A laboratory multi-partition air volume dynamic scheduling method based on edge computing

Abstract

The present application relates to the technical field of ventilation control, in particular to a laboratory multi-partition air volume dynamic scheduling method based on edge computing. The present application generates a virtual differential pressure traveling wave vector by collecting equipment electrical sequence data for phase space reconstruction; during system initialization and background asynchronous update, the cloud node extracts a sparse impedance discriminant matrix with topological isomorphism and issues it; the edge computing node inputs a space-time graph attention network with the traveling wave vector as the node feature, the sparse impedance matrix as the edge feature, and the correlation matrix as the adjacency matrix, extracts a flow field prediction vector; a divergence zero constraint rigid projection layer based on graph Laplace operator is constructed at the output end to convert the fluid continuity equation into a solution space matrix, and the flow field prediction vector is subjected to orthogonal projection and boundary clipping to map out a target control vector and format it for writing to the underlying communication interface. The present application significantly shortens the scheduling response delay and guarantees the stability of the multi-partition fluid.

Description

Technical Field

[0001] This invention relates to the field of ventilation control technology, specifically to a method for dynamic scheduling of air volume in multiple zones of a laboratory based on edge computing. Background Technology

[0002] High-level biosafety laboratories are primarily used for experiments involving highly pathogenic pathogens. The key technological defense against the leakage of pathogenic microorganism aerosols lies in establishing and maintaining a strict negative pressure gradient and absolutely directional airflow organization. Among these, the chemical shower rooms and airlocks within the core protected area, due to their complex operating procedures, represent the areas with the most drastic changes in operating conditions within the entire dynamic airflow control system.

[0003] Currently, laboratory variable air volume (VAV) control systems mainly employ single-loop PID closed-loop control based on pipeline differential pressure sensors, or centralized HVAC control systems relying on IoT architectures. However, existing technologies exhibit significant limitations when applied to the dynamic scheduling of multiple zones in high-level biosafety laboratories.

[0004] First, existing systems generally employ single-node passive feedback control logic that monitors differential pressure deviation and then adjusts the valve opening. When faced with multi-source transient physical disturbances such as transient aerodynamic shocks from opening and closing airtight doors, high-frequency fluctuations in space volume caused by the supply and exhaust of positive-pressure protective clothing, and multiphase fluid dynamic mutations induced by high-pressure spraying of chemical reagents, the post-event feedback mechanism inevitably suffers from response lag, failing to effectively intervene in the initial stage of the disturbance. This easily leads to instantaneous differential pressure runaway and airflow reversal. Second, laboratory zones are typically connected to the same exhaust duct network, exhibiting strong coupling relationships among multiple variables in fluid dynamics. When a zone significantly adjusts its exhaust valve to cope with local transient disturbances, it causes a sharp drop in static pressure within the shared main duct. Because existing centralized control models lack a horizontal real-time coordination mechanism between nodes, and the main fan's frequency conversion adjustment has mechanical inertia, static pressure fluctuations propagate along the duct network, forcing adjacent zones to drastically reduce their exhaust volume, creating a snatching effect and disrupting the stepped negative pressure gradient between multiple zones. Furthermore, existing centralized data processing architectures aggregate data to a central controller or cloud platform for processing. The data packaging, network transmission, and command issuance processes introduce significant latency in control data flow. This latency not only violates the physical constraints of control systems requiring extremely short response times, but also introduces a single point of failure risk into the centralized architecture. Under conditions of network fluctuations or central processing unit failure, the entire system will lose its dynamic scheduling capabilities.

[0005] In summary, the problem of global dynamic scheduling imbalance caused by the delay in control data flow and the limitations of passive feedback data from single nodes when facing multi-source transient physical disturbances in multi-partition strongly coupled systems is a technical problem that urgently needs to be solved in this field.

[0006] To address this, a dynamic scheduling method for air volume in multiple laboratory zones based on edge computing is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic airflow scheduling method for multi-zone laboratory environments based on edge computing. This invention reconstructs the phase space from collected electrical sequence data of equipment to generate a virtual differential pressure traveling wave vector. During system initialization and asynchronous background updates, cloud nodes extract and distribute a topologically isomorphic sparse impedance discrimination matrix. Edge computing nodes input the traveling wave vector as node features, the sparse impedance matrix as edge features, and the correlation matrix as the adjacency matrix into a spatiotemporal graph attention network to extract flow field prediction vectors. At the output end, a zero-constraint rigid projection layer based on the graph Laplacian operator is constructed to transform the fluid continuity equation into a solution space matrix. Orthogonal projection and boundary clipping are performed on the flow field prediction vectors to map out the target control vector and format it for writing to the underlying communication interface. This invention significantly reduces scheduling response latency and ensures multi-zone fluid stability.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A dynamic airflow scheduling method for multiple zones in a laboratory based on edge computing, comprising: Collect electrical operation sequence data of ventilation-related equipment; construct a quasi-static pressure prediction phase space matrix based on the electrical operation sequence data, and generate a virtual pressure difference traveling wave vector through mapping operation; During system initialization and background asynchronous updates, the correlation matrix characterizing the laboratory physical topology is obtained, topology correlation pruning is performed on cloud nodes, the fluid dynamics prediction model is refined into a sparse impedance discrimination matrix isomorphic to the correlation matrix, and then distributed to edge computing nodes. Using the virtual pressure difference traveling wave vector as node features, the sparse impedance discrimination matrix as edge weight features, and combined with the graph topology represented by the correlation matrix, feature extraction is performed at the edge computing node input spatiotemporal graph attention network to obtain the flow field prediction vector. A zero-constraint rigid projection layer based on the graph Laplacian operator is constructed at the output of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix; the flow field prediction vector is orthogonally projected to map the target control vector in the digital space; the target control vector is formatted and written into the underlying communication interface of the ventilation actuator.

[0009] Preferably, the electrical operation sequence data of ventilation-related equipment is collected, including: acquiring the three-phase stator current time-series data and output frequency change rate time-series data of the exhaust fan inverter; acquiring the electromagnetic torque pulse sequence data of the airtight door drive motor; aligning the three-phase stator current time-series data, the output frequency change rate time-series data and the electromagnetic torque pulse sequence data according to a preset time-series sampling step size and performing tensor splicing, and extracting the time sequence with a set sliding time window to generate electrical operation sequence data.

[0010] Preferably, constructing a quasi-static pressure prediction phase space matrix and generating a virtual pressure difference traveling wave vector through mapping operations includes: performing electromagnetic empirical mode decomposition and mechanical inertial masking on electrical operation sequence data to extract fundamental envelope sequence data; performing phase space reconstruction processing on the fundamental envelope sequence data to construct the quasi-static pressure prediction phase space matrix; extracting the pipeline reference operating state vector for the current scheduling cycle and calling a cross-media nonlinear mapping function with embedded aerodynamic boundary conditions; using the pipeline reference operating state vector as modulation constraints, performing nonlinear adaptive modulation on the static electrical propagation constant and hydrodynamic constant to calculate the state-dependent dynamic transfer coefficient tensor; using the dynamic transfer coefficient tensor to perform mapping operations on the quasi-static pressure prediction phase space matrix to calculate the transfer coefficient matrix between the electrical propagation constant and the hydrodynamic constant; using the transfer coefficient matrix, inversely mapping the electrical evolution characteristics in the quasi-static pressure prediction phase space matrix to the aerodynamic flow field space, and generating a virtual pressure difference traveling wave vector through discrete-time integration.

