High-precision wind tunnel body measurement method

Abstract

The invention provides a high-precision wind tunnel body measurement method, which belongs to the technical field of wind tunnel body measurement, and comprises the following steps of: starting CUDA (Compute Unified Device Architecture) flows matched with the number of sensors, establishing three layers of sub-thread grids, and respectively executing data preprocessing, space geometric reconstruction and geometric parameter calculation; a three-dimensional point cloud reconstruction network based on a ViT architecture is adopted to realize accurate hole body three-dimensional modeling, a hole body three-dimensional coordinate matrix is processed into a geometric stable matrix, a geometric first-level change matrix and a geometric second-level change matrix in a layered mode according to deformation amplitudes, and a mixed storage strategy of GPU video memory buffering and CPU memory asynchronous transmission is adopted. And finally, generating a hole body measurement report containing the maximum deviation position, deviation statistical distribution and deformation trend analysis. The technical problems of system response delay and insufficient processing capability, which are easily caused by the fact that mass data generated by high-precision measurement needs to be processed in real time, are solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wind tunnel body measurement, and particularly relates to a high-precision wind tunnel body measurement method. BACKGROUND

[0002] As an important test link in the field of aerospace, the traditional technology mainly uses laser displacement sensors and three-coordinate measuring machines to detect the geometric accuracy of the tunnel body. In the existing wind tunnel body measurement system, a single processor is usually used for data acquisition and processing, and the traditional serial calculation method is used to realize the three-dimensional reconstruction of the tunnel body. With the improvement of measurement accuracy, modern wind tunnel measurement systems need to arrange a large number of high-precision sensors on the inner wall surface of the tunnel, and the sampling frequency is usually above 1000Hz, and the measurement accuracy requirement reaches 0.1mm level. In the current wind tunnel body measurement technology, due to the large number of sensors and the extremely high sampling frequency, the data generated by the system increases exponentially, and the traditional serial processing method cannot meet the real-time requirement. The data processing flow in the prior art is usually to collect all sensor data first, and then process in batches through a single processor. This method is prone to cause system response delay when facing massive data, and the processing efficiency is low, and even the processing capacity is seriously insufficient. The traditional technology faces serious computing bottleneck when processing large-scale sensor array data, and cannot realize real-time processing and analysis. That is to say, in the prior art, there is a technical problem that the massive data generated by high-precision measurement need real-time processing, which is prone to cause system response delay and insufficient processing capacity. SUMMARY

[0003] Therefore, the present application provides a high-precision wind tunnel body measurement method, which can solve the technical problem that the massive data generated by high-precision measurement need real-time processing in the prior art, which is prone to cause system response delay and insufficient processing capacity.

[0004] This invention is implemented as follows: A method for high-precision wind tunnel measurement includes: arranging multiple laser displacement sensors on the inner wall surface of the wind tunnel to form a three-dimensional measurement grid; initiating a sufficient number of CUDA streams to match the number of sensors, with each CUDA stream initiating a three-layer sub-thread grid; the laser displacement sensors acquiring distance data from the inner wall of the wind tunnel in real time; the first-layer sub-thread grid performing noise filtering and temperature compensation processing on the raw distance data; the second-layer sub-thread grid inputting the processed distance data and sensor spatial position information into a spatial geometric reconstruction model; the spatial geometric reconstruction model automatically adjusting the feature fusion weights based on three parameters: the degree of wind tunnel deformation, the density of measurement points, and data stability, to reconstruct the three-dimensional spatial coordinates of the wind tunnel; the third-layer sub-thread grid... The system calculates the internal diameter deviation, ellipticity, and surface roughness of the tunnel using the reconstructed 3D spatial coordinates, and calculates the tunnel deformation level using a deformation evaluation function. The CPU receives the processing results from each CUDA stream, spatially stitches all measurement data at the same time, establishes a complete 3D model of the tunnel, constructs a 3D coordinate matrix of the tunnel based on the 3D spatial coordinates in the 3D model, and obtains the stability threshold, first variation threshold, and second variation threshold through game theory optimization calculation. The 3D coordinate matrix of the tunnel is then processed in layers according to the deformation amplitude through multi-level decomposition to obtain the geometric stability matrix, first-level geometric variation matrix, and second-level geometric variation matrix, and the overall geometric accuracy parameters of the tunnel are calculated. The 3D model of the tunnel is compared with the standard design model, and a tunnel measurement report is output.

[0005] Specifically, the arrangement steps of the laser displacement sensor involve setting up multiple measurement sections along the axial direction of the tunnel, with multiple sensors evenly distributed around the circumference of each measurement section. The laser displacement sensor uses the triangulation principle, projecting a laser beam onto the inner wall surface of the tunnel and receiving the reflected light signal to calculate the distance.

[0006] Specifically, the CUDA stream startup step involves using multi-batch processing if the number of sensors exceeds the number of CUDA streams. This means grouping all sensors according to their spatial location, with each group containing no more than the number of CUDA streams, and processing the sensor data of each group sequentially in chronological order to ensure that all measurement data is processed completely.

[0007] Specifically, the step of the laser displacement sensor acquiring distance data in real time involves the first sub-thread grid performing noise filtering and temperature compensation on the raw distance data, and then transmitting the processed data to the second sub-thread grid.

[0008] The specific structure of the spatial geometric reconstruction model is a three-dimensional point cloud reconstruction network based on the ViT architecture, which includes a point cloud segmentation module, a multi-head attention mechanism, and a spatial coordinate regression module. The point cloud segmentation module divides the measured point cloud data into multiple blocks according to spatial location. The multi-head attention mechanism calculates the spatial correlation between different measurement points. The spatial coordinate regression module outputs accurate three-dimensional coordinate values.

[0009] The deformation evaluation function is used to calculate the deformation level value of the tunnel. The inputs include the degree of tunnel deformation, the density of measurement points, the data stability, the temperature compensation coefficient, and the noise filtering coefficient. The output is the deformation level value of the tunnel. When the deformation level value of the tunnel is in different ranges, the spatial geometric reconstruction model adopts different processing methods.

[0010] Specifically, the step of the CPU receiving the processing results of each CUDA stream involves the GPU memory first storing the processing results of each CUDA stream in the memory buffer, the CPU copying the data in the memory to the system memory through asynchronous memory transfer, and the CPU sorting and splicing the measurement data from different CUDA streams in the system memory according to spatial coordinates to form a complete three-dimensional coordinate dataset of the cave.

[0011] The game theory model includes an upper-level model with the goal of optimizing geometric accuracy and a lower-level model with the goal of optimizing computational efficiency. The objective function of the upper-level model is to minimize the product of the sum of squares of the geometric deviation of the hole and the variance of the measurement noise plus a geometric consistency constraint term. The objective function of the lower-level model is to minimize the product of the computation time and memory usage minus the logarithmic function of the data processing speed. The two objective functions influence each other through a threshold optimization coupling term.

[0012] The objective function inputs of the upper-level model include the sum of squares of the geometric deviations of the cavity, the variance of the measurement noise, the geometric consistency constraint, the data sampling density, and the sensor calibration accuracy. The objective function inputs of the lower-level model include the computation time, memory usage, data processing speed, parallel computing efficiency, and data transmission bandwidth.

[0013] The geometric stability matrix is ​​used to represent the stable part of the cave's geometry, and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is less than the stability threshold. The first-order geometric variation matrix is ​​used to represent the main variation part of the cave's geometry, and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between the stability threshold and the first variation threshold. The second-order geometric variation matrix is ​​used to represent the secondary variation part of the cave's geometry, and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between the first variation threshold and the second variation threshold.

[0014] Optionally, the thresholds can also be determined empirically: the stability threshold is used to distinguish between the geometrically stable matrix and the first-order geometric variation matrix, and its value is 0.8 times the average value of the cave deformation; the first variation threshold is used to distinguish between the first-order geometric variation matrix and the second-order geometric variation matrix, and its value is 1.5 times the average value of the cave deformation; the second variation threshold is used to distinguish between the second-order geometric variation matrix and the noise component, and its value is 2.5 times the average value of the cave deformation.

[0015] The degree of deformation of the cave body is obtained by calculating the root mean square error between the actual measured coordinates and the theoretical design coordinates, with a value range of 0 to 10 mm. The density of measurement points is obtained by calculating the number of effective measurement points per unit area, with a value range of 1 to 100 per square meter. The data stability is obtained by calculating the standard deviation of 10 consecutive measurement results, with a value range of 0 to 1 mm.

