Time history deep learning method for flow field measuring point position selection

Abstract

The invention relates to the field of flow field measuring point position selection, in particular to a time history deep learning method for flow field measuring point position selection, which comprises the following steps of: arranging flow field monitoring points according to original data of an experimental object; numerical simulation of the flow field is carried out, and flow field time history data at the measuring points are obtained; constructing an FTH-AE deep learning network model, and training the deep learning network model by using the original data; clustering the feature codes to obtain a measuring point set with similar time-varying features; and visualizing the measuring point samples in each region, analyzing sample characteristics in each region, and determining an optimal measuring point arrangement scheme. According to the method, the limitation of traditional experience dependence is broken through, an area with remarkable flow characteristics is accurately positioned, and scientific arrangement of measuring points is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of flow field measurement point position selection, and particularly relates to a time series deep learning method for flow field measurement point position selection. BACKGROUND

[0002] Flow field experiments are an important research method in multidisciplinary fields, but due to the continuous deformation of fluid under the action of small shear force, it is difficult to observe and extract the characteristics of the flow field physical quantity. At present, the most common measurement method is to observe the flow physical quantity through the measurement points at fixed positions in the flow field. However, in actual experiments, the number of sensors arranged is far less than the high-dimensional characteristics of complex flow. This is mainly due to two reasons: first, complex flow covers multiple scales, and the number of sensors required for full resolution measurement is huge; second, the presence of sensors will interfere with the surrounding flow field, and their size and other factors will also limit the arrangement position of the measurement points. Therefore, how to use limited measurement point sensors to obtain more valuable flow field information has become a key problem in fluid experiments.

[0003] At present, there is no accurate and unified method for selecting the arrangement position of measurement points in flow field experiments. For example, in the classic problems of cylinder flow and square cylinder flow, although flow velocity sensors are often arranged in the wake flow to collect time series signals for calculating the Strouhal number or studying the wake distribution characteristics, there is little research on the arrangement position of measurement points, and the arrangement positions of measurement points in different literatures are quite different. In existing research, the arrangement position of flow field measurement points is mostly determined by experience, and the main difficulties are: on the one hand, the flow characteristics at different positions of the flow field are significantly different, and the amount of data is huge; on the other hand, the flow characteristics of the same measurement point change complexly with time.

[0004] In addition, with the increasing demand for accuracy and efficiency of fluid mechanics research in modern engineering technology, the traditional method of arranging measurement points by experience cannot meet the needs of complex flow field experiments, which may not only lead to incomplete experimental data and biased feature extraction, but also cause waste of sensor resources or prolong the experimental period. Moreover, in some cutting-edge research fields with high experimental accuracy requirements, such as hypersonic aircraft aerodynamic characteristic research and deep sea complex flow field environment simulation, the lack of a scientific method for arranging measurement points has become a technical bottleneck restricting the in-depth development of research, and innovative solutions are urgently needed to meet the increasingly complex needs of flow field experiments. SUMMARY

[0005] The technical problems existing in the prior art are solved according to the technical problems proposed above, and a time-history deep learning method for flow field monitoring point position selection is provided.The present application can accurately obtain time-history data of a flow field by reasonably arranging flow field monitoring points, construct an FTH-AE deep learning network model and train the same, cluster low-dimensional representations of time-varying characteristics at different positions in the flow field, select a plurality of monitoring points in each region for time-history display, observe time-varying characteristics of the samples, and determine an optimal monitoring point arrangement scheme.

[0006] The technical means adopted by the present application are as follows: A time-history deep learning method for flow field monitoring point position selection, comprising: arranging flow field monitoring points for original data of an experimental object; carrying out numerical simulation of a flow field to obtain time-history data of the flow field at monitoring points; constructing an FTH-AE deep learning network model and training the deep learning network model using the original data; clustering feature codes to obtain a set of monitoring points having similar time-varying characteristics; visualizing monitoring point samples in each region, analyzing sample characteristics in each region, and determining an optimal monitoring point arrangement scheme.

[0007] Further, the arrangement of flow field monitoring points specifically comprises: selecting a flow range for arranging flow field monitoring points, arranging a plurality of monitoring points in a flow direction structure-2D~+8D and a transverse direction-3D~+3D range, wherein D is the side length of a square column, the x-axis is the flow direction, the corresponding fluid velocity component is represented by u, the y-axis is the transverse flow direction, and the corresponding fluid velocity component is represented by v; and randomly generating monitoring point coordinates in a determined region.

