Pneumatic data intelligent analysis method based on pneumatic rule mode matching

Abstract

According to the pneumatic data intelligent analysis method based on pneumatic rule mode matching, pneumatic characteristics are added to wind tunnel historical data, the problem of data normalization is avoided, the problem of data length inconsistency caused by wind tunnel test state differences is solved, the range of available historical pneumatic data is widened, and the reliability of the wind tunnel test is improved. A self-organizing neural network intelligent algorithm is adopted to mine potential aerodynamic laws, similarity measurement of the aerodynamic laws is achieved through an Euclidean distance method, the problem that the complex, changeable and highly-coupled aerodynamic laws cannot be manually represented and defined is solved, and an aerodynamic single-component and multi-component analysis method is designed. Wind tunnel historical test data with similar aerodynamic laws are intelligently matched, and a data auxiliary analysis effect is provided for a wind tunnel test which is being carried out.

Description

Technical Field

[0001] The present invention relates to an intelligent analysis method for pneumatic data based on pneumatic law pattern matching, belonging to the interdisciplinary field of artificial intelligence and aerodynamics. Background Art

[0002] First, briefly introduce the basic knowledge and technical terms of the industry involved in this patent. The types of wind tunnel tests include force measurement tests, pressure measurement tests, inlet tests, capture trajectory tests, heat ablation tests, free flight tests, etc. Among them, the force measurement test is a basic and large-scale wind tunnel test, resulting in the largest amount of historical data accumulated in force measurement tests, which is a treasure trove waiting to be developed. However, there has been no effective method to fully explore and utilize this treasure trove.

[0003] The state of a wind tunnel force measurement test generally refers to the angle of attack α, sideslip angle β, roll angle , pitch rudder deflection d p , yaw rudder deflection d y , roll rudder deflection d r , and flight Mach number Ma.

[0004] The pneumatic data obtained from a wind tunnel force measurement test generally refers to the six-component force and moment coefficients obtained by converting the electrical signals measured by a balance through a formula under different wind tunnel force measurement test states, specifically the forces in the X, Y, and Z directions and the moments in the Mx, My, and Mz directions.

[0005] Secondly, as a scarce resource, wind tunnels can no longer meet the needs of aircraft development. It is urgent to improve the efficiency of wind tunnel tests, and the rapid analysis of wind tunnel test data is the key to improving the efficiency of wind tunnel tests. At present, the analysis of wind tunnel test data mainly relies on pneumatic experts with many years of wind tunnel work experience. With the sharp increase in the amount of wind tunnel tests, the purely manual method is inefficient, and there are problems such as large workload, insufficient personnel, and analysis misjudgments caused by fatigue operations, which seriously restrict the production efficiency of wind tunnels. Therefore, there is an urgent need for a fast and effective analysis method for the pneumatic data obtained from wind tunnel tests.

[0006] Traditional pneumatic data analysis methods usually conduct wind tunnel tests on a new aircraft shape and obtain pneumatic data. When analyzing the pneumatic data, it is hoped to find similar shapes in historical tests. The pneumatic data of these similar shapes can be used to evaluate the correctness and reliability of the laws of newly obtained pneumatic data, and at the same time, it is also an important resource for tracing complex pneumatic laws. This analysis method greatly improves the data analysis efficiency and supports the smooth progress of a large number of new shape tests.

[0007] However, since the mapping relationship between aircraft shape and aerodynamic data is a high-dimensional (aerodynamic shape belongs to high-dimensional geometric features) to low-dimensional (usually six aerodynamic forces and moment coefficients) mapping, even if there are thousands of aerodynamic data of aircraft shapes, it still belongs to the category of "small sample". This leads to the limitation of the application scenarios of the above analysis method, and in most cases similar shapes cannot be found.

[0008] In fact, there are many cases where aircraft with very different appearances have similar aerodynamic laws. That is, searching for historical aerodynamic data with similar aerodynamic laws through the aerodynamic laws presented by the aerodynamic data is an analysis method purely based on aerodynamic laws. This is because the analysis and mining of common aerodynamic laws is not only beneficial for wind tunnel operators to conduct discriminant analysis on newly generated aerodynamic data, but also can guide the overall design department of the aircraft to optimize the design of the aircraft shape, and even to mine potential physical laws to achieve more valuable knowledge discovery. However, this data-based analysis method cannot be completed manually because the aerodynamic data and the aerodynamic laws it presents are massive and high-dimensional. This is also an important reason why historical wind tunnel test data, as a treasure, has not been fully mined and utilized.

[0009] The artificial intelligence technology, which is booming at this stage, is good at processing massive high-dimensional data. Therefore, it is a task with great development potential to develop data mining methods based on artificial intelligence to acquire and utilize the complex aerodynamic laws buried in the data stream, and to use the continuous execution capabilities of machine learning to summarize and utilize the "deep laws" and "complex laws" that are difficult for human power to summarize.

[0010] The key to the wind tunnel test data analysis method based on similar aerodynamic laws is how to characterize the aerodynamic laws and measure the similarity of the aerodynamic laws. The characterization method of the aerodynamic laws contained in the traditional aerodynamic data generally takes one wind tunnel test train as the minimum research unit. The horizontal axis is generally the angle of attack α of the aircraft, and the vertical axis is the force and moment coefficient of a certain component. The other test states of the aircraft are the sideslip angle β, the roll angle , pitch rudder deflection d p 、Yaw rudder deflection d y 、Roll rudder deflection c r , the influence of flight Mach number Ma on aerodynamic force cannot be reflected in one wind tunnel test train, that is, when studying aerodynamic data analysis based on data-driven and similar aerodynamic law matching and recommendation, the traditional aerodynamic law characterization method cannot fully characterize the aerodynamic law, and there is randomness in a single wind tunnel test train. Therefore, there is an urgent need for an effective aerodynamic law characterization and similarity measurement method, which is an aerodynamic data analysis method based on the mining of historical wind tunnel big data and the matching and recommendation of similar aerodynamic laws. Summary of the invention

[0011] The technical problem solved by the present invention is: aiming at the problem that in the current existing technology, the complex and highly coupled aerodynamic laws cannot be characterized and defined manually, an intelligent aerodynamic data analysis method based on aerodynamic law pattern matching is proposed.

