A building wind pressure aerodynamic force prediction method based on PIV flow field space-time coding and physical enhancement

Abstract

This invention discloses a method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement. The method includes the following steps: calculating the temporal difference based on particle image velocity measurement flow field at adjacent times around the building, and using a channel dimension stitching method to combine the velocity field components at adjacent times with the temporal difference to obtain a multi-channel input tensor; using a state space model to perform spatiotemporal coding on the multi-channel input tensor, and using one-dimensional depthwise separable convolutions performed along the height and width directions to serialize and mix the input features to obtain flow field features; performing decoding and mapping processing based on the flow field features to obtain the wind pressure coefficient at several discrete pressure measurement points on the building cross section; and using an aerodynamic mapping module with learnable mapping parameters to perform affine transformation processing on the wind pressure coefficient to obtain the overall lift coefficient and drag coefficient of the building.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of wind engineering and intelligent computing, and in particular relates to a method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement. Background Technology

[0002] In fields such as building construction, bridge engineering, and marine engineering, the aerodynamic characteristics of bluff structures under wind or water flow are key issues in structural wind-resistant design and safety assessment. For high-rise buildings, square column structures, and similar components, there is a close physical relationship between the surface wind pressure distribution and the overall aerodynamic forces (lift and drag). Accurately obtaining the time history of these aerodynamic forces is of great significance for wind-induced response analysis, fatigue assessment, and structural reliability design. Currently, the acquisition of building aerodynamic forces mainly relies on two methods: one is to arrange a large number of pressure measurement holes on the model surface through wind tunnel tests to measure the surface wind pressure coefficient, and then calculate the overall aerodynamic forces by integrating the pressure across the entire surface; the other is to perform high-precision transient numerical simulation of the flow field based on computational fluid dynamics (CFD) methods, and then extract the aerodynamic data through post-processing. In addition, with the development of flow field measurement technologies such as particle image velocimetry (PIV), it has become more convenient to obtain high-resolution transient flow field information around buildings. Some researchers have begun to try to use data-driven methods to directly establish the mapping relationship between flow field characteristics and aerodynamic forces in order to achieve rapid prediction.

[0003] However, existing technologies still face many challenges in practical applications and urgently need further improvement. First, although traditional wind tunnel pressure measurement integration methods and CFD numerical simulations are theoretically sound, they generally suffer from problems such as long test cycles, complex measurement point layouts, large data processing volumes, or high computational costs, making it difficult to meet the needs of rapid multi-condition evaluation and parametric analysis in the early stages of engineering projects. Second, existing flow field data-driven prediction methods mostly treat the flow field as a black box input, relying mainly on convolutional neural networks (CNNs) or recurrent neural networks (RNNs) for modeling. Their core shortcomings are: on the one hand, they ignore the crucial physical mediating role of the wind pressure coefficient between the flow field and the overall aerodynamic forces, resulting in poor model interpretability and the introduction of unnecessary nonlinear complexity, affecting the stability and generalization ability of the predictions; on the other hand, existing methods do not explicitly design dedicated coding structures for the long-range spatiotemporal dependence of unsteady flow fields, making it difficult to efficiently capture large-scale dynamic features such as wake evolution and flow separation, and they do not embed the differential / integral physical quantities of the flow field as differentiable constraints into the learning objectives, resulting in model prediction results that contradict the basic principles of fluid mechanics and have weak physical consistency. Furthermore, traditional data-driven models typically rely on additional measured aerodynamic data as supervisory signals, which not only increases experimental and calibration costs but also limits their widespread application in engineering scenarios lacking complete aerodynamic labels. Therefore, this invention proposes a method for predicting building wind pressure and aerodynamic forces based on PIV flow field spatiotemporal coding and physical enhancement. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for predicting building wind pressure and aerodynamic forces based on PIV flow field spatiotemporal coding and physical enhancement, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement, comprising: The temporal difference is calculated based on the particle image velocity flow field of adjacent moments around the building, and the channel dimension stitching method is used to combine the velocity field components of adjacent moments with the temporal difference to obtain a multi-channel input tensor. The multi-channel input tensor is spatiotemporally encoded using a state-space model. The flow field features are obtained by serializing and mixing the input features through one-dimensional depthwise separable convolutions performed along the height and width directions, respectively. Based on the aforementioned flow field characteristics, a decoding and mapping process is performed to obtain the wind pressure coefficient at several discrete pressure measurement points on the building cross section. The wind pressure coefficient is subjected to affine transformation using an aerodynamic mapping module with learnable mapping parameters to obtain the overall lift coefficient and drag coefficient of the building.

