Data-driven method for predicting flow field and performance of compressor

Abstract

This invention belongs to the field of aero-engine performance simulation technology, specifically disclosing a data-driven method for predicting compressor flow field and performance. The method includes: acquiring operational data of the target equipment under different operating conditions using a volumetric force model and forming a snapshot matrix; decomposing the snapshot matrix and extracting the spatial mode matrix and modal coefficient matrix of the inlet and outlet flow fields, and constructing a training set using these matrices; inputting the training set into a BP neural network to continuously update the mapping relationship between the inlet and outlet flow field modal coefficient vectors; inputting the sampled data into the BP neural network to perform flow field reconstruction and predict the performance parameters of the target equipment. The method has the following advantages: the volumetric force model reduces data acquisition costs, and the POD method performs data dimensionality reduction and feature extraction, enabling accurate prediction of complex distorted flow fields and performance parameters even under small sample conditions; the data-driven flow field prediction method accurately captures high-dimensional flow field features, overcomes the problem of predicting nonlinear characteristics of the flow field under inlet distortion, improves prediction accuracy, reduces computational costs, and optimizes performance prediction efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engine performance simulation, in particular to a compressor flow field and performance prediction method based on data driving. BACKGROUND

[0002] The overall performance simulation of an aero-engine is an important technology in various stages of the life cycle of an aero-engine, such as development, test, production, delivery and maintenance. Considering the high cost of tests, the development of numerical simulation methods and calculation tools, the research on the overall performance simulation of an engine based on numerical methods is increasingly in-depth, which greatly shortens the development cycle and cost of an aero-engine. In order to solve the demand for selecting simulation dimensions according to the precision and cost of each component in the framework of the overall model, the existing mainstream performance evaluation method of turbomachinery introduces high-dimensional and high-precision calculation results of components, corrects low-dimensional models, and improves the simulation precision and reduces the development cost of the overall machine. Although the existing rapid evaluation method of turbomachinery has high prediction accuracy, there are still significant limitations in dealing with inlet distortion, complex inflow and strong coupling between components. There is a certain loss in the information transmission process between the high-dimensional flow field characteristics and the low-dimensional performance model of the existing method, especially when facing complex inlet conditions, the traditional low-dimensional data transmission cannot accurately express the high-dimensional flow field characteristics, thereby affecting the performance prediction of the overall system. With the gradual deepening of the integrated design of engines and aircraft, the influence of inlet distortion on the compressor is increasingly significant, and the existing prediction method cannot effectively capture the nonlinear characteristics of the flow field and the complex interaction.

[0003] Therefore, a compressor flow field and performance prediction method based on data driving is proposed to solve the above problems. SUMMARY

[0004] The present application aims to provide a compressor flow field and performance prediction method based on data driving, which efficiently predicts high-dimensional flow fields to solve or improve the problems of information transmission loss between components when dealing with complex inlet distortion and high-dimensional flow field characteristics, inability to accurately capture the nonlinear characteristics of the flow field and complex interaction between components, and limitation of the accuracy of system performance prediction.

[0005] Therefore, the first aspect of the present application is to provide a compressor flow field and performance prediction method based on data driving.

[0006] The second aspect of the present application is to provide a system.

[0007] The third aspect of the present application is to provide an electronic device.

[0008] The fourth aspect of the present application is to provide a computer readable storage medium.

[0009] The first aspect of the present application provides a data-driven compressor flow field and performance prediction method, based on a body force model, obtaining operation data of an inlet flow field and an outlet flow field of a target device under different working conditions and corresponding component performance parameters under each physical flow rate; forming a snapshot matrix through the operation data; decomposing the snapshot matrix through a POD method to extract a spatial mode matrix and a mode coefficient matrix of the inlet flow field and the outlet flow field, respectively; constructing a training set through all mode coefficient matrices, the physical flow rate and the component performance parameters; inputting the training set into a BP neural network to continuously update the mapping relationship between the physical flow rate and the mode coefficient vector of the inlet flow field and the mode coefficient vector of the outlet flow field and the component performance parameters until the BP neural network is trained; inputting sampling data of the inlet flow field of the target device into the trained BP neural network to generate prediction data of the outlet flow field of the target device and corresponding component performance parameters; and reconstructing the device flow field according to the prediction data and the spatial mode matrix.

[0010] The second aspect of the present application provides a system, comprising: a data acquisition module for acquiring operation data and forming a snapshot matrix according to the operation data; a snapshot matrix decomposition module for POD method analysis of the snapshot matrix, extracting a spatial mode matrix and a mode coefficient matrix, and providing required data for training set construction through dimension reduction and feature extraction of the snapshot matrix; a neural network training module for inputting a training set into a BP neural network and continuously updating the mapping relationship between the inlet flow field and the outlet flow field in the mode coefficient vector until the neural network is trained; and a flow field reconstruction and performance prediction module for inputting sampling data of the inlet flow field of a target device into the trained BP neural network to predict the flow field and performance parameters of the target device.

[0011] The third aspect of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0012] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the above method.

