Self-adaptive control method and system for air tightness of wind measurement laser radar shell

Abstract

The invention provides an air tightness self-adaptive control method and system for a wind measurement laser radar shell, and relates to the technical field of air tightness control, and the method comprises the steps: reconstructing the three-dimensional shape of a sealing surface of the wind measurement laser radar shell through a machine vision method, and building a mapping model of a multi-scale shape characteristic parameter and a gas leakage rate; obtaining the position coordinates of a sensitive area, and carrying out the dynamic pressure regulation of the sealing surface of the wind measurement laser radar housing according to the position coordinates of the sensitive area; the method comprises the following steps: acquiring air tightness control parameters under different working conditions, performing self-learning optimization on the air tightness control parameters to obtain an optimal control parameter matrix, storing the control parameter matrix to a controller, and predicting the air tightness performance of a wind measurement laser radar shell by adopting a pre-trained air tightness prediction model to obtain the air tightness performance of the wind measurement laser radar shell. And when the predicted gas leakage rate exceeds a preset early warning threshold value, outputting a gas tightness alarm signal, and starting a gas tightness self-adaptive control flow according to the stored optimal control parameter matrix.

Description

Technical Field

[0001] The present invention relates to airtightness control technology, and in particular to an airtightness adaptive control method and system for a wind laser radar housing. Background Art

[0002] Wind LiDAR is an important atmospheric detection equipment, widely used in meteorological monitoring, wind energy resource assessment and other fields. The air tightness of its casing directly affects the measurement accuracy and service life of the equipment. The traditional air tightness control method of the wind LiDAR casing mainly relies on manual detection and fixed parameter control. The air tightness detection method is single and it is difficult to accurately reflect the microscopic morphological characteristics of the sealing surface. The existing technology mainly detects air tightness through the pressure decay method or the tracer gas method. It is impossible to obtain the three-dimensional morphological information of the sealing surface and it is difficult to analyze the microscopic mechanism of gas leakage. The air tightness control parameters are fixed and difficult to adapt to different working conditions. The current air tightness control mostly uses fixed parameters, which cannot be adjusted in real time according to changes in the external environment and the use status of the equipment, resulting in poor control effect.

[0003] Therefore, it is urgent to develop a method and system that can realize adaptive airtightness control of the wind laser radar housing to improve the airtightness performance and reliability of the equipment. The method should have the functions of sealing surface three-dimensional morphology reconstruction, multi-scale feature extraction, airtightness performance prediction, active deformation compensation, parameter self-learning optimization, etc., to realize intelligent and adaptive airtightness control. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for adaptively controlling the air tightness of a wind laser radar housing, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provided is a method for adaptively controlling air tightness of a wind laser radar housing, comprising:

[0007] The three-dimensional morphology of the sealing surface of the wind laser radar shell is reconstructed through machine vision methods, and the multi-scale morphological feature parameters reflecting the airtight performance are extracted. At the same time, the air pressure and temperature signals of the inner cavity of the wind laser radar shell are obtained. The gas leakage rate of the inner cavity of the shell is calculated based on the air pressure and temperature signals, and a mapping model between the multi-scale morphological feature parameters and the gas leakage rate is established.

[0008] According to the mapping model, the variation law of characteristic parameters under different leakage rates is obtained. Combined with the mechanical properties of the wind laser radar housing material, the deformation and force distribution of the sealing surface are predicted through airtightness simulation analysis, and the position coordinates of the sensitive area are obtained. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar housing is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements;

[0009] The airtightness control parameters under different working conditions are obtained, and the optimal control parameter matrix is ​​obtained by self-learning optimization of the airtightness control parameters. The control parameter matrix is ​​stored in the controller, and the airtightness prediction model that has been pre-trained is used to predict the airtightness performance of the wind measuring lidar casing. When the predicted gas leakage rate exceeds the preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

[0010] In an optional embodiment,

[0011] The three-dimensional morphology of the sealing surface of the wind laser radar shell is reconstructed by machine vision methods, and the multi-scale morphological feature parameters reflecting the airtight performance are extracted. At the same time, the air pressure and temperature signals of the inner cavity of the wind laser radar shell are obtained, and the gas leakage rate of the inner cavity of the shell is calculated according to the air pressure and temperature signals. The mapping model between the multi-scale morphological feature parameters and the gas leakage rate is established, including:

[0012] Acquire multi-view images of the sealing surface of the wind laser radar housing, reconstruct a three-dimensional point cloud of the sealing surface by a stereo matching method, and project the three-dimensional point cloud onto a two-dimensional plane to obtain a depth map of the sealing surface;

[0013] Performing multi-scale wavelet transform on the sealing surface depth map, extracting frequency domain characteristics at different scales, determining the deformation area at each scale by using the maximum suppression algorithm, calculating the geometric features and spectral features in each deformation area, and obtaining the topological invariant features at different scales by using persistent coherence, and constructing a multi-scale morphological feature set reflecting the local deformation of the sealing surface;

[0014] The features in the multi-scale morphological feature set are screened using a maximum correlation minimum redundancy algorithm, and the correlation between each feature and the gas leakage rate is calculated by mutual information measurement. The optimal features are selected one by one with the goal of maximizing the correlation between the feature and the gas leakage rate and minimizing the redundancy between the features, and a feature subset whose correlation with the gas leakage rate is higher than a preset correlation threshold is screened out;

[0015] The support vector machine algorithm is used to construct a nonlinear regression mapping model between gas leakage rate and feature subsets. The Gaussian radial basis kernel function is selected as the kernel function of the support vector machine. The kernel function width parameters and model hyperparameters are optimized through grid search and cross validation to obtain a mapping model that can estimate the airtight performance within a specified leakage rate range.

[0016] In an optional embodiment,

[0017] Obtain multi-view images of the sealing surface of the wind laser radar housing, and reconstruct the three-dimensional point cloud of the sealing surface by stereo matching method, including:

[0018] The three-dimensional point cloud of the sealing surface is constructed as an undirected graph, wherein the nodes represent the points in the three-dimensional point cloud, and the edges represent the spatial constraint relationship between the points in the three-dimensional point cloud;

[0019] Obtaining the Euclidean distance between any two adjacent points in the undirected graph in the three-dimensional point cloud, and constructing a global error energy function based on the coordinate difference and the Euclidean distance between any two adjacent points, wherein the global error energy function is used to measure the optimization degree of the three-dimensional point cloud;

[0020] An incremental equation is constructed based on the global error energy function to solve the increment of the three-dimensional point cloud coordinates, and the three-dimensional point cloud coordinates are updated using the increment of the three-dimensional point cloud coordinates. The iteration is repeated until the preset convergence condition is met to obtain the reconstructed three-dimensional point cloud of the sealing surface;

[0021] Among them, the calculation formula of the global error energy function is as follows:

[0022]

[0023] Where E(V) represents the global error energy function, n represents the number of points in the 3D point cloud, and v j represents the jth point, N(v i ) represents the set of neighboring points of the i-th point, ρ represents the robust kernel function, v i represents the i-th point, d ij Represents the Euclidean distance between the i-th point and the j-th point.

[0024] In an optional embodiment,

[0025] According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar housing is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model. The active compensation for the deformation of the sealing surface includes:

[0026] The sealing surface is divided into multiple sub-areas, each of which is equipped with an independent controller and pneumatic actuator, and the position coordinates and deformation data of the sensitive area in each sub-area are obtained;

[0027] According to the position coordinates and deformation data of the sensitive areas in each sub-area, combined with the geometric characteristics and environmental conditions of the sub-area, a local deformation control model of each sub-area is established;

[0028] Based on the local deformation control model of each sub-region, the local control law is designed, the local compensation pressure of each sub-region is calculated, and the local compensation pressure, deformation data and key status information of each sub-region are exchanged and shared through the preset communication protocol;

[0029] The local compensation pressure of each sub-region and the exchanged local compensation pressure, deformation data and key state information of other sub-regions are input into the distributed collaborative control algorithm. Through distributed model predictive control and consistency optimization, the coordinated compensation pressure of each sub-region is optimized and calculated in real time under the constraints of meeting the global deformation performance index and local accuracy requirements.