[0006] Optionally, the process of combining the velocity field components at adjacent time points with the temporal difference using a channel-dimensional stitching method to obtain a multi-channel input tensor includes: The first and second velocity fields obtained from the particle image velocimetry flow field are processed at the corresponding spatial grid positions using point-by-point difference operation to obtain temporal difference; A multi-channel input tensor is obtained by combining two components of the velocity field at the first time step, two components of the velocity field at the second time step, and at least one component of the temporal difference using a channel dimension stacking method.

[0007] Optionally, the process of using a state-space model to perform spatiotemporal encoding on the multi-channel input tensor to obtain flow field features includes: A convolutional downsampling layer is used to process the multi-channel input tensor to obtain a dimension-reduced expanded feature map; The cascaded two-dimensional state-space coding module is used to process the dimensionality-reduced extended feature map. The input features are serialized and mixed by one-dimensional depthwise separable convolution along the height direction and one-dimensional depthwise separable convolution along the width direction, respectively. The mixing results of the two directions are fused and output through residual connection to obtain the flow field features characterizing the unsteady evolution of the wake.

[0008] Optionally, the process of using cascaded two-dimensional state-space coding modules to perform serialized state mixing of input features includes: Normalized features are obtained by processing the input features using a group normalization layer; The normalized features are processed by a channel-expanding convolutional layer to obtain expanded features; The extended features are processed using a gated branch structure. First branch features and second branch features are generated by parallel pointwise convolution. The second branch features are processed by a nonlinear activation function to obtain gate weights. The first branch features and the gate weights are multiplied element-wise to obtain gated modulation features. Based on the gated modulation features, one-dimensional depthwise separable convolutions are performed along the height direction and along the width direction to obtain mixed features in the two directions. The mixed features in the two directions are then fused and processed by a channel compression convolutional layer, and finally added to the input features through residual connections to obtain the output features.

[0009] Optionally, the one-dimensional depthwise separable convolution along the height direction is implemented by merging the width dimension and batch dimension of the input features and then performing a one-dimensional depthwise convolution operation with a kernel length of not less than 17 along the height direction, and the one-dimensional depthwise separable convolution along the width direction is implemented by merging the height dimension and batch dimension of the input features and then performing a one-dimensional depthwise convolution operation with a kernel length of not less than 17 along the width direction.

[0010] Optionally, the process of obtaining the wind pressure coefficient at several discrete pressure measurement points on a building cross-section by decoding and mapping based on the flow field characteristics includes: A global average pooling layer is used to compress the flow field features in both the spatial height and spatial width dimensions to obtain a feature vector. A multi-layer fully connected network is used to perform layer-by-layer nonlinear mapping on the feature vector to obtain a wind pressure coefficient vector with the number of output nodes equal to the number of discrete pressure measurement points.

[0011] Optionally, the process of using an aerodynamic mapping module with learnable mapping parameters to perform an affine transformation on the wind pressure coefficient to obtain the overall lift coefficient and drag coefficient of the building includes: The wind pressure coefficients at each discrete pressure measurement point are constructed into a wind pressure coefficient vector. The wind pressure coefficient vector is subjected to matrix multiplication using a learnable mapping coefficient matrix to obtain a weighted result; The weighted results are summed using a learnable bias vector to obtain a two-dimensional output vector containing the lift coefficient and drag coefficient.

[0012] Optionally, the size of the mapping coefficient matrix is ​​two rows multiplied by the number of discrete pressure measurement points, where the first row of elements corresponds to the contribution weight of the wind pressure coefficient of each pressure measurement point to the lift coefficient, the second row of elements corresponds to the contribution weight of the wind pressure coefficient of each pressure measurement point to the drag coefficient, and the bias vector has a length of two, corresponding to the bias term of the lift coefficient and the bias term of the drag coefficient, respectively.