[0011] Preferably, the correlation matrix is ​​obtained, and topology correlation pruning is performed on the cloud node to refine the fluid dynamics prediction model into a sparse impedance discrimination matrix isomorphic to the correlation matrix. This includes: during the system initialization phase and during the background asynchronous update cycle independent of the airflow dynamic scheduling cycle, reading the physical cascaded topology structure composed of multiple laboratory zones, multi-level airlocks, and ventilation ducts, and constructing a correlation matrix characterizing the connectivity state of each ventilation node and the spatial geometric distance between each zone; in the cloud node, calculating the correlation coefficient of the weight parameters in each network layer of the fluid dynamics prediction model; zeroing out weight parameters that do not reach the preset pruning ratio, removing non-isomorphic weight connections that do not have a topological mapping relationship with the physical cascaded topology structure of the ventilation ducts; and sparsely reconstructing the pruned model weight matrix to obtain a sparse impedance discrimination matrix, making the coordinate space distribution of the non-zero elements in the sparse impedance discrimination matrix equivalent to the physical topological connectivity state of the correlation matrix.

[0012] Preferably, the flow field prediction vector is obtained by using the virtual pressure difference traveling wave vector as node features and the sparse impedance discrimination matrix as edge weight features, and extracting features from the spatiotemporal graph attention network at the edge computing node. This includes: configuring a spatiotemporal graph attention network containing a temporal convolutional layer and a graph attention layer in the edge computing node; inputting the virtual pressure difference traveling wave vector as the input node feature matrix into the temporal convolutional layer for temporal dimension feature compression, and outputting a temporally compressed feature vector; using the correlation matrix as the topological adjacency matrix of the graph structure to define the edge connectivity boundary of the multi-partition graph network; inputting the temporally compressed feature vector, the topological adjacency matrix, and the sparse impedance discrimination matrix together into the graph attention layer, and outputting a flow field prediction vector containing the airflow change trend and gas velocity distribution of each partition within a preset time window through adaptive weighted calculation.

[0013] Preferably, a divergence-constrained rigid projection layer based on the graph Laplace operator is constructed at the output end of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix. This includes: calculating the corresponding graph Laplace matrix based on the correlation matrix to characterize the discrete operator structure of airflow between multiple zones; extracting the spatial volume parameters and fluid isothermal compressibility of each zone in the laboratory, and calculating the transient pressure-capacity buffer source term vector for the graph vertices in combination with the prediction time step; transforming the mass conservation boundary conditions in the fluid dynamics continuity equation into a non-dispersion numerical constraint equation containing the transient pressure-capacity buffer source term vector; using the graph Laplace matrix to perform discretization algebraic reconstruction of the non-dispersion numerical constraint equation, transforming the transient airflow dynamic balance law of multiple zones into a non-homogeneous linear equation system; and extracting the general solution structure matrix of the non-homogeneous linear equation system as the solution space matrix satisfying the transient tolerance physical constraints.

[0014] Preferably, performing an orthogonal projection transformation on the flow field prediction vector to the solution space matrix to map the target control vector in the digital space includes: performing singular value decomposition and orthogonalization on the solution space matrix to calculate the orthogonal projection operator parameters of the solution space matrix; performing matrix projection transformation on the flow field prediction vector and the orthogonal projection operator parameters to generate multi-zone airflow manifold projection vectors; trimming and filtering out the non-dispersion data in the multi-zone airflow manifold projection vectors that do not satisfy the homogeneous linear equations corresponding to the solution space matrix in the digital space to obtain calibration manifold feature vectors; and performing boundary alignment and nonlinear mapping processing on the calibration manifold feature vectors within the algebraic manifold to map the target control vector.

[0015] Preferably, formatting the target control vector and writing a set of digital control vectors to the underlying communication interface of the ventilation actuator includes: decomposing the execution dimension of the target control vector and parsing it into digital execution data pairs corresponding to each partition; encapsulating the digital execution data pairs according to the industrial control protocol frame structure to generate digital control data packets; and synchronously writing the digital control data packets to the ventilation valve servo actuators and fan controller frequency converters of each partition through the industrial Ethernet interface and fieldbus interface of the edge computing node.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts the electrical operation sequence data of ventilation-related equipment and reverse maps it to the quasi-static pressure phase space of the fluid, transforming it into a virtual differential pressure traveling wave vector that characterizes the trend of spatial pressure fluctuation. This enables advanced data-driven prediction of the flow field evolution trend at the initial stage of the feedforward disturbance source's action, shortens the time delay of control command generation, and improves the timing response matching degree of the scheduling data stream when dealing with sudden transient disturbances.

[0017] 2. This invention utilizes a sparse impedance discrimination matrix, which is isomorphic to the cascaded topology of laboratory physical pipelines and is extracted during initialization and background asynchronous updates, as the edge weight feature of the graph network. The spatial coupling fluid dynamics constraints of the multi-zone pipeline network are directly embedded in the graph topology data structure. Local graph attention space aggregation calculation is performed at the edge computing nodes, which reduces the complexity of full matrix solution and the consumption of computing resources when performing multi-zone collaborative prediction at the edge.

[0018] 3. This invention constructs a zero-constraint rigid projection layer based on the graph Laplacian operator at the output of the spatiotemporal graph attention network. It performs an orthogonal projection transformation on the preliminary flow field prediction vector through the solution space matrix of the non-homogeneous linear equation system, filtering out and pruning numerical divergent solutions that do not satisfy the law of mass conservation from the algebraic geometric manifold structure. This ensures that the final output set of digital control vectors strictly conforms to the physical boundary constraints of fluid continuity, thereby improving the system-level stability of the multi-partition distributed collaborative scheduling data flow. Attached Figure Description

[0019] Figure 1 A flowchart of a laboratory multi-zone dynamic air volume scheduling method based on edge computing is provided for an embodiment of the present invention; Figure 2 This is a topology diagram of cloud-edge dual-track asynchronous collaboration and data flow provided in an embodiment of the present invention; Figure 3 The flowchart of data flow between the spatiotemporal graph attention network and the rigid projection layer is provided for embodiments of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1 to 3 This invention provides a method for dynamic scheduling of air volume in multiple zones of a laboratory based on edge computing, the technical solution of which is as follows: A dynamic airflow scheduling method for multiple zones in a laboratory based on edge computing, comprising: The electrical operation sequence data of ventilation-related equipment is collected and phase space reconstruction is performed to construct a quasi-static pressure prediction phase space matrix, and a virtual pressure difference traveling wave vector is generated through mapping operation. During system initialization and background asynchronous updates, the correlation matrix characterizing the laboratory physical topology is obtained, topology correlation pruning is performed on cloud nodes, the fluid dynamics prediction model is refined into a sparse impedance discrimination matrix isomorphic to the correlation matrix, and then distributed to edge computing nodes. Using the virtual pressure difference traveling wave vector as node features, the sparse impedance discrimination matrix as edge weight features, and combined with the graph topology represented by the correlation matrix, feature extraction is performed at the edge computing node input spatiotemporal graph attention network to obtain the flow field prediction vector. A zero-constraint rigid projection layer based on the graph Laplacian operator is constructed at the output end of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix; the flow field prediction vector is orthogonally projected to map the target control vector in the digital space; the target control vector is formatted and written into the underlying communication interface of the ventilation actuator.

[0022] Example 1: This embodiment applies to a multi-zone dynamic airflow scheduling scenario within a biosafety laboratory protected area. This scenario typically includes a chemical shower room, multi-stage airlock rooms, and a main laboratory room sharing a main exhaust duct network. During operation, frequent opening and closing of airtight doors generates extreme transient aerodynamic impacts from physical displacement and the high-pressure spray of the chemical shower. Due to the strong hydrodynamic coupling between different zones within the same duct network, if the exhaust valve of a certain zone adjusts its opening in response to local pressure changes, the static pressure in the main duct will experience a chain reaction, triggering a sudden drop in exhaust volume in adjacent zones, creating a "stealing air" effect.