[0012] The present invention solves the above technical problem through the following technical solutions:

[0013] An intelligent aerodynamic data analysis method based on aerodynamic law pattern matching, comprising:

[0014] Collecting historical wind tunnel aerodynamic data and appending aerodynamic characteristics;

[0015] Analyzing and obtaining aerodynamic laws based on the historical wind tunnel aerodynamic data;

[0016] Setting the research object aircraft, collecting the wind tunnel test data of the research object aircraft and appending aerodynamic characteristics;

[0017] Analyzing all the wind tunnel test data of the research object aircraft to obtain aerodynamic laws;

[0018] According to the current aerodynamic law, selecting any wind tunnel test data of the research object aircraft to perform single-component aerodynamic data analysis;

[0019] Traversing all components to perform multi-component aerodynamic data analysis of the research object aircraft.

[0020] The method for collecting historical wind tunnel aerodynamic data and appending aerodynamic characteristics is:

[0021] Taking the wind tunnel test vehicle data as the unit, appending aerodynamic characteristics to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz respectively on each wind tunnel test vehicle data. The aerodynamic characteristics include fitting coefficient characteristics and statistical quantity characteristics, which are respectively constructed according to the six-component aerodynamic forces.

[0022] The construction method of the fitting coefficient characteristics is:

[0023] For any wind tunnel test vehicle data, respectively fitting the six-component aerodynamic forces by the fourth-order Chebyshev method to obtain the Chebyshev coefficients c of each component i ;

[0024] The Chebyshev coefficient c i The subscript i represents the order of the Chebyshev coefficient, c 0 is the constant term, c 1 is the first-order term coefficient, c 2 is the second-order term coefficient, c 3 is the third-order term coefficient, c 4 is the fourth-order term coefficient.

[0025] The construction method of the statistic features is as follows:

[0026] For the data of any wind tunnel test vehicle, calculate the mean, variance, number of peaks num_u, and number of valleys num_l for the six-component aerodynamic forces respectively.

[0027] The method for obtaining the aerodynamic law based on the historical aerodynamic data of the wind tunnel is as follows:

[0028] Construct neural networks net-X, net-Y, net-Z, net-Mx, net-My, net-Mz based on the six-component aerodynamic forces X, Y, Z, Mx, My, Mz respectively;

[0029] Set the parameters of the neighborhood function neighborhood_function, neighborhood radius sigma, and learning rate learning_rate;

[0030] Arbitrarily take the aerodynamic feature of any aerodynamic component of the data of one wind tunnel test vehicle as the input sample X of the neural network j , traverse each node in the competitive layer of the neural network, and calculate X j and the Euclidean distance between the nodes, and select the node with the smallest distance as the winning node;

[0031] Take the winning node as the center, determine the nodes to be updated according to the neighborhood radius sigma, and determine the update amplitude of the nodes to be updated according to the neighborhood function neighborhood_function;

[0032] Taking the completion of the calculation of all historical aerodynamic data as one round of iteration, set the number of iteration rounds M as the stop condition for the neural network training to be completed;

[0033] In the trained neural network of any aerodynamic component, use the number of nodes in the competitive layer to characterize the aerodynamic law of the corresponding aerodynamic component.

[0034] All the neural networks have a competitive layer topological structure. Set the competitive layer topological structure as a two-dimensional plane structure rectangular, and set the side length dimension to and round up. Set N as the number of vehicle trips of all historical aerodynamic data, and the number of nodes in the competitive layer is dimension * dimension;

[0035] In the input sample X j , the value range of j is [1, N], and N is the number of vehicle trips of all historical aerodynamic data;

[0036] The principle for determining the update amplitude of the nodes to be updated is: the update amplitude of the nodes to be updated closer to the winning node is greater than that of the nodes to be updated farther from the winning node.

[0037] For any neural network of a trained pneumatic component, the pneumatic law is represented by the number of nodes dimension*dimension in the competitive layer, and all N pieces of historical pneumatic data are classified into the nodes of the competitive layer according to the pneumatic law.

[0038] The method for single-component pneumatic data analysis using the pneumatic characteristics of the wind tunnel test data of the research object aircraft is as follows:

[0039] Take any one of the six-component data X, Y, Z, Mx, My, and Mz of the wind tunnel test data of the research object aircraft as the research object component W;

[0040] In any wind tunnel test run CC, calculate the Euclidean distance between the pneumatic characteristics of the W component in the CC run and the nodes of the competitive layer of the trained neural network of the W component. The node with the closest distance is the node with successful pattern matching, which is used to represent the highest similarity between the pneumatic law and the pneumatic law of the W component in the CC run;

[0041] After training with the W component, obtain one of the neural networks net-X, net-Y, net-Z, net-Mx, net-My, and net-Mz;

[0042] According to the pneumatic law, classify the nodes of the competitive layer. Calculate the Euclidean distance between all the historical wind tunnel pneumatic data classified into the nodes with successful pattern matching and the pneumatic characteristics of the W component in the CC run, and find the top TOP run data sorted from smallest to largest Euclidean distance as the run data with the highest similarity to the pneumatic law of the W component in the CC run;

[0043] Based on the top TOP run data, determine the corresponding aerodynamic shape and all the historical pneumatic data corresponding to the aerodynamic shape, and analyze the aerodynamic characteristics of the research object aircraft based on the historical data analysis.