[0013] Optionally, the encoding parameters, decoding mapping parameters, and mapping parameters of the aerodynamic mapping module of the state space model are optimized and determined by a joint loss function. The joint loss function is composed of a weighted sum of the error term between the predicted and measured values ​​of the wind pressure coefficient, and a physical consistency constraint term constructed based on the root mean square value of vorticity and the wake momentum deficit value obtained by online calculation of the input velocity field.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a multi-channel input structure for flow field and temporal difference between adjacent time steps, enabling the model to simultaneously perceive the spatial distribution and temporal evolution trend of the flow field, thus providing a rich dynamic information foundation for subsequent spatiotemporal coding. A state-space model is used to perform one-dimensional depthwise separable convolutions along the height and width directions for sequential state mixing, effectively capturing the large-scale unsteady evolution characteristics and flow separation features of the wake region with linear computational complexity, overcoming the limitation of the receptive field in traditional convolutional networks. A two-level prediction architecture is employed, first predicting the wind pressure coefficient at discrete pressure measurement points on building sections and then converting it into the overall lift and drag coefficient through a learnable mapping module. This explicitly introduces the physical mediator relationship between wind pressure and aerodynamic forces, avoiding the black-box drawback of directly predicting aerodynamic forces end-to-end from flow field images, significantly improving the physical interpretability and result stability of the prediction process. The mapping parameters in the aerodynamic mapping module are automatically learned by the model, eliminating the need for manual pre-setting of the integral weights for each pressure measurement point, enhancing adaptability to different building section forms and inflow conditions. The entire prediction process only requires input of particle image velocity flow field data to quickly output the overall lift coefficient and drag coefficient of the building. It does not require full wall pressure integration or high-cost numerical simulation, which greatly reduces the time and economic cost of aerodynamic assessment. It is suitable for rapid assessment and parametric analysis of multiple working conditions in the early stage of engineering. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the overall process of a building wind pressure and aerodynamics prediction method based on PIV flow field spatiotemporal coding and physical enhancement according to an embodiment of the present invention; Figure 2 This is a block diagram of the system structure modules according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a neural network according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the input data organization format in an embodiment of the present invention; Figure 5 This is a schematic diagram showing the connection relationship between the wind pressure prediction subnetwork and the aerodynamic mapping module in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the calculation of physical proxy quantities based on the PIV velocity field in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the joint loss structure during the training phase of an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a method for predicting building wind pressure and aerodynamic forces based on PIV flow field spatiotemporal coding and physical enhancement. This method can achieve high-precision prediction of the overall lift coefficient and drag coefficient of a building by using only the flow field information around the building and a limited number of wind pressure coefficient data without the need for full-surface pressure integration or high-cost numerical simulation. This significantly improves the efficiency and engineering applicability of aerodynamic assessment.

[0019] This method takes particle image velocimetry (PIV) flow field data around a building model and flow field change information at adjacent time points as input, and uses the PIV velocity field and its time difference between two adjacent time points as multi-channel input. First, it extracts local vortex structure and shear layer features through convolutional downsampling. Then, it introduces a two-dimensional state-space spatiotemporal coding module (SSM2D Block) to perform long-kernel state mixing along the height and width directions to model the wake evolution and unsteady separation process. Based on this, it outputs wind pressure coefficients at multiple pressure measurement points, and obtains lift and drag coefficients through a trainable mapping module. The method uses a spatiotemporal coding structure constructed by a state-space model (SSM) (including one-dimensional depthwise separable convolution and residual connections along the H / W dimension) to efficiently extract long-range spatiotemporal dependent features of the flow field. During training, a ternary joint loss function (wind pressure error + aerodynamic error + physical surrogate consistency constraint) is designed, where the physical constraint term achieves self-supervision by online calculation of vorticity RMS and momentum deficit, significantly improving physical interpretability and generalization ability. Compared with existing technologies, this invention achieves high-precision and rapid prediction of building aerodynamics without requiring full-wall pressure integration or high-cost transient numerical simulation. It boasts advantages such as strong physical interpretability, high prediction efficiency, and strong adaptability to engineering applications. Specifically, it includes the following steps: 1) Obtain particle image velocity measurement flow field data of the building model under the influence of incoming flow. The particle image velocity measurement flow field data is two-dimensional or quasi-two-dimensional flow field data, and the flow field data includes at least the velocity component distribution information of two adjacent time points; the velocity component distribution information of two adjacent time points is the point-by-point difference of the flow field at the two adjacent time points. Input data is provided by and A three-channel input tensor is constructed by concatenating the data along the channel dimension. Specifically, the temporal difference is calculated based on the particle image velocity flow field at adjacent time points around the building, and the velocity field components at adjacent time points are combined with the temporal difference using the channel dimension concatenation method to obtain a multi-channel input tensor.

[0020] The flow field data is velocity field data obtained through PIV technology. The velocity field includes at least velocity components along the direction of incoming flow and perpendicular to the direction of incoming flow.

[0021] As a specific implementation of this embodiment, the process of obtaining the multi-channel input tensor includes: using point-by-point difference operation to process the velocity field at the first moment and the velocity field at the second moment obtained based on the particle image velocimetry flow field at the corresponding spatial grid positions to obtain the temporal difference; and using a channel dimension stacking method to combine the two components of the velocity field at the first moment, the two components of the velocity field at the second moment, and at least one component of the temporal difference to obtain the multi-channel input tensor.

[0022] 2) Input the flow field data and their difference information at adjacent time points into the flow field feature extraction module of the deep neural network to extract flow field features characterizing flow separation, shear layer development, and wake evolution. That is, use a state-space model to perform spatiotemporal encoding on the multi-channel input tensor, and use one-dimensional depthwise separable convolutions performed along the height and width directions to serialize and mix the input features to obtain the flow field features.