[0013] The present application has the following beneficial effects compared with the prior art: By means of the data-driven flow field prediction method, the intake distortion and the nonlinear characteristics of the flow field can be efficiently captured, precise performance prediction of turbomachinery under complex inflow conditions can be realized, and the deficiency that the traditional interpolation fitting method cannot effectively deal with high-dimensional complex conditions is overcome; and by means of the neural network to build the compressor proxy model, the calculation time and resource consumption required for high-dimensional simulation of the complex flow field are avoided, and the early design cost of the compressor is reduced.

[0014] By constructing a volume force source library and adopting a volume force model to perform geometric simulation, large-scale training sample libraries can be quickly generated under the premise of significantly reducing the calculation amount, the volume force model simulates the disturbance of the blade row to the flow field through a volume force source term, fine modeling of the blade geometry is not needed, therefore the network scale is greatly reduced, the calculation time of a single example is much lower than that of the Reynolds average method, the database construction cost is effectively reduced while the simulation accuracy of the turbomachinery is maintained; By means of POD feature extraction and dimension reduction based on the flow characteristics of turbomachinery, the high-dimensional flow field is decomposed into a linear combination of main modes, the network parameter scale is significantly reduced, the prediction ability under the condition of small samples is improved, and the prediction error is reduced compared with the direct image prediction method; meanwhile, the model integration idea is introduced in combination with the independence between the prediction targets, a plurality of small-scale single-target networks are used to replace a single large-scale multi-target network, the robustness and accuracy under complex working conditions are improved, and high-precision prediction of the outlet flow field and performance parameters caused by complex distorted inflow under the condition of small samples is realized.

[0015] Additional aspects and advantages of embodiments according to the present application will become apparent from the following description with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 The method step flowchart of the present application is shown in Fig. 1; Figure 2 The prediction model construction flowchart of the present application is shown in Fig. 2; Figure 3 The compressor rotor and meridian flow passage schematic diagram of the present application is shown in Fig. 3; Figure 4 The non-uniform inlet flow field constructed by the partition matrix of the present application is shown in Fig. 4; Figure 5 The volume force model example setting schematic diagram of the present application is shown in Fig. 5; Figure 6 The 98% distortion intensity characteristic comparison diagram of the present application with the hub as the starting point and the radial 30% distortion range is shown in Fig. 6; Figure 7 The neuron schematic diagram of the present application is shown in Fig. 7; Figure 8 This is a schematic diagram of the early stopping method training process of the present invention; Figure 9 This is a schematic diagram of the distributed ensemble neural network prediction model of the present invention; Figure 10 This is the first-order mode prediction result of the present invention; Figure 11 The prediction models of this invention for different datasets of various sizes provide the restoration results for each test sample. Figure 12 This is the result of restoring the number of different modes under a certain working condition according to the present invention; Figure 13 This is a system logic block diagram of the present invention; Figure 14 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] Please see Figures 1-14 The following describes some embodiments of the data-driven compressor flow field and performance prediction method, system, electronic device, and computer-readable storage medium of the present invention.

[0020] Regarding the above-mentioned appendix Figure 2 - Appendix Figure 12 , Figure 2 This is a flow diagram of the prediction model construction, which is mainly divided into three parts: database construction, feature extraction and dimensionality reduction, and prediction model training. Figure 3 It shows the actual compressor rotor and a schematic diagram of the meridional flow channel, demonstrating the geometric features of the test object; Figure 4 The method for traversing the inlet distortion form of a complex incoming flow field under a turbomachinery is demonstrated. Figure 5 This is a schematic diagram of the setup for the volume force model calculation example; Figure 6 The comparison between the volumetric force model and the RANS calculation results confirmed the reliability of the data source; Figure 7 This is a schematic diagram of the smallest unit of the prediction model, the neuron. Figure 8 This is a schematic diagram of the early stopping training process, which can effectively avoid overfitting in neural network training. Figure 9This is a schematic diagram of a distributed ensemble neural network prediction model; Figure 10 These are the coefficient prediction results for the first-order mode. The flow field is reconstructed by weighted summation of the coefficients of several main modes. Figure 11 These are the prediction models' reconstruction results for each test sample based on datasets of different sizes. The more samples, the higher the accuracy. Figure 12 It is the result of restoring different modal numbers under a certain working condition. The more modes selected, the richer the details.

[0021] An embodiment of the first aspect of the present invention proposes a data-driven method for predicting compressor flow field and performance. In some embodiments of the present invention, such as... Figures 1-12 As shown, the compressor flow field and performance prediction method includes the following steps: S100 divides the inlet flow field and / or outlet flow field in space along the axial direction to construct a matrix region, and sets sampling points for collecting operational data according to the matrix region.

[0022] Define the form and intensity of the flow field distortion in the inlet flow field, and set the sample patterns in the training set according to the form and intensity.

[0023] Here, the flow field cross-section of the target device is spatially divided to facilitate subsequent data acquisition and flow field analysis. Spatial division of the flow field is intended to address the inhomogeneity of the flow field under different operating conditions, especially for accurate modeling and prediction of distorted flow fields. The division process helps the system analyze and record in detail the distribution of physical quantities in the flow field at different locations.