[0044] The method for multi-component pneumatic data analysis for all pneumatic components of the wind tunnel test data is as follows:

[0045] For a wind tunnel test run CC, calculate the Euclidean distance between the pneumatic characteristics of the six components in the CC run and the nodes of the competitive layer of the trained neural network of the corresponding components respectively. The node with the closest distance is the node with successful pattern matching for each component. The nodes with successful pattern matching for the six components are used to represent the highest similarity between the pneumatic law and the six-component pneumatic law corresponding to the CC run;

[0046] Obtain the historical wind tunnel pneumatic data classified into the nodes with successful pattern matching, denoted as data-X, data-Y, data-Z, data-Mx, data-My, and data-Mz;

[0047] Calculate the intersections Set1, Set2, Set3, Set4, and Set5 of the historical wind tunnel data corresponding to each aerodynamic component;

[0048] Based on the aerodynamic shape of the research object aircraft corresponding to each intersection and all the historical aerodynamic data corresponding to the aerodynamic shape, as the highest - order data of the six - component aerodynamic law similarity, it is used to analyze the aerodynamic data of the research object aircraft.

[0049] The calculation method for the intersections of the historical wind tunnel data corresponding to each aerodynamic component is as follows:

[0050] Determine the analysis order of the aerodynamic components X, Y, Z, Mx, My, and Mz;

[0051] Use the intersection of data - X and data - Y as intersection Set1;

[0052] Use the intersection of Set1 and data - Z as intersection Set2;

[0053] Use the intersection of Set2 and data - Mx as intersection Set3;

[0054] Use the intersection of Set3 and data - My as intersection Set4;

[0055] Use the intersection of Set4 and data - Mz as intersection Set5;

[0056] When the intersection of any two data sets is empty:

[0057] If Set5 is not empty, Set = Set5;

[0058] Otherwise, if Set4 is not empty, Set = Set4;

[0059] Otherwise, if Set3 is not empty, Set = Set3;

[0060] Otherwise, if Set2 is not empty, Set = Set2;

[0061] Otherwise, if Set1 is not empty, Set = Set1.

[0062] The advantages of the present invention compared with the prior art are as follows:

[0063] (1) An intelligent pneumatic data analysis method based on pneumatic law pattern matching provided by the present invention mines potential pneumatic laws by designing pneumatic characteristics of aerodynamic forces and moments and combining self-organizing neural network intelligent algorithms, solves the problem that complex and highly coupled pneumatic laws cannot be characterized and defined manually, and can intelligently match the historical wind tunnel test data with similar pneumatic laws, providing data-assisted analysis for the ongoing wind tunnel test, thereby improving the efficiency of the wind tunnel test and providing guarantee for the rapid development of aerospace aircraft in China;

[0064] (2) The present invention does not need to regularize and label the existing wind tunnel historical data. The application of artificial intelligence methods can be realized only by appending data after the existing data, saving manpower and material resources. At the same time, due to the wind tunnel tests of different aircraft configurations, the length of the angle of attack sequence and the minimum and maximum angles of attack may be different, resulting in the inability to perform mathematical operation comparisons on the pneumatic data of different aircraft configurations. The present invention selects the method of appending pneumatic characteristics. The appended pneumatic characteristics include Chebyshev fitting coefficient characteristics and statistic characteristics, solving the problem of unified measurement of pneumatic data and enabling the value of a large amount of wind tunnel test data to be utilized;

[0065] (3) The present invention selects the Chebyshev fitting method. Compared with polynomial and other fitting methods, it can effectively reduce the fluctuation of high-order polynomial fitting, effectively avoid the Runge phenomenon, and increase stability; at the same time, the selected Chebyshev fitting method can better control the distribution of errors in the entire angle of attack interval compared with polynomial and other fitting methods, avoid locally excessive errors, and obtain the effect of minimizing the maximum error;

[0066] (4) The present invention selects the fourth-order Chebyshev fitting coefficient as the appended pneumatic characteristic, which can better characterize the pneumatic data. Because the pneumatic data can be roughly divided into a small angle of attack linear region and a large angle of attack non-linear region according to the curve shape. For the linear region, the constant term and the first-order slope trend term can characterize the main pneumatic laws, and the second-order coefficient can describe the pneumatic laws more precisely. For the non-linear region, the second-order coefficient and the third-order coefficient can characterize the main pneumatic laws, and the fourth-order coefficient can describe them more precisely;

[0067] (5) The present invention uses a simple Euclidean distance to measure the similarity of pneumatic laws. The smaller the distance, the higher the similarity of the pneumatic laws. By mining historical data through algorithms for pneumatic laws, the problem of manually defining and characterizing pneumatic laws is solved because the pneumatic laws are along different wind tunnel test state variables, the angle of attack α, sideslip angle β, roll angle pitch rudder deflection d p 、yaw rudder deflection d y 、roll rudder deflection d r, the flight Mach number Ma has different forms of expression. Manual experience can only characterize the rules that occur frequently and have obvious patterns. It is very difficult for humans to summarize, generalize, and characterize the "deep rules" and "complex rules".

[0068] (6) The pattern matching based on aerodynamic rules proposed by the present invention can perform matching analysis on both single-component and multi-component aerodynamics, and can also focus on observing different flight vehicle test states by adjusting the order of multi-component matching. The historical data obtained by the method of the present invention can provide a more instructive auxiliary analysis function. Description of the Drawings

[0069] Figure 1 is the flow chart of the intelligent aerodynamic data analysis method based on aerodynamic rule pattern matching provided by the present invention;

[0070] Figure 2 is the flow chart of mining the aerodynamic rules of the Y unit by the self-organizing neural network provided by the present invention;

[0071] Figure 3 is the flow chart of the intelligent aerodynamic data analysis of the single-component Mz unit provided by the present invention;

[0072] Figure 4 is the flow chart of the intelligent aerodynamic data analysis of the multi-component provided by the present invention. Detailed Embodiments

[0073] An intelligent aerodynamic data analysis method based on aerodynamic rule pattern matching adds aerodynamic features to the historical wind tunnel data, avoiding the problem of data regularization and solving the problem of inconsistent data lengths caused by differences in wind tunnel test states, broadening the range of available historical aerodynamic data. The self-organizing neural network intelligent algorithm is used to mine potential aerodynamic rules, and the similarity measurement of aerodynamic rules is realized by the Euclidean distance method, solving the problem that complex and highly coupled aerodynamic rules cannot be characterized and defined manually. The single-component and multi-component aerodynamic analysis methods are designed to intelligently match the historical wind tunnel test data with similar aerodynamic rules, providing data auxiliary analysis for the ongoing wind tunnel test.