[0023] Furthermore, the process of obtaining flow field features by spatiotemporally encoding the multi-channel input tensor using a state-space model includes: processing the multi-channel input tensor using a convolutional downsampling layer to obtain a dimension-reduced extended feature map; processing the dimension-reduced extended feature map using a cascaded two-dimensional state-space encoding module; performing serialized state mixing of the input features by one-dimensional depth-separable convolution along the height direction and one-dimensional depth-separable convolution along the width direction respectively; and fusing the mixing results of the two directions and outputting them through residual connection to obtain flow field features characterizing the unsteady evolution of the wake.

[0024] Furthermore, the process of serializing and mixing the input features using a cascaded two-dimensional state-space coding module includes: processing the input features with a group normalization layer to obtain normalized features; processing the normalized features with a channel-expanding convolutional layer to obtain expanded features; processing the expanded features with a gated branch structure, generating first branch features and second branch features through parallel point-by-point convolution, and processing the second branch features with a nonlinear activation function to obtain gate weights, multiplying the first branch features and gate weights element-wise to obtain gated modulation features; performing one-dimensional depthwise separable convolution along the height direction and one-dimensional depthwise separable convolution along the width direction based on the gated modulation features to obtain mixed features in the two directions, fusing the mixed features in the two directions and processing them with a channel-compressing convolutional layer, and adding them to the input features through residual connections to obtain the output features.

[0025] As a specific implementation of this embodiment, the flow field feature extraction module adopts a spatiotemporal coding structure constructed by a state-space model (SSM): the input is downsampled and then fed into a cascaded SSM2D Block. Each Block contains group normalization, channel expansion, one-dimensional depthwise separable convolution along the height / width direction, channel compression, and residual connections to explicitly model the unsteady spatiotemporal evolution of the flow field. The one-dimensional depthwise separable convolution along the height / width direction is achieved by merging the two-dimensional features in batches according to the width dimension to obtain (B·W, C, H) and convolving them along the H direction, and then merging the batches according to the height dimension to obtain (B·H, C, W) and convolving them along the W direction (the convolution kernel length is ≥17, preferably 31) to extract long-range dependencies.

[0026] The SSM2D Block includes a gated state hybrid structure, specifically: mapping the input to the features of the first branch. Features of the second branch The second branch features are processed using a nonlinear activation function to obtain the gating weights, and then... Gating is performed to obtain gated modulation features; then long-kernel one-dimensional depthwise separable convolutional mixing is performed along the height and width directions respectively, and channel compression is performed through 1×1 convolution after each mixing, and finally residual connection is formed with the input to obtain the output features.

[0027] Furthermore, the spatiotemporal coding structure is implemented using a two-dimensional state hybrid approach with linear complexity: the downsampled features... F is obtained by convolution expansion and gating. g Then, a one-dimensional depthwise separable convolution is performed with the width dimension as the sequence axis to achieve state mixing across the H direction, and then a one-dimensional depthwise separable convolution is performed with the height dimension as the sequence axis to achieve state mixing across the W direction. The convolution kernel is a long kernel to cover the tail scale. Finally, the convolution is compressed and output in the form of residuals.

[0028] 3) Based on flow field characteristics, the wind pressure coefficient at multiple discrete pressure measurement points on the building's cross-section is predicted using a wind pressure prediction module. That is, the wind pressure coefficient at several discrete pressure measurement points on the building's cross-section is obtained through decoding and mapping based on flow field characteristics.

[0029] Furthermore, the process of obtaining the wind pressure coefficient at several discrete pressure measurement points on the building section includes: using a global average pooling layer to compress the flow field features in the spatial height and spatial width dimensions to obtain a feature vector; and using a multi-layer fully connected network to perform layer-by-layer nonlinear mapping processing on the feature vector to obtain a wind pressure coefficient vector with the number of output nodes equal to the number of discrete pressure measurement points.

[0030] The wind pressure prediction module is used to predict the wind pressure coefficient at multiple discrete pressure measurement points located circumferentially arranged in the cross section of a building, with no fewer than four pressure measurement points.

[0031] Assuming that a total of [number] sections are arranged at the cross-section of the building. If there are 10 pressure measurement points, the predicted wind pressure coefficient vector is represented as: ; in, This represents a wind pressure prediction model based on the spatiotemporal feature encoding of the flow field. The two-dimensional velocity field obtained for PIV; 4) The predicted wind pressure coefficient is used as input, and the wind pressure coefficient is converted into the overall lift and drag coefficient through the aerodynamic mapping module.

[0032] Furthermore, the process of obtaining the overall lift coefficient and drag coefficient of the building includes: constructing a wind pressure coefficient vector from the wind pressure coefficients at each discrete pressure measurement point; performing matrix multiplication on the wind pressure coefficient vector using a learnable mapping coefficient matrix to obtain a weighted result; and adding the weighted result using a learnable bias vector to obtain a two-dimensional output vector containing the lift coefficient and drag coefficient.

