00085, the system state returned to safety, the steady-state convergence time was recorded as 42s, and the temperature overshoot was 0.12℃.

[0041] In the above operating conditions, the physical constraint deduction operator constrains the three-dimensional twin prediction vector by using the residual penalty of the fluid and energy conservation equations to reduce the degree of deviation of the predicted state from the physical process when the load changes abruptly. After the measured state vector and the three-dimensional twin prediction vector form a residual tensor, a residual functional is constructed so that the precursor anomaly caused by local airflow obstruction can enter the early warning judgment tens of seconds before the physical measurement point reaches the static temperature control over-limit threshold. When the electrical power undergoes a step change, the orthogonal projection basis vector is updated online through the perturbation snapshot matrix and Givens rotation transformation to reduce the projection mismatch of the fixed basis after the operating condition changes. In the reverse linkage control process, the temperature adjustment amount is superimposed with first-order differential advance compensation to offset the heat transfer inertia of the terminal controlled unit and the pipeline delivery time lag, so that the continuous space temperature field is restored to the safe operating range under non-steady-state disturbance conditions.

[0042] Example 3: When deploying a modular closed hot aisle power environment digital twin system for a data center, the orthogonal modal decomposition model is initially constructed using a historical full-condition finite element simulation snapshot dataset obtained through numerical heat transfer and computational fluid dynamics. The physical server room space is divided into a spatial topological coordinate system containing 10,000 discrete fluid mesh nodes. During the initial system calibration phase, parameters covering continuous variations in rack electrical power (0-20.0 kW), cold aisle air velocity (0.50-5.00 m / s), and supply air set temperature (15.0-25.0℃) are extracted. A snapshot of the simulated spatial flow field, with assembled components having dimensions of [missing information]. High-dimensional physical state matrix ,in Carrying historical full-condition finite element simulation snapshot data, during actual system operation, the data reduction module collects multi-dimensional data through the sensor network inside the computer room. After processing, it transmits the main mode coefficient vector to the physical twin inference module via shared memory. The basis update module monitors the bus electrical parameters in real time. When the electrical power jumps beyond the preset threshold, it updates the orthogonal projection basis vector online through Givens rotation transformation and synchronizes the updated orthogonal projection basis vector to the data reduction module and the physical twin inference module. The physical twin inference module pushes the three-dimensional twin prediction vector to the anomaly diagnosis and early warning module in each control cycle. When the residual functional exceeds the early warning threshold, the anomaly diagnosis and early warning module locates the coordinates of the anomaly source and outputs the early warning level. At the same time, it sends the coordinates of the anomaly source and the early warning event to the sensitivity control linkage module. After calculating the optimal control increment vector and completing the first-order differential lead compensation, the sensitivity control linkage module distributes control messages to the controlled units through the standard industrial communication protocol, forming a closed-loop control architecture.

[0043] For high-dimensional physical state matrix Solving the singular value decomposition yields , where the matrix The matrix of orthogonal eigenvectors in space, and the diagonal matrix Singular values ​​in Sort by descending order and calculate cumulative energy contribution rate; select the top performers whose cumulative energy contribution rate reaches 95% of the preset threshold. Orthogonal basis vectors in a space of order , constructing a space of size Reduced-order projection matrix After acquiring dynamic environmental sensing data including inlet air temperature, outlet air temperature, supply air velocity, and bus electrical parameters, the sensor array maps this data to 10,000 discrete grid nodes using spatial three-dimensional spline interpolation, and constructs a multi-dimensional state matrix based on the spatial node positions. This multi-dimensional state matrix is ​​then input into an orthogonal mode decomposition model with orthogonal projection basis vectors. By subtracting the steady-state reference operating point vector from the state column corresponding to the current sampling period, the global state deviation vector is obtained. Then, the inner product operation is performed through orthogonal projection. The high-dimensional physical field is linearly mapped to the low-dimensional coefficient space, and the output is a principal mode coefficient vector consisting of four scalar elements. This preserves the main spatial characteristics of the flow field energy distribution and reduces the computational cost caused by 10,000 discrete grid nodes directly participating in subsequent deductions.

