Equipment multi-source asynchronous data alignment method based on space-time soft decoupling twin registration

Abstract

The application discloses an equipment multi-source asynchronous data alignment method based on space-time soft decoupling twin registration, and relates to the field of data processing. The method comprises the following steps: configuring a physical perception network on equipment, and collecting multi-source asynchronous sensing data; constructing a digital twin model of the equipment, and performing quasi-real-time mapping and state synchronization with the multi-source asynchronous sensing data; constructing a space-time soft decoupling coding model, and performing space-time soft decoupling, feature aggregation and twin registration by using the space-time soft decoupling coding model to obtain unified intermediate frequency registration data; feeding back the unified intermediate frequency registration data to the digital twin model for multi-physical field simulation, constructing a joint loss function optimization model parameter, and obtaining a trained space-time soft decoupling coding model; and completing equipment multi-source asynchronous data alignment. The application can improve the acquisition quality and dynamic fidelity of multi-source asynchronous sensing data, and provide high-precision data support for equipment performance prediction and state monitoring.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method for aligning asynchronous multi-source data of equipment based on spatiotemporal soft decoupling twin registration. Background Technology

[0002] In the operational status monitoring of equipment such as high-end CNC machine tools and high-power wind turbines, the significant differences in the evolution mechanisms of different physical properties (such as heat, vibration, pressure, and strain) lead to huge differences in the sampling frequencies of sensors located at different physical positions on the equipment. For example, temperature sensors describing thermal behavior typically sample at lower frequencies, while accelerometers capturing transient dynamic characteristics require extremely high frequencies. This significant frequency difference results in severe asynchronous and heterogeneous characteristics of multi-source sensor data in the time domain, making high-fidelity alignment impossible and thus limiting the subsequent high-precision modeling and physical simulation of digital twins. Existing data-driven alignment methods have the following significant drawbacks: Loss of physical characteristics: Mathematical interpolation that relies solely on low-frequency data (such as temperature completion) is highly susceptible to introducing artificial noise, failing to reflect the true physical mechanism of the gradual evolution or abrupt changes in the physical field, resulting in distortion of strong nonlinear signals.

[0003] Over-matching and noise sensitivity: Methods such as dynamic time warping are too sensitive to local random noise. The alignment process is based entirely on mathematical distance metrics rather than the physical topology of the equipment, which can easily produce alignment results that contradict the actual physical state.

[0004] Lack of closed-loop mechanism constraints: Existing technologies often only perform synchronous conversion at the pure data level, without introducing the physical evolution mechanism of the digital twin model for verification, and cannot guarantee the high fidelity of the registered asynchronous data throughout its entire life cycle service process. Summary of the Invention

[0005] The purpose of this application is to provide a method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration, which can improve the acquisition quality and dynamic fidelity of multi-source asynchronous sensor data, and provide high-precision data support for equipment performance prediction and status monitoring.

[0006] To achieve the above objectives, this application provides a method for aligning asynchronous multi-source data of equipment based on spatiotemporal soft decoupling twin registration, comprising: Configure physical sensing networks on the equipment and collect multi-source asynchronous sensing data; Construct a digital twin model of the equipment and perform near real-time mapping and state synchronization with the multi-source asynchronous sensing data; A spatiotemporal soft decoupling coding model is constructed; the spatiotemporal soft decoupling coding model includes a two-stream recurrent neural network, a physically enhanced graph attention network, and a decoding network; The dual-stream recurrent neural network is used to perform spatiotemporal soft decoupling on the multi-source asynchronous sensing data. The physical augmentation graph attention network is used to aggregate the decoupled features along the physical transmission path to obtain aggregated features. The decoding network is used to perform twin registration on the aggregated features to obtain unified intermediate frequency registration data. The unified intermediate frequency registration data is fed back to the digital twin model for multiphysics simulation. A joint loss function including data reconstruction error and residual of the physical mechanism equation of the digital twin model is constructed. The parameters of the spatiotemporal soft decoupling coding model are optimized to obtain the trained spatiotemporal soft decoupling coding model. The alignment of multi-source asynchronous data for equipment is completed based on the trained spatiotemporal soft decoupling coding model.