[0024] This provides pattern calibration for training set construction, i.e., by setting the distortion type and intensity in the inlet flow field, generating data patterns for neural network training. The setting of distortion type and intensity directly affects subsequent training set organization and the neural network learning process. By controlling the form and intensity of distortion, the neural network can learn the patterns of flow field changes under different operating conditions, thereby improving prediction accuracy.

[0025] As mentioned above, the inlet and outlet flow fields are first divided according to their axial positions. Axial direction refers to the mainstream direction of the airflow, with the upstream being the inlet and the downstream the outlet. For both the inlet and outlet flow fields, a uniform division method is used to spatially divide their space into multiple matrix regions. Specifically, the inlet flow field can be divided into several regions along the axial direction, and each region can be further subdivided according to the circumferential direction. In this way, the flow field can be refined into small regions, providing high-resolution spatial data for subsequent data acquisition and analysis. The matrix region division method can be implemented by defining different division units to effectively capture the spatial variations of the flow field. Within each divided matrix region, a specific sampling density needs to be set according to the required distortion resolution of the model. These sampling points will be used to collect flow field data and serve as the basis for subsequent processing.

[0026] The form and intensity of flow field distortion in the inlet flow field are defined. Flow field distortion refers to the non-uniformity of the flow field caused by factors such as equipment structure and airflow disturbance. Common forms of distortion include radial distortion, circumferential distortion, and combined distortion, which manifest as non-uniform distribution of physical quantities such as pressure and velocity at different locations and directions in the flow field. To effectively capture and analyze these non-uniformities, it is necessary to clearly classify the distortion forms before data acquisition and set different intensity levels for each distortion to quantify its impact on the flow field. The intensity of flow field distortion is usually quantified by the air pressure ratio in the distorted region, specifically by comparing the total air pressure in the distorted region with the standard air pressure in the undistorted region. Based on the magnitude of the air pressure ratio, distortion can be divided into multiple intensity levels, such as mild distortion, moderate distortion, and severe distortion. These intensity levels reflect the degree of influence of different distortions on the flow field, thus affecting subsequent performance predictions. Therefore, setting an appropriate distortion intensity is crucial to ensuring that the training data accurately reflects the changes in the flow field under different operating conditions. Once the form and intensity of the distortion are determined, the next step is to construct the training example patterns based on these settings. The generation of these example patterns involves combining different distortion forms and intensities to create different training data labels. These labels will serve as the basis for the neural network's learning, enabling the network to fit and learn different distortion scenarios.

[0027] The intake distortion of the target equipment is generally defined by four parameters: circumferential distortion range, radial distortion range, distortion intensity, and distortion direction. The distortion direction distinguishes whether radial distortion begins at the hub or the rim. For example: radial distortion 30% of blade height, circumferential distortion range 120°, 95% intensity, rim distortion. Each distortion type, after supplementing with flow parameters, constitutes a complete definition of the corresponding equipment operating condition. Numerical simulation methods are used to set up calculation examples based on the sample definitions to obtain flow field data of the target equipment under specified operating conditions.

[0028] Furthermore, an improved distributed force source volumetric force model is adopted to significantly reduce sample acquisition costs, thereby generating more samples under limited resource conditions, constructing a more complete training database, and further improving the generalization ability and prediction performance of the data-driven prediction network. To simulate the impact of non-uniform distorted incoming flow on the flow field, this paper uses a partitioned matrix to model the inlet flow field. For example... Figure 4 As shown, its radial direction is evenly divided into ten regions, and its circumferential direction is evenly divided into 36 regions at 10° intervals, thus dividing the inlet circular surface into... The import matrix region.

[0029] The entire dataset is composed of three distortion types: radial distortion, circumferential distortion, and combined distortion, encompassing four distortion intensities. For simplicity, the distortion intensity D in this paper is defined as:

[0030] Each distortion range includes four cases with distortion intensities of 94%, 96%, 97%, and 98%. Five different flow rate conditions are calculated for each distortion intake condition. Statistically, there are 380 cases for radial distortion, 360 cases for circumferential distortion, 360 cases for combined distortion, and 5 cases for uniform intake, totaling 1105 cases in the dataset. The calculation model and inlet boundary conditions are set as follows. Figure 5 As shown. In any of the above embodiments, the simulation pattern is generated by combining categories based on the form and intensity of the flow field distortion.

[0031] The forms include radial distortion, circumferential distortion, and combined distortion occurring in the matrix region.

[0032] The intensity is determined based on the ratio between the total air pressure in the distorted region and the atmospheric pressure in the undistorted region of the matrix area.

[0033] In this embodiment, the simulation examples are generated by combining categories based on the form and intensity of flow field distortion. This aims to classify and label flow field samples under different distortion conditions in a unified and standardized manner, enabling simulation examples under the same distortion condition to be organized and used as a group of samples with similar characteristics within the training set. Specifically, each simulation example is a category label obtained by combining two dimensions: distortion form and distortion intensity. Different distortion forms correspond to different flow field distortion distribution patterns, and different distortion intensities correspond to varying degrees of distortion. By discretizing and combining these two dimensions, a limited set of simulation example categories covering typical operating conditions can be constructed. When constructing the training set and calibrating flow field features, this simulation example is directly used as the category identifier for the samples, thereby achieving structured management of the training set.