[0023] As one embodiment of the present invention, refer to Figure 1 A flowchart of a dynamic airflow scheduling method for multi-zone laboratory spaces based on edge computing, referencing... Figure 2Cloud-edge dual-track asynchronous collaboration and data flow topology diagram.

[0024] Furthermore, electrical operation sequence data of ventilation-related equipment is collected, including: acquiring three-phase stator current time-series data and output frequency change rate time-series data of the exhaust fan inverter; acquiring electromagnetic torque pulse sequence data of the airtight door drive motor; aligning the three-phase stator current time-series data, the output frequency change rate time-series data, and the electromagnetic torque pulse sequence data with time steps and tensor splicing according to a preset time sampling step size, and extracting the time sequence with a set sliding time window to generate electrical operation sequence data.

[0025] Specifically, the edge computing nodes establish high-frequency communication with the main exhaust fan inverter and the airtight door servo drive via a fieldbus. The system continuously reads the three-phase stator current timing data collected by the inverter's internal current transformer according to a preset timing sampling step size (the timing sampling step size is adaptively set based on the fieldbus communication bandwidth and the node's hardware interrupt handling capability; a recommended value range is 1 to 50 milliseconds, preferably 10 milliseconds). This data is stored as a continuous one-dimensional floating-point array to characterize transient changes in motor load. Simultaneously, the derivative of the control board's output frequency is acquired to extract the output frequency change rate timing data. For the airtight door, the system reads the servo drive's torque register to extract the electromagnetic torque pulse sequence data at the moment the drive motor overcomes static friction during startup. To eliminate asynchronous interference from different communication protocols, the nodes establish a unified timestamp alignment framework in memory. The current, frequency change rate, and torque pulse data are linearly interpolated and resampled according to the time sequence sampling step size. After tensor splicing in the spatial dimension, the sequence is dynamically reassembled with a set sliding time window (such as the time sequence interception interval of 100 milliseconds before the current moment), and the electrical operation sequence data in a unified format is output into the circular buffer.

[0026] This invention bypasses traditional differential pressure sensors, which suffer from mechanical lag, and instead acquires electrical operating state parameters at the speed of light in the initial stage of physical disturbances, thereby obtaining prior features characterizing mechanical displacement and fluid fluctuations. By employing a unified timestamp for interpolation and stitching, it eliminates timing misalignments caused by asynchronous hardware clocks in heterogeneous underlying devices, improving the fidelity of multidimensional input data and providing a stable and sensitive data source for cross-media extrapolation.

[0027] Further, a quasi-static pressure prediction phase space matrix is ​​constructed, and a mapping operation is performed on the quasi-static pressure prediction phase space matrix to generate a virtual pressure difference traveling wave vector. This includes: performing phase space reconstruction processing on electrical operation sequence data to construct a quasi-static pressure prediction phase space matrix characterizing the nonlinear changes in fluid pressure state; extracting the pipeline reference operating state vector for the current scheduling cycle and calling a cross-media nonlinear mapping function with embedded aerodynamic boundary conditions; using the pipeline reference operating state vector as modulation constraints, performing nonlinear adaptive modulation on the static electrical propagation constant and fluid dynamic constant to calculate the state-dependent dynamic transfer coefficient tensor; using the dynamic transfer coefficient tensor, performing a mapping operation on the quasi-static pressure prediction phase space matrix to calculate the transfer coefficient matrix between the electrical propagation constant and the fluid dynamic constant; using the transfer coefficient matrix, inversely mapping the electrical evolution characteristics in the quasi-static pressure prediction phase space matrix to the aerodynamic flow field space, and generating a virtual pressure difference traveling wave vector through discrete-time integration.

[0028] Specifically, in order to resolve the physical contradiction between high-frequency continuous sampling data and the limited computing power of edge nodes, and to ensure the real-time performance of dynamic scheduling, the processor first performs sliding window truncation and dynamic downsampling processing on the electrical operation sequence data based on 1 milliseconds, compressing the original high-frequency data into a low-frequency feature sequence that is adapted to the computing power load of the edge.

[0029] Subsequently, to eliminate time-scale misalignment across physical fields and simulate the mechanical inertia filtering effect of the actuator, the edge computing node invokes its internally integrated dedicated digital signal processor or asynchronous hardware acceleration module to run an electromagnetic empirical mode decomposition algorithm matrix on the downsampled low-frequency feature sequence in digital space, decomposing it into a feature matrix containing multiple intrinsic mode function components. The system retrieves the mechanical inertia envelope threshold matrix related to the physical quality of ventilation-related equipment, and performs feature masking operations on the feature matrix using the mechanical inertia envelope threshold matrix; in digital space, transient electromagnetic harmonic components whose duration does not reach the mechanical response dead zone of the actuator are filtered out, and the fundamental envelope sequence data that can effectively overcome rotational inertia and be transformed into the macroscopic displacement trend of the entity is extracted.

[0030] The process of obtaining the mechanical inertial envelope threshold matrix is ​​as follows: During the offline calibration stage of the system, transient electromagnetic step excitations with different pulse widths and amplitudes are injected into the servo motor of the target ventilation equipment; whether the mechanical output end of the equipment generates an actual physical angular displacement is monitored synchronously; the maximum electromagnetic excitation peak value that cannot cause effective physical displacement and its corresponding pulse width time are recorded as mechanical response dead zone parameters, thereby constructing a mechanical inertial envelope threshold matrix specifically for this entity.

[0031] Subsequently, the edge computing node applies a dynamic phase space reconstruction method, using the extracted fundamental envelope sequence data as the input source. The delay time is determined according to a preset mutual information method, such as setting it to 5 sampling periods. In discrete computing, the mutual information function curve reaches its first local minimum at 5 sampling periods (e.g., a 50-millisecond time span in conventional bus communication). The embedding dimension is calculated using the spurious nearest neighbor method (e.g., set to 4 dimensions, this value is obtained offline from the spurious nearest neighbor method). Experimental tests show that when the embedding dimension gradually increases from 1 to 4 dimensions, the proportion of spurious nearest neighbors in the one-dimensional electrical wave sequence directly converges to below 1%, and the dynamic topological trajectory is fully unfolded in geometric space without overlap. Further increasing the dimension not only fails to bring more explicit dynamic features but also exponentially increases the phase space dot product operation complexity of the edge computing gateway. The one-dimensional sequence data is mapped into a multi-dimensional topological trajectory matrix, constructing a quasi-static pressure prediction phase space matrix. This matrix describes the geometric characteristics of the system's evolution from an effective disturbance to a non-steady state.

[0032] Subsequently, the processor extracts the pipeline reference operating state vector for the current scheduling cycle. The pipeline reference operating state vector includes the initial opening degree code of each zone's air valve, the reference frequency data of the main exhaust fan, and the calculated value of the background flow velocity characterizing the Reynolds number of the flow field.

[0033] Based on a pre-established nonlinear regression model (e.g., a fully connected neural network with three hidden layers) fitted with experimental data and using the pipeline network's baseline operating state vector as input variables, the model's weight parameters encode the empirical correlation between the electrical system's time constant and the fluid pipeline network's impedance coefficient. The processor calls this model as a cross-medium nonlinear mapping function to construct the Jacobian transformation matrix. The specific process is as follows: First, based on the empirical laws of classical fluid dynamics pipeline network impedance, a mechanistic mathematical function of nonlinear fluid damping and equivalent volume is established, with the initial opening encoding, the main and exhaust fan reference frequencies, and the measured background velocity as independent variables. Second, the first-order partial differential operators are solved for each independent variable in the mechanistic mathematical function, and the obtained partial differential values ​​are arranged according to the variable dimensions to construct the Jacobian transformation matrix reflecting the real-time physical response gradient of the local pipeline.