[0030] According to the optimized coordinated compensation pressure of each sub-area and the pneumatic actuator characteristics of each sub-area, the corresponding aerodynamic force is applied to each sub-area synchronously, and the actual deformation data of the sealing surface is measured in real time. The actual deformation data is compared with the target deformation data, and the deformation control error is calculated. The deformation control error is fed back to the distributed collaborative control algorithm for closed-loop iterative optimization until the actual deformation of the sealing surface meets the preset deformation control accuracy requirement, thereby realizing active compensation for the deformation of the sealing surface.

[0031] In an optional embodiment,

[0032] The calculation formula of the distributed cooperative control algorithm is as follows:

[0033]

[0034] Where U(t) represents the control input vector of all sub-regions, J(·) represents the global optimization objective function, X(t) represents the deformation state vector of all sub-regions, G represents the number of sub-regions, and X a (t) represents the deformation state vector of the a-th sub-region at time t, r a (t) represents the expected state of the a-th sub-region at time t, Q a represents the weight matrix for measuring the state tracking error of the a-th sub-region, u a (t) represents the control input vector of the a-th sub-region at time t, R a represents the weight matrix for measuring the control input of the a-th sub-region, x b (t) represents the deformation state vector of the b-th sub-region at time t, S ab Represents a matrix that measures the consistency of deformation data between the a-th sub-region and the b-th sub-region.

[0035] In an optional embodiment,

[0036] Acquiring airtightness control parameters under different working conditions, and performing self-learning optimization on the airtightness control parameters to obtain an optimal control parameter matrix includes:

[0037] An airtightness data acquisition system based on a heterogeneous sensor array is designed. The heterogeneous sensor array includes a gas leakage rate sensor, a pressure sensor, a temperature sensor, and a vibration sensor, which are respectively used to collect airtightness-related information inside the shell of a wind-measuring lidar. In view of the differences in sampling frequency, accuracy, and resolution of different sensors in the heterogeneous sensor array, a multi-sensor data synchronization and fusion algorithm based on Kalman filtering is introduced to perform spatiotemporal alignment and feature extraction on the collected airtightness-related information to obtain airtightness-related data features.

[0038] Preprocess and clean the features of airtightness-related data, obtain standard airtightness-related data through data quality assessment and anomaly detection, and design a self-learning optimization algorithm that integrates reinforcement learning and evolutionary algorithms based on the standard airtightness-related data;

[0039] The deep deterministic policy gradient algorithm is used to model the airtightness control problem of the wind lidar housing as a Markov decision process in a continuous action space. The airtightness performance represented by standard airtightness-related data is used as the state, the airtightness control parameters are used as the action, and the airtightness index is used as the reward function. Through interactive learning between the agent and the environment, the optimized airtightness control strategy and optimal airtightness control parameters are obtained.

[0040] The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met, and the optimal control parameter matrix covering the optimal control strategy of the wind measuring lidar shell under different working conditions is obtained.

[0041] In an optional embodiment,

[0042] The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met. The optimal control parameter matrix covering the optimal control strategy of the wind laser radar shell under different working conditions is obtained, including:

[0043] The policy function parameters of the deep deterministic policy gradient algorithm are used as the individual encoding of the genetic algorithm. Each individual represents a set of possible policy function parameter combinations, and the population size is set to ensure the diversity and coverage of the exploration.

[0044] A certain number of individuals are generated through random sampling or heuristic methods to form the initial population of the genetic algorithm. In each iteration of the genetic algorithm, the individuals in the current population are input into the deep deterministic policy gradient algorithm. The cumulative reward of each individual corresponding to the policy function is obtained through interaction with the environment. At the same time, the diversity measure of the individual and other individuals in the population is calculated, and the fitness function is constructed to comprehensively evaluate the quality of the individual.

[0045] According to the evaluation result of the fitness function, the father is selected from the current population, and the selected father individuals are paired with each other, and the encoding information of the father individuals is exchanged through the crossover operator to generate new offspring individuals;

[0046] The encoded information of the new generation of individuals is mutated, and the resulting offspring individuals are merged with the parent population to form a new generation population, which is then input into the next round of iteration. At the same time, in each iteration, the actor network and critic network of the deep deterministic policy gradient algorithm are used to update the parameters of the individual's corresponding policy function based on the empirical data obtained from the interaction between the individual and the environment.

[0047] Repeat the iteration until the preset number of iterations is met, select the individual with the highest fitness function value from the last generation of the population, extract the corresponding strategy function parameters, and use them as the optimal strategy parameters for airtightness control of the wind laser radar casing. Apply them to the wind laser radar system, and adaptively adjust the control parameters according to the characteristics of airtightness-related data. Finally, the optimal control parameter matrix covering the optimal control strategy of the wind laser radar casing under different working conditions is obtained.

[0048] According to a second aspect of the embodiments of the present invention,

[0049] Provided is a wind measuring laser radar housing air tightness adaptive control system, comprising:

[0050] The first unit is used to reconstruct the three-dimensional morphology of the sealing surface of the wind laser radar shell through machine vision methods, extract multi-scale morphological feature parameters reflecting the airtight performance, and simultaneously obtain the air pressure and temperature signals of the inner cavity of the wind laser radar shell. The gas leakage rate of the inner cavity of the shell is calculated based on the air pressure and temperature signals, and a mapping model between multi-scale morphological feature parameters and gas leakage rate is established;

[0051] The second unit is used to obtain the characteristic parameter variation law under different leakage rates according to the mapping model, and predict the deformation and force distribution of the sealing surface through airtightness simulation analysis in combination with the mechanical properties of the wind laser radar shell material, and obtain the position coordinates of the sensitive area. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar shell is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements;

[0052] The third unit is used to obtain the airtightness control parameters under different working conditions, perform self-learning optimization on the airtightness control parameters to obtain the optimal control parameter matrix, store the control parameter matrix in the controller, and simultaneously use a pre-trained airtightness prediction model to predict the airtightness performance of the wind measuring laser radar casing. When the predicted gas leakage rate exceeds a preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] According to a fourth aspect of the embodiments of the present invention,

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0060] In this embodiment, by establishing a mapping model between multi-scale morphological feature parameters and gas leakage rate, the airtightness of the wind laser radar housing can be evaluated with high precision. By dynamically regulating the pressure and real-time compensation of these sensitive areas, the deformation of the sealing surface can be actively compensated to ensure that the airtightness meets the use requirements. By obtaining the airtightness control parameters under different working conditions, the optimal control parameter matrix is ​​obtained through self-learning optimization and stored in the controller. At the same time, the pre-trained airtightness prediction model is used to predict the future airtightness. When the predicted leakage rate exceeds the threshold, the airtightness adaptive control process is automatically started, and intelligent compensation is performed according to the optimal control parameter matrix to achieve adaptive airtightness control. The pre-trained airtightness prediction model can predict the airtightness in the future. When the predicted gas leakage rate exceeds the preset warning threshold, the system will output an airtightness alarm signal, thereby warning of possible faults in advance, which is conducive to taking measures in advance to avoid serious faults. High-precision evaluation, active compensation control, adaptive intelligent control and early fault warning of the airtightness of the wind laser radar housing are realized, which can greatly improve the airtightness and reliability of the wind laser radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1It is a schematic diagram of the flow chart of the method for adaptively controlling the air tightness of the housing of a wind-measuring laser radar according to an embodiment of the present invention;

[0062] Figure 2 It is a schematic diagram of the structure of the air tightness adaptive control system of the wind measuring laser radar housing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of a method for adaptively controlling the air tightness of a housing of a wind laser radar according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Reconstruct the three-dimensional morphology of the sealing surface of the wind measuring laser radar shell through machine vision methods, extract multi-scale morphological feature parameters that reflect the airtight performance, and simultaneously obtain the air pressure and temperature signals of the inner cavity of the wind measuring laser radar shell. Calculate the gas leakage rate of the inner cavity of the shell based on the air pressure and temperature signals, and establish a mapping model between multi-scale morphological feature parameters and gas leakage rate.