[0044] Obtain the principal mode coefficient vector Then, the heat load vector of the computer room in the current sampling period is obtained, and the principal mode coefficient vector is concatenated with the heat load vector of the computer room to form the input tensor; in this embodiment, the heat load component corresponding to the target rack is the rack heat load power. The physical constraint derivation operator employs a differentiable neural network consisting of a four-layer feedforward network and a skip residual layer. The number of hidden layer nodes is set to 64, 128, 64, and 4 respectively. The inter-layer activation function is a continuously second-differentiable arctangent activation function. The loss function includes unsteady Navier-Stokes equation discrete residual penalty terms and convective heat transfer energy balance residual penalty terms. For spatial discrete fluid grid nodes, the mass continuity residual penalty term is determined by the sum of the partial derivatives of the airflow velocity vector along the three coordinate axes of the Cartesian coordinate system and is used to constrain the mass continuity flux at each discrete grid node. The energy balance residual penalty term is determined by the air density... Specific heat capacity at constant pressure is 1.2 kg / m³ The dot product residual of 1 kJ / (kg·℃) grid node velocity vector and local convective heat transfer temperature rise gradient is determined and used to constrain heat flux convergence.

[0045] When the physical constraint extrapolation operator forward extrapolates the low-dimensional modal evolution state within a 30-second time window, it calculates the mass continuity flux loss and heat flux convergence loss at each grid node using an automatic differentiation framework, and uses these as gradient penalty terms to update the network weight parameters. Within the discrete extrapolation step, it compares the future modal coefficient evolution state output by the network with the reduced-order projection matrix. Multiplication and inversion yield the three-dimensional temperature field, flow field, pressure field, and three-dimensional twin prediction vectors corresponding to key heat-generating nodes in the entire spatial domain. This enables the predicted physical field to converge under the residual penalty of the fluid and energy conservation equations. Specifically, during the derivation, the feedforward neural network receives input data consisting of a principal mode coefficient vector composed of four scalars and a heat load vector of the computer room. This data undergoes nonlinear feature transformation through hidden layers containing 64, 128, 64, and 4 nodes. In the loss function calculation stage, automatic differentiation is used to calculate the first and second order partial derivatives of the velocity field and temperature field of each discrete grid node in various spatial directions, forming the mass divergence residual and the heat convection-diffusion balance residual of the grid nodes. The physical conservation residual is then incorporated into the backpropagation gradient according to the penalty weight. Subsequently, the four updated mode coefficients obtained from the derivation are linearly reconstructed with a 10,000-row, 4-column reduced-order projection matrix, and the three-dimensional temperature, pressure, and velocity tensors of the 10,000 grid nodes in the entire domain are obtained point-by-point.

[0046] During continuous monitoring of the power environment, the bus current sensor and power transmitter monitor the rack electrical power at a 10ms cycle. The fluctuation state is monitored. When the detected jump amplitude of electrical power within a 10s sliding time window exceeds the 30% steady-state disturbance threshold, the basis update module triggers online update of the orthogonal projection basis vectors. It extracts the time-sensing vector sequences of the 5 sampling periods before and after the mutation, constructing a vector array with a size of [size missing]. Perturbation snapshot matrix To address the truncation mismatch caused by the perturbation snapshot matrix acting on the original reduced-order projection matrix, a sequence of Givens plane rotation matrices is constructed. For the ... Rank and number The plane containing the orthogonal projection basis vectors is used to determine the off-diagonal perturbation components based on the off-diagonal covariance components of the perturbation snapshot matrix in the corresponding projection direction, and the rotation angle is calculated accordingly. Construct a planar rotation matrix Then apply it to the original reduced-order projection matrix. Refresh the spatial orientation of the orthogonal projection basis vectors and maintain The orthogonal normalization constraint makes the updated orthogonal projection basis vector follow the change in thermal convection flow field caused by the jump in electrical power, and limits the mode reconstruction truncation error under transient impact to within 0.5℃ tolerance.