[0074] The steps of the intelligent aerodynamic data analysis method based on aerodynamic rule pattern matching are as follows:

[0075] Collect historical wind tunnel aerodynamic data and add aerodynamic features;

[0076] Obtain aerodynamic rules according to the analysis of historical wind tunnel aerodynamic data;

[0077] Set the research object flight vehicle, collect the wind tunnel test data of the research object flight vehicle and add aerodynamic features;

[0078] Analyze all the wind tunnel test data of the research object aircraft to obtain the aerodynamic laws;

[0079] According to the current aerodynamic laws, select the wind tunnel test data of any research object aircraft for single-component aerodynamic data analysis;

[0080] Traverse all components for multi-component aerodynamic data analysis of the research object aircraft;

[0081] The method for collecting historical wind tunnel aerodynamic data and appending aerodynamic characteristics is as follows:

[0082] Taking the wind tunnel test vehicle data as the unit, append aerodynamic characteristics to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz respectively on each wind tunnel test vehicle data. The aerodynamic characteristics include fitting coefficient characteristics and statistical characteristics, which are constructed respectively according to the six-component aerodynamic forces.

[0083] The construction method of the fitting coefficient characteristics is as follows:

[0084] For any wind tunnel test vehicle data, fit the six-component aerodynamic forces respectively by the fourth-order Chebyshev method to obtain the Chebyshev coefficients c of each component i ;

[0085] The Chebyshev coefficient c i The subscript i represents the order of the Chebyshev coefficient, c 0 is the constant term, c 1 is the first-order coefficient, c 2 is the second-order coefficient, c 3 is the third-order coefficient, c 4 is the fourth-order coefficient.

[0086] The construction method of the statistical characteristics is as follows:

[0087] For any wind tunnel test vehicle data, calculate the mean, variance, number of peaks num_u, and number of valleys num_l of the six-component aerodynamic forces respectively.

[0088] The method for obtaining aerodynamic laws based on historical wind tunnel aerodynamic data analysis is as follows:

[0089] Construct neural networks net-X, net-Y, net-Z, net-Mx, net-My, and net-Mz respectively according to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz;

[0090] Set the parameters of the neighborhood function neighborhood_function, neighborhood radius sigma, and learning rate learning_rate;

[0091] Arbitrarily take the aerodynamic characteristics of any aerodynamic component of the vehicle data in a wind tunnel test as the input sample X of the neural network j Traverse each node in the competitive layer of the neural network and calculate X j The Euclidean distance between and the nodes, and select the node with the smallest distance as the winning node;

[0092] With the winning node as the center, determine the nodes to be updated according to the neighborhood radius sigma, and determine the update amplitude of the nodes to be updated according to the neighborhood function neighborhood_function;

[0093] Taking the completion of the calculation of all historical aerodynamic data of all the numbers as the completion of one round of iteration, and setting the number of iteration rounds M as the stopping condition for the neural network training to be completed;

[0094] In the trained neural network of any aerodynamic component, use the number of nodes in the competitive layer to characterize the aerodynamic law of the corresponding aerodynamic component.

[0095] The neural network has a competitive layer topology structure. Set the competitive layer topology structure as a two-dimensional plane structure rectangular, and set the side length dimension as And round up. Set N as the number of vehicle trips of all historical aerodynamic data, and the number of nodes in the competitive layer is dimension * dimension;

[0096] Input sample X j In, the value range of j is [1, N], and N is the number of vehicle trips of all historical aerodynamic data;

[0097] The principle for determining the update amplitude of the nodes to be updated is that the update amplitude of the nodes to be updated closer to the winning node is greater than that of the nodes to be updated farther from the winning node.

[0098] For the neural network of any aerodynamic component after training, the aerodynamic law is represented by the number of nodes dimension * dimension in the competitive layer, and all N historical aerodynamic data are classified into the nodes of the competitive layer according to the aerodynamic law.

[0099] The method for performing single-component aerodynamic data analysis using the aerodynamic characteristics of the wind tunnel test data of the research object aircraft is:

[0100] Take any one of the six-component X, Y, Z, Mx, My, Mz data of the wind tunnel test data of the research object aircraft as the research object component;

[0101] In any wind tunnel test run CC, calculate the Euclidean distance between the aerodynamic characteristics of the W component of run CC and the nodes of the competitive layer of the neural network trained for the W component. The node with the closest distance is the node where the pattern matching is successful, which is used to represent the highest similarity of the aerodynamic law to the aerodynamic law of the W component of run CC;

[0102] After training with the W component, obtain one of the neural networks net-X, net-Y, net-Z, net-Mx, net-My, net-Mz;

[0103] Classify the nodes of the competitive layer according to the aerodynamic law. Calculate the Euclidean distance between all the historical wind tunnel aerodynamic data classified as the nodes where the pattern matching is successful and the aerodynamic characteristics of the W component of run CC, and find the top TOP train data sorted from smallest to largest Euclidean distance as the train data with the highest similarity of the aerodynamic law to the aerodynamic law of the W component of run CC;

[0104] According to the top TOP train data, determine the corresponding aerodynamic shape and all the historical aerodynamic data corresponding to the aerodynamic shape, and analyze the aerodynamic characteristics of the research object aircraft based on the historical data.