[0008] Further, obtaining the time-history data of the flow field specifically comprises: calculating the Reynolds number based on the structure side length, classifying and saving flow field velocity components at each time step in the calculation to form a flow direction velocity time-history and a transverse velocity time-history parameter set of each monitoring point; and selecting typical monitoring points on the side of the structure, in the backflow area and in the wake area for observation to obtain the flow direction velocity and the transverse velocity time-history.

[0009] Further, the training of the deep learning network model specifically comprises: dividing the time-history data of the flow field into a training set and a test set; inputting the training set into a time-history input layer of the deep learning network model, training the deep learning network model to obtain a trained deep learning network model; testing the trained deep learning network model through the test set; if the model output result reaches a convergence state, the model is determined as an optimal deep learning network model; if the model does not converge, the training set and the test set are re-divided, and the training is re-performed; The loss function L of the deep learning network model is represented as: L=

[0010] Wherein, n represents the sample number, s represents the time of the time history signal in the sample, represents the number of samples in the sample set, represents the length of the time history signal; is the output data of the deep network model at time s at the n-th sample measuring point, is the real sample time history signal at time s at the n-th sample measuring point.

[0011] The dimensionless relative error formula is represented as:

[0012] Wherein, is the original sample curve; is the reconstructed sample curve; is the incoming flow velocity; the physical coordinates of the sample are combined with the dimensionless relative error, different colors are used to divide the error range, and a relative error scatter plot is drawn.

[0013] Further, the feature encoding is clustered, specifically including: The original flow field time history data is compressed and reconstructed by the self-encoding deep learning network, the high-dimensional flow field data is mapped to the low-dimensional space, and the feature encoding of each measuring point in the low-dimensional space is obtained; the K-Means clustering algorithm is used to analyze the feature encoding in the low-dimensional space, and the number of clustering categories K is pre-set; Randomly select data points from the data set as initial cluster centers, calculate the distance of each feature encoding to the K cluster centers, and assign each feature encoding to the category where the nearest cluster center is located; for each cluster, calculate the mean of all feature encodings in the cluster, and take the mean as the new cluster center; repeat the steps of data point assignment and cluster center update, adjust the attribution category of each feature encoding and the position of the cluster center, until the cluster center no longer changes significantly; the clustering result is displayed according to the spatial position of the flow field measuring point, and the clustering result is distinguished by different colors.

[0014] Further, the optimal measuring point arrangement scheme is determined, specifically including: Draw the flow velocity feature distribution graph and the transverse velocity feature distribution graph of each class of feature classification result respectively, and observe the distribution of measuring points with similar flow time-varying characteristics; randomly select several measuring points in each region for time history display, and observe the time-varying characteristics of the time history samples in different regions, including mean, root mean square and phase; The time-varying feature list of the time schedule in each region is selected, including the average value and standard deviation of the transverse velocity and flow direction velocity of the sample of each type of region; the region with larger standard deviation of the fluid velocity component u and larger standard deviation of the fluid velocity component v is selected as the optimal measurement point arrangement scheme.

[0015] Compared with the prior art, the present application has the following advantages: The time schedule deep learning method for flow field measurement point position selection in the present application breaks through the limitations of traditional experience dependence, accurately locates the flow feature significant region, realizes scientific arrangement of measurement points, and greatly improves the effectiveness and integrity of experimental data compared with the traditional method, which can significantly improve the experimental efficiency.

[0016] The time schedule deep learning method for flow field measurement point position selection in the present application takes low Reynolds number square cylinder flow as the research object to construct the method system, which can be applied to complex flow field experiments in shipbuilding, aviation, machinery and other fields, and provides a unified and reliable solution for measurement point arrangement in different scenarios, which can effectively solve the problem of lack of universality of existing measurement point arrangement. The time schedule deep learning method for flow field measurement point position selection in the present application optimizes the measurement point layout through deep learning, which can reasonably plan the number and position of sensors, avoid resource waste caused by redundant measurement points or improper arrangement, reduce experimental equipment investment and debugging time, and significantly reduce experimental cost while ensuring experimental accuracy. Based on the above reasons, the present application can be widely popularized in the field of flow field measurement point position selection. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 The flow chart of the flow field measurement point position selection learning method in the present application.