[0034] Based on this, the system performs nonlinear adaptive modulation operation with the pipeline network reference operating state vector as the modulation constraint condition. The specific logic is as follows: extract the static electrical response time constant on the motor design nameplate and the static aerodynamic time constant pre-calibrated according to the pipeline geometry; use the Jacobian transformation matrix as the dynamic weight coefficient matrix to perform matrix multiplication and division proportional conversion operations on the above static aerodynamic time constant and static electrical response time constant, thereby solving the state-dependent dynamic transfer coefficient tensor that is strictly constrained by the current pipeline network physical boundary.

[0035] Using the state-dependent dynamic transfer coefficient tensor, a discrete-time integral transformation is performed on the quasi-static pressure prediction phase space matrix to deduce a virtual pressure difference traveling wave vector that precedes the actual physical airflow changes in the digital space, describing the transient drop trajectory of static pressure in each zone within a future time period under the current pipeline boundary constraints.

[0036] This invention introduces electromagnetic empirical mode decomposition and feature extraction steps of the mechanical inertial envelope threshold matrix between electrical sequence data and phase space reconstruction, thereby equivalently simulating the inertial filtering effect of a physical mechanical rotor in digital space. Under steady-state reference load and conventional start-up and shutdown physical boundary conditions, the method successfully converged and aligned cross-scale temporal features, as verified by sensor-measured waveform alignment, ensuring the engineering accuracy of the causal mapping between electrical and fluid systems.

[0037] Further, the correlation matrix is ​​obtained, and topology correlation pruning is performed on the cloud node to refine the fluid dynamics prediction model into a sparse impedance discrimination matrix isomorphic to the correlation matrix. This includes: during the system initialization phase and in the background asynchronous update cycle independent of the air volume dynamic scheduling cycle, reading the physical cascaded topology structure composed of multiple laboratory zones, multi-level airlocks, and ventilation ducts, and constructing a correlation matrix characterizing the connectivity state of each ventilation node and the spatial geometric distance between each zone; in the cloud node, calculating the correlation coefficient of the weight parameters in each network layer of the fluid dynamics prediction model; zeroing out weight parameters that do not reach the preset pruning ratio, removing non-isomorphic weight connections that do not have a topological mapping relationship with the physical cascaded topology structure of the ventilation ducts; and sparsely reconstructing the pruned model weight matrix to obtain a sparse impedance discrimination matrix, making the coordinate space distribution of the non-zero elements in the sparse impedance discrimination matrix equivalent to the physical topological connectivity state of the correlation matrix.

[0038] Specifically, during the asynchronous computing cycle of laboratory debugging and subsequent operation and maintenance, the cloud server retrieves the digital twin building model and resolves the cascaded topology, which includes, for example, five independent experimental zones, three levels of airlocks, and twelve main and secondary air ducts. A symmetric matrix of correlation degrees is generated based on spatial connectivity and pipe impedance characteristics. A large neural network for fluid prediction runs in the cloud, and the system calculates the correlation coefficients between the weight parameters of the hidden layer neurons in the model and the geometric connections of the physical pipes. A fixed network pruning threshold of 70% was set (this value is based on the convergence result of multiple rounds of offline pruning tests on a large-scale fluid dynamics deep learning model in the cloud. Experimental data shows that due to the inherent sparsity of the geometric cascading relationship between laboratory physical pipes and partitions, a large number of weight parameters in the model correspond to "non-isomorphic connections" that do not physically exchange airflow. At a pruning ratio of 70%, the spatial distribution of non-zero elements in the model weight matrix can achieve the greatest possible topological equivalence with the physical correlation matrix of the physical pipe network). The system executes a pruning algorithm to forcibly reset non-isomorphic network weight data with correlation below this threshold and no substantial fluid exchange path in physical space to 0. After multiple rounds of iterative compression, a sparse impedance discrimination matrix with an element distribution completely consistent with the actual pipe network connectivity is output. This matrix extracts key resistance information along the fluid transmission path and distributes it to the edge computing gateway corresponding to each partition through a dedicated encryption protocol.

[0039] This invention utilizes a correlation matrix equivalent to a real pipeline network to perform targeted weight pruning in the cloud, refining the computationally intensive global black-box model into a lightweight sparse matrix with physical space interpretability. This step significantly reduces unnecessary computational redundancy, lowers the storage and memory requirements after the model is deployed, and is well-suited to the compact computing resources of edge gateways. Combined with the timing design of asynchronous updates in the background, it ensures that the model reconstruction and update process does not block the local high-frequency closed-loop scheduling thread, improving the execution efficiency of the system's edge inference.

[0040] During the system initialization phase and in the background asynchronous update cycle independent of the airflow dynamic scheduling cycle, the step of performing topology correlation pruning on the cloud node to generate a sparse impedance discrimination matrix and then distributing it further includes: pre-calculating the structural resistance deviation of the airtight door in the open, closed, and half-open states on the cloud node, generating a transient impedance perturbation tensor library, and distributing it to the edge computing node; the method further includes: when a jump in the airtight door state signal is detected, the edge computing node extracts a matching transient impedance perturbation tensor from the transient impedance perturbation tensor library; performing tensor addition operation on the transient impedance perturbation tensor and the sparse impedance discrimination matrix to dynamically update the edge weight features of the graph network.

[0041] Specifically, while the cloud server performs full model pruning, the system offline analyzes the changes in the cross-sectional area of ​​the airtight door in three discrete physical states: open, closed, and half-open. The cloud calculates the corresponding structural drag deviations for each of these three states, generating a transient impedance perturbation tensor library containing multiple drag change tensors. This library, along with the sparse impedance discrimination matrix, is then sent to the memory of the edge computing gateway in each partition. During the dynamic scheduling cycle, the edge computing nodes continuously read the level signals from the airtight door position sensors. When a state transition occurs in this signal, the edge computing node does not initiate a network communication request to the cloud for re-pruning; instead, it directly triggers local memory data retrieval, retrieving the transient impedance perturbation tensor corresponding to the current physical state. The processor calls a matrix addition instruction to add the original sparse impedance discrimination matrix to the extracted impedance perturbation tensor. Through discrete tensor addition operations at the local memory level, the system reconstructs the edge weight parameters of the graph network within a very short computation cycle, achieving digital spatial synchronization of abrupt changes in the connectivity of the physical pipeline.

[0042] This invention pre-constructs a transient impedance perturbation tensor library containing different physical states in the cloud, and directly performs local tensor addition operations when the edge computing node detects a state signal jump. This avoids the communication network time overhead caused by re-requesting the cloud for full network model pruning and parameter distribution when the physical space topology changes. This technique enables the graph network model of the edge node to follow the deformation of the physical geometric space in real time with extremely short digital cycles, reducing the computational latency during topology state updates and improving the data processing consistency and collaborative scheduling accuracy of the fluid dynamics prediction model in dealing with frequent access control opening and closing conditions.

[0043] Furthermore, referring to Figure 3 The data flow diagram of the spatiotemporal graph attention network and the rigid projection layer is as follows: using the virtual pressure difference traveling wave vector as node features and the sparse impedance discrimination matrix as edge weight features, the spatiotemporal graph attention network is input at the edge computing node for feature extraction to obtain the flow field prediction vector. This includes: configuring a spatiotemporal graph attention network containing a temporal convolutional layer and a graph attention layer in the edge computing node; inputting the virtual pressure difference traveling wave vector as the input node feature matrix into the temporal convolutional layer for temporal dimension feature compression, outputting a temporally compressed feature vector; using the correlation matrix as the topological adjacency matrix of the graph structure to define the edge connectivity boundary of the multi-partition graph network; inputting the temporally compressed feature vector, the topological adjacency matrix, and the sparse impedance discrimination matrix together into the graph attention layer, and outputting a flow field prediction vector containing the airflow change trend and gas velocity distribution of each partition within a preset time window through adaptive weighted calculation.