[0016] The temperature compensation coefficient is calculated by linearly relating the difference between the ambient temperature and the standard temperature of 20°C, and its value ranges from 0.9 to 1.1. The noise filtering coefficient is calculated by logarithmically relating the signal-to-noise ratio, and its value ranges from 0.1 to 1.0.

[0017] The steps for establishing the training dataset for the spatial geometry reconstruction model specifically include collecting actual measurement data from 1000 wind tunnels of different sizes as input samples, obtaining corresponding standard three-dimensional coordinates as label data using a high-precision coordinate measuring machine, adding different levels of noise and temperature drift to the measurement data to simulate actual working conditions, and establishing a dataset containing 500,000 training samples. The training steps for the spatial geometry reconstruction model specifically include end-to-end training using the Adam optimizer, with a learning rate of 0.001, a batch size of 64, 200 training epochs, and a weighted combination of mean squared error loss and geometric consistency loss with a weight ratio of 3:1. Data augmentation techniques, including random rotation, translation, and scaling transformations, are used during training to improve the model's generalization ability.

[0018] The memory allocation strategy between the CPU and GPU is as follows: 80% of the GPU memory is allocated to store real-time measurement data and spatial geometric reconstruction model parameters, and 20% is used for intermediate calculation result caching. 60% of the CPU memory is allocated to store historical measurement data, and 40% is used for the final 3D model and analysis results. Data transmission adopts an asynchronous copy method to avoid synchronization waiting between the CPU and GPU.

[0019] This invention, by employing a CUDA streaming parallel processing architecture, transforms the traditional serial data processing method into a multi-stream parallel processing mode, effectively solving the computational bottleneck problem of real-time processing of massive data. This invention overcomes the inefficiency caused by the serialization of data processing flows in traditional technologies by establishing a three-layer sub-thread mesh processing structure, executing data preprocessing, 3D reconstruction, and geometric analysis in parallel at different levels. Simultaneously, through the asynchronous transmission mechanism between GPU memory and CPU memory, the risks of system response latency and insufficient processing capacity in traditional technologies are avoided. In summary, this invention, through multi-stream parallel processing and layered asynchronous transmission technology, solves the technical problems mentioned in the background art, such as the need for real-time processing of massive amounts of data generated by high-precision measurements, which easily leads to system response latency and insufficient processing capacity. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a schematic diagram of the neural network structure of the spatial geometry reconstruction model involved in the present invention.

[0022] Figure 3 This is a diagram showing the distribution of cavity deviation in the embodiment.

[0023] Figure 4 This is a diagram showing the axial deformation trend of the cavity in the embodiment.

[0024] Figure 5 This is a multi-level geometric accuracy distribution diagram in the embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 The diagram shown is a flowchart of a high-precision wind tunnel measurement method provided by the present invention. This method includes the following steps:

[0027] S01. Multiple laser displacement sensors are arranged on the inner wall surface of the wind tunnel. Each laser displacement sensor has a measurement section set every 50mm along the tunnel axis. Eight sensors are evenly distributed in 360° within each measurement section to form a three-dimensional measurement grid.

[0028] S02. Start a sufficient number of CUDA streams, matching the number of sensors. Start a three-layer sub-thread mesh within each CUDA stream. If the number of sensors exceeds the number of CUDA streams, use multiple batch processing.

[0029] S03. Each laser displacement sensor collects distance data from the inner wall of the cave to the sensor in real time, with a sampling frequency of 1000Hz. The accuracy of the distance data reaches 0.01mm. The first layer of sub-thread mesh performs noise filtering and temperature compensation processing on the original distance data.

[0030] S04. The second-layer sub-thread mesh inputs the processed distance data and sensor spatial location information into the spatial geometric reconstruction model. The spatial geometric reconstruction model automatically adjusts the feature fusion weights according to three parameters: the degree of cave deformation, the density of measurement points, and the stability of data, and reconstructs the three-dimensional spatial coordinates of the cave.

[0031] S05. The third-layer sub-thread mesh calculates the internal diameter deviation, ellipticity, and surface roughness of the cavity based on the reconstructed three-dimensional spatial coordinates, and calculates the cavity deformation level through the deformation evaluation function. When the cavity deformation level belongs to different ranges, different weight adjustment functions are used to adjust the feature fusion weight parameters of the spatial geometric reconstruction model.

[0032] S06. The CPU receives the processing results of each CUDA stream, spatially stitches together all the measurement data at the same time, establishes a complete three-dimensional model of the cave, constructs a three-dimensional coordinate matrix of the cave based on the three-dimensional spatial coordinates in the three-dimensional model of the cave, and obtains the stability threshold, the first variation threshold and the second variation threshold through game theory optimization calculation. The three-dimensional coordinate matrix of the cave is processed in layers according to the deformation amplitude through multi-level decomposition to obtain the geometric stability matrix, the first-level geometric variation matrix and the second-level geometric variation matrix, and calculates the overall geometric accuracy parameters of the cave.

[0033] S07. Compare the 3D model of the tunnel with the standard design model, calculate the deviation distribution map, and output the tunnel measurement report, including the location of the maximum deviation, the statistical distribution of the deviation, and the results of the deformation trend analysis.

[0034] The laser displacement sensor uses the triangulation principle. It projects a laser beam onto the surface of the inner wall of the cave and receives the reflected light signal to calculate the distance. The measurement range is from 10mm to 500mm, and the resolution is 0.001mm.

[0035] The multi-batch processing is used to solve the data processing problem when the number of sensors exceeds the number of CUDA streams. All sensors are grouped according to their spatial location, with the number of sensors in each group not exceeding the number of CUDA streams. The data of each group of sensors is processed sequentially according to time order to ensure that all measurement data is processed completely.

[0036] During the process of the CPU receiving the processing results of each CUDA stream, the GPU memory first stores the processing results of each CUDA stream in the memory buffer. The CPU copies the data in the memory to the system memory through asynchronous memory transfer. The CPU sorts and splices the measurement data from different CUDA streams in the system memory according to the spatial coordinates to form a complete three-dimensional coordinate dataset of the cave.

[0037] The game theory model includes an upper-level model that aims for optimal geometric accuracy and a lower-level model that aims for optimal computational efficiency. The objective function of the upper-level model is to minimize the product of the sum of squares of the geometric deviation of the hole and the variance of the measurement noise plus a geometric consistency constraint term. The objective function of the lower-level model is to minimize the product of the computation time and memory usage minus the logarithmic function of the data processing speed. The two objective functions influence each other through a threshold optimization coupling term.

[0038] The objective function input of the upper-level model includes the sum of squares of the geometric deviations of the cavity, the variance of the measurement noise, the geometric consistency constraint term, the data sampling density, and the sensor calibration accuracy. The constraint condition is that the sum of squares of the geometric deviations is less than 0.1 mm. 2 Furthermore, the measurement noise variance is less than 0.01 mm. 2 The objective function input of the lower-level model includes computation time, memory usage, data processing speed, parallel computing efficiency, and data transmission bandwidth. The constraints are that the computation time is less than 10 seconds and the memory usage is less than 8 GB.

[0039] The deformation evaluation function is used to calculate the cave deformation level value. The inputs include the cave deformation degree, measurement point density, data stability, temperature compensation coefficient, and noise filtering coefficient. The output is the cave deformation level value. When the cave deformation level value is in the range of 0 to 0.3, the spatial geometric reconstruction model is processed by a single CPU thread. When the cave deformation level value is in the range of 0.3 to 0.7, the spatial geometric reconstruction model is processed by a first-level CUDA sub-thread. When the cave deformation level value is in the range of 0.7 to 1.0, the spatial geometric reconstruction model is processed by a second-level CUDA sub-thread.

[0040] The specific structure of the spatial geometric reconstruction model is a three-dimensional point cloud reconstruction network based on the ViT architecture, which includes a point cloud segmentation module, a multi-head attention mechanism, and a spatial coordinate regression module. The point cloud segmentation module divides the measured point cloud data into 16×16 blocks according to spatial location. The multi-head attention mechanism calculates the spatial correlation between different measurement points, and the spatial coordinate regression module outputs accurate three-dimensional coordinate values.

[0041] The specific steps for establishing the training dataset of the spatial geometric reconstruction model include collecting actual measurement data from 1,000 wind tunnels of different sizes as input samples, obtaining the corresponding standard three-dimensional coordinates as label data using a high-precision coordinate measuring machine, adding different levels of noise and temperature drift to the measurement data to simulate actual working conditions, and establishing a dataset containing 500,000 training samples.