[0105] The method for multi-component aerodynamic data analysis for all aerodynamic components of wind tunnel test data is as follows:

[0106] For a wind tunnel test run CC, calculate the Euclidean distance between the aerodynamic characteristics of the six components of run CC and the nodes of the competitive layer of the neural network trained for the corresponding components respectively. The node with the closest distance is the node where the pattern matching is successful for each component. The nodes where the pattern matching is successful for the six components are used to represent the highest similarity of the aerodynamic law to the six-component aerodynamic law of run CC;

[0107] Obtain the historical wind tunnel aerodynamic data classified as the nodes where the pattern matching is successful, denoted as data-X, data-Y, data-Z, data-Mx, data-My, data-Mz;

[0108] Calculate the intersections Set1, Set2, Set3, Set4, Set5 of the historical data corresponding to each aerodynamic component;

[0109] According to the aerodynamic shape of the research object aircraft corresponding to each intersection and all the historical aerodynamic data corresponding to the aerodynamic shape, as the data with the highest similarity of the six-component aerodynamic law, it is used to analyze the aerodynamic data of the research object aircraft.

[0110] The calculation method for the intersections of the historical data corresponding to each aerodynamic component is as follows:

[0111] Determine the analysis order of the aerodynamic components X, Y, Z, Mx, My, Mz;

[0112] Use the intersection of data-X and data-Y as the intersection Set1;

[0113] Use the intersection of Set1 and data-Z as the intersection Set2;

[0114] Use the intersection of Set2 and data-Mx as the intersection Set3;

[0115] Use the intersection of Set3 and data-My as the intersection Set4;

[0116] Use the intersection of Set4 and data-Mz as the intersection Set5;

[0117] When the intersection of any two data sets is empty:

[0118] If Set5 is not empty, Set = Set5;

[0119] Otherwise, if Set4 is not empty, Set = Set4;

[0120] Otherwise, if Set3 is not empty, Set = Set3;

[0121] Otherwise, if Set2 is not empty, Set = Set2;

[0122] Otherwise, if Set1 is not empty, Set = Set1.

[0123] The following is a further description in conjunction with the accompanying drawings of the specification and preferred embodiments:

[0124] In the current embodiment, an intelligent pneumatic data analysis method based on pneumatic law pattern matching is characterized in that, as Figure 1 shown, it includes the following steps:

[0125] The first step: Append pneumatic features to the historical wind tunnel pneumatic data

[0126] Without changing the data format of the existing historical wind tunnel data, only a part of the data is added to the existing data; compared with the mainstream method of usually regularizing and annotating the data of a new field when applying artificial intelligence to a new field, the advantage of this design is that there is no need to invest manpower and material resources in regularizing the original data, because the historical wind tunnel data is in a non-standard format, and with the optimization and improvement of the wind tunnel equipment and the increase in the types of wind tunnel tests, the data format saved by the wind tunnel over the years has also been changing, resulting in the regularization of the historical wind tunnel data over the years being a very large task. The design of the present invention saves the workload and reduces the threshold for applying artificial intelligence to the pneumatic field, and can focus on the main work to be solved;

[0127] S1. For a large amount of historical wind tunnel aerodynamic data, taking the wind tunnel test vehicle data as a unit, on the basis of each wind tunnel test vehicle data, aerodynamic characteristics are appended to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz respectively.

[0128] The historical aerodynamic data generated by the wind tunnel is saved according to the wind tunnel blowing vehicle numbers. The historical aerodynamic data of one vehicle number includes the wind tunnel test state information, the aircraft model state information during the test of this vehicle number, and the aerodynamic data obtained from the wind tunnel test. Among them, the aerodynamic data generally refers to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz.

[0129] The wind tunnel historical data is saved according to the vehicle numbers. Its abscissa is generally the angle of attack attitude of the aircraft model, and the lengths and values of the angle of attack sequences of different aircraft shapes are different, resulting in difficulty in unifying the data formats.

[0130] S2. The aerodynamic characteristics include fitting coefficient characteristics, and the construction method is as follows:

[0131] For any wind tunnel test vehicle data, the six-component aerodynamic forces are respectively fitted by the fourth-order Chebyshev method to obtain the Chebyshev coefficients c of each component. i , which are the aerodynamic characteristics of their respective components.

[0132] Among them, the Chebyshev coefficient c i The subscript i represents the order of the Chebyshev coefficient, c 0 is the constant term, c 1 is the first-order term coefficient, c 2 is the second-order term coefficient, c 3 is the third-order term coefficient, c 4 is the fourth-order term coefficient;

[0133] The first four-order expressions of Chebyshev are:

[0134] T 0 (x) = 1

[0135] T 1 (x) = x

[0136] T 2 (x) = 2x 2 - 1

[0137] T 3 (x) = 4x 3 - 3x

[0138] T 4 (x) = 8x 4 - 8x 2 + 1

[0139] Taking the My unit aerodynamic component as an example, the abscissa is the angle of attack α, and the aerodynamic value is denoted as fMy (α), the fitting form using the fourth-order Chebyshev method is:

[0140] f My (α) = c 4 *T 4 (α) + c 3 *T 3 (α) + c 2 *T 2 (α) + c 1 *T 1 (α) + c 0 *T 0 (α)

[0141] S3. The aerodynamic characteristics also include statistical characteristics, and the construction method is:

[0142] For the data of any wind tunnel test vehicle, calculate the mean, variance, number of peaks num_u, and number of valleys num_l for the six-component aerodynamic forces respectively;

[0143] The method for obtaining the above statistical characteristics can use ordinary mathematical methods, aiming at implementation, without being limited to a specific method;

[0144] The second step: Mine the aerodynamic laws contained in the historical wind tunnel aerodynamic data

[0145] As Figure 2 shown is the flowchart of the self-organizing neural network for mining the aerodynamic laws of the Y unit in the embodiment of the present invention, and the specific steps are as follows:

[0146] S4. Conduct research on the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz of all historical aerodynamic data respectively, and use the self-organizing neural network algorithm to mine the aerodynamic laws of the six-component aerodynamic forces in an unsupervised manner, that is, train the neural networks net-X, net-Y, net-Z, net-Mx, net-My, and net-Mz according to the aerodynamic components respectively. The processing methods and steps for each aerodynamic component are the same;

[0147] Figure 2 It is to conduct research on the Y component, and finally train the net-Y network;

[0148] S5. The self-organizing neural network needs to set parameters such as the neighborhood function neighborhood_function, neighborhood radius sigma, and learning rate learning_rate;

[0149] The neighborhood function is used to determine the strength of the influence of the winning node on its neighboring nodes, that is, the update amplitude of the nodes near the winning node. In the present invention, a Gaussian function is selected, which can characterize the relationship between the strength of influence and the distance within the winning neighborhood. The principle is that the closer to the winning node, the greater the update amplitude, and the farther from the winning node, the smaller the update amplitude.