[0044] The specific architecture of the spatiotemporal graph attention network is as follows: After receiving the virtual pressure difference traveling wave vector, the input layer first connects to a 1D temporal causal convolutional layer (containing 3 residual blocks, with dilation coefficients set to 1, 2, and 4 respectively, and a convolutional kernel size of 3) to extract temporal local dependency features; the output sequence of this temporal convolutional layer is nonlinearly mapped by the ReLU activation function and then used as node features input to a multi-head graph attention layer (the number of attention heads is set to 4, and the dimension of the single-head hidden layer is 64). At the same time, the sparse impedance discrimination matrix is ​​used as the edge weight input to perform multi-topological spatial feature aggregation; finally, a flow field prediction vector with a prediction time window of 500 milliseconds is output through a fully connected layer.

[0045] In the HVAC architecture of a high-level biosafety laboratory, the total physical, mechanical, and fluid dynamic response time of the entire chain, from the issuance of digital control commands by edge computing nodes to the action of servo motors of venturi valves overcoming static friction, the redistribution of airflow organization in the pipeline, and finally the establishment of a stable negative pressure ladder at the high-precision differential pressure sensor, is typically between 350ms and 450ms.

[0046] To enable the distributed collaborative scheduling strategy to possess "forward-looking defense," the prediction step size of the spatiotemporal graph attention network must be slightly larger than and perfectly cover this end-to-end physical lag time. Therefore, setting the prediction time window to 500 milliseconds ensures that the valve opening command calculated within the current 10-millisecond scheduling cycle can reserve a digital-dimensional physical compensation margin for the "preemptive pressure wave" that will not actually be transmitted to adjacent partitions until 500 milliseconds later.

[0047] Before system deployment, the network needs to be trained offline. Historical electrical operation sequences and corresponding real-time measurement data from high-frequency differential pressure sensors from multiple high-level laboratories under conditions such as sudden opening / closing of airtight doors and sudden changes in exhaust ventilation are collected as training sets. A joint loss function is constructed, including the mean square error between predicted and actual measured airflow, as well as a penalty term for the amount of airflow interception by adjacent nodes.

[0048] The constructed joint loss function L is specifically as follows: ; in, This represents the predicted wind volume matrix output by the network (N is the number of partitions, and T is the prediction time step). This represents the measured airflow matrix from the high-frequency differential pressure sensor, and... Same dimension, This represents the slope of the predicted air volume change (m³ / s) for partition i within the prediction window. 2 E represents the set of neighboring edges corresponding to the non-zero elements of the sparse impedance discrimination matrix, and |E| is the total number of edges (normalization avoids the influence of the number of partitions). This represents the permissible threshold for cooperative variation difference between adjacent partitions (recommended to be calibrated in units of slope, such as 0.05 m³ / s). 2 ), λ is the balance coefficient (recommended initial value 0.1, annealed to 0.05 with each training round), max(0,·) is the Hinge structure, which only penalizes differences exceeding the threshold δ, and the symmetrical form covers the bidirectional wind-stealing scenario.

[0049] Backpropagation is performed using the Adam optimizer (with an initial learning rate of 0.001) to update the network weights until the model's flow field prediction slope error on the independent validation set is below the set threshold for 50 consecutive epochs, at which point the model is solidified.

[0050] Specifically, edge nodes deploy a spatiotemporal graph attention network within their computing units, consisting of a temporal convolution module and a spatial graph attention module connected in series. When a dynamic disturbance is detected, the node adapts the acquired virtual pressure difference traveling wave vector to the input temporal convolutional layer, uses a dilated convolution kernel with a causal structure to perform dimensionality reduction and compression of the wave sequence features within 100 milliseconds along the time dimension, extracts the temporal dependency correlation of the disturbance, and outputs a temporal compressed feature vector.

[0051] (Based on offline fluid dynamics simulations and on-site physical calibration, the core high-frequency energy clusters of the electromagnetic-aerodynamic multi-scale fluctuations in an airtight door, from the moment the door lock is released to the first wave of transient air impact caused by air leakage, are typically concentrated within a dynamic window of 80ms to 120ms. If the sliding time window is set too short (e.g., less than 50ms), the temporal causal convolutional layer cannot capture the complete temporal waveform of the "disturbance intent," leading to distorted feedforward predictions. If it is set too long (e.g., greater than 200ms), it will introduce a large amount of steady-state redundant data unrelated to the current impact, wasting extremely valuable computational resources at the edge and diluting the feature weights of transient disturbance characteristics. Therefore, 100ms is preferred as the truncation and compression boundary of the time dimension.)

[0052] In the topology calculation phase, the network directly reads the correlation matrix as the graph's topological adjacency matrix, using digital logic to define which two laboratory zones have direct airflow crosstalk boundaries; simultaneously, it reads the sparse impedance discrimination matrix as edge weight parameters. The graph attention layer, based on a multi-head attention mechanism, uses temporally compressed feature vectors as node information and calculates the interference influence coefficients of adjacent zone features on local nodes along the boundaries of the adjacency matrix. After aggregation, it outputs a multi-zone collaborative flow field prediction vector, representing the slope of airflow change in each space within the next 500 milliseconds (its physical dimension is m). 3 / s 2 ).

[0053] This invention establishes fluid interference boundaries by specifying a correlation matrix and configures a sparse matrix to represent pipe resistance coefficients, transforming the complex, strongly coupled fluid dynamics problem in high-level laboratories into a standardized graph node information aggregation task. This feature extraction method enables individual sub-partitions to adaptively quantify and absorb the fluctuation intentions of adjacent sub-partitions and the main air duct when generating scheduling schemes, significantly improving the overall system's lateral anti-wind-stealing coordination performance when handling concurrent disturbances.

[0054] The temporal compression feature vector, topological adjacency matrix, and sparse impedance discrimination matrix are jointly input into the graph attention layer. The flow field prediction vector is output through adaptive weighted calculation, including: determining whether the slope of the flow field prediction vector output by the edge computing node exceeds a preset slope threshold; when the slope of the flow field prediction vector exceeds the preset slope threshold, the flow field prediction vector is converted into an action intent penalty coefficient, and the action intent penalty coefficient is sent to the edge computing nodes of adjacent partitions under the same main exhaust duct network through a multicast communication protocol; the edge computing nodes of the adjacent partitions use the received action intent penalty coefficient as an external bias and inject it into the weight normalization calculation step of the local graph attention layer to generate a flow field prediction vector that includes air volume compensation margin.

[0055] Specifically, when the spatiotemporal graph attention network performs cross-region feature aggregation, the edge computing node monitors the slope value of its output flow field prediction vector in real time. When the slope value exceeds the preset slope threshold defined by safety specifications, the node determines that a pumping behavior exceeding the normal load is about to occur in this region. Before the physical action mapped to this behavior is executed, the edge computing node extracts the core parameters of the prediction vector and transforms them into a non-negative action intent penalty coefficient through a mapping function.

[0056] When the edge computing node detects that the local zone airflow prediction increment ΔQ in the locally output flow field prediction vector exceeds the preset slope threshold... At that time, the node generates a scalar form of the action intent penalty coefficient according to the following steps. And broadcast to neighboring nodes: ; in The function is the Sigmoid function, k is the gain coefficient (calibrated according to the pipeline network scale during engineering deployment), and ΔQ is the scalar of the air volume change amplitude of the flow field prediction vector in this partition within the current scheduling cycle (normalized after taking the L2 norm of the multidimensional prediction vector). To preset the slope threshold, offline calibration was performed during the system commissioning phase as follows: Using 10% to 15% of the rated airflow of the main exhaust fan as an initial reference value, the static pressure drop in the pipeline was measured under the condition of an emergency opening of the airtight door. If the static pressure drop exceeded 50% of the design negative pressure gradient of adjacent sections, the threshold was lowered. The final value was determined by iterating until the pipeline static pressure stabilized. In this embodiment, for a typical pipeline network containing 5 zones and 12 ducts, the calibration value was 0.08m. 3 / s 2 (Based on the slope of airflow change within a 500-millisecond prediction window).