[0042] The training steps for the spatial geometric reconstruction model specifically include end-to-end training using the Adam optimizer, with a learning rate of 0.001, a batch size of 64, 200 training epochs, and a weighted combination of mean squared error loss and geometric consistency loss with a weight ratio of 3:1. Data augmentation techniques, including random rotation, translation, and scaling transformations, are used during training to improve the model's generalization ability.

[0043] The degree of cave deformation is obtained by calculating the root mean square error between the actual measured coordinates and the theoretical design coordinates, with a value range of 0 to 10 mm. The density of measurement points is obtained by calculating the number of effective measurement points per unit area, with a value range of 1 to 100 per square meter. The data stability is obtained by calculating the standard deviation of 10 consecutive measurement results, with a value range of 0 to 1 mm. The temperature compensation coefficient is obtained by calculating the difference between the ambient temperature and the standard temperature of 20°C according to a linear relationship, with a value range of 0.9 to 1.1. The noise filtering coefficient is obtained by calculating the signal-to-noise ratio according to a logarithmic relationship, with a value range of 0.1 to 1.0.

[0044] The geometric stability matrix represents the stable part of the cave's geometry and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is less than a stability threshold. The first-order geometric variation matrix represents the main variation part of the cave's geometry and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between the stability threshold and a first variation threshold. The second-order geometric variation matrix represents the secondary variation part of the cave's geometry and is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between a first variation threshold and a second variation threshold.

[0045] The stability threshold is used to distinguish between the geometrically stable matrix and the first-order geometric variation matrix, and its value is 0.8 times the average value of the cave deformation. The first variation threshold is used to distinguish between the first-order geometric variation matrix and the second-order geometric variation matrix, and its value is 1.5 times the average value of the cave deformation. The second variation threshold is used to distinguish between the second-order geometric variation matrix and the noise component, and its value is 2.5 times the average value of the cave deformation.

[0046] The memory allocation strategy between the CPU and GPU is as follows: 80% of the GPU memory is allocated to store real-time measurement data and spatial geometric reconstruction model parameters, and 20% is used for intermediate calculation result caching. 60% of the CPU memory is allocated to store historical measurement data, and 40% is used for the final 3D model and analysis results. Data transmission adopts an asynchronous copy method to avoid synchronization waiting between the CPU and GPU.

[0047] The three-dimensional coordinate matrix of the cave is a matrix composed of the three-dimensional spatial coordinates in the three-dimensional model of the cave arranged in spatial order. The rows represent different measurement points, and the columns represent the three coordinate components of x, y, and z. The game model is used to optimize the threshold parameters of the multi-level decomposition. The optimal stable threshold, the first variable threshold, and the second variable threshold are determined by iterative solution of the upper-level model and the lower-level model.

[0048] Specifically, the sum of squares of the geometric deviation of the tunnel body is obtained by calculating the sum of squares of the differences between the actual measured coordinates and the theoretical design coordinates; the measurement noise variance is obtained by calculating the variance of continuous measurement data; the geometric consistency constraint term is obtained by calculating the geometric continuity between adjacent measurement points; the data sampling density is obtained by calculating the ratio of the total number of measurement points to the surface area of ​​the tunnel body; the sensor calibration accuracy is obtained by the measurement error of standard parts; the parallel computing efficiency is obtained by the GPU utilization rate; and the data transmission bandwidth is obtained by the data transmission rate between the CPU and the GPU.

[0049] The specific implementation methods of the above steps are described in detail below.

[0050] The specific implementation of step S01 involves using a laser displacement sensor array based on the triangulation principle to achieve full-coverage measurement of the tunnel's inner wall surface. First, the number of measurement sections is determined based on the axial length of the wind tunnel. These sections are then evenly spaced, with one section every 50mm, ensuring high-precision measurement density along the tunnel's axial direction. Within each measurement section, eight laser displacement sensors are arranged at 360° intervals, with each sensor spaced 45° apart, forming an octagonal measurement array using a circumferential uniform distribution algorithm. The laser displacement sensors use infrared lasers as the light source, emitting laser beams onto the tunnel's inner wall surface. The distance between the laser beam projection point and the sensor is calculated using the triangulation principle. The sensor's measurement range is set from 10mm to 500mm, with a resolution of 0.001mm and a measurement accuracy of 0.01mm. The sensor's installation angle must be perpendicular to the tunnel's inner wall surface, with the deviation angle controlled within ±2° to ensure the accuracy of the measurement data. This arrangement forms a three-dimensional measurement grid covering the entire tunnel's inner wall, providing a sufficient data foundation for subsequent geometric reconstruction.

[0051] The specific implementation of step S02 is to launch multiple unified computing device architecture streams for data processing based on a parallel computing architecture of the graphics processor. First, the total number of laser displacement sensors is counted, and the number of unified computing device architecture streams that can be launched simultaneously is determined based on the number of stream processors and memory capacity of the graphics processor. When the number of sensors is less than or equal to the maximum number of streams supported by the graphics processor, an independent unified computing device architecture stream is allocated to each sensor. When the number of sensors exceeds the number of unified computing device architecture streams, a multi-batch processing strategy is adopted, grouping the sensors according to their spatial location, with the number of sensors in each group not exceeding the number of available unified computing device architecture streams, ensuring that each sensor is effectively processed. Within each unified computing device architecture stream, a three-layer sub-thread mesh structure is launched. The first layer of sub-thread mesh is responsible for data preprocessing, including noise filtering and temperature compensation, with the number of sub-threads set to twice the number of sensors. The second layer of sub-thread mesh is responsible for spatial geometry reconstruction, with the number of sub-threads set to four times the number of sensors. The third layer of sub-thread mesh is responsible for geometric parameter calculation, with the number of sub-threads set to eight times the number of sensors. This parallel computing architecture can fully utilize the computing resources of the graphics processor, significantly improving data processing speed.

[0052] The specific implementation of step S03 involves obtaining accurate distance measurement data through high-frequency sampling and real-time data processing. A laser displacement sensor continuously collects distance data from the inner wall of the cave to the sensor at a sampling frequency of 1000Hz, ensuring that minute deformations of the cave are captured. The first layer of sub-thread mesh preprocesses the collected raw distance data. First, a Kalman filter algorithm is applied to filter noise in the measurement data, with the process noise covariance and observation noise covariance set to 0.001 and 0.01 respectively in the filtering parameters. Temperature compensation processing uses a linear compensation model, calculating the temperature compensation coefficient based on the difference between the ambient temperature and the standard temperature of 20℃. The compensation coefficient ranges from 0.9 to 1.1, with a change in the compensation coefficient of 0.005 for every 1℃ change in temperature. A sliding window algorithm is used during data processing, with a window length of 10 sampling points. The stability of the data is determined by calculating the mean and standard deviation of the data within the window. When the standard deviation of the continuous measurement data is less than 0.05mm, the data is considered stable; otherwise, outlier detection and removal are performed. The purpose of this step is to obtain high-quality raw measurement data to provide reliable input for subsequent geometric reconstruction.

[0053] The specific implementation of step S04 is to achieve accurate reconstruction of the three-dimensional spatial coordinates of the cave through a three-dimensional point cloud reconstruction network based on a vision transformer architecture. The second-layer sub-thread mesh inputs preprocessed distance data and sensor spatial location information into the spatial geometric reconstruction model. The model first divides the measured point cloud data into 16×16 block regions according to spatial location using a point cloud segmentation module. Each block contains 256 measurement points, facilitating parallel processing. A multi-head attention mechanism calculates the spatial correlation between different measurement points, with 8 attention heads and each attention head having a dimension of 64. The feature fusion weights are automatically adjusted based on three parameters: the degree of cave deformation, the density of measurement points, and data stability. The deformation degree accounts for 40% of the weight, the measurement point density accounts for 35%, and the data stability accounts for 25%. When the cave deformation is large, the weight of the deformation degree is increased; when the measurement point density is high, the weight of the point density is increased. The spatial coordinate regression module adopts a multilayer perceptron structure, containing three hidden layers with 512, 256, and 128 neurons per layer, respectively, ultimately outputting accurate three-dimensional coordinate values. This reconstruction process can effectively handle uncertainties and noise in sensor data and generate high-precision three-dimensional spatial coordinates.