[0067] Among them, machine vision three-dimensional morphology reconstruction technology refers to the technology of using computer vision algorithms to restore the three-dimensional geometric shape of an object from a two-dimensional image sequence. Specifically, structured light three-dimensional scanning technology is used to obtain high-precision three-dimensional point cloud data of the sealing surface of the wind laser radar housing, and then the point cloud processing algorithm (such as point cloud filtering, splicing, gridding, etc.) is used to pre-process the point cloud data, and finally a three-dimensional digital model of the sealing surface is obtained. By analyzing the geometric shape characteristics of the three-dimensional digital model, multi-scale morphological feature parameters that reflect the airtight performance can be extracted, such as surface roughness, flatness, normal vector variance, etc. These characteristic parameters can quantitatively describe the microscopic morphology and macroscopic geometric characteristics of the sealing surface, and provide important morphological information for subsequent airtight performance analysis.

[0068] Gas leakage rate is an important physical quantity that characterizes airtightness performance. It reflects the ratio of the mass of gas leaking from a closed cavity to the external environment per unit time to the total mass of gas in the cavity. By adopting machine learning methods, a large number of wind laser radar shell samples and corresponding gas leakage rate data are collected to train and establish a mapping model between multi-scale morphological feature parameters and gas leakage rate. The model can be a classic machine learning model such as a linear regression model, support vector machine, random forest, or a deep learning model such as a deep neural network. The input of the model is multi-scale morphological feature parameters, and the output is the predicted value of gas leakage rate. By comparing the error between the predicted value and the measured value, the model parameters are optimized using the back propagation algorithm to continuously improve the prediction accuracy of the mapping model. Based on the trained mapping model, the variation law of characteristic parameters under different leakage rates can be predicted, providing important boundary conditions for subsequent airtightness simulation analysis.

[0069] In an optional embodiment,

[0070] The three-dimensional morphology of the sealing surface of the wind laser radar shell is reconstructed by machine vision methods, and the multi-scale morphological feature parameters reflecting the airtight performance are extracted. At the same time, the air pressure and temperature signals of the inner cavity of the wind laser radar shell are obtained, and the gas leakage rate of the inner cavity of the shell is calculated according to the air pressure and temperature signals. The mapping model between the multi-scale morphological feature parameters and the gas leakage rate is established, including:

[0071] Acquire multi-view images of the sealing surface of the wind laser radar housing, reconstruct a three-dimensional point cloud of the sealing surface by a stereo matching method, and project the three-dimensional point cloud onto a two-dimensional plane to obtain a depth map of the sealing surface;

[0072] Performing multi-scale wavelet transform on the sealing surface depth map, extracting frequency domain characteristics at different scales, determining the deformation area at each scale by using the maximum suppression algorithm, calculating the geometric features and spectral features in each deformation area, and obtaining the topological invariant features at different scales by using persistent coherence, and constructing a multi-scale morphological feature set reflecting the local deformation of the sealing surface;

[0073] The features in the multi-scale morphological feature set are screened using a maximum correlation minimum redundancy algorithm, and the correlation between each feature and the gas leakage rate is calculated by mutual information measurement. The optimal features are selected one by one with the goal of maximizing the correlation between the feature and the gas leakage rate and minimizing the redundancy between the features, and a feature subset whose correlation with the gas leakage rate is higher than a preset correlation threshold is screened out;

[0074] The support vector machine algorithm is used to construct a nonlinear regression mapping model between gas leakage rate and feature subsets. The Gaussian radial basis kernel function is selected as the kernel function of the support vector machine. The kernel function width parameters and model hyperparameters are optimized through grid search and cross validation to obtain a mapping model that can estimate the airtight performance within a specified leakage rate range.

[0075] For example, the sealing surface is first photographed using cameras with multiple viewing angles to obtain a set of parallax images. Then, a stereo matching algorithm is used to calculate the parallax information between the image pairs, and the three-dimensional coordinates of each pixel point on the sealing surface are restored through the triangulation principle to obtain the three-dimensional point cloud data of the sealing surface. Since the subsequent processing of point cloud data requires a large amount of calculation, the three-dimensional point cloud is projected onto a two-dimensional plane to obtain a sealing surface depth map with depth information. The grayscale value of each pixel in the depth map represents the vertical distance from the point to the projection plane, which can reflect the three-dimensional morphological characteristics of the sealing surface.

[0076] In order to comprehensively characterize the morphological features of the sealing surface at different scales, the sealing surface depth map is subjected to multi-scale wavelet transform. Wavelet transform can describe the local detail features of the image at multiple scales by performing multi-level decomposition of the image to obtain sub-band coefficients at different frequencies and directions. By analyzing the statistical characteristics of the sub-band coefficients at each scale, such as mean, variance, skewness, etc., the frequency domain features reflecting the morphology of the sealing surface are extracted. At the same time, the maximum suppression algorithm is used to process the wavelet coefficients at each scale to determine the local deformation area. The maximum suppression algorithm obtains a stable local deformation area by marking the local maximum points of the wavelet coefficients as points of interest and suppressing the non-maximum points in its neighborhood. In each deformation area, the geometric features of the area, such as area, perimeter, roundness, etc., and the spectral features, such as spectral energy, spectral entropy, etc., are calculated. In addition, the topological invariant features of the deformation area at different scales are obtained by the persistent homology method. Persistent homology is an algebraic topological tool that extracts scale-independent topological invariants, such as Betti number and persistent diagram, by analyzing the topological properties of the deformation region at different scales, such as connectivity and number of holes. Finally, various geometric features, spectral features and topological invariants are combined to form a multi-scale morphological feature set to comprehensively characterize the local deformation characteristics of the sealing surface.

[0077] Since the multi-scale morphological feature set has a high dimension, directly using all features for modeling may have the risk of redundancy and overfitting, so feature selection is required. The maximum relevance minimum redundancy algorithm is used to screen out the optimal feature subsets one by one based on the correlation between the feature and the gas leakage rate and the redundancy between the features. First, the mutual information between each morphological feature and the gas leakage rate is calculated. The mutual information is a nonlinear correlation measure that can characterize the statistical dependence between the feature and the target variable. Then, with the goal of maximizing the mutual information between the feature and the gas leakage rate and minimizing the redundancy between the selected feature and the candidate feature, the optimal feature is iteratively selected from the candidate features until the correlation of all features is lower than the preset threshold. Through the maximum relevance minimum redundancy algorithm, a feature subset with high correlation with the gas leakage rate and low redundancy is obtained, which reduces the complexity and uncertainty of subsequent modeling.

[0078] After obtaining the preferred feature subset, the support vector machine algorithm is used to construct a nonlinear regression model between the features and the gas leakage rate. Support vector machine is a small sample machine learning method based on statistical learning theory. It realizes high-dimensional mapping and classification prediction of nonlinear data by finding the optimal classification hyperplane that minimizes structural risk. The support vector machine is extended to the regression field. By introducing the ε-insensitive loss function, the model is allowed to predict within a certain error range, which improves the generalization performance of the model. Considering the complex nonlinear relationship between gas leakage rate and morphological features, the Gaussian radial basis kernel function is selected as the kernel function of the support vector machine to map the features to a high-dimensional space, thereby enhancing the nonlinear fitting ability of the model. In order to obtain the optimal model parameters, a combination of grid search and cross-validation is used to search for different kernel function widths and model penalty coefficients within the preset parameter range. The generalization performance of the model is evaluated by k-fold cross-validation, and the parameter combination with the smallest average validation error is selected as the final model parameters.