[0047] When constructing the Givens plane rotation matrix, first calculate the perturbation snapshot matrix at the th... Rank and number The off-diagonal covariance components along the directions of the orthogonal projection basis vectors are determined, along with the autovariance difference between the two basis vector directions. The off-diagonal covariance is divided by the autovariance difference, and the arctangent of the quotient is multiplied by one-half to determine the rotation angle of the current plane. Then, cosine values ​​are filled into the corresponding positions on the main diagonal of the rotation matrix, sine and negative sine values ​​are filled into the off-diagonal positions, and 1 is filled into the remaining diagonal positions. The sequence of rotation matrices corresponding to each order of orthogonal projection basis vector is applied sequentially to the original reduced-order projection matrix until the absolute value of each off-diagonal covariance decays below the convergence threshold of 0.001, resulting in the updated orthogonal projection basis vectors.

[0048] During the continuous simulation, the anomaly diagnosis and early warning module calculates the residual tensor between the measured state vector and the three-dimensional twin prediction vector, and constructs a residual functional based on the sum of squared temperature residuals and the sum of squared flow velocity residuals. When the residual functional exceeds the first-level warning threshold When the value is 0.05, the system locates the coordinates of the anomaly source and triggers reverse control; the sensitivity control linkage module approximates the safety reference target temperature using the three-dimensional twin prediction vector at the anomaly source coordinates. The target functional is defined as follows: 35℃, with constraints including reducing the increment of valve opening and fan speed in the controlled unit. The objective functional is then constructed to combine comprehensive energy consumption and temperature rise penalty. The objective functional consists of a tracking deviation term between the predicted temperature of the hotspot in the controlled area and the target temperature of the safety baseline, and a weighted energy consumption penalty term consisting of the square of the increment of the chilled water regulating valve opening and the square of the increment of the variable frequency fan speed.

[0049] When calculating the Jacobian sensitivity matrix, the terminal variable frequency fan speed and chilled water regulating valve opening are used as controlled unit parameters, the space supply air velocity and supply air temperature are used as intermediate heat transfer parameters, and the objective functional is composed of hot spot temperature tracking deviation and energy consumption penalty. As output, a forward computation directed acyclic graph is formed within the computational framework. An automatic differentiation algorithm is used to traverse the first-order partial derivatives of the infinitesimal perturbations of each controlled unit parameter with respect to the target functional, and these are assembled to form a Jacobian sensitivity matrix. In the temperature transfer branch of this computational graph, the first partial derivative of the hotspot temperature with respect to the cold aisle airflow velocity is extracted using the Jacobian sensitivity matrix. And the second partial derivative of the hot spot temperature with respect to the supply air temperature The corresponding values ​​under the current operating conditions are -4.5 and 0.9, respectively; then, the current control cycle is calculated along the negative gradient direction of the Jacobian sensitivity matrix. The air supply speed adjustment increment below With temperature regulation The two together form the optimal control increment vector.

[0050] To address the inherent heat transfer lag in the terminal cold water regulating valve and the pipeline transport inertia, a time constant is preset. Read the temperature adjustment amount calculated in the previous control cycle. Calculate the quotient of the difference between the temperature adjustment in the current control cycle and the previous control cycle, and then combine this quotient with the time constant. The product of these factors is used as the advance damping amount, which is then added to the temperature adjustment amount in the current control cycle to obtain the actual control increment. The control increment after first-order differential lead compensation is encapsulated into a Modbus-RTU industrial communication protocol control message and sent to the variable frequency precision air conditioner and electric regulating valve controller; after the control message is sent, it is compared with the time constant. A new round of measured data is collected from each physical measurement point during a 45-second alignment period. The temperature drop convergence slope in the controlled area is compared with that of the three-dimensional twin prediction vector. When the relative deviation of the temperature drop convergence slope remains within the preset tolerance range of 5%, the current closed-loop adjustment is maintained. When the deviation of the temperature drop convergence slope exceeds the preset tolerance range of 5%, the adaptive correction process of the state space covariance of the digital twin model is triggered. The state space covariance of the digital twin model is corrected based on the state residual covariance information formed by the new round of measured state vector and the three-dimensional twin prediction vector. The Jacobian sensitivity matrix is ​​recalculated based on the corrected prediction state for closed-loop adjustment in subsequent control cycles.