[0033] The aerodynamic mapping module is a linear mapping module (non-linear is optional). Its input is a wind pressure coefficient vector, and its output is the overall building lift coefficient and drag coefficient. The mapping coefficients in the linear mapping relationship are automatically learned by the model during training. Structurally, the aerodynamic mapping module includes transformation units and bias terms to characterize the linear relationship between the wind pressure coefficient and the overall aerodynamic forces.

[0034] The mapping relationship between wind pressure coefficient and overall aerodynamic force is a linear mapping relationship. The mapping coefficient of this relationship is automatically learned by the network during training, rather than being preset by humans.

[0035] The overall lift coefficient and drag coefficient are calculated from the wind pressure coefficient vector using an aerodynamic mapping operator: ; in, For having trainable parameters The mapping module. Preferably, the mapping operator This can be represented in affine form as follows: ; in, The mapping coefficient matrix, Let be the bias vector, and It can be learned automatically from training data or updated adaptively by the model during training.

[0036] Furthermore, as a specific implementation of this embodiment, a linear affine mapping is preferred based on the predicted wind pressure coefficient; in another embodiment, a shallow mapping network with nonlinear activation can be used to achieve robust mapping. This model automatically learns the mapping coefficients between the wind pressure coefficient and the overall building lift and drag coefficients, thereby outputting the overall building aerodynamic prediction results.

[0037] 5) During model training, a physical enhancement constraint module is introduced, which uses physical surrogate quantities such as vorticity intensity and momentum deficit in the wake region, calculated online based on flow field data, as auxiliary constraints to enhance the physical consistency of the prediction results.

[0038] The physical surrogate quantity of vorticity intensity is obtained by spatial finite difference calculation of the velocity field, and its expression is: ; In this process, spatial partial derivatives are discretized using central difference and first-order difference discretization is used at the boundary of the computational domain; data containing invalid measurement points are statistically analyzed using a valid point mask, and the root mean square value is calculated on the set of valid points.

[0039] The instantaneous vorticity of a two-dimensional flow field is defined as: ; Where the spatial partial derivatives It is preferable to use central difference and first-order difference for discrete calculation at the boundary.

[0040] And its effective region of the flow field is used The root mean square value within the range is used as a characterization of vorticity intensity: ; The momentum loss proxy is obtained by integrating the velocity components in the downstream region of the building model. It is used to characterize the degree of energy loss in the wake region and serves as a physical auxiliary constraint for drag prediction, thereby characterizing the building drag characteristics.

[0041] The momentum loss proxy in the wake region is defined as: ; in, The reference velocity for the incoming flow can be obtained from the velocity field average of the area upstream of the building. This indicates the downstream wake area of ​​a building.

[0042] The vorticity is obtained by differentiating the velocity field using central and first-order boundary differences, and invalid measurement points are statistically analyzed using effective point masks. The momentum deficit is estimated from the upstream region mask mean to measure the reference velocity of the flow. The result is obtained by integration / summation in the downstream wake region and combined with the corresponding agent quantity prediction value output by the network to form the physical consistency loss term.

[0043] The physical surrogate quantity does not need to be used as the final prediction output; it is only introduced into the loss function as an auxiliary constraint during the model training phase. The auxiliary constraint is achieved by minimizing the error between the physical surrogate quantity predicted by the model and the corresponding physical surrogate quantity calculated based on the flow field data.

[0044] Training uses a joint loss function: ; in, The mean square error loss of the wind pressure coefficient at the pressure measurement point; Monitoring loss for lift coefficient and drag coefficient (optional); For physical consistency loss, satisfying , and Physical self-supervision without aerodynamic labels is achieved through online differential / integral calculation of the input flow field; These are the weighting coefficients.

[0045] The physical consistency constraint term is defined as follows: ; The overall aerodynamic prediction process can complete model training without directly measuring or inputting overall lift or drag data.

[0046] like Figure 1The diagram illustrates the overall flow of a building wind pressure and aerodynamics prediction method based on PIV flow field spatiotemporal coding and physical enhancement, as described in this invention. The process begins by acquiring particle image velocity measurement flow field data around a building model. Velocity fields and their differences between adjacent time points are extracted and input into the PIV processing module of a deep neural network for flow field feature extraction. Subsequently, the spatiotemporal evolution characteristics of the flow field are modeled using an SSM spatiotemporal encoder to predict the wind pressure coefficients at multiple pressure measurement points across the building's cross-section. Based on this, an aerodynamic mapping module transforms the wind pressure coefficients into overall lift and drag coefficients. A physical enhancement constraint module is introduced during the training phase, using physical surrogate quantities such as vorticity intensity and momentum deficit as auxiliary supervision signals to improve the model's physical consistency and generalization ability. The final output includes complete aerodynamic prediction results, including wind pressure coefficient distribution, lift coefficient, and drag coefficient.