[0054] The specific implementation of step S05 is based on calculating the key geometric parameters and deformation level of the cavity using reconstructed three-dimensional spatial coordinates. The third-layer sub-thread mesh receives the three-dimensional coordinate data output by the spatial geometric reconstruction model. First, it calculates the deviation of the cavity's inner diameter by fitting the circular contour of each measurement section to calculate the deviation between the actual and designed inner diameter. Ellipticity is calculated by fitting an ellipse using the least squares method, calculating the ratio difference between the major and minor axes. Surface roughness is obtained by calculating the root mean square deviation of the measurement points relative to the fitted surface. The deformation evaluation function uses a weighted comprehensive evaluation method. Input parameters include the degree of cavity deformation, measurement point density, data stability, temperature compensation coefficient, and noise filtering coefficient. After normalization, each parameter is mapped to the range of 0 to 1, and then weighted and summed according to a weight ratio of 3:2:2:1:1 to obtain the cavity deformation level value. When the deformation level value is in the range of 0 to 0.3, it indicates that the cavity has almost no deformation, and the spatial geometric reconstruction model adopts a single-threaded processing mode of the central processing unit. When the deformation level value is between 0.3 and 0.7, it indicates that the cavern has slight deformation, and the first-level unified computing device architecture sub-thread processing mode is adopted. When the deformation level value is between 0.7 and 1.0, it indicates that the cavern has severe deformation, and the second-level unified computing device architecture sub-thread processing mode is adopted to improve calculation accuracy and efficiency.

[0055] The specific implementation of step S06 involves establishing a complete 3D model of the cave through asynchronous memory transfer and game theory model optimization. The graphics processing unit (GPU) first stores the processing results of each unified computing architecture stream in a memory buffer, with the buffer size set to 20% of the total GPU memory. The central processing unit (CPU) copies the data from the GPU to system memory using asynchronous memory transfer technology, employing a double-buffering mechanism to avoid data transfer blocking the computation process. In system memory, the CPU sorts and concatenates the measurement data from different unified computing architecture streams according to spatial coordinates, using a quicksort algorithm to arrange them in priority order of x, y, and z coordinates, forming a complete 3D coordinate dataset of the cave. A 3D coordinate matrix of the cave is constructed based on the 3D spatial coordinates in the 3D model, where rows represent different measurement points and columns represent the x, y, and z coordinate components. The game theory model employs a two-layer optimization structure. The upper-layer model aims for optimal geometric accuracy, with the objective function being the minimization of the product of the sum of squared geometric deviations of the cave and the variance of the measurement noise, plus a geometric consistency constraint term. The constraint condition is that the sum of squared geometric deviations is less than 0.1 mm. 2 Furthermore, the measurement noise variance is less than 0.01 mm. 2 The lower-level model aims for optimal computational efficiency. The objective function is the minimization of the product of computation time and memory usage minus the logarithm of data processing speed. The constraints are that the computation time is less than 10 seconds and the memory usage is less than 8GB. The optimal stable threshold, the first variable threshold, and the second variable threshold are obtained through iterative solutions. The stable threshold is set to 0.8 times the average deformation degree of the cave, the first variable threshold is set to 1.5 times the average deformation degree of the cave, and the second variable threshold is set to 2.5 times the average deformation degree of the cave.

[0056] The specific implementation of step S07 involves generating a tunnel measurement report through model comparison and deviation analysis. The reconstructed 3D tunnel model is geometrically compared with a pre-established standard design model. A nearest-point iteration algorithm is used for model registration, ensuring the two models are compared within the same coordinate system. The Euclidean distance between each measurement point and its corresponding standard design point is calculated, generating deviation value distribution data. Statistical analysis methods are used to determine the location of the maximum deviation, and the mean, standard deviation, maximum, and minimum deviation values ​​are calculated to generate a deviation statistical distribution map. Deformation trend analysis employs time series analysis, performing regression fitting on historical measurement data to predict the development trend of tunnel deformation. The final output is a comprehensive measurement report including measurement accuracy assessment, geometric parameter statistics, deviation distribution visualization, and deformation trend prediction, providing a scientific basis for tunnel maintenance and quality control.

[0057] The detailed structure of the spatial geometric reconstruction model is based on a 3D point cloud reconstruction network with a visual transformer architecture. The overall architecture consists of four main parts: an input encoding layer, a feature extraction layer, an attention fusion layer, and a coordinate regression layer. The input encoding layer receives distance data and spatial position information from a laser displacement sensor, converting the original distance values ​​and position coordinates into high-dimensional feature vectors with an encoding dimension of 256. The point cloud segmentation module normalizes the input measurement point cloud data according to spatial position, using voxelization to divide the 3D space into a 16×16×16 grid. Each grid cell contains an average of 16 measurement points, forming a total of 4096 feature blocks. The feature extraction layer employs a multi-layer convolutional neural network structure, containing four convolutional layers with kernel sizes of 3×3×3, 5×5×5, 3×3×3, and 1×1×1, all with a stride of 1. Symmetrical padding is used, and the activation function is a modified linear unit function. The multi-head attention mechanism is the core component of the model, comprising eight parallel attention heads. Each attention head independently computes the query, key-value, and numerical matrices. The query and key-value matrices are 64×64 in dimension, and the numerical matrix is ​​64×256 in dimension. Attention weights are calculated using a scaled dot product attention mechanism, with the scaling factor set to the reciprocal of the square root of the query matrix dimension. The outputs of the multi-head attention are fused through linear transformation and residual connections to obtain the fused feature representation. The spatial coordinate regression module employs a fully connected neural network structure, containing three hidden layers with 512, 256, and 128 neurons respectively. Batch normalization layers and random deactivation layers are added between each layer, with the random deactivation probability set to 0.2. The final output layer contains three neurons, corresponding to the regression values ​​of the x, y, and z coordinate components, respectively.

[0058] The detailed steps for establishing the training dataset began with the collection of 1000 wind tunnels of different sizes and types through field surveys. These included low-speed, high-speed, and supersonic wind tunnels, with diameters ranging from 0.5m to 10m and lengths from 2m to 50m. A high-precision coordinate measuring machine (CMM) was used to perform standard measurements on each wind tunnel, achieving an accuracy of 0.001mm, obtaining the standard three-dimensional coordinates of the tunnel's inner wall as training labels. For each wind tunnel, multiple measurements were performed under different environmental conditions, including varying temperatures, humidity, and vibration levels, simulating the actual measurement environment. Data augmentation techniques were used to expand the training samples. Different levels of Gaussian noise were added to the original measurement data, with a standard deviation ranging from 0.001mm to 0.1mm, simulating sensor measurement noise. A temperature drift effect was also added to simulate the impact of environmental temperature changes on the measurement results, with the temperature range from -20℃ to 60℃. The measurement data were subjected to random rotation, translation, and scaling transformations, with rotation angles ranging from -180° to 180°, translation distances ranging from -10mm to 10mm, and scaling ratios ranging from 0.9 to 1.1, to improve the model's adaptability to different installation conditions. Through the above data processing and enhancement techniques, a large-scale dataset containing 500,000 training samples was finally established. Each sample contains the input sensor measurement data and the corresponding standard 3D coordinate labels.

[0059] The model training process employed an end-to-end deep learning approach, using an adaptive moment estimator (IME) optimizer for parameter optimization. The initial learning rate was set to 0.001, and a cosine annealing learning rate scheduling strategy was adopted, with a minimum learning rate of 0.00001. The batch size was set to 64 to ensure the stability and convergence of the training process, with a total of 200 training epochs. The loss function used a weighted combination of mean squared error loss and geometric consistency loss, with a weight ratio of 3:1. The mean squared error loss constrained the difference between the predicted coordinates and the standard coordinates, while the geometric consistency loss constrained the geometric continuity between adjacent measurement points. An early stopping strategy was employed during training; training was stopped when the validation set loss showed no improvement for 10 consecutive epochs to prevent overfitting. Model training was conducted on a high-performance computing platform equipped with eight GPUs, with a total training time of approximately 72 hours. The final model achieved an average measurement error of less than 0.005 mm on the test set, meeting the requirements for high-precision measurement.