[0079] In an optional embodiment,

[0080] Obtain multi-view images of the sealing surface of the wind laser radar housing, and reconstruct the three-dimensional point cloud of the sealing surface by stereo matching method, including:

[0081] The three-dimensional point cloud of the sealing surface is constructed as an undirected graph, wherein the nodes represent the points in the three-dimensional point cloud, and the edges represent the spatial constraint relationship between the points in the three-dimensional point cloud;

[0082] Obtaining the Euclidean distance between any two adjacent points in the undirected graph in the three-dimensional point cloud, and constructing a global error energy function based on the coordinate difference and the Euclidean distance between any two adjacent points, wherein the global error energy function is used to measure the optimization degree of the three-dimensional point cloud;

[0083] An incremental equation is constructed based on the global error energy function to solve the increment of the three-dimensional point cloud coordinates, and the three-dimensional point cloud coordinates are updated using the increment of the three-dimensional point cloud coordinates. The iteration is repeated until the preset convergence condition is met to obtain the reconstructed three-dimensional point cloud of the sealing surface;

[0084] Among them, the calculation formula of the global error energy function is as follows:

[0085]

[0086] Where E(V) represents the global error energy function, n represents the number of points in the 3D point cloud, and v j represents the jth point, N(v i ) represents the set of neighboring points of the i-th point, ρ represents the robust kernel function, v i represents the i-th point, d ij Represents the Euclidean distance between the i-th point and the j-th point.

[0087] Exemplarily, each point in the three-dimensional point cloud is traversed and set as a node in an undirected graph. For each point, other points within a certain spatial range around it are found, and these points are connected with the current point to form edges of the undirected graph. The edges represent the spatial constraint relationship between the points, wherein a distance threshold can be set to control the number of neighboring points connected to each point.

[0088] Traverse each edge in the undirected graph, obtain the three-dimensional coordinates of the two points connected by the edge, and calculate the Euclidean distance between the two points. Traverse each edge in the undirected graph, obtain the two points connected by the edge and their Euclidean distance, calculate the coordinate difference between the two points in three-dimensional space, substitute this coordinate difference and Euclidean distance into the energy term function, and accumulate them into the global error energy function.

[0089] By taking the partial derivatives of the global error energy function with respect to the coordinates of each point and setting these partial derivatives to 0, we can obtain a set of equations for the coordinate increments of each point. We use numerical methods (such as the Gauss-Newton method) to solve the incremental equations to obtain the coordinate increments of each point. We add the increments to the current coordinates to update the new coordinates of each point. We repeat the iterations until the preset convergence conditions are met (such as the energy function value is less than a threshold, or the increment is less than a threshold) to obtain a reconstructed three-dimensional point cloud of the sealing surface.

[0090] In this embodiment, the three-dimensional point cloud of the sealing surface is reconstructed by machine vision multi-view image acquisition and stereo matching, and the point cloud is optimized by an iterative method based on error energy optimization, so that high-precision and low-noise three-dimensional morphological data of the sealing surface can be obtained. The geometric, spectral and topological invariant features at different scales are extracted from the sealing surface depth map by wavelet transform, maximum suppression and other methods, and a morphological feature set reflecting the local multi-scale deformation of the sealing surface is constructed, which can comprehensively describe the morphological information of the sealing surface. Through the maximum correlation minimum redundancy algorithm, a subset of morphological features with high correlation with the gas leakage rate is automatically screened out, avoiding the subjectivity of manually selecting features and improving the representativeness of features. Using the support vector machine algorithm, a nonlinear regression mapping model between multi-scale morphological features and gas leakage rate is constructed. The optimized model can accurately estimate the airtight performance within the specified leakage rate range. Combined with air pressure, temperature measurement and leakage rate estimation, the intelligent and automated evaluation of the airtight performance of the wind measuring lidar casing can be realized. When the predicted leakage rate exceeds the threshold, an early warning can be issued, realizing the intelligent and automated quality inspection and evaluation of the airtight performance of the wind measuring lidar casing, and improving the accuracy of airtightness detection.

[0091] S102. The characteristic parameter variation pattern under different leakage rates is obtained according to the mapping model, and combined with the mechanical properties of the wind laser radar shell material, the deformation and force distribution of the sealing surface are predicted through airtightness simulation analysis, and the position coordinates of the sensitive area are obtained. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar shell is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-constructed control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements.

[0092] Among them, the airtightness simulation analysis technology refers to the technology that uses numerical simulation methods to analyze the stress deformation and contact state of the sealing structure and predict the sealing performance. Based on the mechanical parameters of the wind laser radar shell material, such as elastic modulus, Poisson's ratio, yield strength, etc., a finite element analysis model of the sealing surface is established. By applying loads and boundary conditions corresponding to the actual working conditions on the model, such as internal and external pressure difference, temperature load, bolt preload, etc., the deformation and stress distribution of the sealing surface under different leakage rates are simulated and analyzed. Focus on the uniformity of the stiffness of the sealing surface and the contact stress distribution, and identify the key areas that affect the airtightness performance, that is, the sensitive areas. According to the position coordinates of the sensitive area obtained by the airtightness simulation analysis, an air pressure cavity corresponding to the sensitive area is set on the back of the sealing surface. By filling or releasing compressed gas into the air pressure cavity, the deformation compensation of the local area is realized, so that the shape of the sealing surface is kept in the optimal state. The dynamic pressure regulation process is feedback controlled by the pre-constructed control model, and the adjustment amount of the compensation pressure is calculated in real time according to the deformation of the sensitive area and the pressure signal in the air pressure cavity. The control model adopts advanced control theories such as PID control, adaptive control and robust control, and has strong robustness and adaptability to the nonlinear and time-varying characteristics of the system.

[0093] In an optional embodiment,

[0094] According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar housing is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model. The active compensation for the deformation of the sealing surface includes:

[0095] The sealing surface is divided into multiple sub-areas, each of which is equipped with an independent controller and pneumatic actuator, and the position coordinates and deformation data of the sensitive area in each sub-area are obtained;

[0096] According to the position coordinates and deformation data of the sensitive areas in each sub-area, combined with the geometric characteristics and environmental conditions of the sub-area, a local deformation control model of each sub-area is established;

[0097] Based on the local deformation control model of each sub-region, the local control law is designed, the local compensation pressure of each sub-region is calculated, and the local compensation pressure, deformation data and key status information of each sub-region are exchanged and shared through the preset communication protocol;

[0098] The local compensation pressure of each sub-region and the exchanged local compensation pressure, deformation data and key state information of other sub-regions are input into the distributed collaborative control algorithm. Through distributed model predictive control and consistency optimization, the coordinated compensation pressure of each sub-region is optimized and calculated in real time under the constraints of meeting the global deformation performance index and local accuracy requirements.

[0099] According to the optimized coordinated compensation pressure of each sub-area and the pneumatic actuator characteristics of each sub-area, the corresponding aerodynamic force is applied to each sub-area synchronously, and the actual deformation data of the sealing surface is measured in real time. The actual deformation data is compared with the target deformation data, and the deformation control error is calculated. The deformation control error is fed back to the distributed collaborative control algorithm for closed-loop iterative optimization until the actual deformation of the sealing surface meets the preset deformation control accuracy requirement, thereby realizing active compensation for the deformation of the sealing surface.