[0057] Edge computing nodes in adjacent partitions receive action intent penalty coefficients Then, this is injected as an external bias into the attention score normalization step of the local graph attention layer. Let the original attention score calculated for node i with respect to its neighboring node j be... The corrected normalized attention weights are: ; ; in The directional mask corresponding to the broadcast source node j ( (This indicates that the penalty comes from node j, and the rest are 0). Let i be the set of neighbors of node i. This operation applies a negative bias to the attention score of the broadcast source node before softmax normalization, reducing its aggregation weight, thereby prompting neighboring nodes to pre-calculate and reserve the corresponding wind volume compensation margin, achieving feedforward suppression of the wind-stealing effect at the algebraic level.

[0058] in, Physical meaning: This indicates no penalty (predicted airflow is normal). This indicates that a significant pumping action is about to occur in the partition, and adjacent nodes should minimize the aggregation weights on their features and reserve compensation margins. The Sigmoid mapping smoothly compresses the continuous change of ΔQ to the [0,1) interval, avoiding weight collapse caused by exponential forms and ensuring the stability of normalized values.

[0059] The node invokes the multicast communication protocol of Industrial Ethernet to package and send the action intent penalty coefficient to the adjacent partition edge computing nodes associated with the same main exhaust duct network. Upon receiving this data packet, the graph attention calculation module of the adjacent nodes does not change the original node feature input. Instead, it uses the action intent penalty coefficient as an external bias and directly injects it into the weight normalization step of the graph attention calculation process. This step adjusts the attention weight allocation of adjacent partitions at the digital algebra level, enabling the graph network prediction output of adjacent nodes to pre-calculate and generate the corresponding airflow compensation margin before the occurrence of physical duct network fluid dynamic phenomena.

[0060] This invention introduces a digital multicast mechanism with an action intent penalty coefficient to convert local predicted fluctuations into external biases in the network layers of adjacent nodes before mechanical displacement occurs in the physical actuator. This technique leverages the data transmission speed advantage of communication networks to overcome the propagation delay of aerodynamic pressure waves in physical pipelines, allowing the graph attention models of adjacent nodes to reserve digital dimension compensation margins during the attention weight normalization stage. This mechanism breaks the physical chain of fluid disturbances in the pipeline network, enhancing the spatial collaborative anti-interference capability of multi-zone control systems when dealing with concurrent interference.

[0061] Furthermore, a divergence-constrained rigid projection layer based on the graph Laplace operator is constructed at the output end of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix. This includes: calculating the corresponding graph Laplace matrix based on the correlation matrix to characterize the discrete operator structure of airflow between multiple zones; extracting the spatial volume parameters and isothermal compressibility of the fluid in each zone of the laboratory, and calculating the transient pressure-capacity buffer source term vector for the graph vertices in combination with the prediction time step; transforming the mass conservation boundary conditions in the fluid dynamics continuity equation into a non-dispersion numerical constraint equation containing the transient pressure-capacity buffer source term vector; using the graph Laplace matrix to perform discretization algebraic reconstruction of the non-dispersion numerical constraint equation, transforming the transient airflow dynamic balance law of multiple zones into a non-homogeneous linear equation system; and extracting the general solution structure matrix of the non-homogeneous linear equation system as the solution space matrix that satisfies the transient tolerance physical constraints.

[0062] Specifically, to ensure that pure data computation follows objective fluid laws and accurately matches transient aerodynamic characteristics, the system embeds algebraic constraint mapping logic at the network endpoints. Edge computing nodes read the correlation matrix, calculate the corresponding vertex degree matrix, and use the difference between the two to generate a graph Laplace matrix representing the multi-partition discrete flow topology.

[0063] To address the volumetric effects under transient disturbances in high-level biosafety laboratories, the system no longer employs the rigid assumption of incompressible dispersion. The processor extracts the spatial volume parameters of each independent zone of the laboratory from the building information model, and, combined with the isothermal compressibility of air and the system's prediction time step, calculates and obtains the transient pressure-capacity buffer source term vector characterizing the room's gas storage buffer capacity.

[0064] Based on the transient mass conservation equation in HVAC physics, the system maps the continuous equation to the fluid divergence operation of the discrete graph vertices, establishing a non-zero divergence numerical constraint equation with the transient pressure-capacity buffer source term vector as the right-hand side of the equation. Combining this constraint equation with the aforementioned graph Laplace operator, matrix multiplication is used to reconstruct the equation, transforming the multivariable differential conservation conditions into a non-homogeneous linear equation system with multi-zone airflow prediction increments as unknowns. The general solution structure matrix of this non-homogeneous linear equation system is extracted as the solution space matrix containing transient physical tolerance stiffness, and stored in memory for subsequent boundary verification.

[0065] This invention introduces spatial volume parameters and fluid isothermal compressibility to calculate the transient pressure-capacity buffer source term vector, upgrading the steady-state dispersion constraint to a transient tolerance non-dispersion constraint that includes volume effects. This mechanism reproduces the air-capacity buffering capacity of a room at the algebraic constraint layer, correcting the idealization deviation of fluid incompressibility. Under transient shocks such as the opening and closing of airtight doors, this constraint allows for legitimate transient inflow-outflow differences at graph vertices while maintaining macroscopic mass conservation, avoiding overly aggressive control commands and flow field oscillations, and improving the physical rationality of the multi-zone dynamic scheduling model under transient conditions.

[0066] Further, the flow field prediction vector is subjected to an orthogonal projection transformation to the solution space matrix to map the target control vector in the digital space. This includes: performing singular value decomposition and orthogonalization on the solution space matrix to calculate the orthogonal projection operator parameters of the solution space matrix; performing matrix projection transformation on the flow field prediction vector and the orthogonal projection operator parameters to generate multi-zone airflow manifold projection vectors; trimming and filtering out the non-dispersion data in the multi-zone airflow manifold projection vectors that do not satisfy the homogeneous linear equations corresponding to the solution space matrix in the digital space to obtain calibration manifold feature vectors; and performing boundary alignment and nonlinear mapping processing on the calibration manifold feature vectors within the algebraic manifold to map the target control vector.

[0067] Specifically, firstly, the matrix factorization algorithm module of the edge computing node reads the solution space matrix of the homogeneous linear equation system and performs singular value decomposition and Schmitt orthogonalization on it to obtain the corresponding null basis, thereby extracting the orthogonal projection operator parameters that can force any high-dimensional features to be compressed onto a legal physical plane.

[0068] Subsequently, the first stage of vector evolution is entered: the scheduling loop program executes matrix multiplication, and performs orthogonal projection transformation on the original flow field prediction vector output by the spatiotemporal graph attention network and the orthogonal projection operator parameters, filtering out the algebraic components in the vector that violate the plane constraints of the fluid macroscopic dynamics, and generating multi-zone airflow manifold projection vectors on the algebraic manifold space.

[0069] Next, the second stage of vector evolution begins: the processor initiates a boundary value verification and retrieval program on the multi-zone airflow manifold projection vector, and performs zero-reduction pruning or forcibly rewrites transient numerical divergence data in the multi-zone airflow manifold projection vector that exceeds the limit of single airflow adjustment and causes the balance of the dispersion numerical constraint equation to be destroyed, thereby converging the data within the safe pressure manifold of the laboratory enclosure structure and solving to obtain the calibration manifold feature vector.

[0070] Finally, the system enters the third stage of vector evolution: the system performs dimensional alignment of the calibrated manifold feature vector within the manifold subspace that satisfies the conservation law, and calls a nonlinear activation function to perform a smooth transformation, mapping it into a target control vector that conforms to the electrical characteristics of each partition actuator.