[0060] The key technical ideas of this invention are mainly reflected in four aspects: arrayed arrangement of laser sensors and parallel data processing, 3D geometric reconstruction based on vision transformers, multi-level adaptive deformation analysis, and parameter optimization driven by game theory models. The arrayed arrangement of laser sensors and parallel data processing technology establishes a high-density 3D measurement grid on the inner wall surface of the wind tunnel, combined with the large-scale parallel computing capabilities of graphics processors, to achieve real-time measurement of the entire wind tunnel. Compared with traditional point-by-point measurement methods, this technology can simultaneously acquire data from thousands of measurement points, improving measurement efficiency by two orders of magnitude. Simultaneously, parallel data processing significantly reduces data processing time, providing a technical foundation for real-time monitoring. The 3D geometric reconstruction technology based on vision transformers introduces the attention mechanism from deep learning into the 3D point cloud reconstruction task. Through a multi-head attention mechanism, it automatically learns the spatial correlation between different measurement points, effectively handling noise and uncertainty in sensor data. Compared with traditional geometric fitting methods, this technology has stronger generalization ability and noise resistance, maintaining high-precision reconstruction results in complex measurement environments. Multi-level adaptive deformation analysis technology employs different calculation strategies based on the degree of cave deformation. It automatically determines the deformation level of the cave through a deformation evaluation function and adjusts the allocation of computational resources and algorithm parameters accordingly. This technology achieves a dynamic balance between computational accuracy and efficiency, avoiding the waste of computational resources or insufficient accuracy caused by fixed algorithm parameters in traditional methods. Game-theoretic model-driven parameter optimization technology considers both measurement accuracy and computational efficiency simultaneously through a two-layer optimization structure. It finds the optimal system parameter configuration through iterative solutions of the upper and lower-level models, achieving an effective solution to multi-objective optimization problems.

[0061] The synergistic effect of these four key technological approaches forms a complete high-precision wind tunnel measurement system, with each technology supporting and promoting the others. The arrayed arrangement of laser sensors provides high-quality input data for the vision transducer's reconstructed model, significantly improving reconstruction accuracy. Geometric reconstruction technology based on the vision transducer provides accurate geometric parameters for multi-level deformation analysis, making deformation level judgment more reliable. Multi-level adaptive deformation analysis provides important constraints for the game theory model, making parameter optimization more precise. The optimization results of the game theory model, in turn, guide the adjustment of sensor placement and data processing strategies, forming a closed-loop optimization system. This synergy enables the entire measurement system to maintain high precision while possessing good real-time performance and adaptability. Compared with existing technologies, it has significantly improved in measurement accuracy, measurement efficiency, and system stability, providing advanced technical means for precision measurement and quality control in wind tunnel engineering.

[0062] It should be noted that this invention also solves the following three technical problems: First, the technical problem of insufficient accuracy in the three-dimensional reconstruction of wind tunnels in existing technologies. Traditional tunnel measurement techniques mainly rely on single-point measurements or interpolation calculations using a small number of measurement points. Due to the sparse measurement points and lack of effective spatial correlation analysis, the accuracy of the reconstructed three-dimensional model is limited and cannot accurately reflect the true geometric shape of the tunnel. This invention establishes a three-dimensional point cloud reconstruction network based on the ViT architecture, divides the measurement point cloud data into 16×16 blocks according to spatial location, uses a multi-head attention mechanism to calculate the spatial correlation between different measurement points, and outputs accurate three-dimensional coordinate values ​​through a spatial coordinate regression module, effectively improving the accuracy and reliability of the three-dimensional reconstruction of the tunnel. Second, the technical problem of wasted computing resources due to the lack of adaptive processing strategies in existing technologies. Traditional tunnel measurement systems usually adopt a fixed processing flow, using the same computing resource configuration regardless of the degree of tunnel deformation. This approach wastes computing resources when processing tunnels with mild deformation, while it may result in insufficient computing power when processing tunnels with severe deformation. This invention establishes a deformation evaluation function to calculate the deformation level value of the cave based on parameters such as the degree of cave deformation, the density of measurement points, and data stability. When the deformation level value falls within different ranges, it automatically selects the corresponding processing strategy, moving from single-threaded CPU processing to multi-layered CUDA sub-thread processing, thus achieving intelligent allocation and optimized utilization of computing resources. Third, it addresses the technical problem of existing cave geometric analysis methods lacking hierarchical processing capabilities. Traditional cave analysis methods typically treat all measurement data as equally important, lacking the ability to process regions with different deformation levels in a hierarchical manner. This approach cannot effectively distinguish between stable and deformable regions of the cave, affecting the accuracy and reliability of the measurement results. This invention optimizes the calculation of a stability threshold, a first variation threshold, and a second variation threshold using a game theory model. It then processes the cave's three-dimensional coordinate matrix into a geometrically stable matrix, a first-level geometric variation matrix, and a second-level geometric variation matrix according to the deformation amplitude, achieving refined analysis of regions with different deformation levels and improving the accuracy of cave geometric precision assessment.

[0063] Specifically, the principle of this invention is as follows: The core principle behind the technical solution of this invention in solving the problem of real-time processing of massive amounts of data lies in the adoption of a hybrid computing mode combining GPU parallel computing architecture and CPU collaborative processing. Traditional serial processing methods, when faced with high-frequency data generated simultaneously by a large number of sensors, require the processor to process the data from each sensor one by one, creating a severe computational bottleneck. This invention, however, achieves parallel data processing by initiating CUDA streams matching the number of sensors. Each CUDA stream independently processes the data from its corresponding sensor, thereby transforming the processing time from the traditional accumulation mode to a parallel mode. The hierarchical processing architecture of this invention further improves processing efficiency. The first-layer sub-thread is responsible for basic preprocessing such as noise filtering and temperature compensation; the second-layer sub-thread performs spatial geometry reconstruction; and the third-layer sub-thread performs geometric parameter calculations. This pipeline-like hierarchical processing ensures the continuity and efficiency of data processing. Simultaneously, through threshold settings optimized by a game theory model, the system can adaptively adjust the processing strategy according to the degree of cave deformation, optimizing the allocation of computing resources while ensuring accuracy. In terms of memory management, this invention employs a GPU memory buffer mechanism and asynchronous memory transfer technology, avoiding the system response latency problem caused by insufficient CPU processing power in traditional technologies. GPU memory is used to store real-time data and model parameters, while CPU memory is used to store historical data and final results. This separate storage strategy effectively avoids memory contention and processing bottleneck risks. The asynchronous transfer mechanism ensures efficient data flow between the GPU and CPU, avoiding performance bottlenecks caused by synchronous waiting.

[0064] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0065] The specific implementation of step S01 is to achieve full coverage measurement of the inner wall surface of the tunnel by arranging a laser displacement sensor array based on the triangulation principle. The formula for calculating the sensor arrangement density is:

[0066]

[0067] In the formula, N sensor L represents the total number of sensors, in units. tunnel d represents the axial length of the wind tunnel body, in mm. interval For measuring the cross-sectional spacing, a value of 50 mm is used; N circle The number of sensors for each measurement section is set to 8. The sensor spatial coordinates are represented as follows:

[0068]

[0069] In the formula, P i，j r represents the spatial coordinates of the j-th sensor at the i-th cross section, in mm. tunnelWhere is the radius of the tunnel, in mm; i is the section number. j is the sensor number, j = 0, 1, 2, ..., 7.

[0070] The specific implementation of step S02 is to initiate multiple unified computing device architecture streams for data processing based on a parallel computing architecture of a graphics processor. The stream allocation strategy calculation formula is as follows:

[0071] N stream =min(N) sensor N max_stream );

[0072] In the formula, N stream N represents the actual number of streams started. max_stream This represents the maximum number of streams supported by the graphics processor. When N... sensor >N max_stream At that time, multiple batch processing was used, with the following batch number:

[0073]

[0074] In the formula, N batch To determine the number of batches to process; This is the floor function.

[0075] The specific implementation of step S03 involves obtaining accurate distance measurement data through high-frequency sampling and real-time data processing. The temperature compensation calculation formula is as follows:

[0076] D compensated =D raw ×(1+α temp ×(T env -T ref ));

[0077] In the formula, D compensated This is the distance value after temperature compensation, in mm; D raw The original measured distance value is in mm; α temp The temperature compensation coefficient is 0.005 / ℃; T env The ambient temperature is expressed in °C (°C); T ref The reference temperature is set to 20℃. The state update equation for the Kalman filter is:

[0078]

[0079] In the formula, Here is the optimal estimate at time k, in mm; Here is the predicted value at time k, in mm; K k is the Kalman gain matrix, in mm / mm, representing the ratio of estimation error to observation error; zk Let be the observation value at time k, in mm; H is the observation matrix, in mm / mm, used to map the state vector to the observation space. The data stability assessment formula is:

[0080]

[0081] In the formula, σ stability n represents the standard deviation of data stability, in mm. window D is the length of the sliding window, with a value of 10; i This is the i-th measurement value within the window, in mm; This represents the average value measured within the window, in mm. The formula for calculating the degree of deformation of the tunnel is:

[0082]

[0083] In the formula, ξ deformation The degree of deformation of the tunnel body is expressed in mm, with a value range of 0 to 10 mm; n total P represents the total number of measurement points. measured，k P represents the k-th actual measured coordinate, in mm. design，k The coordinates are the corresponding theoretical design coordinates, in mm. The formula for calculating the density of the measurement points is:

[0084]

[0085] In the formula, ρ density The density of the measurement points is expressed in units of points per m. 2 The numerical range is 1 to 100 per m 2 ;n effective Number of valid measurement points; A surface The surface area of ​​the cave's internal walls, in meters. 2 .