[0051] Example 4: This example combines Figures 1 to 2 This paper describes the method and system for linking multimodal perception of the environment with digital twin AI early warning, such as... Figure 1As shown, step S1 involves spatial interpolation of the dynamic environmental sensing data, including temperature and flow velocity, to construct a multidimensional state matrix. This multidimensional state matrix is ​​then input into an orthogonal modal decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors. Step S2 involves extracting sensing data to construct a perturbation snapshot matrix when the electrical power jump exceeds a preset threshold. The Givens rotation transformation is then used to update the orthogonal projection basis vectors online. Step S3 involves inputting the updated principal mode coefficient vectors constrained by the orthogonal projection basis vectors and the acquired computer room heat load vector into the physical constraint derivation operator. This process is then applied to the fluid and energy conservation equations. The three-dimensional twin prediction vector is solved under differential penalty. Step S4 calculates the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the warning threshold, the coordinates of the anomaly source are located and the warning level is output. Step S5 constructs the target functional with the goal of eliminating the thermal accumulation at the coordinates of the anomaly source. The Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit is calculated. The optimal control increment vector containing the temperature regulation amount is solved along the negative gradient direction. The temperature regulation amount is superimposed with the first-order differential advance compensation in combination with the preset time constant to generate the control message and send it to the controlled unit.

[0052] like Figure 2 As shown, the dynamic environment multimodal perception and digital twin early warning linkage system includes a data reduction module, a basis update module, a physical twin inference module, an anomaly diagnosis and early warning module, and a sensitivity control linkage module. The data reduction module constructs a multidimensional state matrix from the dynamic environment perception data and inputs it into the orthogonal mode decomposition model. The basis update extracts the perturbation snapshot matrix and updates the orthogonal projection basis vector through Givens rotation transformation. The physical twin inference inputs the computer room heat load vector into the physical constraint inference operator to solve the three-dimensional twin prediction vector. The anomaly diagnosis and early warning calculates the residual between the measured state vector and the three-dimensional twin prediction vector and constructs the residual functional to locate the coordinates of the anomaly source. The sensitivity control linkage constructs a target functional for the coordinates of the anomaly source to solve the Jacobian sensitivity matrix and output the optimal control increment vector.

[0053] Example 5: After the modular high-density data center power and environmental monitoring system is deployed on-site, the zero-point temperature drift and installation geometric offset of the multi-source physical sensors are first statically baseline-calibrated. The on-site monitoring host is connected to the inlet air temperature sensor, the outlet air infrared temperature measurement array, the hot-wire anemometer, and the bus power transmitter. Under the initial preheating steady-state conditions of no load in the data center, supply air velocity maintained at 1.5 m / s, and supply air temperature maintained at 20℃, each sensor channel continuously collects 120 sets of multi-dimensional sensing data sequences with a sampling period of 5 seconds. Abnormal disturbance measurement points that deviate from the mean of each channel by more than 3 times the standard deviation are removed by the Laida criterion. Then, the time average value of the effective sampling sequence of each physical channel is calculated to obtain the spatial steady-state reference working point vector. Simultaneously, the initial calibration mean of each physical channel is subtracted from the preset design reference value to obtain the static hardware temperature drift bias and zero-point error of each sensing channel; subsequent real-time sampling forms a state space vector. At that time, the corresponding static hardware temperature drift offset and zero-point error are first deducted, and then used for state deviation calculation, with temperature normalization reference constant. Calibrated at 10℃, normalized flow rate reference constant The value is set to 1 m / s, which serves as a unified benchmark for subsequent orthogonal modal projection of state deviation and residual functional calculation.