[0071] This invention utilizes orthogonal projection transformation to forcibly filter eigenvalues ​​that do not satisfy fluid dynamic continuity, effectively preventing data divergence illusions that may occur in neural networks when dealing with extreme and sudden disturbances. Data pruning and boundary verification operations limit the incremental anomaly control to within the safe pressure-bearing range of the laboratory enclosure structure. This rigorous pure algebraic projection transformation logic completely eliminates the potential for out-of-bounds bias in command flow, ensuring the absolute rigidity and conservation of multi-zone airflow distribution commands in both mathematical and physical dimensions, and enhancing the underlying safety control of the system operation.

[0072] Further, formatting the target control vector and writing a set of digital control vectors to the underlying communication interface of the ventilation actuator includes: decomposing the execution dimension of the target control vector and parsing it into digital execution data pairs corresponding to each partition; encapsulating the digital execution data pairs into digital control data packets according to the industrial control protocol frame structure; and synchronously writing the digital control data packets to the ventilation valve servo actuators and fan controller frequency converters of each partition through the industrial Ethernet interface and fieldbus interface of the edge computing node.

[0073] Specifically, the edge computing node enters the communication distribution process, invoking the dimensionality reduction slicing program. The continuous high-dimensional target control vector is decomposed into digital execution data pairs specifically corresponding to the specific electromechanical entities, based on the hardware mapping table of the laboratory equipment. For example, it extracts the 16-bit integer value used to control the opening of the Venturi valve actuator for the airlock exhaust, and the 0Hz to 50Hz floating-point frequency value for setting the target speed of the main exhaust fan driver. Then, it enters the protocol stack packet encapsulation stage. Following the standard fieldbus frame format, a frame header containing the start address and node code is appended to the front of the data pair, and a cyclic redundancy check code obtained through cyclic division is added to the back end, generating a consistent digital control data packet. Finally, the node calls its Ethernet interface chip and underlying communication peripherals to initiate an asynchronous non-blocking concurrent writing program, broadcasting multiple data packets in parallel to the valve and fan registers of each partition, completing the distribution and driving of digital scheduling instructions to the underlying execution end.

[0074] This invention establishes a reliable information transmission bridge between the continuous multidimensional feature array of the algorithm model and the independent and heterogeneous underlying electromechanical actuators through data decomposition and control frame repackaging. A cyclic redundancy check mechanism is embedded in the message to resist signal pollution of the data link caused by strong electromagnetic pulses commonly found in complex industrial equipment environments, ensuring the integrity of command transmission. The concurrent write communication mode ensures the timing consistency of action command acquisition by ventilation actuators in each zone, avoiding the serial response time difference caused by queuing, and consolidating the collaborative effect of dynamic decoupling.

[0075] Example 2: This embodiment describes the core computation and data streaming conversion process of the edge computing-based laboratory multi-zone dynamic air volume scheduling method within the digital space.

[0076] As one embodiment of the present invention, refer to Figure 1 A flowchart of a dynamic airflow scheduling method for multi-zone laboratory spaces based on edge computing, referencing... Figure 2 Cloud-edge dual-track asynchronous collaboration and data flow topology diagram, refer to Figure 3 Flowchart of data flow between spatiotemporal graph attention network and rigid projection layer.

[0077] After system startup, edge computing nodes continuously acquire electrical operation sequence data of ventilation-related equipment via a data acquisition bus. To eliminate time-scale misalignment across physical fields and simulate the mechanical inertia filtering effect of actuators, the processor performs electromagnetic empirical mode decomposition and mechanical inertia masking on the electrical operation sequence data in digital space. This decomposes the data into a feature matrix containing multiple intrinsic mode function components. A preset mechanical inertia envelope threshold matrix is ​​then used to filter out transient electromagnetic harmonic components, extracting the fundamental wave envelope sequence data. Subsequently, the system performs dynamic phase space reconstruction processing, transforming the one-dimensional fundamental wave envelope sequence data into a multi-dimensional topological trajectory matrix, thus constructing the quasi-static pressure prediction phase space matrix.

[0078] The system invokes a pre-defined cross-media nonlinear mapping function with embedded aerodynamic boundary conditions at the edge side. In this calculation step, the processor extracts the pipeline reference operating state vector for the current scheduling cycle and uses it as a partial derivative variable input to construct a Jacobian transformation matrix reflecting the current fluid compression characteristics and the real-time damping characteristics of the pipeline. Using the pipeline reference operating state vector as a modulation constraint, the system performs nonlinear adaptive modulation operations on the static electrical propagation constant and the hydrodynamic constant to calculate the state-dependent dynamic transfer coefficient tensor, which is strictly constrained by the current pipeline physical boundary. Using the dynamic transfer coefficient tensor, the processor performs mapping operations and discrete-time integration on the quasi-static pressure prediction phase space matrix, mapping the electrical evolution characteristics inversely to the aerodynamic flow field space without delay, generating a multi-dimensional virtual pressure difference traveling wave vector.

[0079] During system initialization and asynchronous background updates, cloud nodes retrieve the physical cascaded topology to construct a correlation matrix representing the connectivity state. The cloud nodes calculate the correlation coefficients of the fluid prediction model parameters and run a pruning algorithm to extract a sparse impedance discrimination matrix isomorphic to the correlation matrix, which is then distributed. Simultaneously, the resistance deviations of the airtight door under different states are pre-calculated, generating a transient impedance perturbation tensor library, which is also distributed synchronously. During dynamic scheduling, when an edge computing node detects a jump in the airtight door state signal, it directly retrieves the matching transient impedance perturbation tensor from its local machine and performs tensor addition with the pre-stored sparse impedance discrimination matrix, enabling dynamic updates of the graph network edge weights.

[0080] The edge computing node is configured with a spatiotemporal graph attention network including a temporal convolutional layer and a graph attention layer. The processor inputs the virtual pressure difference traveling wave vector into the temporal convolutional layer for temporal feature compression, outputting a temporally compressed feature vector. Subsequently, the correlation matrix is ​​used as the topological adjacency matrix of the graph structure, and the updated sparse impedance discrimination matrix is ​​used as the edge weight parameter, both input to the graph attention layer. In this step, the edge computing gateway executes a multi-agent intent broadcast mechanism. When it detects that the slope of the locally output flow field prediction vector exceeds a preset slope threshold, it converts it into an action intent penalty coefficient and sends it to associated neighboring nodes via a multicast protocol. The neighboring nodes inject this intent penalty coefficient as an external bias into the normalization calculation step of their local attention weights, adaptively outputting a flow field prediction vector that includes airflow compensation margin.

[0081] After obtaining the initial flow field prediction vector, the system constructs a divergence-constrained rigid projection layer based on the graph Laplace operator at the output of the spatiotemporal graph attention network. The processor calculates the graph Laplace matrix using the correlation matrix. To accurately reflect the volume effect of a high-level biosafety laboratory under transient disturbances, the system extracts the spatial volume parameters and fluid isothermal compressibility of each zone in the laboratory, and calculates the transient pressure-capacity buffer source term vector for the graph vertices based on the prediction time step. Accordingly, the system transforms the mass conservation boundary conditions in the fluid dynamics continuity equation into a non-dispersion numerical constraint equation containing the transient pressure-capacity buffer source term vector. The graph Laplace matrix is ​​used to discretize and algebraically reconstruct this non-dispersion numerical constraint equation, transforming the transient airflow dynamic balance law of multiple zones into a non-homogeneous linear equation system, and extracting its general solution structure matrix as the solution space matrix that satisfies the transient tolerance physical constraints.