[0034] Regarding distortion forms, these include radial distortion, circumferential distortion, and combined distortion occurring within the matrix region. All are defined based on the aforementioned matrix regions obtained by dividing the inlet flow field radially and circumferentially. Radial distortion refers to a significant distortion distribution in the radial direction of the matrix region; specifically, several radial bands are designated as distorted regions, while the remaining radial bands are undistorted regions. Circumferential distortion refers to distorted partitions in the circumferential direction of the matrix region; specifically, several circumferential sectors are distorted regions, while the remaining sectors are undistorted regions. Combined distortion refers to the simultaneous existence of radial and circumferential distortion distributions within the same matrix region, meaning that some small areas exhibit distortion in both radial and circumferential dimensions. By arranging the spatial positions of distorted and undistorted regions within the matrix region, different forms of distortion field distribution can be numerically constructed, and the corresponding distortion form can then be labeled for each calculation example.

[0035] Regarding distortion intensity, the intensity is defined based on the ratio between the total air pressure in the distorted area and the atmospheric pressure in the undistorted area within the matrix region. That is, the degree of distortion is quantified by statistically analyzing the total air pressure in both distorted and undistorted areas and using the pressure difference or ratio between them. In practice, the total air pressure of the small area defined as the distorted area within the matrix region is first averaged to obtain the average total air pressure of the distorted area; then, the total air pressure of the small area within the undistorted area is averaged, using atmospheric pressure or its equivalent reference pressure as the reference pressure for the undistorted area. The distortion intensity is then defined by the ratio or difference between the two, and based on a pre-set intensity grading threshold, continuous intensity values ​​are divided into several discrete intensity levels, specifically corresponding to different pressure difference percentage intervals. Therefore, after determining the distortion form of each flow field case, it is further classified into a certain intensity level according to the intensity range to which the ratio of the total air pressure in the distorted region to the undistorted region belongs. Thus, the case pattern corresponding to the case is finally generated in the form of distortion form × intensity level, which is used to classify and calibrate the flow field characteristics under different distortion conditions in the training set.

[0036] S101, based on the volume force model, obtains the operational data of the inlet and outlet flow fields of the target equipment under different working conditions, as well as the corresponding component performance parameters under each physical flow rate; and forms a snapshot matrix through the operational data.

[0037] Here, operational data for the inlet and outlet flow fields need to be collected under different operating conditions. The characteristics of the inlet and outlet flow fields of target equipment, such as turbomachinery, compressors, and fans, will change depending on the operating conditions. Therefore, the operating status of the equipment under different conditions must be considered when collecting data. The flow field characteristics under each operating condition may include physical quantities such as flow rate, temperature, and pressure; these data form the basis for describing flow field behavior. Data acquisition for the inlet flow field is usually carried out at the equipment's inlet, with collection points located in different areas of the equipment, such as the center, edge, and axial direction, to ensure data comprehensiveness. Data acquisition equipment can be pressure sensors, flow velocity sensors, or data extracted from numerical simulation examples. These sensors record various physical quantities of the inlet flow field in real time. Data acquisition for the outlet flow field is mainly carried out at the equipment's outlet, recording the final state of the airflow after the equipment has operated. Similar to the inlet flow field, the data collection points for the outlet flow field should also cover different areas of the equipment outlet to ensure that the flow field characteristics are comprehensively and accurately reflected.

[0038] The volumetric force model uses volumetric force source terms to replace the work done by the actual blades on the airflow, performing steady-state numerical solutions for the target equipment. After simulation convergence, flow field physical quantities such as total pressure are extracted from the full-field calculation results at the inlet and outlet sections according to pre-arranged sampling points. Based on the same calculation example, the component performance parameters under that operating condition, including flow rate, total pressure ratio, and efficiency, are directly calculated. This allows for the simultaneous acquisition of inlet and outlet flow field operational data and corresponding component performance parameters under multiple operating conditions without relying on experimental testing and complex full 3D geometric modeling.

[0039] By collecting data from the inlet and outlet flow fields, a series of flow field data snapshots can be obtained. Each snapshot represents the state of the flow field at a certain moment or under a certain operating condition. To perform eigenorthogonal decomposition, these snapshots need to be organized into a matrix form, which is the snapshot matrix. Definition and organization of snapshots: Each snapshot contains the physical quantity data of all sampling points of the flow field at a certain moment. These sampling points are typically flow field data measured by sensors placed at different locations on the equipment. The data from each sampling point is converted into column vectors in a certain order, and each row represents the flow field state under different operating conditions. Specifically, assuming multiple sensors are set up in the inlet and outlet flow fields of the equipment, data such as pressure and flow velocity can be obtained through these sensors. This data will be recorded as column vectors, each column vector corresponding to the flow field values ​​of all sampling points under a certain operating condition. The data for all operating conditions are organized in the matrix according to the spatial order of the sampling points, ultimately forming a complete snapshot matrix. Each column of this matrix represents the flow field data under a specific operating condition, while each row represents the flow field value of a sampling point under different operating conditions. In other words, the number of rows in the matrix corresponds to the number of sampling points, and the number of columns corresponds to the number of operating conditions. The snapshot matrix provides the necessary input data for subsequent POD analysis. By decomposing the above data, the main features of the flow field can be extracted. It can not only describe the state of the flow field at a certain moment, but also reveal the evolution law of the flow field by integrating data from multiple operating conditions.