[0054] During long-term continuous operation of the monitoring system and the degradation of equipment performance, after the terminal chilled water regulating valve and fan execute discrete control messages, and after the time constant of the chilled water regulating valve... With an aligned time delay window, a multi-source sensor array acquires a new round of measured temperature sequences for the controlled area and calculates the relative deviation between the measured temperature drop convergence slope and the temperature drop convergence slope corresponding to the 3D twin prediction vector. When the relative deviation of the temperature drop convergence slope exceeds the preset tolerance range of 5% in three consecutive linkage control cycles, the state space covariance adaptive correction process of the digital twin model is triggered. The measured state vectors and 3D twin prediction vectors of the most recent 20 consecutive control cycles are extracted to form a state residual sequence and calculate the state residual covariance matrix. A recursive update method with a forgetting factor of 0.95 is adopted to increase the update weight of residual data close to the current control cycle, and the original state space covariance of the digital twin model is recursively corrected accordingly, so that the state space covariance is updated with the current equipment performance and pipeline resistance changes.

[0055] After completing the state-space covariance correction, the Jacobian sensitivity matrix is ​​recalculated based on the corrected digital twin prediction state, and the first partial derivative of the hot spot temperature with respect to the cold aisle airflow velocity is extracted from this matrix. And the second partial derivative of the hot spot temperature with respect to the supply air temperature Simultaneously, a recursive least squares algorithm with a forgetting factor of 0.95 is employed to apply exponentially decaying weights to the measured and predicted residual sequences of the most recent 20 consecutive control periods, and to correct them online using weighted residual covariance information. and The gain step size is adjusted to update as heat exchanger performance deteriorates and pipe resistance changes; during the control cycle... Next, the incremental adjustment of the supply air velocity is recalculated along the negative gradient direction of the Jacobian sensitivity matrix. With temperature regulation Then, by combining first-order differential advance compensation, the actual control increment is obtained. The actual control increment will be issued. Encapsulated into industrial communication messages and sent to the controlled unit, the subsequent control cycle continues to perform closed-loop regulation based on the updated digital twin prediction state and Jacobian sensitivity matrix to adapt to the time delay changes caused by fouling of the terminal heat exchange pipeline and mechanical transmission wear, and drive the continuous three-dimensional temperature field and flow field to converge to the target safe operating range.

[0056] Example 6: The first rack is currently under full electrical load during the collaborative operation of the multi-rack array, and its power input has reached the upper limit. The power is 20kW, and the adjacent second rack is in standby power-off state. The power is 0kW; at this time, the heat load of adjacent racks exhibits an asymmetric limiting distribution, and the total air supply velocity in the cold aisle is 0kW. Under the lower limit operating condition of 0.5 m / s, the inlet reference temperature At a temperature of 22℃, a multi-channel sensor array deployed in the computer room synchronously collects multidimensional discrete sequences of the temperature and velocity fields of the incoming and outgoing air. After interpolation using grid spatial splines, a multidimensional state matrix is ​​constructed according to the spatial node positions. The multidimensional state matrix is ​​then input into an orthogonal mode decomposition model, and the steady-state reference operating point vector is subtracted from the state column corresponding to the current sampling period to obtain the global state deviation vector. Then use the reduced-order projection matrix Project it onto a low-dimensional coefficient space composed of four principal modes, extract the principal mode coefficient vector, and form the room heat load vector for the current sampling period according to the rack spatial position of the heat load components corresponding to the first rack and the second rack. This vector is then input into the physical constraint deduction operator along with the principal mode coefficient vector.