[0007] According to the specific embodiments provided in this application, this application has the following technical effects: (1) Using a dual-stream recurrent neural network to perform spatiotemporal soft decoupling on multi-source asynchronous sensing data not only preserves the slow variation trend of the low-frequency temperature field, but also accurately captures the transient nonlinear characteristics of the high-frequency vibration field, effectively avoiding the mathematical oscillations and errors introduced by traditional interpolation.

[0008] (2) By using the physical augmented graph attention network to aggregate the decoupled features along the physical transmission path, the blindness of pure data-driven approaches is eliminated, and the data registration process follows the physical transmission law, making the spatial feature aggregation more scientific.

[0009] (3) The registration data is fed back to the digital twin model for simulation. A joint loss function of the reconstruction error of the fused data and the residual of the physical mechanism equation is constructed. The model parameters are optimized in a closed loop. This forces the unified intermediate frequency registration data to not only approximate the observed value in numerical terms, but also to satisfy the energy conservation and dynamic balance in terms of physical evolution laws. This overcomes the problem of asynchronous data fusion distortion and improves the dynamic fidelity and physical consistency of the data acquired by the equipment under varying working conditions and strong noise environments. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating a method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration, provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a two-stream recurrent neural network. Figure 3 A schematic diagram of the process of a physics-enhanced graph attention network. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] In one exemplary embodiment, such as Figure 1 As shown, a method for aligning asynchronous multi-source data of equipment based on spatiotemporal soft decoupling twin registration is provided. This method is executed by a computer device, specifically by a computer device such as a terminal or a server alone, or by a terminal and a server together. In this embodiment, the method is described using a server as an example, including the following steps S1 to S6.

[0015] S1: Configure a physical sensing network on the equipment and collect multi-source asynchronous sensing data.

[0016] S2: Construct a digital twin model of the equipment and perform near real-time mapping and state synchronization with the multi-source asynchronous sensing data.

[0017] S3: Construct a spatiotemporal soft decoupling coding model; the spatiotemporal soft decoupling coding model includes a two-stream recurrent neural network, a physically enhanced graph attention network, and a decoding network.

[0018] S4: The dual-stream recurrent neural network is used to perform spatiotemporal soft decoupling on the multi-source asynchronous sensing data. The physical augmentation graph attention network is used to aggregate the decoupled features along the physical transmission path to obtain aggregated features. The decoding network is used to perform twin registration on the aggregated features to obtain unified intermediate frequency registration data.

[0019] S5: Feedback the unified intermediate frequency registration data to the digital twin model for multiphysics simulation, construct a joint loss function including data reconstruction error and residual of the physical mechanism equation of the digital twin model, optimize the parameters of the spatiotemporal soft decoupling coding model, and obtain the trained spatiotemporal soft decoupling coding model.

[0020] S6: Alignment of multi-source asynchronous data for equipment is completed based on the trained spatiotemporal soft decoupling coding model.

[0021] By implementing steps S1 to S6 above, a soft decoupling of temporal features is achieved through a dual-stream recurrent neural network, avoiding signal distortion caused by mathematical interpolation. Physical topological constraints are introduced through a physically enhanced graph attention network, solving the problem of lack of physical correlation in purely data-driven methods. A closed-loop optimization mechanism is constructed through residual feedback of the physical mechanism equations of the digital twin model, overcoming the deficiency of lack of mechanism verification in existing technologies. This application solves the problem of spatiotemporal fusion distortion caused by large step frequency differences and lack of physical constraints in multi-source asynchronous sensor data of equipment, providing high-fidelity, high-quality data and methodological support for equipment monitoring, maintenance, and performance prediction.

[0022] Furthermore, step S1 specifically includes: Based on the equipment's geometric topology and physical conduction characteristics, a physical sensing network comprising multiple source sensors for temperature, vibration, and pressure is configured to collect multi-source asynchronous sensing data under all service conditions of the equipment, and then perform dimensionless normalization processing. The multi-source asynchronous sensing data includes temperature data and vibration data.