[0047] like Figure 2 The diagram shown is a structural block diagram of the system of the present invention, clearly illustrating the logical relationships and data flow between the various functional modules. The system mainly consists of the following five core modules: Flow field data acquisition module: Used to acquire PIV flow field data around the building model. It collects instantaneous velocity field data obtained from PIV measurements around the building model, including the first... Frame and the frame velocity field and generate its time difference. ; Data construction module: Groups and synthesizes the velocity fields of two frames and their difference into multi-channel input data to form spatiotemporal features containing dynamic information of flow. SSM / Spatiotemporal Coding Module: Employs a State-Space Model (SSM) to spatiotemporally encode the flow field and capture its unsteady evolution patterns; Physical Enhancement Constraint Module: Online Calculation of Eddy Power With momentum loss These physical agent quantities are then fed back into the training process as auxiliary supervisory signals. Prediction and Output Modules: The wind pressure prediction subnetwork outputs the wind pressure coefficients from 36 pressure measurement points. Then, through a trainable linear mapping module, it automatically learns the conversion relationship between wind pressure and lift and drag, and finally outputs the overall aerodynamic force. and .

[0048] The entire system achieves end-to-end prediction from the original flow field to the building aerodynamics, without relying on full-wall pressure integration or high-cost numerical simulation.

[0049] like Figure 3As shown, the SSM2D Block is the core spatiotemporal coding unit: after the input features are processed by GroupNorm, the channels are expanded through convolution, and one-dimensional depthwise separable convolutions with kernel sizes of 17 (or 31) are performed along the H / W dimensions to achieve cross-scale spatiotemporal information fusion; residual connections and FFN modules further enhance feature representation, significantly outperforming the computational efficiency and long-range modeling capabilities of ordinary 3D convolutions. This network uses PIV flow field images and Wind field grid data is used as a dual input: Input path 1: The PIV flow field image is first processed by the UVE encoder for feature extraction, and then the feature expression capability is enhanced by the FILM feature modulation module. FiLM modulation occurs after the second stage of spatiotemporal coding and before further downsampling, so as to introduce the incoming flow state conditions while maintaining the resolution of the local vortex structure.

[0050] Input path two: The wind field grid data is fed into a feature extraction branch consisting of multiple convolutional layers, and simultaneously connected to a regularization layer and a wind pressure coefficient prediction network; The two path features are fused in the SSM2D Block, which contains gated pointwise convolution, depth-separable long-kernel one-dimensional convolution hybrid units along the height / width direction, and residual connections, to realize the state-space spatiotemporal encoding of two-dimensional flow field features; The output consists of two prediction heads: Prediction head: Directly predicts the wind pressure coefficient at 36 pressure measurement points; Upward and resistance prediction head: combining intermediate variables and Perform aerodynamic prediction; In addition, the network introduces the SiLU activation function to enhance nonlinear fitting capabilities and ensures effective collaboration of information from different sources through a feature fusion mechanism.

[0051] This structure takes into account both the spatial details and temporal evolution characteristics of the flow field, enhancing the model's ability to perceive complex flow structures.

[0052] like Figure 4 The diagram illustrates the input data organization format of this invention, showing how to organize the velocity fields of two adjacent frames. and its difference Construct a unified multichannel input tensor. Specifically: Each frame's velocity field contains direction and direction Two components; Calculate the time difference between two frames , used to characterize the trend of flow field changes; Simultaneously calculate the vorticity for each frame. And take the root mean square value within the effective region. ; Calculate the momentum loss in the downstream wake region This reflects the physical characteristics related to resistance; Finally, the above five types of information Stacked along the channel dimension, forming The five-channel input tensor is used for neural network processing.

[0053] This data organization method fully integrates the spatiotemporal evolution of the flow field with key physical properties, enhancing the model's ability to understand the nature of the flow. In addition to the three-way temporal differential inputs, the system also receives a two-dimensional velocity grid at the same time as a conditional input, which is encoded into a conditional vector cond by a UV encoder. The intermediate layer features are then modulated at the feature level using FiLM, and simultaneously used as additional inputs to the aerodynamic mapping head and the physical surrogate quantity prediction head.

[0054] like Figure 5 The diagram illustrates the connection between the wind pressure prediction subnetwork and the aerodynamic mapping module in this invention. This diagram reveals the mapping mechanism from local wind pressure to global aerodynamic forces: The wind pressure prediction subnetwork outputs a wind pressure coefficient vector from 36 discrete pressure measurement points. ; all The values ​​are fed into the Aerodynamic Mapping Module, which uses an affine transformation: ; Where the matrix With bias vector These are trainable parameters that are automatically learned during training. The mapping relationship is established based on the linear physical relationship between wind pressure and overall aerodynamic force, avoiding complex nonlinear black box fitting; This module supports automatic learning of mapping parameters without the need for manual preset, significantly improving the model's adaptability and generalization performance.