[0086] The specific implementation of step S04 is to achieve accurate reconstruction of the three-dimensional spatial coordinates of the cave by using a 3D point cloud reconstruction network based on a vision transformer architecture. The adaptive adjustment formula for feature fusion weights is as follows:

[0087]

[0088] In the formula, w1, w2, and w3 are the dimensionless weights of the cave deformation degree, measurement point density, and data stability, respectively; η1, η2, and η3 are the corresponding weight adjustment factors, dimensionless, with default values ​​of 0.4, 0.35, and 0.25, respectively. The regression equation for spatial coordinate reconstruction is:

[0089]

[0090] In the formula, The reconstructed 3D coordinate vector, in mm; f ViT Reconstruction function for the visual transformer; W is the input feature vector, in mm. attention This is a multi-head attention weight matrix, which is dimensionless.

[0091] The specific implementation of step S05 is based on calculating the key geometric parameters and deformation level of the cavity using reconstructed three-dimensional spatial coordinates. The formula for calculating the internal diameter deviation of the cavity is:

[0092] δ diameter =|D measured -D design |;

[0093] In the formula, δ diameter This refers to the inner diameter deviation, in mm; D measured The actual measured inner diameter is in mm; D design The inner diameter is for design purposes, and the unit is mm. The formula for calculating ellipticity is:

[0094]

[0095] In the formula, ε ellipticity ellipticity, dimensionless; a major The length of the major axis of the fitted ellipse, in mm; h minor The minor axis length of the fitted ellipse is given in mm. The surface roughness calculation formula is:

[0096]

[0097] In the formula, R roughness Surface roughness, in mm; n points P represents the total number of measurement points. k Here are the coordinates of the kth measurement point, in mm; Here are the corresponding fitted surface coordinates, in mm. The formula for calculating the noise filtering coefficient is:

[0098]

[0099] In the formula, κ filter , where is the noise filtering coefficient, dimensionless, with a value range of 0.1 to 1.0; SNR is the signal-to-noise ratio, dimensionless; SNR max The maximum signal-to-noise ratio is set to 100. The deformation evaluation function is:

[0100] G deformation =β1·γ1+γ2·γ2+β3·γ3+β4·γ4+β5·γ5;

[0101] In the formula, G deformationβ1 represents the deformation level of the cave, dimensionless, ranging from 0 to 1; β2, β3, β4, and β5 are weighting coefficients, with values ​​of 0.3, 0.2, 0.2, 0.1, and 0.1 respectively; γ1 represents the normalized degree of cave deformation. Dimensionless; γ2 is the normalized density of the measurement points. Dimensionless; γ3 represents the stability of normalized data. Dimensionless; γ4 is the normalized temperature compensation coefficient. Dimensionless; γ5 is the normalized noise filter coefficient, γ5 = k filter , dimensionless.

[0102] The specific implementation of step S06 involves establishing a complete 3D model of the cave through asynchronous memory transfer and game theory model optimization. The 3D coordinate matrix of the cave is constructed as follows:

[0103]

[0104] In the formula, M coordinate The three-dimensional coordinate matrix of the cave; x i y i , z i Let n be the three-dimensional coordinates of the i-th measurement point, in mm; points This represents the total number of measurement points. The formula for calculating data sampling density is:

[0105]

[0106] In the formula, ψ sampling This represents the data sampling density, expressed in units per m. 2 A surface The surface area of ​​the cave is expressed in m². 2 The formula for calculating sensor calibration accuracy is:

[0107]

[0108] In the formula, ∈ calibration For sensor calibration accuracy, the unit is mm; n standard Number of measurement points for standard parts; D standard，j D represents the j-th standard value, in mm. measured，j The corresponding measured value is in mm. The formula for calculating parallel computing efficiency is:

[0109]

[0110] In the formula, η parallel For parallel computing efficiency, dimensionless; T used_cores T represents the number of GPU cores already used. total_coresThis represents the total number of GPU cores. The formula for calculating data transfer bandwidth is:

[0111]

[0112] In the formula, B bandwidth Data transmission bandwidth, measured in GB / s; S data The size of the transmitted data is expressed in GB; T transfer The data transmission time is expressed in seconds (s). The objective function of the upper-level game theory model is:

[0113]

[0114] In the formula, f upper The objective function is the function of the upper layer. Let be the coordinate vector of the i-th measurement point, in mm; This is the corresponding design coordinate vector, in mm; To measure noise variance, the unit is mm. 2 ;λ cons The constraint weighting coefficients are dimensionless; C consistency For geometric consistency constraints, The unit is mm, where Let i be the coordinate vector of the (i+1)th measurement point. This represents the corresponding design coordinate vector. The lower-level objective function is:

[0115] minf lower =T compute ·M memory -ln(V process );

[0116] In the formula, f lower T is the lower-level objective function; compute For time calculation, the unit is seconds (s); M memory Memory usage, in GB; V process The data processing speed is expressed in points per second (bps). The threshold optimization formula is:

[0117] θ stable =0.8·μ deformation θ first =1.5·μ deformation θ second =2.5·μ deformation ;

[0118] In the formula, θ stable θ first θ second These are the stable threshold, the first variable threshold, and the second variable threshold, respectively, in mm; μ deformationThis represents the average degree of deformation of the cave. The unit is mm; ξ deformation，i Let ξ be the deformation amplitude at the i-th measurement point. deformation，i =‖P measured，i -P design，i ‖, the unit is mm, where P measured，i P represents the i-th actual measured coordinate, in mm. design，i The coordinates are the corresponding theoretical design coordinates, in mm. The formula for constructing the geometric stability matrix is:

[0119] M stable ={(x i y i , z i )|ξ deformation，i <θ stable};

[0120] In the formula, M stable The geometrically stable matrix is ​​composed of coordinate points whose deformation amplitude is less than the stability threshold; (x i y i , z i ) represents the three-dimensional coordinates of the i-th measurement point, in mm; ξ deformation，i Let be the deformation amplitude at the i-th measurement point, in mm. The formula for constructing the first-order geometric deformation matrix is:

[0121] M level1 ={(x i y i , z i )|θ stable ≤ξ deformation，i <0 first};

[0122] In the formula, M level1 The first-order geometrical variation matrix is ​​composed of coordinate points whose deformation amplitudes fall between the stability threshold and the first variation threshold. The formula for constructing the second-order geometrical variation matrix is:

[0123] M level2 ={(x i y i , z i )|θ first ≤ξ deformation，i <θ second};

[0124] In the formula, M level2 It is a geometric second-order variation matrix, consisting of coordinate points whose deformation amplitude is between the first variation threshold and the second variation threshold.

[0125] The specific implementation of step S07 involves generating a tunnel measurement report through model comparison and deviation analysis. The deviation calculation formula is as follows:

[0126]

[0127] In the formula, Δ deviation Geometric deviation, in mm; The measurement is of coordinate vectors, in mm. The standard design coordinate vector is given, with units in mm. The mean and standard deviation of the deviation statistical distribution are calculated as follows:

[0128]

[0129] In the formula, μ deviation σ is the mean deviation, in mm. deviation Δ represents the standard deviation of the deviation, in mm. deviation，i Let be the deviation value at the i-th measurement point, in mm. The formula for calculating the overall geometric accuracy parameter is:

[0130]

[0131] In the formula, Φ accuracy This refers to the overall geometric accuracy parameter, in mm.