[0023] Taking high-end CNC machine tools as a typical example, this study addresses the thermo-dynamic coupling behavior of high-end CNC machine tools during high-speed cutting by deploying a multi-source sensor network on key structural components such as the spindle system, guideways, and lead screws. Specifically: Eight thermocouple temperature sensors (sampling frequency of 1Hz) are deployed to capture the slowly changing characteristics of the temperature field; a three-axis accelerometer (sampling frequency of 10kHz) is deployed to capture the transient and severe vibration characteristics during the high-speed rotation of the spindle and the cutting process.

[0024] After collecting multi-source asynchronous sensing data, the method further includes: performing dimensionless processing on the multi-source asynchronous sensing data using the max-min normalization method.

[0025] In the formula, For the m-th sensing channel at time... The normalized observations; For the m-th sensing channel at time... The original observations; This represents the minimum observed value of the m-th sensing channel; This represents the maximum observed value of the m-th sensing channel.

[0026] Furthermore, step S2 specifically includes: A virtual mapping base is constructed based on the three-dimensional CAD geometric model of a high-end CNC machine tool, and the thermoelastic dynamic equilibrium equation and the temperature field heat conduction equation of the high-end CNC machine tool are embedded into the digital twin model.

[0027] The thermoelastic dynamic equilibrium equation after finite element discretization can be expressed as: in, , and These represent the mass matrix, damping matrix, and stiffness matrix of the machine tool structure, respectively. , and These represent the machine tool structure at time [time]. The displacement, velocity, and acceleration vectors; Indicates the machine tool structure at time External load vectors such as cutting force and driving force; Indicates at time The thermal load vector caused by the temperature field T.

[0028] The heat conduction equation for the temperature field can be expressed as: in, Indicates spatial location At any moment Temperature; , and These represent the material density, specific heat capacity, and thermal conductivity, respectively. Indicates spatial location At any moment Heat source per unit volume; and These represent the gradient operator and the divergence operator, respectively.

[0029] The digital twin model receives control commands (such as spindle speed and feed rate) from OPC UA or fieldbus in real time, as well as collected asynchronous sensor data, to ensure that the machine tool model in the virtual space keeps its evolutionary behavior synchronized with the actual equipment.

[0030] Furthermore, step S4 specifically includes: (1) Temporal soft decoupling.

[0031] Specifically, the dual-stream recurrent neural network is used to perform spatiotemporal soft decoupling on the multi-source asynchronous sensing data. The temperature data is input into the low-frequency recurrent network layer for feature extraction to obtain low-frequency time-series features; the vibration data is input into the high-frequency recurrent network layer for feature extraction to obtain high-frequency time-series features; the low-frequency time-series features and the high-frequency time-series features are the decoupled features.

[0032] Specifically, such as Figure 2As shown, the dual-stream recurrent neural network includes a low-frequency recurrent network layer and a high-frequency recurrent network layer. Low-frequency temperature data and high-frequency vibration data are fed into the low-frequency recurrent network layer and the high-frequency recurrent network layer, respectively. The low-frequency recurrent network layer uses a large time step to extract the long-period trend of heat conduction (i.e., long time step / slow-change feature extraction), while the high-frequency recurrent network layer uses a small time step to extract the transient high-frequency time-correlation features of vibration (i.e., short time step / fast-change feature extraction). This achieves feature decoupling in the time domain and prevents fast-change signals from overwhelming slow-change signals.

[0033] by This represents the low-frequency branch and the high-frequency branch. For the k-th target time with a unified intermediate frequency... Input windows corresponding to the target time are extracted from both low-frequency and high-frequency time-series features. The low-frequency branch uses a window covering a longer physical time period to extract slow-changing processes such as heat conduction; the high-frequency branch uses a window covering a shorter physical time period and containing more sampling points to extract fast-changing processes such as vibration. Both branches use GRU (Gated Recurrent Unit) encoders with independent parameters. Figure 2 The low-frequency GRU encoder and the high-frequency GRU encoder are recursively calculated according to their original sampling time order. For any branch... For hidden layer temporal feature fusion, the update gate, reset gate, candidate hidden state, and hidden state are calculated sequentially using the following formula: in, and These are the update gate and reset gate for branch s at time step t, respectively; Let be the input vector of branch s at time step t; and Branch s is in the th The hidden state at time step t and the time step t; Let be the candidate hidden state of branch s at time step t; , and Update the input weight matrices for the gate, reset, and hidden states respectively; , and These are the cyclic weight matrices for updating the gate, resetting, and hiding the state, respectively; , and These are the bias vectors for updating the gate, resetting, and hiding the state, respectively; This represents the Sigmoid activation function; Represents the hyperbolic tangent function; This indicates element-wise multiplication.