[0100] Exemplarily, first, the sealing surface is divided into multiple sub-areas according to its geometric shape and deformation characteristics. The division of sub-areas needs to comprehensively consider factors such as the geometric characteristics of the area, deformation mode, and control requirements to ensure that the deformation characteristics in each sub-area are relatively consistent and easy to control. Each sub-area is equipped with an independent controller and pneumatic actuator. The controller is responsible for the deformation modeling, control decision-making and actuator control of the area. The actuator applies aerodynamic force to the area according to the instructions of the controller to achieve local deformation regulation. After the division is completed, the position coordinates and deformation data of the sensitive areas in each sub-area are obtained. Sensitive areas refer to key areas that have a greater impact on the airtight performance of the sealing surface, such as the edges and interfaces of the sealing surface.

[0101] For each sub-region, a local deformation control model of the region is established based on the position coordinates and deformation data of the sensitive region inside it, combined with the geometric characteristics and environmental conditions of the sub-region. The local deformation control model describes the dynamic relationship between the deformation inside the sub-region and the control input and environmental disturbance, and is the basis for achieving deformation control. Common modeling methods include finite element analysis, data-driven modeling, and physical mechanism modeling. Finite element analysis obtains the static or dynamic relationship between deformation response and control force by meshing the sub-region and solving the control equation; data-driven modeling uses machine learning methods to fit the mapping relationship between deformation and control force and environmental factors through test data or simulation data; physical mechanism modeling derives the analytical model of deformation control based on physical laws such as material mechanics and aeroelasticity. In the modeling process, it is necessary to comprehensively consider the performance indicators such as model accuracy, computational efficiency, and robustness, select appropriate modeling methods and parameter identification algorithms, and verify the effectiveness of the model through experiments.

[0102] Design a corresponding local control law for each sub-region. The local control law calculates the control input required for the region, that is, the compensation pressure, based on the current deformation state and the target deformation state, so that the deformation of the sub-region gradually approaches the target value. The design of the control law needs to comprehensively consider factors such as the control objective, actuator characteristics, and algorithm performance. Commonly used methods include PID control, optimal control, robust control, and adaptive control. For example, when PID control is used, the control quantity is calculated based on the deformation error and its differential and integral; when optimal control is used, the deformation control problem is converted into a trajectory tracking problem to solve the optimal control sequence; when robust control is used, model uncertainty and external disturbances are considered to design a controller with robust stability and robust performance; when adaptive control is used, the controller parameters are adjusted in real time according to the deformation response to achieve adaptation to dynamic changes and unknown disturbances.

[0103] In order to achieve coordination among sub-regions, a distributed cooperative control algorithm is proposed. The algorithm realizes the exchange and sharing of local compensation pressure, deformation data and key state information among sub-regions through a preset communication protocol. Each sub-region broadcasts its own local compensation pressure, deformation data and other information to other sub-regions, and receives the broadcast information of other sub-regions at the same time. After obtaining the global information, each sub-region inputs the local compensation pressure, deformation data and other information of all sub-regions into the distributed cooperative control model. Through distributed model predictive control and consistency optimization, the coordinated compensation pressure of the region is optimized and calculated in real time under the constraints of meeting the global deformation performance index and local accuracy requirements. Distributed model predictive control predicts the deformation response in the future period and considers the predicted values ​​of other sub-regions to obtain the optimal control sequence in the long time domain; consistency optimization makes the control decisions of each sub-region gradually tend to be consistent through iterative calculation to achieve the global optimum. In the optimization process, it is necessary to reasonably set the weight coefficient to balance the global and local objectives, and adopt distributed optimization algorithms such as distributed gradient descent, distributed sub-gradient, ADMM, etc. to achieve distributed computing and fast convergence.

[0104] According to the coordinated compensation pressure of each sub-region obtained by distributed collaborative control, combined with the characteristics of the pneumatic actuator, such as response time, pressure range, flow characteristics, etc., the actuator is controlled to synchronously apply the required pneumatic force. During the execution process, the deformation data of each part of the sealing surface is measured in real time, the actual deformation data is compared with the target deformation data, and the deformation control error is calculated. The deformation control error reflects the gap between the current deformation state and the expected state, and is an important basis for evaluating control performance and performing feedback correction. The deformation control error is fed back to the distributed collaborative control algorithm as the input for the next optimization step. Through closed-loop iterative optimization, the control decision is continuously corrected until the actual deformation of the sealing surface meets the preset deformation control accuracy requirements.

[0105] In an optional embodiment,

[0106] The calculation formula of the distributed cooperative control algorithm is as follows:

[0107]

[0108] Where U(t) represents the control input vector of all sub-regions, J(·) represents the global optimization objective function, X(t) represents the deformation state vector of all sub-regions, G represents the number of sub-regions, and X a (t) represents the deformation state vector of the a-th sub-region at time t, r a (t) represents the expected state of the a-th sub-region at time t, Q a represents the weight matrix for measuring the state tracking error of the a-th sub-region, u a (t) represents the control input vector of the a-th sub-region at time t, R a represents the weight matrix for measuring the control input of the a-th sub-region, x b (t) represents the deformation state vector of the b-th sub-region at time t, S ab Represents a matrix that measures the consistency of deformation data between the a-th sub-region and the b-th sub-region.

[0109] In this embodiment, deformation sensors are arranged in sensitive areas to measure their three-dimensional spatial coordinates and deformation in real time as input data for subsequent modeling and control. Through distributed collaborative control, information interaction and decision coordination between sub-areas are realized, which not only ensures global deformation performance but also takes into account local control accuracy. Closed-loop control can effectively suppress the influence of environmental disturbances and modeling errors, improve the robustness and accuracy of deformation control, and realize accurate compensation and collaborative control of sealing surface deformation. Compared with traditional centralized control methods, this method has the advantages of strong adaptability, efficient calculation, and scalability, and provides a new solution for the precise deformation control of wind laser radar housing.

[0110] S103. Obtain the airtightness control parameters under different working conditions, perform self-learning optimization on the airtightness control parameters to obtain the optimal control parameter matrix, store the control parameter matrix in the controller, and simultaneously use a pre-trained airtightness prediction model to predict the airtightness performance of the wind measuring laser radar casing. When the predicted gas leakage rate exceeds the preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

[0111] Among them, the air tightness prediction model is trained by collecting state parameters such as air pressure and temperature in the inner cavity of the shell, as well as interference factors such as ambient temperature and vibration. The prediction model takes the current state parameters as input and outputs the predicted value of the gas leakage rate in the future. When the predicted gas leakage rate exceeds the pre-set warning threshold, the system will issue an air tightness alarm signal to remind the user to take corresponding maintenance measures. At the same time, the air tightness adaptive control process is triggered, and the compensation pressure of the sensitive area is automatically adjusted according to the control parameter matrix obtained in advance to achieve rapid recovery of air tightness performance. The air tightness prediction model adopts time series deep learning models such as long short-term memory networks, which can explore the long-term dependence between state parameters and air tightness performance, and improve the timeliness and accuracy of the prediction.

[0112] The self-learning optimization algorithm autonomously explores the control parameter space through continuous trial and error and learning, and searches for the parameter combination that optimizes the system performance indicators. Commonly used self-learning optimization algorithms include heuristic search algorithms such as genetic algorithms, particle swarm optimization algorithms, and simulated annealing algorithms.

[0113] In an optional embodiment,

[0114] Acquiring airtightness control parameters under different working conditions, and performing self-learning optimization on the airtightness control parameters to obtain an optimal control parameter matrix includes:

[0115] An airtightness data acquisition system based on a heterogeneous sensor array is designed. The heterogeneous sensor array includes a gas leakage rate sensor, a pressure sensor, a temperature sensor, and a vibration sensor, which are respectively used to collect airtightness-related information inside the shell of a wind-measuring lidar. In view of the differences in sampling frequency, accuracy, and resolution of different sensors in the heterogeneous sensor array, a multi-sensor data synchronization and fusion algorithm based on Kalman filtering is introduced to perform spatiotemporal alignment and feature extraction on the collected airtightness-related information to obtain airtightness-related data features.