[0132] Explanation of Formula Principle and Effect: Temperature Compensation Formula D compensated =D raw ×(1+α temp ×(T env -T ref By correcting for the influence of ambient temperature on laser ranging through a linear relationship, this formula maintains a measurement accuracy of 0.01 mm within a temperature range of -20℃ to 60℃, significantly improving measurement reliability in harsh environments compared to traditional methods that only consider ambient temperature measurements. Kalman filter state update equation By eliminating measurement noise through optimal estimation theory, compared to simple moving average filtering, this equation can reduce measurement noise from 0.1 mm to 0.01 mm, thus improving data quality. (Formula for calculating the degree of tunnel deformation) By quantifying the degree of deformation using the root mean square error, this formula provides a more comprehensive reflection of the overall deformation state compared to the maximum deviation method, offering an accurate basis for deformation level assessment. (Noise filtering coefficient calculation formula) By mapping the signal-to-noise ratio (SNR) to the filter coefficients using a logarithmic relationship, this formula exhibits higher filtering performance under low SNR conditions compared to a linear mapping, effectively suppressing strong noise interference. (Data transmission bandwidth calculation formula) Transmission efficiency is quantified by the ratio of data volume to transmission time. Compared to a fixed bandwidth setting, this formula can monitor the transmission status in real time, ensuring that data transmission does not become a bottleneck for system performance. The feature fusion weight adaptive adjustment formula dynamically allocates weights based on the actual state of the cavity. Compared to a fixed weight method, this formula can automatically optimize reconstruction accuracy based on the degree of deformation, reducing reconstruction error by 30% in cases of severe deformation. Deformation evaluation function G. deformation =β1·γ1+β2·γ2+β3·γ3+β4·γ4+β5·γ5 This multi-parameter weighted comprehensive evaluation of the tunnel's condition, compared to single-parameter assessment, can accurately identify minute deformations at the 0.1mm level, providing a precise basis for maintenance decisions. Geometric consistency constraint term. Geometric continuity is constrained by calculating the difference between the spacing between adjacent measurement points and the design spacing, ensuring spatial consistency of the reconstructed model and avoiding the impact of local outliers on overall accuracy. The game theory model's objective function considers both measurement accuracy and computational efficiency through two-layer optimization. Compared to traditional single-objective optimization, this model reduces computation time to less than 10 seconds while maintaining 0.01mm accuracy, achieving a balance between accuracy and efficiency. The threshold optimization formula determines the layered threshold based on the multiple relationship of the mean deformation of the cavity. Compared to the fixed threshold method, this formula can adapt to the deformation characteristics of different cavities, improving layered accuracy by 20%. Matrix factorization formula M... stable M level1 M level2 By processing measurement data in layers based on deformation amplitude, this decomposition strategy, compared to a uniform processing method, can employ differentiated algorithms for different degrees of deformation, thereby improving overall processing efficiency and accuracy. (Overall geometric accuracy parameter formula) The overall accuracy is evaluated by comprehensively calculating the mean, standard deviation, and roughness. Compared to evaluation using a single index, this formula can comprehensively reflect the overall performance of the measurement system and provide a quantitative basis for system optimization. The deviation calculation formula quantifies the measurement error using Euclidean distance. Compared to simple coordinate difference, this formula can accurately reflect the geometric deviation in three-dimensional space and provide a scientific basis for quality assessment.

[0133] To better understand and implement this invention, the following is a specific application scenario of embodiment 2: The technical team received a task to perform high-precision geometric measurements on a wind tunnel. The team first arranged the sensors. Based on the tunnel's axial length of 12000mm, and following the standard of setting a measurement section every 50mm, a total of 240 measurement sections were set up. Each measurement section had 8 laser displacement sensors evenly distributed at 360° intervals, with a 45° spacing between the sensors, forming an octagonal measurement array. The total number of sensors was 1920, using infrared laser displacement sensors with a measurement range of 10mm to 500mm and a resolution of 0.001mm. The sensors were mounted on a ring-shaped support, which moved along the tunnel's axial direction via a precision guide rail system, ensuring that each sensor was perpendicular to the tunnel's inner wall surface, with the deviation angle controlled within ±1.5°.

[0134] The system is configured with a parallel computing platform consisting of eight high-performance graphics processors (GPUs), each supporting 256 unified device architecture (UDA) computing streams. Since the number of sensors (1920) exceeds the total number of streams (2048), a single-batch processing strategy is employed. Within each DUA computing stream, a three-layer sub-thread mesh is initiated: the first layer of 3840 sub-threads handles data preprocessing, the second layer of 7680 sub-threads handles spatial geometry reconstruction, and the third layer of 15360 sub-threads handles geometric parameter calculations.

[0135] During the measurement, the laser displacement sensor continuously acquired distance data at a sampling frequency of 1000Hz. The ambient temperature during the measurement was 26℃, a deviation of 6℃ from the standard temperature of 20℃, with a calculated temperature compensation coefficient of 1.03. The first-layer sub-thread mesh performed Kalman filtering on the raw data, with the process noise covariance set to 0.001 and the observation noise covariance set to 0.01. The sliding window length was set to 10 sampling points, and the data was considered stable when the standard deviation of continuous measurement data was less than 0.05mm. After preprocessing, the noise level of the measurement data was reduced from the original 0.08mm to 0.015mm.

[0136] Spatial geometric reconstruction employs a 3D point cloud reconstruction network based on a visual transformer architecture. The measured point cloud data is divided into 16×16 block regions according to spatial location, with each block containing 256 measurement points. A multi-head attention mechanism is used, employing 8 attention heads, each with a dimension of 64. Based on the current state of the cavity, the deformation degree is 2.3 mm, the measurement point density is 45 points per square meter, and the data stability standard deviation is 0.025 mm. The feature fusion weights are automatically adjusted as follows: deformation degree weight 0.42, measurement point density weight 0.33, and data stability weight 0.25. The spatial coordinate regression module contains three hidden layers with 512, 256, and 128 neurons respectively, outputting accurate 3D coordinate values.

[0137] As shown in Table 1, the key system parameters are configured as follows:

[0138] Table 1 System Key Parameter Configuration Table

[0139]

[0140]

[0141] After the second-layer sub-thread mesh completed the spatial geometric reconstruction, the third-layer sub-thread mesh calculated the key geometric parameters of the cavity. By fitting the circular contour of each measured section, the average inner diameter of the cavity was found to be 2401.2 mm, with a deviation of 1.2 mm from the design inner diameter. Ellipticity calculation results showed that 38 out of 240 sections had an ellipticity exceeding 0.002, mainly concentrated in the 0-3000 mm area at the front and 9000-12000 mm areas at the rear. The surface roughness calculation result was 0.035 mm, meeting the design requirement of within 0.05 mm.

[0142] The deformation assessment function calculated a cavern deformation level of 0.65, falling within the range of 0.3 to 0.7, indicating slight deformation. The normalized cavern deformation degree was 0.23, the normalized measurement point density was 0.45, the normalized data stability was 0.975, the normalized temperature compensation coefficient was 0.97, and the normalized noise filtering coefficient was 0.82. Since the deformation level was within a moderate range, the spatial geometric reconstruction model adopted a first-layer computational unified equipment architecture sub-thread processing mode.

[0143] The graphics processing unit (GPU) memory stores the processing results of various computing unified device architecture streams in a buffer, and the central processing unit (CPU) copies the data to system memory via asynchronous memory transfer. The transfer bandwidth reaches 45 GB / s, and the entire data transfer process takes 3.2 seconds. The CPU quickly sorts the data according to the priority of x, y, and z coordinates to form a complete three-dimensional coordinate dataset of the cave, constructing a 1920×3 three-dimensional coordinate matrix of the cave.

[0144] The game theory model optimization calculation adopts a two-layer structure. The upper-layer model aims for optimal geometric accuracy, with the sum of squares of the geometric deviations of the cavity being 0.078 mm. 2 The measured noise variance was 0.008 mm. 2 The geometric consistency constraint term is 0.012 mm, which is satisfied. The lower-level model aims for optimal computational efficiency, with a computation time of 8.5 seconds, memory usage of 230 GB, data processing speed of 226 points per second, and parallel computing efficiency of 0.89, all of which satisfy the constraints. Through iterative solutions, a stable threshold of 1.84 mm, a first variable threshold of 3.45 mm, and a second variable threshold of 5.75 mm were obtained.

[0145] Based on a threshold-based multi-level decomposition, the geometric stability matrix contains 1456 measurement points, accounting for 75.8% of the total, mainly distributed in the 3000-9000mm region of the middle section of the tunnel. The first-level geometric variation matrix contains 328 measurement points, accounting for 17.1% of the total, mainly distributed in the front and rear sections of the tunnel. The second-level geometric variation matrix contains 136 measurement points, accounting for 7.1% of the total, concentrated in the transition area between the tunnel entrance and exit. The overall geometric accuracy parameter calculation result is 0.042mm.