[0034] After completing the recursion of their respective time windows, the terminal hidden states of the low-frequency branch and the high-frequency branch are retrieved respectively. and As the target moment Temporal representation: in, and Representing the target time respectively The corresponding low-frequency timing characteristics and high-frequency timing characteristics; and These represent the number of time steps within the low-frequency window and the high-frequency window, respectively.

[0035] By decoupling the time scale in the hidden feature layer rather than the original data layer, the direct interference of high-frequency amplitude fluctuations on the low-frequency trend representation can be avoided.

[0036] (2) Spatial physical enhancement.

[0037] The decoupled features are aggregated along the physical propagation path using the Physically Augmented Graph Attention Network to obtain aggregated features. Specifically, this includes: mapping the low-frequency and high-frequency temporal features to nodes of the Physically Augmented Graph Attention Network; constructing physical prior weights for nodes with physical propagation paths; calculating unnormalized attention coefficients based on the physical prior weights and normalizing them to obtain normalized attention weights; updating node features layer by layer along the physical propagation path based on the normalized attention weights; and performing graph-level aggregation on all node features after multi-layer aggregation to obtain aggregated features.

[0038] Specifically, such as Figure 3 As shown, the low-frequency temporal features and the high-frequency temporal features are mapped to nodes 1, 2, 3, ..., i of the physical augmentation graph attention network. Let the sensor corresponding to node i belong to the branch. Then the initial features of the node are represented as: in, The initial features of node i at the k-th target time point; The hidden state of the branch output corresponding to the k-th target time node i; and Branch roads The mapping matrix and bias vector.

[0039] The graph edges are established based on the actual connection relationships between structural components such as the machine tool spindle, bearings, guide rails, and housing. For nodes i and j with physical transmission paths, physical prior weights are constructed by combining structural connectivity, spatial distance, and conduction impedance. in, For the k-th target time node With nodes Physical prior weights between them; Represents a node With nodes Does a physical transmission path exist? If a physical transmission path exists, the value is 1; otherwise, the value is 0. For nodes With nodes Spatial distance between them; For the k-th target time node With nodes Thermal resistance (or kinetic transmission impedance) between them. This is the distance scale factor; This is the impedance scaling factor; It is an exponentially decaying function.

[0040] In the In the layered graph attention mechanism, the unnormalized attention coefficients between adjacent nodes are calculated: in, For the k-th target time, the Nodes in layer graph attention mechanism With nodes The unnormalized attention coefficients between them; For the first Attention vectors in layer graph attention mechanisms; For the first The node feature transformation matrix of the layer graph attention mechanism; For the k-th target time, the Nodes in layer graph attention mechanism Features; For the k-th target time, the Nodes in layer graph attention mechanism Features; This represents the physical prior influence coefficient. It is the natural logarithm function; A positive number is set to prevent logarithms from being undefined; Indicates a linear unit with leakage correction; This indicates vector concatenation.

[0041] Normalize within the neighborhood of node i: in, For nodes The set of neighboring nodes; For the k-th target time, the Nodes in layer graph attention mechanism With nodes Normalized attention weights between them.

[0042] The formula for updating node features is: in, For the k-th target time, the Nodes in layer graph attention mechanism Features; It is a non-linear activation function; For the k-th target time, the The node feature transformation matrix of the layer graph attention mechanism; For the first Nodes in layer graph attention mechanism Its characteristics.

[0043] go through After aggregation using the layered graph attention mechanism, graph-level convergence is performed on all node features: in, The aggregated features at the k-th target time; For the k-th target time interval The feature set of all nodes after the layer graph attention mechanism is updated; The number of nodes; Pool represents the average pooling, weighted pooling, or attention pooling operation.

[0044] (3) Intermediate frequency conversion and registration output.

[0045] The decoding network outputs registration results at a unified intermediate frequency. Specifically, based on the optimal convergence step size of the digital twin model, a unified registration intermediate frequency of 100Hz is set. Through the decoding network, the spatiotemporal aggregation features are mapped into registration data with continuous physical meaning at 100Hz (including aligned temperature data at 100Hz and vibration data at 100Hz), achieving lossless time registration of different frequency sensor data.