[0040] As described above, operational data under different working conditions were collected from the inlet and outlet flow fields of the target equipment, and this data was organized into a structured snapshot matrix. This snapshot matrix provides a precise data foundation for subsequent flow field analysis, feature extraction, and neural network training, ensuring that flow field characteristics can be accurately captured and predicted.

[0041] The volumetric force model acquires operational data of the inlet and outlet flow fields of the target equipment under different operating conditions, specifically including: The main idea of ​​the volume force model is to assume that the channel blades are infinitely numerous, infinitely thin, and axisymmetrically distributed. The work done by the actual blades on the airflow is replaced by the circumferentially distributed volume force. By adding the volume force as a source term to the governing equations for solving, the influence of the actual blades on the direction deflection, entropy increase, and pressure ratio of the airflow is simulated.

[0042] The core idea of ​​the distributed force source volumetric force model is to divide the known flow field (solved using a RANS solver) into control volumes and extract the forces acting on the fluid within each control volume as input to the volumetric force vector source field. When the control volumes of the blade channel domain are sufficiently finely divided, the model will have high spatial resolution in describing the distribution of blade loads. Based on this, by repeatedly extracting the volumetric force vector field at multiple operating points, a volumetric force matrix containing information on the changes in blade force distribution during compressor blade throttling can be formed. Then, in solving the model's governing equations, this volumetric force matrix is ​​indexed by the spatial location of each infinitesimal control volume and its local aerodynamic parameters through an association function, gradually converging during iterative solving until a stable flow field and volumetric force field are obtained.

[0043] Specifically, the steps of collecting operational data on the inlet and outlet flow fields of the target equipment under different operating conditions and forming a snapshot matrix include: Collect operational data on the inlet and outlet flow fields during operation, including flow field conditions, and generate a total pressure cloud map from the operational data; The total pressure cloud map of each working condition is converted into a column vector according to the sampling points in a predetermined order, and a snapshot matrix is ​​constructed based on the column vectors of all working conditions.

[0044] Regarding the specific description above, operational data containing the flow field state at the inlet and outlet flow fields are collected during operation, and a total pressure contour map is formed from the operational data. In specific implementation, the operating condition cannot be fully defined solely by the total pressure data; it needs to be determined jointly by the total pressure contour map and the corresponding physical flow rate. Therefore, when constructing training data using numerical simulation methods, this invention directly extracts flow field data at the inlet and outlet sections according to pre-set sampling point locations based on the full-field results obtained from numerical simulation calculations. The extracted data includes at least the total pressure distribution and simultaneously acquires the physical flow rate parameters used to define the operating condition. For each preset operating condition, after the numerical simulation converges to a steady state, the total pressure distribution at the inlet and outlet sections is read from the continuous flow field of the simulation results, and the above data is imported into the post-processing environment. Since numerical simulation can directly output a continuously visible total pressure distribution field, the corresponding total pressure contour map can be directly generated at the inlet and outlet sections, allowing the spatial distribution of the flow field to be obtained in a continuous field form without relying on experimental sampling, thereby providing complete operating condition input data for subsequent snapshot matrix construction. After obtaining the total pressure contour map, it is transformed into a column vector according to the sampling points in a predetermined order. This transformation converts a two-dimensional or multi-dimensional spatial distribution into a flattened data structure. In practice, the sampling points on the inlet and outlet flow field sections are pre-numbered and a fixed traversal order is determined. Specifically, the total pressure value of each sampling point on the contour map is read sequentially according to a nested loop order in the radial and circumferential directions, or according to the row and column index order of the matrix region. The data is then arranged strictly according to this predetermined order, and the total pressure data of each sampling point is sequentially filled into a one-dimensional array. This one-dimensional array constitutes the column vector representation of the total pressure contour map for the corresponding section under the current operating condition. This method ensures that elements at the same position in the column vectors generated under different operating conditions always correspond to sampling points at the same physical spatial location. This maintains a one-to-one correspondence between spatial location and data index in subsequent matrix operations, avoiding calculation errors caused by position mismatches.