[0019] Figure 2 The measurement point position layout schematic diagram in the specific embodiment of the present application.

[0020] Figure 3 The FTH-AE model encoding and decoding schematic diagram in the embodiment of the present application.

[0021] Figure 4 The model training loss value schematic diagram in the embodiment of the present application.

[0022] Figure 5This is a scatter plot of the relative error of the flow velocity in an embodiment of the present invention.

[0023] Figure 6 This is a scatter plot of the relative error of lateral velocity in an embodiment of the present invention.

[0024] Figure 7 This is a schematic diagram illustrating the feature classification principle in an embodiment of the present invention.

[0025] Figure 8 This is the classification result of the flow velocity feature in the embodiment of the present invention.

[0026] Figure 9 This is the classification result of the lateral velocity feature in the embodiment of the present invention.

[0027] Figure 10 The feature selection result for the large velocity fluctuation in the embodiment of the present invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0032] like Figure 1 As shown, this invention provides a time-history deep learning method for selecting flow field measurement point locations, including: Based on the raw data of the experimental object, flow field monitoring points were arranged. The tail flow induced by a low Reynolds number square prism was selected as the flow field to be processed. Specifically, in a preferred embodiment of this invention, the flow range for arranging the monitoring points was selected, with multiple measuring points arranged within a range of -2D to +8D in the flow direction and -3D to +3D in the transverse direction. Here, D is the side length of the square prism; the x-axis represents the downstream direction, with the corresponding fluid velocity component denoted by u; and the y-axis represents the transverse direction, with the corresponding fluid velocity component denoted by v. The coordinates of the measuring points were randomly generated within the defined area. The specific locations of the measuring points are as follows: Figure 2 As shown, a total of 5900 measuring points were randomly arranged.

[0033] Numerical simulation of the flow field is carried out to obtain the flow field time history data at the measuring points. In a preferred embodiment of the present invention, the Reynolds number is calculated based on the side length of the structure, and the flow field velocity components at each time step in the calculation are classified and saved to form a set of flow velocity time history and lateral velocity time history parameters for each measuring point. Typical measuring points are selected on the side of the structure, in the recirculation region, and in the wake region for observation to obtain the flow velocity and lateral velocity time history.

[0034] In the implementation, the Reynolds number calculated using a square column was 100, water was selected as the simulation medium, and the flow velocity was 1 m / s; the steady-state flow time range was selected as 100 s to 200 s, the sampling interval was 0.05 s, and a total of 6 typical measuring points were selected (see...). Figure 2 Observe the red dots; An FTH-AE deep learning network model is constructed, and trained using the original data. A schematic diagram of the FTH-AE model encoding and decoding is shown below. Figure 3 As shown.

[0035] In implementation, the first step is to construct an autoencoder deep learning network, with the following structure: Figure 3 As shown; the model input layer is a one-dimensional flow field component time history signal; convolution operation is performed on the time history signal of the input layer to obtain the model's convolutional layer 1; the output data of convolutional layer 1 is convolved again to obtain the model's convolutional layer 2; the output data of convolutional layer 2 is convolved again to obtain the model's convolutional layer 3; the multi-dimensional feature data output by convolutional layer 3 is "flattened" into a one-dimensional vector for easy processing by the fully connected layer. The flattened one-dimensional vector is subjected to fully connected computation to integrate features and generate a low-dimensional code, resulting in the fully connected layer 1 of the model. The low-dimensional code generated by the fully connected layer 1 is then subjected to fully connected computation again to expand it into a high-dimensional feature vector, resulting in the fully connected layer 2 of the model. The high-dimensional vector generated by the fully connected layer 2 is adjusted to a multi-dimensional structure suitable for deconvolution, resulting in the deformed layer of the model. The output data of the deformed layer is subjected to deconvolution computation, resulting in the deconvolution layer 1 of the model. The output data of the deconvolution layer 1 is then subjected to deconvolution computation again, resulting in the deconvolution layer 2 of the model. The output data of the deconvolution layer 2 is then subjected to deconvolution computation again, resulting in the deconvolution layer 3 of the model. The output data of the deconvolution layer 3 is then subjected to deconvolution computation, resulting in the output layer of the model.