[0046] The formula for calculating unified intermediate frequency registration data is: Registered data for the unified intermediate frequency at the k-th target time; The weight matrix for decoding the network; This is the bias vector for the decoding network.

[0047] Furthermore, step S5 specifically includes: For the k-th target time under a unified intermediate frequency, a dual-stream recurrent neural network extracts the temporal features of low-frequency temperature data and high-frequency vibration data, respectively. These features are then aggregated using a physical augmentation graph attention network, and finally, a decoding network is used to obtain the unified intermediate frequency registration data. The relationship between these features is expressed as follows: in, and These are low-frequency input data (i.e., temperature data) and high-frequency input data (i.e., vibration data), respectively. and These are low-frequency recurrent network layers and high-frequency recurrent network layers, respectively. For physical augmented graph attention networks; For decoding the network.

[0048] Data reconstruction loss is calculated based on unified intermediate frequency registration data and corresponding real sensor observation data. : in, This represents the actual sensor observation data corresponding to the k-th target time. The target number of time points.

[0049] Further from Temperature data is extracted and fed back into the digital twin model to solve for the temperature field, thus obtaining the predicted temperature field. : in, Unify the temperature component in intermediate frequency registration data. This represents the temperature field solution function for the digital twin model.

[0050] Based on the heat conduction equation, a physical constraint loss is constructed. The residuals of the physical mechanism equation of the digital twin model are as follows: : in, The density of the material; Specific heat capacity; Thermal conductivity; For the Laplace operator; Intensity of the internal heat source; It is the square of the L2 norm.

[0051] The expression for the joint loss function is: in, For joint losses; These are the weight coefficients for data constraint terms; This represents the weighting coefficient for the physical constraint term. In this embodiment, Set to 0.6, Set it to 0.4.

[0052] During model training, the underlying thermophysical mechanism of the machine tool is transformed into a loss function. With the joint loss function as the optimization objective, gradient descent algorithm and backpropagation mechanism are used to jointly update the trainable parameters in the spatiotemporal soft decoupled coding model, forming a closed-loop optimization mechanism driven by both mechanism and data.

[0053] After model convergence, the output is a high-quality, fully synchronized 100 Hz dataset of equipment, with high-fidelity spatiotemporal registration. This data can be directly used as input to the thermal deformation prediction model of CNC machine tools, or for high-precision unsupervised anomaly detection of machine tool service status, significantly enhancing the monitoring, early warning, and maintenance capabilities of equipment status.

[0054] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0055] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0056] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0058] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for aligning asynchronous multi-source data of equipment based on spatiotemporal soft decoupling twin registration, characterized in that, include: Configure physical sensing networks on the equipment and collect multi-source asynchronous sensing data; Construct a digital twin model of the equipment and perform near real-time mapping and state synchronization with the multi-source asynchronous sensing data; A spatiotemporal soft decoupling coding model is constructed; the spatiotemporal soft decoupling coding model includes a two-stream recurrent neural network, a physically enhanced graph attention network, and a decoding network; The dual-stream recurrent neural network is used to perform spatiotemporal soft decoupling on the multi-source asynchronous sensing data. The physical augmentation graph attention network is used to aggregate the decoupled features along the physical transmission path to obtain aggregated features. The decoding network is used to perform twin registration on the aggregated features to obtain unified intermediate frequency registration data. The unified intermediate frequency registration data is fed back to the digital twin model for multiphysics simulation. A joint loss function including data reconstruction error and residual of the physical mechanism equation of the digital twin model is constructed. The parameters of the spatiotemporal soft decoupling coding model are optimized to obtain the trained spatiotemporal soft decoupling coding model. The alignment of multi-source asynchronous data for equipment is completed based on the trained spatiotemporal soft decoupling coding model.