[0150] The sigma parameter of the neighborhood radius determines the range of the winning node's neighborhood. A large value indicates a large influence range. The value must be greater than 0; otherwise, it will not affect any competing nodes, that is, the nodes will not update and learn. The value cannot be greater than the side length of the two-dimensional output plane.

[0151] Since the neighborhood function is a continuous Gaussian function, the effective value range of the neighborhood radius sigma is also continuous. Sigma essentially controls the decay degree with distance. When sigma takes a very small value, such as 0.1, only the update amplitude of the winning node is 1, and the other nodes hardly update. When sigma is set to 1, all nodes have a certain update amplitude. The update amplitude of the central winning node is 1, and the farther away from the winning node, the smaller the update amplitude. When sigma takes a large value, the decay speed is slow, and nodes far away also have a relatively large update amplitude. Through experiments, setting sigma to [1, 2.5] has good results.

[0152] The topological structure of the competitive layer selects a two-dimensional plane structure, rectangular, and the side length dimension is set to and rounded up. Where N is the number of train trips of all historical aerodynamic data. From the network structure, the number of nodes in the competitive layer is dimension * dimension.

[0153] The side length of the competitive layer determines the number of nodes in the competitive layer and the upper limit of the number of aerodynamic laws to be mined. The value of the side length of the competitive layer is related to the data volume and data laws. When the data volume is large and the data laws are relatively complex, a relatively large number of competitive nodes need to be set. It is set to to be more reasonable.

[0154] The weights of the network structure are initialized to very small random numbers.

[0155] Since the number of nodes in the competitive layer is reasonably selected, initializing the weights of the weights using a simple random initialization method can meet the requirements.

[0156] S6. Take the aerodynamic characteristics of a certain aerodynamic component of the data of any wind tunnel test train trip as the input sample X of the neural network j , traverse each node in the competitive layer, and calculate X jThe Euclidean distance between nodes is calculated, and the node with the minimum distance is selected as the winning node;

[0157] Among them, j ranges from [1, N], and N is the number of train trips in all historical aerodynamic data;

[0158] The aerodynamic law represented by the winning node has the highest similarity with the aerodynamic law of the input sample X j ;

[0159] S7. Taking the winning node as the center, determine the nodes to be updated according to the neighborhood radius sigma, and determine the update amplitude of the nodes to be updated according to the neighborhood function neighborhood_function. The principle is that the closer to the winning node, the greater the update amplitude, and the farther from the winning node, the smaller the update amplitude;

[0160] S8. Completing the calculation of all N historical aerodynamic data is considered as one round of iteration. Set the number of iteration rounds M as the stopping condition for the neural network training to be completed;

[0161] The stopping condition of the neural network can be further optimized. For example, calculate the adjustment amplitude of the competitive layer nodes in the recent few rounds. If the amplitude change is less than a certain threshold, it is considered that the neural network has been trained well. However, this method reduces the operation efficiency of the neural network. Considering that the task is to mine classification rather than high-precision tasks, simply setting the number of iteration rounds is sufficient. Through experiments, setting M greater than 2000 can obtain good results;

[0162] S9. For the neural network of any trained aerodynamic component, the nodes in the competitive layer can represent the aerodynamic laws mined for the corresponding aerodynamic component. At most, dimension*dimension kinds of aerodynamic laws can be mined, and all N historical aerodynamic data will be classified into different competitive nodes according to the aerodynamic laws;

[0163] The existence of a small number of empty nodes is the desired effect, indicating that the classification of aerodynamic laws is fine enough. The side length dimension is set in step S5 and is verified to be reasonable;

[0164] The third step: Intelligent analysis of single-component aerodynamic data

[0165] As Figure 3 shown is the intelligent analysis flow chart of the single-component Mz unit aerodynamic data of the embodiment of the present invention

[0166] S10. For the new-shaped aircraft, append aerodynamic features to the newly generated wind tunnel test data according to the method described in the first step of the present invention. Take the W component as the research object, and W is one of the six components X, Y, Z, Mx, My, and Mz;

[0167] Figure 3 Select W as the Mz unit;

[0168] S11. For a wind tunnel test vehicle CC, calculate the Euclidean distance between the aerodynamic characteristics of the W component of the CC vehicle and the nodes of the competitive layer of the trained neural network for the W component. The node with the closest distance is the node where the pattern matching is successful, and the aerodynamic law represented by this node has the highest similarity to the aerodynamic law of the W component of the CC vehicle;

[0169] Among them, the trained neural network for the W component is one of net-X, net-Y, net-Z, net-Mx, net-My, net-Mz described in step S4;

[0170] As Figure 3 shown, select the trained net-Mz network corresponding to step S4;

[0171] S12. Find all the historical aerodynamic data of the wind tunnel classified into the node where the pattern matching is successful, calculate the Euclidean distance between these historical data and the aerodynamic characteristics of the W component of the CC vehicle, and find the top TOP vehicle data sorted from small to large in terms of the Euclidean distance, which are the vehicle data with the highest similarity to the aerodynamic law of the W component of the CC vehicle;

[0172] S13. Based on the top TOP vehicle data, the corresponding aerodynamic shape and all the historical aerodynamic data corresponding to the aerodynamic shape can be traced, and the aerodynamic data of the new shape aircraft can be assisted in analysis based on the historical data;