[0036] In a specific implementation, as a preferred embodiment of the present invention, the flow field time history data is divided into a training set and a test set; the training set is input into the time history input layer of the deep learning network model to train the deep learning network model and obtain the trained deep learning network model; the trained deep learning network model is tested using the test set; if the model output results reach a convergence state, the model is considered the optimal deep learning network model; if it does not converge, the training set and test set are re-divided, and the model is retrained.

[0037] The ADAM method is used to accelerate model convergence, and the difference between input and output is measured by the MSE loss function; the loss function L of the deep learning network model is expressed as: L =

[0038] Where n represents the sample number, and s represents the time of the time history signal in the sample. This indicates the number of samples in the sample set. Indicates the length of the time-history signal; This represents the output data of the deep network model at time s at the nth sample measurement point. This represents the actual sample time history signal at time s for the nth sample measurement point. For example... Figure 4 As shown, the loss function is small enough to meet the accuracy requirements, thus completing the training of the deep learning network model. In this embodiment, the final loss values ​​for the flow velocity and the lateral velocity are 7e-6 and 2e-4, respectively.

[0039] The dimensionless relative error formula can be expressed as:

[0040] in, The original sample curve; To reconstruct the sample curve; For the incoming flow velocity, The length of the time history signal is represented; the physical coordinates of the sample are combined with the dimensionless relative error, and different colors are used to divide the error range to draw a relative error scatter plot.

[0041] In practice, scatter plots with relative errors less than 1% are represented in green, those between 1% and 3% in blue, those between 3% and 5% in yellow, those between 5% and 8% in orange, and those greater than 8% in red. A scatter plot of relative error for flow direction velocity is shown below. Figure 5 As shown, the scatter plot of the relative error of the lateral velocity is as follows: Figure 6 As shown.

[0042] Clustering the feature codes yields a set of measurement points with similar time-varying features; the principle of feature classification is as follows: Figure 7 As shown, in a specific implementation, as a preferred embodiment of the present invention, the original flow field time history data is compressed and reconstructed through an autoencoder deep learning network, mapping the high-dimensional flow field data to a low-dimensional space, and obtaining the feature code of each measuring point in the low-dimensional space; the K-Means clustering algorithm is used to perform cluster analysis on the feature code in the low-dimensional space, and the number of cluster categories K is preset. In the implementation, the number of cluster categories is 7.

[0043] Randomly select data points from the dataset as initial cluster centers. Choose 7 data points and calculate the distance from each feature code to the K cluster centers. Assign each feature code to the category of the nearest cluster center. For each cluster, calculate the mean of all feature codes in that cluster and use the mean as the new cluster center. Repeat the steps of assigning data points and updating cluster centers, adjusting the category of each feature code and the location of the cluster centers, until the cluster centers no longer change significantly. Display the clustering results according to the spatial location of the flow field measurement points, and use different colors to distinguish the clustering results. Feature classification results are as follows: Figure 8 , Figure 9 The corresponding part in.

[0044] The measurement point samples in each region are visualized, the sample characteristics in each region are analyzed, and the optimal measurement point layout scheme is determined.

[0045] In specific implementation, as a preferred embodiment of the present invention, determining the optimal measurement point layout scheme specifically includes: Figure 4 The model loss result obtained by the method of the present invention, combined with Figure 7 and Figure 8 It can be observed that the model has small errors and high accuracy.

[0046] Figure 8 and Figure 9 For feature classification results, several measuring points were randomly selected in each region for time history display, and the features of time history samples in different regions were statistically analyzed. Regions with a standard deviation of fluid velocity component u greater than 0.1 and a standard deviation of fluid velocity component v greater than 0.2 were selected as the optimal measuring point layout scheme, i.e., regions with larger fluctuation values ​​in both directional and lateral velocities. The feature selection results with larger velocity fluctuations are listed below. Figure 10 middle.