[0082] Subsequently, the processor performs singular value decomposition and orthogonalization on the solution space matrix to obtain orthogonal projection operator parameters. It then performs matrix projection transformation on the predicted flow field vector and the orthogonal projection operator parameters to generate multi-zone airflow manifold projection vectors. The processor executes a boundary value pruning procedure to filter out scattered divergent data in the manifold projection vectors that do not satisfy the corresponding non-homogeneous linear equation system within the digital space, obtaining calibrated manifold feature vectors. Finally, boundary alignment and nonlinear mapping are performed within the algebraic manifold to generate the target control vector. Edge computing nodes perform dimensionality decomposition on the target control vector and encapsulate it according to industrial control protocols, synchronously and concurrently writing it into the underlying communication interfaces of each ventilation actuator to complete the dynamic distributed collaborative scheduling of multi-zone airflow.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic scheduling of air volume in multiple zones of a laboratory based on edge computing, characterized in that, include: Collect electrical operation sequence data of ventilation-related equipment; A quasi-static pressure prediction phase space matrix is ​​constructed based on the electrical operation sequence data, and a virtual pressure difference traveling wave vector is generated through mapping operations. Specifically, electromagnetic empirical mode decomposition and mechanical inertial masking are performed on the electrical operation sequence data to extract the fundamental envelope sequence data. Phase space reconstruction processing is then performed on the fundamental envelope sequence data to construct the quasi-static pressure prediction phase space matrix. The pipeline reference operating state vector for the current scheduling cycle is extracted, and a cross-media nonlinear mapping function with embedded aerodynamic boundary conditions is invoked. Using the pipeline reference operating state vector as modulation constraints, nonlinear adaptive modulation is performed on the static electrical propagation constant and hydrodynamic constant to calculate the state-dependent dynamic transfer coefficient tensor. The dynamic transfer coefficient tensor is then used to perform mapping operations on the quasi-static pressure prediction phase space matrix to calculate the transfer coefficient matrix between the electrical propagation constant and the hydrodynamic constant. Finally, using the transfer coefficient matrix, the electrical evolution characteristics in the quasi-static pressure prediction phase space matrix are inversely mapped to the aerodynamic flow field space, and a virtual pressure difference traveling wave vector is generated through discrete-time integration. During system initialization and background asynchronous updates, a correlation matrix representing the laboratory's physical topology is obtained. Topology correlation pruning is performed on cloud nodes to refine the fluid dynamics prediction model into a sparse impedance discrimination matrix isomorphic to the correlation matrix, which is then distributed to edge computing nodes. Specifically, during system initialization and in the background asynchronous update cycle independent of the airflow dynamic scheduling cycle, the physical cascaded topology of the laboratory's multi-zone, multi-level airlock, and ventilation ducts is read, and a correlation matrix representing the connectivity state of each ventilation node and the spatial geometric distance between each zone is constructed. On cloud nodes, the correlation coefficients of the weight parameters in each network layer of the fluid dynamics prediction model are calculated. Weight parameters that do not reach the preset pruning ratio are zeroed out, and non-isomorphic weight connections that do not have a topological mapping relationship with the physical cascaded topology of the ventilation ducts are removed. The pruned model weight matrix is ​​sparsified and reconstructed to obtain a sparse impedance discrimination matrix, making the coordinate space distribution of the non-zero elements in the sparse impedance discrimination matrix equivalent to the physical topology connectivity state of the correlation matrix. Using the virtual pressure difference traveling wave vector as node features, the sparse impedance discrimination matrix as edge weight features, and combined with the graph topology represented by the correlation matrix, feature extraction is performed at the edge computing node input spatiotemporal graph attention network to obtain the flow field prediction vector. A zero-constraint rigid projection layer based on the graph Laplacian operator is constructed at the output of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix; the flow field prediction vector is orthogonally projected to map the target control vector in the digital space; the target control vector is formatted and written into the underlying communication interface of the ventilation actuator.

2. The method for dynamic scheduling of air volume in a multi-zone laboratory based on edge computing according to claim 1, characterized in that, The process involves collecting electrical operation sequence data of ventilation-related equipment, including: acquiring the three-phase stator current time-series data and output frequency change rate time-series data of the exhaust fan inverter; acquiring the electromagnetic torque pulse sequence data of the airtight door drive motor; aligning the three-phase stator current time-series data, the output frequency change rate time-series data, and the electromagnetic torque transient pulse sequence data with time steps according to a preset time sampling step size, and performing time-series truncation with a set sliding time window to generate electrical operation sequence data.

3. The method for dynamic scheduling of air volume in a multi-zone laboratory based on edge computing according to claim 1, characterized in that, Using the virtual pressure difference traveling wave vector as node features and the sparse impedance discrimination matrix as edge weight features, a spatiotemporal graph attention network is input into the edge computing node for feature extraction to obtain a flow field prediction vector. This includes: configuring a spatiotemporal graph attention network containing a temporal convolutional layer and a graph attention layer in the edge computing node; inputting the virtual pressure difference traveling wave vector as the input node feature matrix into the temporal convolutional layer for temporal dimension feature compression, outputting a temporally compressed feature vector; using the correlation matrix as the topological adjacency matrix of the graph structure to define the edge connectivity boundaries of the multi-partition graph network; inputting the temporally compressed feature vector, the topological adjacency matrix, and the sparse impedance discrimination matrix together into the graph attention layer, and outputting a flow field prediction vector containing the airflow change trend and gas velocity distribution of each partition within a preset time window through adaptive weighted calculation.

4. The method for dynamic scheduling of air volume in a multi-zone laboratory based on edge computing according to claim 1, characterized in that, A divergence-constrained rigid projection layer based on the graph Laplace operator is constructed at the output end of the spatiotemporal graph attention network to transform the fluid dynamics continuity equation into a solution space matrix. This includes: calculating the corresponding graph Laplace matrix based on the correlation matrix to characterize the discrete operator structure of airflow between multiple zones; extracting the spatial volume parameters and isothermal compressibility of the fluid in each zone of the laboratory, and calculating the transient pressure-capacity buffer source term vector for the graph vertices in combination with the prediction time step; transforming the mass conservation boundary conditions in the fluid dynamics continuity equation into a non-dispersion numerical constraint equation containing the transient pressure-capacity buffer source term vector; using the graph Laplace matrix to perform discretization algebraic reconstruction of the non-dispersion numerical constraint equation, transforming the transient airflow dynamic balance law of multiple zones into a non-homogeneous linear equation system; and extracting the general solution structure matrix of the non-homogeneous linear equation system as the solution space matrix satisfying the transient tolerance physical constraints.

5. The method for dynamic scheduling of air volume in a multi-zone laboratory based on edge computing according to claim 1, characterized in that, The process of performing an orthogonal projection transformation on the flow field prediction vector to the solution space matrix and mapping the target control vector in the digital space includes: performing singular value decomposition and orthogonalization on the solution space matrix to calculate the orthogonal projection operator parameters of the solution space matrix; performing matrix projection transformation on the flow field prediction vector and the orthogonal projection operator parameters to generate multi-zone airflow manifold projection vectors; trimming and filtering out the non-dispersion data in the multi-zone airflow manifold projection vectors that do not satisfy the homogeneous linear equations corresponding to the solution space matrix in the digital space to obtain calibration manifold feature vectors; and performing boundary alignment and nonlinear mapping processing on the calibration manifold feature vectors to map the target control vector.

6. The method for dynamic scheduling of air volume in a multi-zone laboratory based on edge computing according to claim 1, characterized in that, The process of formatting the target control vector and writing a set of digital control vectors to the underlying communication interface of the ventilation actuator includes: decomposing the execution dimension of the target control vector and parsing it into digital execution data pairs corresponding to each partition; encapsulating the digital execution data pairs into digital control data packets according to the industrial control protocol frame structure; and synchronously writing the digital control data packets to the ventilation valve servo actuators and fan controller frequency converters of each partition through the industrial Ethernet interface and fieldbus interface of the edge computing node.

Images