[0173] Fourth step: Intelligent analysis of multi-component aerodynamic data

[0174] As Figure 4 shown is the flow chart of the intelligent analysis of multi-component aerodynamic data according to the embodiment of the present invention;

[0175] S14. For a new shape aircraft, append the aerodynamic characteristics to the newly generated wind tunnel test data according to the method described in the first step of the present invention, and at the same time analyze the aerodynamic laws of multiple aerodynamic components;

[0176] S15. For a wind tunnel test vehicle CC, calculate the Euclidean distance between the aerodynamic characteristics of the six components of the CC vehicle and the nodes of the competitive layer of the trained neural network for the corresponding components respectively. The node with the closest distance is the node where the respective pattern matching is successful, and the aerodynamic laws represented by the six components' pattern matching successful nodes have the highest similarity to the corresponding six-component aerodynamic laws of the CC vehicle;

[0177] S16. Find the historical aerodynamic data of the wind tunnel classified into the nodes where the pattern matching is successful respectively, denoted as data-X, data-Y, data-Z, data-Mx, data-My, data-Mz;

[0178] S17. Obtain the intersection of the historical wind tunnel data corresponding to each component described in S16. The method is to find the intersection pairwise, and the pairwise combination order can be flexibly adjusted according to research requirements. The operations in the order of X, Y, Z, Mx, My, Mz are as follows:

[0179] Set1 = the intersection of data - X and data - Y;

[0180] Set2 = the intersection of Set1 and data - Z;

[0181] Set3 = the intersection of Set2 and data - Mx;

[0182] Set4 = the intersection of Set3 and data - My;

[0183] Set5 = the intersection of Set4 and data - Mz;

[0184] According to the experience of aerodynamic experts, if the pitch rudder deflection state of the aircraft is the key concern, the recommended pairwise combination order is Y, Mz, Z, My, X, Mx. The specific operations are as follows:

[0185] Set1 = the intersection of data - Y and data - Mz;

[0186] Set2 = the intersection of Set1 and data - Z;

[0187] Set3 = the intersection of Set2 and data - My;

[0188] Set4 = the intersection of Set3 and data - X;

[0189] Set5 = the intersection of Set4 and data - Mx;

[0190] If the yaw rudder deflection information of the aircraft is of more concern, the recommended pairwise combination order is Y, Z, My, Mz, X, Mx. The specific operations are as follows:

[0191] Set1 = the intersection of data - Y and data - Z;

[0192] Set2 = the intersection of Set1 and data - My;

[0193] Set3 = the intersection of Set2 and data - Mz;

[0194] Set4 = the intersection of Set3 and data - X;

[0195] Set5 = the intersection of Set4 and data - Mx;

[0196] S18. For the case where there is no intersection between two data sets, the processing method is as follows:

[0197] If Set5 is not empty, Set = Set5;

[0198] Otherwise, if Set4 is not empty, Set = Set4;

[0199] Otherwise, if Set3 is not empty, Set = Set3;

[0200] Otherwise, if Set2 is not empty, Set = Set2;

[0201] Otherwise, if Set1 is not empty, Set = Set1;

[0202] If Set1 is an empty set, it is considered that the pairwise combination order is inappropriate, and the order needs to be adjusted and recalculated according to the steps of S17 and S18;

[0203] Because the pairwise combination order is set based on expert experience, and on this premise, Set1 is still an empty set, it can be determined that no data with similar aerodynamic laws can be found in the historical data for train CC, that is, the new-shaped aircraft to which train CC belongs has a large difference in shape from all the shapes in the historical wind tunnel tests;

[0204] S19. According to Set, the corresponding aerodynamic shape and all historical aerodynamic data corresponding to the aerodynamic shape can be traced, and the aerodynamic data of the new-shaped aircraft can be assisted in analysis based on the historical data;

[0205] In actual wind tunnel tests, for the new-shaped aircraft test being carried out in the wind tunnel, after obtaining the wind tunnel test data of a small number of train trips, the historical data with similar aerodynamic laws can be matched according to the method of the present invention, so as to trace the historical aircraft shapes and their data with similar aerodynamic laws.

[0206] Although the present invention has been disclosed above with preferred embodiments, it is not used to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solution of the present invention without departing from the spirit and scope of the present invention. Therefore, all contents that do not depart from the technical solution of the present invention, any simple modifications, equivalent changes and decorations made to the above embodiments according to the technical essence of the present invention, all belong to the protection scope of the technical solution of the present invention.

[0207] The content not detailedly described in the specification of the present invention belongs to the well-known technology of those skilled in the art.

Claims

1. A pneumatic data intelligent analysis method based on aerodynamic pattern matching, characterized in that include: Collect historical aerodynamic data of wind tunnel and add aerodynamic characteristics; Obtain aerodynamic laws based on historical aerodynamic data analysis of wind tunnel; Set the research object aircraft, collect wind tunnel test data of the research object aircraft and add aerodynamic characteristics; Analyze all wind tunnel test data of the research object aircraft to obtain aerodynamic laws; According to the current aerodynamic laws, select any wind tunnel test data of the research object aircraft and conduct single-component aerodynamic data analysis; Traverse all components to perform multi-component aerodynamic data analysis of the research object aircraft.

2. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 1, characterized in that: The method of collecting historical aerodynamic data of wind tunnel and adding aerodynamic characteristics is: Taking the wind tunnel test vehicle data as the unit, aerodynamic characteristics are added to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz on each wind tunnel test vehicle data. The aerodynamic characteristics include fitting coefficient characteristics and statistical characteristics, which are constructed and acquired according to the six-component aerodynamic forces.

3. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 2, characterized in that: The construction method of the fitting coefficient characteristics is: For any wind tunnel test vehicle data, the six-component aerodynamic force is fitted using the fourth-order Chebyshev method to obtain the Chebyshev coefficient c of each component. i ; Chebyshev coefficient c i The subscript i represents the order of the Chebyshev coefficients, c0 is the constant term, c1 is the coefficient of the first order term, c2 is the coefficient of the second order term, c3 is the coefficient of the third order term, and c4 is the coefficient of the fourth order term.

4. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 2, characterized in that: The construction method of statistical features is: For any wind tunnel test vehicle data, the mean, variance var, number of peaks num_u, and number of troughs num_l of the six-component aerodynamic forces are calculated respectively.

5. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 2, characterized in that: The method of obtaining aerodynamic laws based on historical aerodynamic data analysis of wind tunnel is: According to the six-component aerodynamic forces X, Y, Z, Mx, My, and Mz, neural networks net-X, net-Y, net-Z, net-Mx, net-My, and net-Mz are constructed respectively; Set the neighborhood function neighborhood_function, neighborhood radius sigma, and learning rate learning_rate parameters; Take any aerodynamic characteristics of any aerodynamic component of the wind tunnel test vehicle data as the input sample X of the neural network. j , traverse each node in the competition layer of the neural network and calculate X j The Euclidean distance between nodes, the node with the smallest distance is selected as the winning node; Taking the winning node as the center, determine the node to be updated according to the neighborhood radius sigma, and determine the update amplitude of the node to be updated according to the neighborhood function neighborhood_function; The calculation of all historical aerodynamic data is completed as one round of iteration, and the number of iterations M is set as the stopping condition for the neural network training; In the trained neural network of any aerodynamic component, the number of nodes in the competition layer is used to represent the aerodynamic law of the corresponding aerodynamic component.

6. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 5, characterized in that: The neural networks are all competitive layer topology structures. The competitive layer topology structure is set to a two-dimensional plane structure rectangular, and the side length dimension is set to And round up, set N to the number of trains in all historical aerodynamic data, and the number of nodes in the competition layer is dimension*dimension; Input sample X j In, j ranges from [1 to N], where N is the number of train trips of all historical pneumatic data; The principle for determining the update amplitude of the nodes to be updated is that the update amplitude of the nodes to be updated that are close to the winning node is greater than that of the nodes to be updated that are far from the winning node.

7. The method for intelligent analysis of pneumatic data based on pneumatic law pattern matching according to claim 6, characterized in that: For any trained aerodynamic component of the neural network, the number of nodes dimension*dimension in the competition layer represents the aerodynamic law, and all N historical aerodynamic data are classified into competition layer nodes according to the aerodynamic law.

8. The method for intelligent analysis of pneumatic data based on pneumatic law pattern matching according to claim 6, characterized in that: The method of analyzing single-component aerodynamic data using the aerodynamic characteristics of the research object aircraft wind tunnel test data is as follows: Any data component W among the six components X, Y, Z, Mx, My, and Mz of the wind tunnel test data of the research object aircraft is used as the research object component; In any wind tunnel test train CC, the Euclidean distance between the aerodynamic characteristics of the CC train W component and the nodes of the neural network competition layer trained for the W component is calculated. The node with the closest distance is the node with successful pattern matching, which is used to represent the highest similarity between the aerodynamic law and the aerodynamic law of the CC train W component. After training with the W component, a neural network is obtained from among the neural networks net-X, net-Y, net-Z, net-Mx, net-My, and net-Mz; Classify the nodes in the competition layer according to the aerodynamic law, calculate the Euclidean distance between all the wind tunnel historical aerodynamic data classified in the pattern matching success node and the aerodynamic characteristics of the W component of CC train, and find the top TOP train data sorted from small to large in Euclidean distance as the train data with the highest similarity to the aerodynamic law of the W component of CC train; According to the top TOP train data, the corresponding aerodynamic shape and all historical aerodynamic data corresponding to the aerodynamic shape are determined, and the aerodynamic characteristics of the research object aircraft are analyzed based on the historical data.

9. The method for intelligent analysis of pneumatic data based on pneumatic law pattern matching according to claim 6, characterized in that: The method for multi-component aerodynamic data analysis for all aerodynamic components of wind tunnel test data is: For a wind tunnel test train CC, the Euclidean distances between the six-component aerodynamic characteristics of CC train and the nodes of the neural network competition layer trained for the corresponding components are calculated respectively. The nodes with the closest distances are the nodes with successful pattern matching. The nodes with successful pattern matching of the six components are used to represent the aerodynamic laws with the highest similarity with the six-component aerodynamic laws corresponding to CC train. Obtain the historical aerodynamic data of the wind tunnel classified into the node with successful pattern matching, recorded as data-X, data-Y, data-Z, data-Mx, data-My, data-Mz; Calculate the intersections Set1, Set2, Set3, Set4, and Set5 of the wind tunnel historical data corresponding to each aerodynamic component; The aerodynamic shapes of the research object aircraft corresponding to each intersection and all historical aerodynamic data corresponding to the aerodynamic shapes are used as the highest-order data of the six-component aerodynamic law similarity to analyze the aerodynamic data of the research object aircraft.

10. The method for intelligent analysis of aerodynamic data based on aerodynamic pattern matching according to claim 6, characterized in that: The intersection calculation method of each aerodynamic component corresponding to the wind tunnel historical data is: Determine the analysis order of aerodynamic components X, Y, Z, Mx, My, Mz; Use the intersection of data-X and data-Y as intersection Set1; Use the intersection of Set1 and data-Z as the intersection Set2; Use the intersection of Set2 and data-Mx as the intersection Set3; Use the intersection of Set3 and data-My as intersection Set4; Use the intersection of Set4 and data-Mz as the intersection Set5; When the intersection of any two data sets is empty: If Set5 is not empty, Set = Set5; Otherwise, if Set4 is not empty, Set = Set4; Otherwise, if Set3 is not empty, Set = Set3; Otherwise, if Set2 is not empty, Set = Set2; Otherwise, if Set1 is not empty, Set = Set1.

Images