[0045] As described above, after converting the total pressure contour map to a column vector under a single operating condition, a snapshot matrix is ​​constructed based on the column vectors of all operating conditions. Specifically, for each operating condition, the column vectors generated in a predetermined order are considered as a snapshot of the flow field corresponding to that operating condition; subsequently, the column vectors corresponding to different operating conditions are arranged sequentially as columns of the matrix according to the operating condition order. Thus, each column of the snapshot matrix represents all sampling point data under a specific operating condition, while each row of the snapshot matrix represents the flow field value of a certain sampling point under all operating conditions, realizing a snapshot matrix organization method of one column for each operating condition and one row for each sampling point. The snapshot matrix constructed in the above manner can organize the total pressure distribution of the inlet flow field and / or outlet flow field under multiple operating conditions with a unified data structure, providing a structured input basis for subsequent dimensionality reduction processing and prediction model training, and significantly improving the efficiency and standardization of batch processing and feature extraction of multi-operating condition flow field data. Assuming the cloud map contains m sampling points, and sampling is performed once under each of n operating conditions, the parameters of all sampling points are flattened into column vectors in a certain order, and different snapshots are arranged in a row, resulting in the flattened flow field snapshot matrix P, as shown in the following formula:

[0046] Where r represents location information, including the location of the sampling point in the cloud. Figure 2 The coordinates in 3D space. t represents the operating condition information of the flow field snapshot, which represents different operating conditions corresponding to the flow field.

[0047] In implementing this invention, it is crucial to focus on the comparative relationships between different operating conditions and the total pressure values ​​at the same sampling point under various operating conditions. Therefore, for flow field snapshots under multiple operating conditions, the total pressure column vectors corresponding to each operating condition are typically arranged column-wise to form a snapshot matrix. Each column of the matrix represents a flow field snapshot under one operating condition, while each row of the matrix represents the total pressure value at a fixed sampling point under different operating conditions. The snapshot matrix constructed in this way can organize the total pressure distribution of the inlet flow field and / or outlet flow field under multiple operating conditions with a unified data structure, providing a structured input basis for subsequent dimensionality reduction processing and prediction model training, and significantly improving the efficiency and standardization of batch processing and feature extraction of multi-condition flow field data.

[0048] S102, the snapshot matrix is ​​decomposed using the POD method to extract the spatial mode matrix and mode coefficient matrix of the inlet flow field and the outlet flow field respectively; a training set is constructed using all mode coefficient matrices, physical flow rates and component performance parameters.

[0049] Here, the acquired snapshot matrix is ​​decomposed to extract the spatial mode matrix and modal coefficient matrix of the inlet and / or outlet flow fields. Based on the application of the POD (Orthogonal Eigenvalue Decomposition) method, the aim is to extract the most representative flow field features from the original flow field data, facilitating subsequent training set construction and neural network learning. Dimensionality reduction and feature extraction reduce computational complexity while preserving the main features of the flow field, ensuring the efficiency and accuracy of subsequent predictions.

[0050] POD analysis is performed on the snapshot matrix, which involves feature extraction for each data point in the snapshot matrix. The snapshot matrix itself consists of multi-dimensional data, with each column representing the physical quantities of the flow field under different operating conditions. The goal of POD analysis is to extract several main modes from the large amount of raw data through orthogonal decomposition. These modes represent the most significant patterns of change in the flow field. Specifically, POD decomposes the raw data into multiple orthogonal basis vectors and corresponding weighting coefficients.

[0051] The spatial modal matrix obtained through POD decomposition contains the main characteristic information of the flow field. Each column represents a spatial mode, used to describe the representative spatial distribution characteristics presented in the snapshot data under different operating conditions. Specifically, one spatial mode may correspond to a typical pressure distribution pattern in the flow field, while another mode may reflect a certain type of local flow structure. Each column of the spatial modal matrix is ​​a static feature vector extracted from snapshot data of multiple operating conditions, and does not involve the flow field changing with operating conditions, but is used to represent the main characteristic structure of the flow field in space. Thus, the spatial modal matrix provides a basic feature expression for subsequent flow field reconstruction and performance prediction based on modal coefficients.

[0052] The modal coefficient matrix represents the intensity and contribution of each spatial mode under different operating conditions. The modal coefficients are obtained by calculating the inner product between the spatial modes and the mean-reduced snapshot data, reflecting the weight of each mode under a specific operating condition. By performing the inner product operation on the flow field data for each operating condition, a matrix is ​​obtained, recording the intensity of different modes under each condition. Each row of the modal coefficient matrix represents the modal coefficient vector under a specific operating condition, indicating the intensity or contribution of different spatial modes under that condition. These coefficients not only describe the main characteristics of the flow field under different operating conditions but also provide input data for subsequent neural network training.

[0053] By constructing the spatial mode matrix and mode coefficient matrix, a training set containing the main flow characteristics within the working range is obtained. This data will be input into a neural network for learning and optimization, thereby constructing a model capable of accurately predicting flow field characteristics and aerodynamic mechanical properties. This entire process not only improves data processing efficiency but also ensures that the network can accurately learn the changing patterns of the flow field under different operating conditions, providing a feasible foundation for subsequent efficient performance prediction and flow field reconstruction.