[0047] The experimental results of the embodiments show that regions with similar signal features of measurement points can be obtained through time-history deep learning, and regions with obvious features can be located. A method for arranging mobile measurement points based on time-history deep learning is proposed, which provides an accurate and universal measurement point arrangement scheme for the selection of measurement point locations in mobile experiments. This method obtains a more refined measurement point arrangement scheme than traditional methods.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 time-history deep learning method for selecting flow field measurement point locations, characterized in that, include: Based on the raw data of the experimental subjects, set up flow field monitoring points; Numerical simulation of the flow field was conducted to obtain the flow field time history data at the measuring points; Construct an FTH-AE deep learning network model and train the deep learning network model using the original data; Clustering of feature codes yields a set of measurement points with similar time-varying features; The measurement point samples in each region are visualized, the sample characteristics in each region are analyzed, and the optimal measurement point layout scheme is determined.

2. The time-history deep learning method for selecting flow field measurement point locations according to claim 1, characterized in that, The arrangement of flow field monitoring points specifically includes: Select the flow range for arranging flow field monitoring points. Arrange multiple measuring points within the range of -2D to +8D in the flow direction and -3D to +3D in the transverse direction. Here, D is the side length of the square column; the x-axis is the downstream direction, and the corresponding fluid velocity component is represented by u; the y-axis is the transverse direction, and the corresponding fluid velocity component is represented by v; and the coordinates of the measuring points are randomly generated within the defined area.

3. The time-history deep learning method for selecting flow field measurement point locations according to claim 1, characterized in that, Obtaining the flow field time history data specifically includes: The Reynolds number is calculated based on the side length of the structure. The velocity components of the flow field at each time step in the calculation are classified and saved to form a set of flow velocity time history and lateral velocity time history parameters for each measuring point. Typical measuring points are selected on the side of the structure, in the recirculation region and in the wake region for observation to obtain the flow velocity and lateral velocity time history.

4. The time-history deep learning method for selecting flow field measurement point locations according to claim 1, characterized in that, The training of the deep learning network model specifically includes: The flow field time history data is divided into a training set and a test set; the training set is input into the time history input layer of the deep learning network model to train the deep learning network model, resulting in a trained deep learning network model; the trained deep learning network model is tested using the test set. If the model output reaches convergence, the model is considered the optimal deep learning network model; if it does not converge, the training set and test set are re-divided, and the model is retrained. The loss function L of the deep learning network model is expressed as: L＝ Where n represents the sample number, and s represents the time of the time history signal in the sample. This indicates the number of samples in the sample set. Indicates the length of the time-history signal; This represents the output data of the deep network model at time s at the nth sample measurement point. The actual sample time history signal at time s for the nth sample measurement point; The dimensionless relative error formula can be expressed as: in, The original sample curve; To reconstruct the sample curve; The incoming flow velocity is used as the reference. The physical coordinates of the sample are combined with the dimensionless relative error, and different colors are used to divide the error range to draw a relative error scatter plot.

5. The time-history deep learning method for selecting flow field measurement point locations according to claim 1, characterized in that, The clustering of feature codes specifically includes: The original flow field time history data is compressed and reconstructed by an autoencoder deep learning network, which maps the high-dimensional flow field data to a low-dimensional space and obtains the feature code of each measurement point in the low-dimensional space. The K-Means clustering algorithm is used to perform cluster analysis on the feature code in the low-dimensional space, and the number of cluster categories K is preset. Randomly select data points from the dataset as initial cluster centers, calculate the distance from each feature code to the K cluster centers, and assign each feature code to the category of the nearest cluster center; for each cluster, calculate the mean of all feature codes in that cluster, and use the mean as the new cluster center; repeat the steps of data point assignment and cluster center update, adjusting the category of each feature code and the location of the cluster center, until the cluster centers no longer change significantly; display the clustering results according to the spatial location of the flow field measurement points, and use different colors to distinguish the clustering results.

6. The time-history deep learning method for selecting flow field measurement point locations according to claim 1, characterized in that, Determining the optimal measurement point layout scheme specifically includes: Draw flow velocity feature distribution maps and lateral velocity feature distribution maps for each type of feature classification result, and observe the distribution of measuring points with similar flow time-varying characteristics; randomly select several measuring points in each region to display time histories, and observe the time-varying characteristics of time history samples in different regions, including mean, root mean square and phase. Select the time-varying characteristics of the time history in each region, including the mean and standard deviation of the lateral velocity and flow velocity of the samples in each region; select the region with a larger standard deviation of fluid velocity component u and fluid velocity component v as the optimal measurement point layout scheme.

Images