[0146] As shown in Table 2, the statistical results of the cave deformation distribution are as follows:

[0147] Table 2. Statistics on the Deformation Distribution of the Tunnel

[0148]

[0149] The reconstructed 3D model of the tunnel was compared with the standard design model, and registration was performed using the nearest-point iterative algorithm. The calculation results show that the maximum deviation occurs at 10850 mm along the tunnel's axial direction, with a deviation value of 5.2 mm. The mean of the deviation statistical distribution is 1.78 mm, and the standard deviation is 1.24 mm. Figure 3 As shown, the deviation distribution exhibits a saddle-shaped pattern, with higher deviations at the front and rear sections and lower deviations in the middle section. Deformation trend analysis uses time-series regression fitting, predicting that the tunnel deformation will continue to develop at a rate of 0.3 mm per year, with the main influencing factors being temperature cyclic stress and aerodynamic loads.

[0150] The technical team developed a maintenance plan based on the measurement results. For geometrically stable areas, the existing condition was maintained, with an annual accuracy re-inspection. For areas with primary geometrical variation, monitoring was recommended every six months, with local repairs as necessary. For areas with secondary geometrical variation, immediate precision machining repair was recommended to control deformation within design tolerances. The entire measurement report includes four parts: measurement accuracy assessment, geometric parameter statistics, deviation distribution visualization, and deformation trend prediction, providing a scientific basis for tunnel maintenance decisions.

[0151] like Figure 4 As shown, the axial deformation trend analysis of the tunnel body indicates that the deformation is mainly concentrated in the structural transition area of ​​the tunnel body, which is consistent with the stress concentration location analyzed theoretically. Figure 5 As shown, the geometric accuracy distribution after multi-level decomposition verifies the effectiveness of the game model optimization, and the differentiated processing strategy for different deformation level regions significantly improves the overall measurement accuracy.

[0152] This invention represents a significant technological advancement over traditional wind tunnel measurement methods. Traditional methods typically employ contact measurement or single-point laser measurement, which suffers from problems such as sparse measurement points, limited accuracy, and low efficiency. This invention achieves full-coverage measurement through a laser displacement sensor array, solving the problem of blind spots in traditional methods. The spatial geometry reconstruction technology based on a vision transformer, compared to traditional interpolation methods, can more accurately handle complex deformation patterns, particularly excelling in complex geometries such as elliptical deformation and local distortion. The game-theoretic model optimization strategy simultaneously optimizes measurement accuracy and computational efficiency through a dual-level objective function, overcoming the inability of traditional single-objective optimization methods to balance accuracy and speed. The multi-level decomposition technology employs differentiated processing strategies based on the degree of deformation, enabling refined analysis of the characteristics of different regions compared to traditional uniform processing methods. The introduction of advanced signal processing techniques such as temperature compensation and Kalman filtering effectively eliminates the influence of environmental factors and measurement noise, significantly improving measurement reliability.

[0153] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0154] Table 3. Variable Explanation Table (Part 1)

[0155]

[0156] Table 4. Variable Explanation Table (Part Two)

[0157]

[0158] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-precision method for measuring the structure of a wind tunnel, characterized in that, include: Multiple laser displacement sensors are arranged on the inner wall surface of the wind tunnel to form a three-dimensional measurement grid; A sufficient number of CUDA streams are launched to match the number of sensors, with three sub-thread meshes launched within each CUDA stream. Laser displacement sensors collect distance data from the inner wall of the cave in real time. The first sub-thread mesh performs noise filtering and temperature compensation on the raw distance data. The second sub-thread mesh inputs the processed distance data and sensor spatial position information into a spatial geometric reconstruction model. This model automatically adjusts the feature fusion weights based on three parameters: cave deformation degree, measurement point density, and data stability, reconstructing the cave's three-dimensional spatial coordinates. The third sub-thread mesh calculates the cave's inner diameter deviation, ellipticity, and surface roughness based on the reconstructed three-dimensional spatial coordinates, and calculates the cave deformation level using a deformation evaluation function. The CPU receives the processing results from each CUDA stream, spatially stitches together all measurement data at the same time, and establishes a complete three-dimensional model of the cave. A three-dimensional coordinate matrix is ​​constructed based on the three-dimensional spatial coordinates in the cave model. A game theory model is used to optimize and calculate the stability threshold, the first variation threshold, and the second variation threshold. The cave's three-dimensional coordinate matrix is ​​then processed hierarchically according to the deformation amplitude through a multi-level decomposition method to obtain a geometric stability matrix, a first-level geometric variation matrix, and a second-level geometric variation matrix. Finally, the overall geometric accuracy parameters of the cave are calculated. The 3D model of the cave is compared with the standard design model, and a cave measurement report is output.

2. The wind tunnel measurement method according to claim 1, characterized in that, The arrangement steps of the laser displacement sensor are as follows: multiple measurement sections are set up along the axial direction of the tunnel, and multiple sensors are evenly distributed around the circumference in each measurement section. The laser displacement sensor adopts the triangulation principle, which projects a laser beam onto the inner wall surface of the tunnel and receives the reflected light signal to calculate the distance.

3. The wind tunnel body measurement method according to claim 2, characterized in that, The CUDA stream startup steps specifically involve using multiple batch processing if the number of sensors exceeds the number of CUDA streams. This means grouping all sensors according to their spatial location, with each group containing no more than the number of CUDA streams, and processing the sensor data of each group sequentially in chronological order to ensure that all measurement data is processed completely.

4. The wind tunnel measurement method according to claim 3, characterized in that, The laser displacement sensor acquires distance data in real time, specifically by performing noise filtering and temperature compensation on the raw distance data in the first sub-thread grid, and then transmitting the processed data to the second sub-thread grid.

5. The wind tunnel measurement method according to claim 4, characterized in that, The specific structure of the spatial geometric reconstruction model is a three-dimensional point cloud reconstruction network based on the ViT architecture, which includes a point cloud segmentation module, a multi-head attention mechanism, and a spatial coordinate regression module. The point cloud segmentation module divides the measured point cloud data into multiple blocks according to spatial location. The multi-head attention mechanism calculates the spatial correlation between different measurement points. The spatial coordinate regression module outputs accurate three-dimensional coordinate values.

6. The wind tunnel measurement method according to claim 5, characterized in that, The deformation evaluation function is used to calculate the deformation level value of the tunnel. The inputs include the degree of tunnel deformation, the density of measurement points, the data stability, the temperature compensation coefficient, and the noise filtering coefficient. The output is the deformation level value of the tunnel. When the deformation level value of the tunnel is in different ranges, the spatial geometric reconstruction model adopts different processing methods.

7. The wind tunnel measurement method according to claim 6, characterized in that, The steps for the CPU to receive the processing results of each CUDA stream are as follows: First, the GPU memory stores the processing results of each CUDA stream in the memory buffer. Then, the CPU copies the data in the memory to the system memory through asynchronous memory transfer. Finally, the CPU sorts and splices the measurement data from different CUDA streams in the system memory according to spatial coordinates to form a complete three-dimensional coordinate dataset of the cave.

8. The wind tunnel measurement method according to claim 7, characterized in that, The game theory model includes an upper-level model that aims for optimal geometric accuracy and a lower-level model that aims for optimal computational efficiency. The objective function of the upper-level model is to minimize the product of the sum of squares of the geometric deviations of the hole and the variance of the measurement noise plus a geometric consistency constraint term. The objective function of the lower-level model is to minimize the product of the computation time and memory usage minus the logarithmic function of the data processing speed. The two objective functions influence each other through a threshold optimization coupling term.

9. The wind tunnel measurement method according to claim 8, characterized in that, The objective function inputs of the upper-level model include the sum of squares of the geometric deviations of the cavity, the variance of the measurement noise, the geometric consistency constraint, the data sampling density, and the sensor calibration accuracy. The objective function inputs of the lower-level model include computation time, memory usage, data processing speed, parallel computing efficiency, and data transmission bandwidth.

10. The wind tunnel body measurement method according to claim 9, characterized in that, The geometric stability matrix is ​​used to represent the stable part of the cave's geometry. It is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is less than the stability threshold. The first-order geometric variation matrix is ​​used to represent the main variation part of the cave's geometry. It is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between the stability threshold and the first variation threshold. The second-order geometric variation matrix is ​​used to represent the secondary variation part of the cave's geometry. It is constructed by extracting coordinate points in the cave's three-dimensional coordinate matrix whose deformation amplitude is between the first variation threshold and the second variation threshold.

Images