2. The method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration according to claim 1, characterized in that, The multi-source asynchronous sensing data includes temperature data and vibration data, and the dual-stream recurrent neural network includes a low-frequency recurrent network layer and a high-frequency recurrent network layer. Specifically, the use of the dual-stream recurrent neural network to perform spatiotemporal soft decoupling of the multi-source asynchronous sensing data includes: The temperature data is input into the low-frequency recurrent network layer for feature extraction to obtain low-frequency time-series features. The vibration data is input into the high-frequency recurrent network layer for feature extraction to obtain high-frequency temporal features; The low-frequency timing features and the high-frequency timing features are decoupled features.

3. The equipment multi-source asynchronous data alignment method based on spatiotemporal soft decoupling twin registration according to claim 2, characterized in that, The decoupled features are aggregated along the physical propagation path using the physical augmentation graph attention network to obtain aggregated features, specifically including: The low-frequency and high-frequency temporal features are mapped to nodes of the physical augmentation graph attention network; For nodes with physical transmission paths, construct physical prior weights; The unnormalized attention coefficients are calculated based on the physical prior weights and then normalized to obtain the normalized attention weights. Based on the normalized attention weights, the node features are updated layer by layer along the physical transmission path. After multi-layer aggregation, all node features are aggregated at the graph level to obtain aggregated features.

4. The equipment multi-source asynchronous data alignment method based on spatiotemporal soft decoupling twin registration according to claim 3, characterized in that, The expression for the physical prior weights is: in, For the k-th target time node With nodes Physical prior weights between them; Represents a node With nodes Does a physical transmission path exist? If a physical transmission path exists, the value is 1; otherwise, the value is 0. For nodes With nodes Spatial distance between them; For the k-th target time node With nodes Thermal resistance between them; This is the distance scale factor; This is the impedance scaling factor; It is an exponentially decaying function.

5. The equipment multi-source asynchronous data alignment method based on spatiotemporal soft decoupling twin registration according to claim 4, characterized in that, The formula for calculating the unnormalized attention coefficient is as follows: in, For the k-th target time, the Nodes in layer graph attention mechanism With nodes The unnormalized attention coefficients between them; For the first Attention vectors in layer graph attention mechanisms; For the first The node feature transformation matrix of the layer graph attention mechanism; For the k-th target time, the Nodes in layer graph attention mechanism Features; For the k-th target time, the Nodes in layer graph attention mechanism Features; This represents the physical prior influence coefficient. It is the natural logarithm function; A positive number is set to prevent logarithms from being undefined; Indicates a linear unit with leakage correction; This indicates vector concatenation.

6. The method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration according to claim 5, characterized in that, The formula for updating node features is: in, For the k-th target time, the Nodes in layer graph attention mechanism Features; It is a non-linear activation function; For the k-th target time, the Nodes in layer graph attention mechanism With nodes Normalized attention weights between them; For the k-th target time, the The node feature transformation matrix of the layer graph attention mechanism; For the first Nodes in layer graph attention mechanism Features; For nodes The set of neighboring nodes.

7. The equipment multi-source asynchronous data alignment method based on spatiotemporal soft decoupling twin registration according to claim 6, characterized in that, The formula for calculating the aggregation feature is: in, The aggregated features at the k-th target time; For the k-th target time interval The feature set of all nodes after the layer graph attention mechanism is updated; The number of nodes; Pool represents the average pooling, weighted pooling, or attention pooling operation; The calculation formula for the unified intermediate frequency registration data is as follows: in, Registered data for the unified intermediate frequency at the k-th target time; The weight matrix for decoding the network; This is the bias vector for the decoding network.

8. The method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration according to claim 1, characterized in that, The expression for the joint loss function is: in, For joint losses; For data reconstruction errors; The residuals of the physical mechanism equations for the digital twin model; These are the weight coefficients for the data constraint terms; These are the weighting coefficients for the physical constraint terms; The density of the material; Specific heat capacity; To predict the temperature field; Thermal conductivity; For the Laplace operator; Intensity of the internal heat source; It is the square of the L2 norm.

9. The method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration according to claim 1, characterized in that, After collecting multi-source asynchronous sensor data, the process also includes: The multi-source asynchronous sensing data is dimensionless by using the max-min normalization method.

10. The method for aligning multi-source asynchronous data of equipment based on spatiotemporal soft decoupling twin registration according to claim 1, characterized in that, The digital twin model incorporates thermoelastic dynamic equilibrium equations and temperature field heat conduction equations.

Images