[0055] This design enables the model to not only predict local wind pressure but also accurately infer overall aerodynamic forces, demonstrating its excellent engineering applicability.

[0056] like Figure 6 The diagram illustrates the calculation of physical proxy quantities based on the PIV velocity field according to this invention, focusing on the division of the calculation regions for vorticity intensity and momentum deficit: In the upstream area of ​​the building model Select an appropriate range to calculate vorticity intensity , used to characterize the intensity of vortex generation and separation; downstream wake region of the architectural model Calculate momentum loss by selecting a specific region. This reflects the energy loss caused by turbulent dissipation; These physical proxy quantities are calculated directly from the flow field, without relying on additional experimental measurements; They are introduced into the training process as auxiliary supervision signals to help the model learn predictive behaviors that conform to physical laws; The diagram clearly marks the boundaries between the upstream and downstream regions, reflecting the criteria for selecting the computational domain and ensuring the effectiveness and representativeness of the proxy quantity.

[0057] By introducing such physics-driven constraints, the model is able to maintain high-accuracy prediction capabilities even in the absence of aerodynamic labels.

[0058] like Figure 7 The diagram illustrates the structure of the joint loss function during the training phase of this invention. This loss function consists of three parts that collectively guide model optimization: : Wind pressure coefficient prediction error term, measuring the accuracy of model predictions The difference between the actual value and the true value; Overall aerodynamic prediction error term, assessment and The accuracy, among which These are the weighting coefficients; The physical consistency constraint term is output from the network's physical proxy prediction header. Physical targets obtained based on online calculations of the input velocity field The mean square error between them constitutes the total loss function, which is then combined with the wind pressure coefficient loss and aerodynamic force loss according to weights. These three loss components together constitute the total loss function, which updates the network parameters through the backpropagation algorithm. Specifically, the physical auxiliary loss... It does not participate in the final output, but it can effectively guide the model to learn an internal representation that conforms to the principles of fluid mechanics, thereby improving the stability and generalization ability of the prediction.

[0059] This invention fully utilizes the physical linear relationship between wind pressure and aerodynamic force. The linear mapping between the wind pressure coefficient and the overall aerodynamic force is explicitly introduced into the model structure, and the mapping coefficient is automatically learned by the model during training. This avoids the unnecessary complexity brought about by traditional nonlinear black-box models and improves the physical interpretability and stability of the prediction results.

[0060] This invention reduces reliance on measured aerodynamic data. By introducing physical surrogate quantities such as vorticity intensity and momentum deficit, which are directly calculated from the flow field, as auxiliary constraints, this invention can achieve high-precision prediction without the need to collect overall aerodynamic forces as the sole monitoring signal, thus significantly reducing experimental costs.

[0061] This invention offers high prediction efficiency and strong engineering applicability. It eliminates the need for full-wall pressure integration or high-resolution numerical simulation, enabling aerodynamic prediction in a short time and making it suitable for rapid evaluation and parametric analysis across multiple engineering conditions.

[0062] The model of this invention has good generalization ability and scalability. The method can be applied to different building cross-sectional forms and incoming flow conditions, and the physical consistency constraint does not depend on specific structural parameters, thus having good potential for promotion.

[0063] This invention proposes a collaborative mechanism of SSM spatiotemporal coding and physical constraint loss: SSM2D Block accurately characterizes the unsteady characteristics of the flow field with linear complexity; the ternary loss function embeds the physical laws of vorticity / momentum loss into the training target, realizing a dual-drive of 'data-driven + physical guidance'. No schemes using flow field differential / integral surrogate quantities as self-supervised constraints for aerodynamic prediction have been found in the comparative documents.

[0064] Innovative spatiotemporal coding structure: SSM2D Block uses a long kernel (31) depthwise separable convolution decoupled by H / W dimensions to accurately model the unsteady evolution of the trail with O(N) complexity, overcoming the shortcomings of CNN's limited receptive field and RNN's inefficient temporal modeling, and providing a new paradigm for flow field spatiotemporal coding.

[0065] Physical constraint mechanism: For the first time, vorticity RMS (differential surrogate quantity) and momentum loss (integral surrogate quantity) are embedded as differentiable self-supervised signals into the loss function to achieve physical consistency training without aerodynamic labels, thus solving the core pain point of weak physical interpretability of existing data-driven methods.