[0057] When a strong convective shear layer forms at the junction of the two racks and local gas vortex shedding and unsteady reverse pressure gradient occur, the differentiable neural network in the physical constraint deduction operator solves the spatial flux constraint based on the discrete residual penalty term of the unsteady Navier-Stokes equation and the residual penalty term of the convective heat transfer energy balance. In the low-dimensional space, it processes the coupled heat transfer boundary formed by the high-temperature hot plume on the outlet side of the first rack and the cold convection entrainment of the second rack, and deduces and outputs the three-dimensional temperature and pressure flow field distribution of 10,000 grid nodes in the whole domain within the prediction time window of the next 30 seconds, suppressing the calculation divergence of the abnormal negative temperature rise at the interface of sudden heat load change.

[0058] When dealing with the convective shear layer at the interface between the two frames, the differential neural network applies mass continuity flux constraints and momentum tensor balance constraints to the adjacent fluid grid cells at the interface in the loss function. Specifically, the normal mass flow rate of the high-temperature hot plume overflowing upward from the outlet side of the first frame at the grid interface and the normal mass flow rate of the low-temperature airflow drawn downward from the second frame at the same interface maintain an algebraic sum conservation, thereby satisfying the mass continuity flux constraint at the interface. At the same time, the shear stress continuity of the velocity gradient is used to suppress numerical dispersion, enabling the low-dimensional mode to handle the two reverse high-shear heat transfer boundaries with a temperature difference of more than 20°C.

[0059] When the infrared temperature measurement array on the air outlet side of the first rack returns the measured temperature The temperature rose to 49.8℃, and the local wind speed transmitter measured the local flow velocity. When the velocity decays to 0.4 m / s, the system obtains the temperature state residual. 3.6℃, ​​velocity state residual The value is -0.1 m / s, and the instantaneous residual functional is calculated according to the given residual functional. The value is 0.09324, which exceeds the warning threshold. The value is 0.05, and the first derivative of the residual functional with respect to time is positive. The system compares the instantaneous value and the first derivative of the residual functional with the multi-level safety boundary, locks the local grid cell with the largest absolute value of the residual at the air outlet of the first rack, uses its physical coordinates as the coordinates of the anomaly source, and outputs the warning level according to the corresponding safety boundary.

[0060] In the reverse closed-loop control process, the upper limit of the variable frequency fan's air delivery speed is 5 m / s, and it is also constrained by the saturation boundary of the chilled water regulating valve opening. The sensitivity control linkage module approximates the preset safety reference temperature using the three-dimensional twin prediction vector at the coordinates of the anomaly source, and constructs a comprehensive energy consumption and temperature rise penalty objective functional as the objective functional, with the constraints of reducing the valve opening increment and fan speed increment of the controlled unit. Among them, the upper limit of the air supply speed and the saturation boundary of the cold water regulating valve opening are used as boundary constraints when solving the control increment.

[0061] When calculating the Jacobian sensitivity matrix, the speed of the terminal variable frequency fan and the opening degree of the chilled water regulating valve are used as parameters of the controlled unit. The infinitesimal disturbances of the parameters of each controlled unit are traversed through the target functional using an automatic differentiation algorithm. The first partial derivatives are obtained and assembled into the Jacobian sensitivity matrix; then, the first partial derivative of the hot spot temperature with respect to the cold aisle airflow velocity is extracted from the temperature transfer branch. And the second partial derivative of the hot spot temperature with respect to the supply air temperature Based on the air volume adjustment weight and the cooling temperature control weight, the temperature adjustment difference of the current controlled area is projected onto the air volume channel and the temperature control channel respectively, and the supply air velocity adjustment increment is calculated along the negative gradient direction. and temperature regulation amount .

[0062] To address the time lag in pipeline thermal resistance transmission of the cold water regulating valve, a time constant is used. The first-order differential lead compensation for 45s is the temperature adjustment amount for the current control cycle. Compensation is performed to obtain the actual control increment issued. Furthermore, by combining the upper limit of the variable frequency fan's air delivery speed with the saturation boundary of the chilled water regulating valve's opening, anti-saturation limiting is implemented. The processed control increment is encapsulated into a discrete control message, which is then distributed to the terminal variable frequency precision air conditioner and electric regulating valve via the main control communication bus, so that the spatial flow field under asymmetric extreme load conditions can be restored to the safe margin range within 48 seconds.