[0116] Preprocess and clean the features of airtightness-related data, obtain standard airtightness-related data through data quality assessment and anomaly detection, and design a self-learning optimization algorithm that integrates reinforcement learning and evolutionary algorithms based on the standard airtightness-related data;

[0117] The deep deterministic policy gradient algorithm is used to model the airtightness control problem of the wind lidar housing as a Markov decision process in a continuous action space. The airtightness performance represented by standard airtightness-related data is used as the state, the airtightness control parameters are used as the action, and the airtightness index is used as the reward function. Through interactive learning between the agent and the environment, the optimized airtightness control strategy and optimal airtightness control parameters are obtained.

[0118] The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met, and the optimal control parameter matrix covering the optimal control strategy of the wind measuring lidar shell under different working conditions is obtained.

[0119] For example, in order to fully perceive and collect information related to the air tightness of the wind laser radar housing, an air tightness data acquisition system based on a heterogeneous sensor array is designed. The heterogeneous sensor array consists of a gas leakage rate sensor, a pressure sensor, a temperature sensor, and a vibration sensor, which are used to collect air tightness related parameters such as gas leakage rate, pressure, temperature, and vibration inside the housing. Different types of sensors are arranged at key positions of the housing to form a distributed perception network, which can monitor the spatial distribution and dynamic changes of the housing air tightness in real time. Considering the differences in sampling frequency, accuracy, and resolution among different sensors, a multi-sensor data synchronization and fusion algorithm based on Kalman filtering is introduced. The algorithm realizes the alignment and association of heterogeneous sensor data in time and space scales through state space modeling and recursive estimation, and utilizes the complementarity and redundancy of multi-sensor information to improve the accuracy and robustness of data fusion. On the basis of data fusion, key parameters reflecting air tightness characteristics, such as gas leakage rate, pressure change rate, temperature gradient, vibration frequency, etc., are extracted to form an air tightness related data feature set, which provides input for subsequent data analysis and control strategy optimization.

[0120] In order to ensure the quality and consistency of airtightness-related data, the collected airtightness-related data features were preprocessed and cleaned. First, through data quality assessment, outliers, missing values, duplicate values ​​and other noise data were detected and eliminated to improve the validity and reliability of the data. Then, data format conversion, unit unification, time synchronization and other processing were carried out for data from different sources and types to eliminate the heterogeneity and inconsistency between data. On this basis, anomaly detection algorithms such as isolated forests and local anomaly factors were used to identify and locate abnormal patterns and events in airtightness-related data, providing a basis for fault diagnosis and predictive maintenance. After preprocessing and cleaning, a standardized airtightness-related data set was obtained with unified data format, controllable quality and clear features, which laid a data foundation for the design and training of self-learning optimization algorithms.

[0121] After obtaining standardized airtightness-related data, a self-learning optimization algorithm integrating reinforcement learning and evolutionary algorithms was designed to adaptively optimize airtightness control strategies and parameters. Based on the deep deterministic policy gradient algorithm, the algorithm models the airtightness control problem of the wind lidar housing as a Markov decision process in a continuous action space. Among them, the state space is represented by standard airtightness-related data, reflecting the real-time state of the housing airtightness performance; the action space is composed of airtightness control parameters, such as inflation pressure, gas flow, temperature control, etc.; the reward function is defined according to airtightness indicators, such as leakage rate, pressure stability, etc., which is used to evaluate the pros and cons of the control strategy. Through interactive learning between the agent and the environment, the agent takes actions according to the current state, and the environment generates the next state and reward according to the action. The agent optimizes the control strategy by maximizing the cumulative reward, and finally obtains the optimal airtightness control strategy and parameters under different working conditions.

[0122] In this embodiment, by designing an airtightness data acquisition system based on a heterogeneous sensor array and introducing a multi-sensor data synchronization and fusion algorithm based on Kalman filtering, the effective fusion of multi-source heterogeneous airtightness related data such as gas leakage rate, pressure, temperature, vibration, etc. can be achieved, and the temporal and spatial consistency and quality of the data can be improved. By integrating reinforcement learning and evolutionary algorithms, the airtightness control problem of the wind laser radar shell is modeled as a Markov decision process, and the airtightness control strategy learning is realized through the interaction between the intelligent agent and the environment. At the same time, the genetic algorithm is introduced to guide and optimize the exploration process, and the optimal airtightness control parameter matrix covering various working conditions can be efficiently obtained. The obtained optimal control parameter matrix contains the optimal control strategy of the wind laser radar under different working conditions, which can realize the intelligent control of the airtightness of the wind laser radar shell, improve the control performance, and enhance the adaptive ability.

[0123] In an optional embodiment,

[0124] The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met. The optimal control parameter matrix covering the optimal control strategy of the wind laser radar shell under different working conditions is obtained, including:

[0125] The policy function parameters of the deep deterministic policy gradient algorithm are used as the individual encoding of the genetic algorithm. Each individual represents a set of possible policy function parameter combinations, and the population size is set to ensure the diversity and coverage of the exploration.

[0126] A certain number of individuals are generated through random sampling or heuristic methods to form the initial population of the genetic algorithm. In each iteration of the genetic algorithm, the individuals in the current population are input into the deep deterministic policy gradient algorithm. The cumulative reward of each individual corresponding to the policy function is obtained through interaction with the environment. At the same time, the diversity measure of the individual and other individuals in the population is calculated, and the fitness function is constructed to comprehensively evaluate the quality of the individual.

[0127] According to the evaluation result of the fitness function, the father is selected from the current population, and the selected father individuals are paired with each other, and the encoding information of the father individuals is exchanged through the crossover operator to generate new offspring individuals;

[0128] The encoded information of the new generation of individuals is mutated, and the resulting offspring individuals are merged with the parent population to form a new generation population, which is then input into the next round of iteration. At the same time, in each iteration, the actor network and critic network of the deep deterministic policy gradient algorithm are used to update the parameters of the individual's corresponding policy function based on the empirical data obtained from the interaction between the individual and the environment.

[0129] Repeat the iteration until the preset number of iterations is met, select the individual with the highest fitness function value from the last generation of the population, extract the corresponding strategy function parameters, and use them as the optimal strategy parameters for airtightness control of the wind laser radar casing. Apply them to the wind laser radar system, and adaptively adjust the control parameters according to the characteristics of airtightness-related data. Finally, the optimal control parameter matrix covering the optimal control strategy of the wind laser radar casing under different working conditions is obtained.

[0130] Exemplarily, in order to further improve the exploration efficiency and global convergence of reinforcement learning, the genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through operations such as population initialization, selection, crossover and mutation, the genetic algorithm guides and optimizes the exploration process of reinforcement learning in a larger search space. First, the policy function parameters of the deep deterministic policy gradient algorithm are used as the individual encoding of the genetic algorithm, and each individual represents a set of possible policy function parameter combinations. Through random sampling or heuristic methods, a certain number of individuals are generated to form the initial population of the genetic algorithm. The setting of the population size needs to take into account the diversity of exploration and computational efficiency. In each iteration of the genetic algorithm, the individuals in the current population are input into the deep deterministic policy gradient algorithm, and the cumulative reward of each individual corresponding to the policy function is obtained through interaction with the environment. At the same time, the individual diversity metric is introduced to evaluate the difference between the individual and other individuals in the population in the policy function parameter space, and a fitness function that comprehensively considers the cumulative reward and diversity is constructed to evaluate the pros and cons of the individual.