[0066] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement, characterized in that, Includes the following steps: The temporal difference is calculated based on the particle image velocity flow field of adjacent moments around the building, and the channel dimension stitching method is used to combine the velocity field components of adjacent moments with the temporal difference to obtain a multi-channel input tensor. The multi-channel input tensor is spatiotemporally encoded using a state-space model. The flow field features are obtained by serializing and mixing the input features through one-dimensional depthwise separable convolutions performed along the height and width directions, respectively. Based on the aforementioned flow field characteristics, a decoding and mapping process is performed to obtain the wind pressure coefficient at several discrete pressure measurement points on the building cross section. The wind pressure coefficient is subjected to affine transformation using an aerodynamic mapping module with learnable mapping parameters to obtain the overall lift coefficient and drag coefficient of the building.

2. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement as described in claim 1, characterized in that, The process of combining the velocity field components at adjacent time points with the temporal difference using the channel-dimensional stitching method to obtain a multi-channel input tensor includes: The first and second velocity fields obtained from the particle image velocimetry flow field are processed at the corresponding spatial grid positions using point-by-point difference operation to obtain temporal difference; A multi-channel input tensor is obtained by combining two components of the velocity field at the first time step, two components of the velocity field at the second time step, and at least one component of the temporal difference using a channel dimension stacking method.

3. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 1, characterized in that, The process of obtaining flow field features by spatiotemporally encoding the multi-channel input tensor using a state-space model includes: A convolutional downsampling layer is used to process the multi-channel input tensor to obtain a dimension-reduced expanded feature map; The cascaded two-dimensional state-space coding module is used to process the dimensionality-reduced extended feature map. The input features are serialized and mixed by one-dimensional depthwise separable convolution along the height direction and one-dimensional depthwise separable convolution along the width direction, respectively. The mixing results of the two directions are fused and output through residual connection to obtain the flow field features characterizing the unsteady evolution of the wake.

4. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 3, characterized in that, The process of serializing and mixing the input features using cascaded two-dimensional state-space coding modules includes: Normalized features are obtained by processing the input features using a group normalization layer; The normalized features are processed by a channel-expanding convolutional layer to obtain expanded features; The extended features are processed using a gated branch structure. First branch features and second branch features are generated by parallel pointwise convolution. The second branch features are processed by a nonlinear activation function to obtain gate weights. The first branch features and the gate weights are multiplied element-wise to obtain gated modulation features. Based on the gated modulation features, one-dimensional depthwise separable convolutions are performed along the height direction and along the width direction to obtain mixed features in the two directions. The mixed features in the two directions are then fused and processed by a channel compression convolutional layer, and finally added to the input features through residual connections to obtain the output features.

5. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 4, characterized in that, The one-dimensional depthwise separable convolution along the height direction is achieved by merging the width dimension and batch dimension of the input features and then performing a one-dimensional depthwise convolution operation with a kernel length of not less than 17 along the height direction. The one-dimensional depthwise separable convolution along the width direction is achieved by merging the height dimension and batch dimension of the input features and then performing a one-dimensional depthwise convolution operation with a kernel length of not less than 17 along the width direction.

6. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 1, characterized in that, The process of obtaining the wind pressure coefficient at several discrete pressure measurement points on a building section based on the aforementioned flow field characteristics includes: A global average pooling layer is used to compress the flow field features in both the spatial height and spatial width dimensions to obtain a feature vector. A multi-layer fully connected network is used to perform layer-by-layer nonlinear mapping on the feature vector to obtain a wind pressure coefficient vector with the number of output nodes equal to the number of discrete pressure measurement points.

7. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 1, characterized in that, The process of obtaining the overall lift and drag coefficients of a building by performing an affine transformation on the wind pressure coefficient using an aerodynamic mapping module with learnable mapping parameters includes: The wind pressure coefficients at each discrete pressure measurement point are constructed into a wind pressure coefficient vector. The wind pressure coefficient vector is subjected to matrix multiplication using a learnable mapping coefficient matrix to obtain a weighted result; The weighted results are summed using a learnable bias vector to obtain a two-dimensional output vector containing the lift coefficient and drag coefficient.

8. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 7, characterized in that, The size of the mapping coefficient matrix is ​​two rows multiplied by the number of discrete pressure measurement points. The first row of elements corresponds to the contribution weight of the wind pressure coefficient of each pressure measurement point to the lift coefficient, and the second row of elements corresponds to the contribution weight of the wind pressure coefficient of each pressure measurement point to the drag coefficient. The bias vector has a length of two, corresponding to the bias terms of the lift coefficient and the drag coefficient, respectively.

9. The method for predicting building wind pressure and aerodynamics based on PIV flow field spatiotemporal coding and physical enhancement according to claim 1, characterized in that, The encoding parameters, decoding mapping parameters, and mapping parameters of the aerodynamic mapping module of the state space model are optimized and determined by a joint loss function. The joint loss function is composed of the error term between the predicted and measured values ​​of the wind pressure coefficient, and the physical consistency constraint term constructed by the root mean square value of vorticity and the wake momentum deficit value obtained by online calculation of the input velocity field.

Images