[0063] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for linking dynamic environmental multimodal perception with digital twin AI early warning, characterized in that, Includes the following steps: Step S1: Spatial interpolation is performed on the dynamic environmental sensing data containing temperature and flow rate to construct a multidimensional state matrix, and the multidimensional state matrix is ​​input into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors. Step S2: When the electrical power jump exceeds the preset threshold, extract the sensing data to construct the perturbation snapshot matrix, and use Givens rotation transformation to complete the online update of the orthogonal projection basis vector; Step S3: Input the updated principal mode coefficient vector under the orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint deduction operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; Step S4: Calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the warning threshold, locate the coordinates of the anomaly source and output the warning level. Step S5: Construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation in combination with a preset time constant, and send it to the controlled unit.

2. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps of inputting the multidimensional state matrix into the orthogonal modal decomposition model with orthogonal projection basis vectors in step S1 include: step S11, obtaining a historical full-condition finite element simulation snapshot dataset; step S12, solving the singular value decomposition on the snapshot dataset and extracting the set of orthogonal spatial basis vectors; step S13, according to the criterion that the cumulative energy contribution rate reaches a preset proportion threshold, extracting the top K order spatial orthogonal basis vectors arranged in the front column to construct a reduced-order projection matrix, where K is a positive integer; step S14, using the reduced-order projection matrix to linearly map the multidimensional state matrix to the low-dimensional coefficient space and output the principal modal coefficient vector.

3. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps of solving the three-dimensional twin prediction vector by the physical constraint deduction operator in step S3 include: step S31, using a differentiable neural network to fit the nonlinear mapping trend of the mode evolution over time; step S32, configuring the non-steady Navier-Stokes equation discrete residual penalty term and the convective heat transfer energy balance residual penalty term in the loss function of the differentiable neural network; step S33, according to the loss function, constraining the convergence of mass continuity flux and heat flux of the physical field of the entire computer room at each discrete grid node within the prediction step.

4. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps for constructing the residual functional in step S4 include: Step S41, extracting the temperature and flow rate measurements from the measured state vector, and the twin predicted temperature and flow rate measurements from the three-dimensional twin predicted vector; Step S42: solving the residual functional according to the following formula. : ,in, For the first Temperature measurements at each measuring point For the first Twin predicted temperature values ​​for each measuring point For the first Flow velocity measurements at each measuring point For the first Twin predicted velocity values ​​for each measuring point The preset temperature normalization constant, The preset flow rate normalization constant, This is the temperature residual weighting coefficient. Let be the weighting coefficient for the velocity residual, and satisfy . , This represents the total number of temperature measurement points. This represents the total number of flow velocity measurement points.

5. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 4, characterized in that, The sub-steps in step S4 for locating the coordinates of the anomaly source and outputting the warning level include: step S43, calculating the first derivative of the residual functional with respect to time; step S44, comparing the instantaneous value and the first derivative of the residual functional with the multi-level safety boundary respectively; step S45, when the instantaneous value exceeds the warning threshold and the first derivative is positive, extracting the physical coordinates corresponding to the local grid cell with the largest absolute value of the residual as the coordinates of the anomaly source.

6. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps of updating the orthogonal projection basis vectors online using Givens rotation transformation in step S2 include: step S21, capturing time-series sensing data and assembling it into a perturbation snapshot matrix during the continuous sliding sampling period after the rack electrical power jump; step S22, constructing a Givens plane rotation matrix sequence to eliminate off-diagonal perturbation components; and step S23, applying the Givens plane rotation matrix sequence to the original orthogonal projection basis vectors, refreshing the spatial orientation and constraining orthogonal normalization, and outputting the updated orthogonal projection basis vectors.

7. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps in step S5 for calculating the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit include: Step S51, constructing a target functional with integrated energy consumption and temperature rise penalty as the target functional, using the three-dimensional twin prediction vector at the coordinates of the anomaly source to approximate the preset safety reference temperature and minimize the valve opening and fan speed increment of the controlled unit as constraints; Step S52, assembling the Jacobian sensitivity matrix by traversing the first-order partial derivatives of the disturbance of the infinitesimal elements of each controlled unit parameter with respect to the target functional through an automatic differentiation algorithm.

8. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 1, characterized in that, The sub-steps for solving the optimal control increment vector in step S5 include: step S53, extracting the adjustment difference between the measured temperature of the current controlled area and the preset safe target temperature; step S54, extracting the first partial derivative of the hot spot temperature with respect to the cold aisle air velocity and the second partial derivative of the hot spot temperature with respect to the supply air temperature by decoupling from the Jacobian sensitivity matrix; step S55, projecting the adjustment difference onto the air volume channel and the temperature control channel respectively according to the preset air volume adjustment weight and the cooling temperature control weight, and calculating the air velocity adjustment increment and the temperature adjustment amount.

9. The method for linkage between dynamic environment multimodal perception and digital twin AI early warning according to claim 8, characterized in that, The sub-step of superimposing first-order differential advance compensation on the temperature regulation amount in step S5 includes: step S56, reading the preset time constant of the cold water regulating valve; step S57, calculating the difference quotient between the temperature regulation amount calculated in the current control cycle and the temperature regulation amount calculated in the previous control cycle; step S58, using the product of the difference quotient and the time constant as the advance damping amount, superimposing it on the temperature regulation amount in the current control cycle to generate the actual control increment that offsets the transmission delay of pipeline thermal resistance, and encapsulating it into the control message; after completing the multi-physics closed-loop control, it also includes the sub-step of digital twin verification: step S59, after the control message is sent and a delay period aligned with the time constant is elapsed, collecting a new round of measured data from each physical measurement point; step S510, comparing the new round of measured data with the temperature drop convergence slope of the three-dimensional twin prediction vector in the controlled area; step S511, when the temperature drop convergence slope deviation exceeds the preset tolerance range, triggering the state space covariance adaptive correction process of the digital twin model.

10. A dynamic environment multimodal perception and digital twin AI early warning linkage system, used to implement the dynamic environment multimodal perception and digital twin AI early warning linkage method as described in claim 1, characterized in that, include: Data reduction module, basis update module, physical twin inference module, anomaly diagnosis and early warning module, and sensitivity control linkage module; The data reduction module is used to perform spatial interpolation on the dynamic environmental sensing data containing temperature and flow rate to construct a multi-dimensional state matrix, and input the multi-dimensional state matrix into an orthogonal mode decomposition model with orthogonal projection basis vectors to extract the principal mode coefficient vectors; The basis update module is used to extract sensing data to construct a perturbation snapshot matrix when the electrical power jump exceeds a preset threshold, and to complete the online update of the orthogonal projection basis vector using Givens rotation transformation; The physical twin inference module is used to input the principal mode coefficient vector under the updated orthogonal projection basis vector constraint and the obtained computer room heat load vector into the physical constraint inference operator, and solve the three-dimensional twin prediction vector under the residual penalty of the fluid and energy conservation equation; The anomaly diagnosis and early warning module is used to calculate the residual tensor between the measured state vector and the three-dimensional twin prediction vector to construct the residual functional. When the residual functional exceeds the early warning threshold, the coordinates of the anomaly source are located and the early warning level is output. The sensitivity control linkage module is used to construct a target functional with the goal of eliminating thermal accumulation at the coordinates of the abnormal source, calculate the Jacobian sensitivity matrix of the target functional with respect to the parameters of the controlled unit, solve the optimal control increment vector containing the temperature regulation amount along the negative gradient direction, and generate a control message by superimposing a first-order differential advance compensation on the temperature regulation amount with a preset time constant and sending it to the controlled unit.

Images