[0131] According to the evaluation results of the fitness function, a certain number of excellent individuals are selected from the current population as parents using selection operators such as roulette or tournaments. Then, the selected parent individuals are paired with each other according to a certain probability, and the encoding information of the parent individuals is exchanged according to certain rules through the crossover operator to generate new offspring individuals. In order to maintain the diversity of the population and escape from the local optimum, the encoding information of the offspring individuals is mutated with a certain probability to introduce new gene combinations. The generated offspring individuals are merged with the parent population to form a new generation of population, which is input into the next round of iteration. In each generation of iteration, the actor network and critic network of the deep deterministic policy gradient algorithm are used to update the parameters of the individual's corresponding policy function based on the empirical data obtained by the interaction between the individual and the environment, so as to realize the online learning and optimization of the policy function.

[0132] The iterative process of genetic algorithm and reinforcement learning is repeated until the preset number of iterations or convergence conditions are met. At the end of the optimization, the individual with the highest fitness function value is selected from the last generation of the population, and its corresponding strategy function parameters are extracted as the optimal strategy parameters for the airtightness control of the wind laser radar shell. The optimal strategy parameters are applied to the control module of the wind laser radar system. According to the characteristics of the airtightness-related data collected in real time, the control variables such as inflation pressure, gas flow, and temperature are adaptively adjusted to keep the airtightness performance of the shell in the optimal state. At the same time, for different working conditions, such as ambient temperature, vibration frequency, working mode, etc., the optimization algorithm gives the corresponding optimal control strategy, forming an optimal control parameter matrix covering the wind laser radar shell under different working conditions, realizing the adaptive and full-condition optimization of airtightness control.

[0133] In this embodiment, the genetic algorithm is introduced into the framework of the deep deterministic policy gradient algorithm, and the global optimization ability of population evolution is utilized to effectively avoid the strategy search process from falling into the local optimum, thereby improving the search ability of the global optimal solution. By setting a reasonable population size and generating the initial population by random sampling or heuristic method, it can be ensured that the exploration of the strategy parameter space has sufficient diversity and coverage, thereby increasing the possibility of finding the global optimal solution. Integrating the diversity metric into the fitness function evaluation can optimize the adaptability of individuals while maintaining population diversity. After genetic operations such as crossover and mutation, the new generation of populations has stronger adaptability on the basis of retaining the excellent genes of the parent generation. During the population evolution process, the coded information of excellent individuals will be diffused and transmitted in the population through crossover and mutation operations, thereby realizing the sharing and migration of knowledge within the population and improving the convergence speed of the algorithm. The obtained optimal control parameter matrix can adaptively adjust the control strategy and parameters online according to the real-time airtightness-related data characteristics, thereby meeting the optimal control requirements of the wind laser radar under various working conditions.

[0134] Figure 2 FIG. 1 is a schematic diagram of the structure of the air tightness adaptive control system of the wind laser radar housing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0135] The first unit is used to reconstruct the three-dimensional morphology of the sealing surface of the wind laser radar shell through machine vision methods, extract multi-scale morphological feature parameters reflecting the airtight performance, and simultaneously obtain the air pressure and temperature signals of the inner cavity of the wind laser radar shell. The gas leakage rate of the inner cavity of the shell is calculated based on the air pressure and temperature signals, and a mapping model between multi-scale morphological feature parameters and gas leakage rate is established;

[0136] The second unit is used to obtain the characteristic parameter variation law under different leakage rates according to the mapping model, and predict the deformation and force distribution of the sealing surface through airtightness simulation analysis in combination with the mechanical properties of the wind laser radar shell material, and obtain the position coordinates of the sensitive area. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar shell is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements;

[0137] The third unit is used to obtain the airtightness control parameters under different working conditions, perform self-learning optimization on the airtightness control parameters to obtain the optimal control parameter matrix, store the control parameter matrix in the controller, and simultaneously use a pre-trained airtightness prediction model to predict the airtightness performance of the wind measuring laser radar casing. When the predicted gas leakage rate exceeds a preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

[0138] According to a third aspect of the embodiments of the present invention,

[0139] An electronic device is provided, comprising:

[0140] processor;

[0141] a memory for storing processor-executable instructions;

[0142] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0143] According to a fourth aspect of the embodiments of the present invention,

[0144] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0145] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adaptively controlling the air tightness of a wind laser radar housing, characterized in that: include: The three-dimensional morphology of the sealing surface of the wind laser radar shell is reconstructed through machine vision methods, and the multi-scale morphological feature parameters reflecting the airtight performance are extracted. At the same time, the air pressure and temperature signals of the inner cavity of the wind laser radar shell are obtained. The gas leakage rate of the inner cavity of the shell is calculated based on the air pressure and temperature signals, and a mapping model between the multi-scale morphological feature parameters and the gas leakage rate is established. According to the mapping model, the variation law of characteristic parameters under different leakage rates is obtained. Combined with the mechanical properties of the wind laser radar housing material, the deformation and force distribution of the sealing surface are predicted through airtightness simulation analysis, and the position coordinates of the sensitive area are obtained. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar housing is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements; The airtightness control parameters under different working conditions are obtained, and the optimal control parameter matrix is ​​obtained by self-learning optimization of the airtightness control parameters. The control parameter matrix is ​​stored in the controller, and the airtightness prediction model that has been pre-trained is used to predict the airtightness performance of the wind measuring lidar casing. When the predicted gas leakage rate exceeds the preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

2. The method according to claim 1, characterized in that The three-dimensional morphology of the sealing surface of the wind laser radar shell is reconstructed by machine vision methods, and the multi-scale morphological feature parameters reflecting the airtight performance are extracted. At the same time, the air pressure and temperature signals of the inner cavity of the wind laser radar shell are obtained, and the gas leakage rate of the inner cavity of the shell is calculated according to the air pressure and temperature signals. The mapping model between the multi-scale morphological feature parameters and the gas leakage rate is established, including: Acquire multi-view images of the sealing surface of the wind laser radar housing, reconstruct a three-dimensional point cloud of the sealing surface by a stereo matching method, and project the three-dimensional point cloud onto a two-dimensional plane to obtain a depth map of the sealing surface; Performing multi-scale wavelet transform on the sealing surface depth map, extracting frequency domain characteristics at different scales, determining the deformation area at each scale by using the maximum suppression algorithm, calculating the geometric features and spectral features in each deformation area, and obtaining the topological invariant features at different scales by using persistent coherence, and constructing a multi-scale morphological feature set reflecting the local deformation of the sealing surface; The features in the multi-scale morphological feature set are screened using a maximum correlation minimum redundancy algorithm, and the correlation between each feature and the gas leakage rate is calculated by mutual information measurement. The optimal features are selected one by one with the goal of maximizing the correlation between the feature and the gas leakage rate and minimizing the redundancy between the features, and a feature subset whose correlation with the gas leakage rate is higher than a preset correlation threshold is screened out; The support vector machine algorithm is used to construct a nonlinear regression mapping model between gas leakage rate and feature subsets. The Gaussian radial basis kernel function is selected as the kernel function of the support vector machine. The kernel function width parameters and model hyperparameters are optimized through grid search and cross validation to obtain a mapping model that can estimate the airtight performance within a specified leakage rate range.

3. The method according to claim 2, characterized in that Obtain multi-view images of the sealing surface of the wind laser radar housing, and reconstruct the three-dimensional point cloud of the sealing surface by stereo matching method, including: The three-dimensional point cloud of the sealing surface is constructed as an undirected graph, wherein the nodes represent the points in the three-dimensional point cloud, and the edges represent the spatial constraint relationship between the points in the three-dimensional point cloud; Obtaining the Euclidean distance between any two adjacent points in the undirected graph in the three-dimensional point cloud, and constructing a global error energy function based on the coordinate difference and the Euclidean distance between any two adjacent points, wherein the global error energy function is used to measure the optimization degree of the three-dimensional point cloud; An incremental equation is constructed based on the global error energy function to solve the increment of the three-dimensional point cloud coordinates, and the three-dimensional point cloud coordinates are updated using the increment of the three-dimensional point cloud coordinates. The iteration is repeated until the preset convergence condition is met to obtain the reconstructed three-dimensional point cloud of the sealing surface; Among them, the calculation formula of the global error energy function is as follows: Where E(V) represents the global error energy function, n represents the number of points in the 3D point cloud, and v j represents the jth point, N(v i ) represents the set of neighboring points of the i-th point, ρ represents the robust kernel function, v i represents the i-th point, d ij Represents the Euclidean distance between the i-th point and the j-th point.