[0054] Furthermore, a training set is constructed by using all the modal coefficient vectors that constitute the modal coefficient matrices of the inlet and outlet flow fields, along with the corresponding component performance parameters. This means that each row of the modal coefficient matrix of the inlet flow field contains modal coefficient vectors of various orders, and the corresponding modal coefficient vectors and calculated component performance parameters in the modal coefficient matrix of the outlet flow field under the same operating condition are matched one-to-one and stored as input features and output targets in the training set. This forms training data consisting of inlet field modal coefficient vectors, outlet field modal coefficient vectors, and performance parameter sample pairs, which is used for subsequent training and calibration of the data-driven turbomachinery performance and flow field prediction model.

[0055] As described above, by performing POD analysis on the snapshot matrix, the spatial mode matrix and mode coefficient matrix are extracted, effectively reducing dimensionality and extracting features, thus constructing a training set containing the main changing features of the flow field. These matrices provide accurate input data for subsequent neural network training, ensuring the efficiency and accuracy of flow field prediction.

[0056] Furthermore, to reduce the dimensionality of the training data and improve the prediction accuracy of the network, this paper uses the POD method to extract features and reduce the dimensionality of the dataset, thereby obtaining a low-dimensional model of the research object to construct the prediction model dataset. The POD method performs data dimensionality reduction and feature extraction on the snapshot matrix, obtains a set of optimal orthogonal bases, selects the first few principal modes, and obtains the main features and simplified description of the original flow field. A snapshot refers to a direct observation or record of the flow field state at a certain moment, usually obtained through experimental or numerical simulation methods. In this paper, a snapshot refers to the total pressure cloud map of the Rotor67 fan rotor outlet field at a specified axial position under specified operating conditions.

[0057] Furthermore, POD separates the temporal and spatial variables in the data using the concept of variable separation, obtaining a set of orthogonal spatial modes φ (basis vectors) and corresponding weights a, so that the flow field at any sampling location can be represented by these modes. Any working condition The parameter values ​​below All can be represented as a weighted sum of the POD bases. To capture the most salient features of the decomposition object, the pulsation parameters of the original matrix are generally... The decomposition yields the result of the flow field snapshot matrix P after removing the averages over the working condition dimension t. The demeaning process and decomposition form are shown in the following equation:

[0058]

[0059] To achieve data dimensionality reduction, the goal is to reconstruct the flow field as accurately as possible using as few spatial modes as possible. This requires selecting spatial modes that encompass the most flow field features. To address this need, the POD algorithm constructs a correlation matrix R, transforming the basis vector orientation optimization problem into an error minimization problem. It calculates the eigenvectors and eigenvalues ​​of the correlation matrix as basis vectors and weights, respectively, and arranges the basis vectors in descending order of weight. Through this method, the basis vector sequence is selected to maximize the variance of the dataset, meaning that basis vectors with larger weights capture the most significant structures and trends in the data.

[0060]

[0061] After the solution is obtained, the first N principal modes are selected to reconstruct the complex original flow field matrix within a certain accuracy loss range, as shown in Equation 5:

[0062] in, This represents the value of the pulsation matrix.

[0063] At this point, the original data has been decomposed into a linear combination of several spatial modes. According to the mathematical principles of POD (Programmable Optimization), lower-order modes contain the main features of the decomposed object. Because the decomposed object shares the same characteristics as the research object, this linear combination of spatial modes can reconstruct the main flow characteristics of the research object under any operating condition within the research scope. Therefore, by simply using a BP neural network to establish the mapping relationship between the input parameters and the weights of each mode, the predicted weights can be used. Spatial modes obtained from POD decomposition The target flow field is then reconstructed.

[0064] Specifically, the steps of decomposing the snapshot matrix using the POD method to extract the spatial mode matrices and modal coefficient matrices of the inlet and outlet flow fields respectively include: The snapshot matrix is ​​subjected to mean-reduction processing, which involves averaging the total pressure data at the same sampling point under all operating conditions and subtracting this average value from the corresponding snapshot data to obtain a mean-reduced snapshot matrix containing only the pulsating component. This mean field can be regarded as the 0th-order mode, used to describe the overall static distribution characteristics of the flow field. Although it does not participate in the POD decomposition process, it will be added back as a basis component in the subsequent flow field reconstruction.

[0065] Using the mean-reduced snapshot matrix as the decomposition object, POD dimensionality reduction and feature extraction are performed. By solving the corresponding eigenvalue problems, spatial modes and modal coefficients that can characterize the main features of the flow field are directly obtained. Mathematically, POD decomposition naturally generates a set of mutually orthogonal spatial mode vectors and simultaneously obtains the modal coefficient array corresponding to each operating condition. Arranging the obtained spatial mode vectors in columns forms a spatial mode matrix, and arranging the modal coefficients of each operating condition in rows forms a modal coefficient matrix, which is used to completely represent the low-dimensional characterization of the flow field under each operating condition.