4. The method according to claim 1, characterized in that: According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar housing is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model. The active compensation for the deformation of the sealing surface includes: The sealing surface is divided into multiple sub-areas, each of which is equipped with an independent controller and pneumatic actuator, and the position coordinates and deformation data of the sensitive area in each sub-area are obtained; According to the position coordinates and deformation data of the sensitive areas in each sub-area, combined with the geometric characteristics and environmental conditions of the sub-area, a local deformation control model of each sub-area is established; Based on the local deformation control model of each sub-region, the local control law is designed, the local compensation pressure of each sub-region is calculated, and the local compensation pressure, deformation data and key status information of each sub-region are exchanged and shared through the preset communication protocol; The local compensation pressure of each sub-region and the exchanged local compensation pressure, deformation data and key state information of other sub-regions are input into the distributed collaborative control algorithm. Through distributed model predictive control and consistency optimization, the coordinated compensation pressure of each sub-region is optimized and calculated in real time under the constraints of meeting the global deformation performance index and local accuracy requirements. According to the optimized coordinated compensation pressure of each sub-area and the pneumatic actuator characteristics of each sub-area, the corresponding aerodynamic force is applied to each sub-area synchronously, and the actual deformation data of the sealing surface is measured in real time. The actual deformation data is compared with the target deformation data, and the deformation control error is calculated. The deformation control error is fed back to the distributed collaborative control algorithm for closed-loop iterative optimization until the actual deformation of the sealing surface meets the preset deformation control accuracy requirement, thereby realizing active compensation for the deformation of the sealing surface.

5. The method according to claim 4, characterized in that The calculation formula of the distributed cooperative control algorithm is as follows: Where U(t) represents the control input vector of all sub-regions, J(·) represents the global optimization objective function, X(t) represents the deformation state vector of all sub-regions, G represents the number of sub-regions, and X a (t) represents the deformation state vector of the a-th sub-region at time t, r a (t) represents the expected state of the a-th sub-region at time t, Q a represents the weight matrix for measuring the state tracking error of the a-th sub-region, u a (t) represents the control input vector of the a-th sub-region at time t, R a represents the weight matrix for measuring the control input of the a-th sub-region, x b (t) represents the deformation state vector of the b-th sub-region at time t, S ab Represents a matrix that measures the consistency of deformation data between the a-th sub-region and the b-th sub-region.

6. The method according to claim 1, characterized in that Acquiring airtightness control parameters under different working conditions, and performing self-learning optimization on the airtightness control parameters to obtain an optimal control parameter matrix includes: An airtightness data acquisition system based on a heterogeneous sensor array is designed. The heterogeneous sensor array includes a gas leakage rate sensor, a pressure sensor, a temperature sensor, and a vibration sensor, which are respectively used to collect airtightness-related information inside the shell of a wind-measuring lidar. In view of the differences in sampling frequency, accuracy, and resolution of different sensors in the heterogeneous sensor array, a multi-sensor data synchronization and fusion algorithm based on Kalman filtering is introduced to perform spatiotemporal alignment and feature extraction on the collected airtightness-related information to obtain airtightness-related data features. Preprocess and clean the features of airtightness-related data, obtain standard airtightness-related data through data quality assessment and anomaly detection, and design a self-learning optimization algorithm that integrates reinforcement learning and evolutionary algorithms based on the standard airtightness-related data; The deep deterministic policy gradient algorithm is used to model the airtightness control problem of the wind lidar housing as a Markov decision process in a continuous action space. The airtightness performance represented by standard airtightness-related data is used as the state, the airtightness control parameters are used as the action, and the airtightness index is used as the reward function. Through interactive learning between the agent and the environment, the optimized airtightness control strategy and optimal airtightness control parameters are obtained. The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met, and the optimal control parameter matrix covering the optimal control strategy of the wind measuring lidar shell under different working conditions is obtained.

7. The method according to claim 6, characterized in that The genetic algorithm is introduced into the deep deterministic policy gradient algorithm. Through population initialization, selection, crossover and mutation operations, the exploration process of the deep deterministic policy gradient algorithm is guided and optimized. The iteration is repeated until the preset number of iterations is met. The optimal control parameter matrix covering the optimal control strategy of the wind laser radar shell under different working conditions is obtained, including: The policy function parameters of the deep deterministic policy gradient algorithm are used as the individual encoding of the genetic algorithm. Each individual represents a set of possible policy function parameter combinations, and the population size is set to ensure the diversity and coverage of the exploration. A certain number of individuals are generated through random sampling or heuristic methods to form the initial population of the genetic algorithm. In each iteration of the genetic algorithm, the individuals in the current population are input into the deep deterministic policy gradient algorithm. The cumulative reward of each individual corresponding to the policy function is obtained through interaction with the environment. At the same time, the diversity measure of the individual and other individuals in the population is calculated, and the fitness function is constructed to comprehensively evaluate the quality of the individual. According to the evaluation result of the fitness function, the father is selected from the current population, and the selected father individuals are paired with each other, and the encoding information of the father individuals is exchanged through the crossover operator to generate new offspring individuals; The encoded information of the new generation of individuals is mutated, and the resulting offspring individuals are merged with the parent population to form a new generation population, which is then input into the next round of iteration. At the same time, in each iteration, the actor network and critic network of the deep deterministic policy gradient algorithm are used to update the parameters of the individual's corresponding policy function based on the empirical data obtained from the interaction between the individual and the environment. Repeat the iteration until the preset number of iterations is met, select the individual with the highest fitness function value from the last generation of the population, extract the corresponding strategy function parameters, and use them as the optimal strategy parameters for airtightness control of the wind laser radar casing. Apply them to the wind laser radar system, and adaptively adjust the control parameters according to the characteristics of airtightness-related data. Finally, the optimal control parameter matrix covering the optimal control strategy of the wind laser radar casing under different working conditions is obtained.

8. An adaptive air tightness control system for a wind laser radar housing, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to reconstruct the three-dimensional morphology of the sealing surface of the wind laser radar shell through machine vision methods, extract multi-scale morphological feature parameters reflecting the airtight performance, and simultaneously obtain the air pressure and temperature signals of the inner cavity of the wind laser radar shell. The gas leakage rate of the inner cavity of the shell is calculated based on the air pressure and temperature signals, and a mapping model between multi-scale morphological feature parameters and gas leakage rate is established; The second unit is used to obtain the characteristic parameter variation law under different leakage rates according to the mapping model, and predict the deformation and force distribution of the sealing surface through airtightness simulation analysis in combination with the mechanical properties of the wind laser radar shell material, and obtain the position coordinates of the sensitive area. According to the position coordinates of the sensitive area, the sealing surface of the wind laser radar shell is dynamically pressure-regulated, and the compensation pressure is adjusted in real time according to the pre-built control model to actively compensate for the deformation of the sealing surface until the airtightness meets the use requirements; The third unit is used to obtain the airtightness control parameters under different working conditions, perform self-learning optimization on the airtightness control parameters to obtain the optimal control parameter matrix, store the control parameter matrix in the controller, and simultaneously use a pre-trained airtightness prediction model to predict the airtightness performance of the wind measuring laser radar casing. When the predicted gas leakage rate exceeds a preset warning threshold, an airtightness alarm signal is output, and the airtightness adaptive control process is started according to the stored optimal control parameter matrix.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images