[0066] Based on the specific description above, dimensionality reduction and feature extraction are performed on the snapshot matrix to obtain orthogonal basis vectors that can characterize the spatial modes. These orthogonal basis vectors are then arranged column-wise to generate a spatial mode matrix. In practice, the snapshot matrix obtained from the multi-condition total pressure cloud map is considered as a dataset containing high-dimensional flow field information. Dimensionality reduction algorithms such as intrinsic orthogonal decomposition are used to perform feature analysis on this snapshot matrix. The goal of this feature analysis is to automatically find a set of mutually orthogonal spatial basis vectors in the flow field data of all conditions that retain as much of the original data variance information as possible. These basis vectors are the spatial modes. Each spatial mode can be understood as a typical flow field distribution pattern, representing a major spatial variation of the flow field. The orthogonality between different spatial modes means that there is no mathematical redundancy, which is beneficial for subsequent linear combination reconstruction and numerical stability. By performing dimensionality reduction and feature extraction on the snapshot matrix, a subset of orthogonal basis vectors with higher energy contributions are selected, and these orthogonal basis vectors are arranged column by column in a predetermined order to form a spatial mode matrix. Each column of this matrix corresponds to a spatial mode, and each row corresponds to the response of a certain sampling point under that mode. Thus, a low-dimensional but physically meaningful matrix replaces the original high-dimensional snapshot data.

[0067] S103, the training set is input into the BP neural network to continuously update the mapping relationship between the physical flow rate and the modal coefficient vector of the inlet flow field and the modal coefficient vector of the outlet flow field and the component performance parameters, until the BP neural network completes training.

[0068] Here, the training set constructed in the previous steps is used to input the input data and corresponding output data into the BP neural network in pairs. Through forward propagation, error calculation and back propagation, the weight parameters and bias parameters inside the network are continuously updated, so that the predicted mode coefficients output by the network continuously approach the true mode coefficients until the preset convergence condition or training round is reached, thereby completing the training of the BP neural network.

[0069] In practice, the BP neural network is first structured and its parameters initialized based on the training set obtained in step S102. The BP neural network includes an input layer, one or more hidden layers, and an output layer. The neurons in the input layer receive the input features of each sample in the training set. These input features can be constructed by concatenating the inlet flow field modal coefficients and their corresponding operating parameters in a predetermined order. The neurons in the output layer output the modal coefficient vector corresponding to the sample. The modal coefficients originate from the row vectors matching the operating condition in the modal coefficient matrix constructed in the previous step. After initialization, each sample in the training set is input into the neural network in batches, and a forward propagation calculation is performed: each layer of neurons weights and sums the input signals according to the current weight and bias parameters, and obtains the output value through an activation function. Finally, the output layer provides the prediction result of the modal coefficients for that sample.

[0070] Subsequently, the predicted modal coefficients output by the neural network are compared with the true modal coefficients recorded in the training set to calculate the error, thus obtaining the loss value for the current sample. The loss function can take the form of mean squared error to measure the deviation between the predicted and true modal coefficients. After the error calculation is completed, the error signal is propagated backward from the output layer along the network layers through the backpropagation mechanism of the BP neural network. The gradient information of each parameter is calculated based on the sensitivity of the error to the weight and bias parameters of each layer. On this basis, the weights and biases in the network are updated using gradient descent or its improved algorithm. Through this cyclical process of forward prediction, error calculation, backpropagation, and parameter update, the internal parameters of the network are continuously adjusted on all samples in the training set. This ensures that, given a spatial modal matrix, the correspondence between the input features learned by the network and the outlet flow field modal coefficients gradually converges to the true physical mapping, which is equivalent to continuously optimizing the mapping relationship between the inlet and outlet flow field modal coefficient matrices.

[0071] Specifically, the modal coefficient vector of the inlet flow field and the physical flow rate under the corresponding operating conditions are used as joint input features, and the modal coefficient vector of the outlet flow field is used as data label. The mapping relationship between the modal features of the inlet flow field and the physical flow rate to the modal coefficient vector of the outlet flow field is trained using a BP neural network.

[0072] Meanwhile, the aforementioned joint input features and the corresponding component performance parameters under the working conditions are used as data labels to train the same BP neural network or a network structure that works in conjunction with it, so that the network can learn the coupling relationship between the inlet flow field modal features, physical flow rate and outlet flow field and component performance parameters in a unified data-driven framework.

[0073] Furthermore, this paper uses a backpropagation (BP) neural network to establish the mapping relationship between input and output parameters. The BP neural network gets its name from its error backpropagation (ERP) training algorithm. Its basic idea is to use gradient descent to minimize the error between the network output and the actual label, thereby adjusting the parameters in the network. Specifically, the algorithm first inputs information from the input layer into the network for forward propagation. During this process, each neuron receives the numerical values ​​output by neurons in the previous layer. Multiply that value by the corresponding weight. In addition to the bias variable Then input activation function The final output layer result is obtained from the process. , as in equation (7). Figure 7 As shown.

[0074]

[0075] After the forward propagation process is completed, the output value is calculated. and the true value Error between It is then backpropagated to each neuron to adjust their respective weights and biases, and the step size is adjusted accordingly. This is also known as the learning rate. This process is called the error-backpropagation training algorithm. This completes one round of training for the neural network. Through repeated training iterations, the results become increasingly accurate.

[0076] The formulas for calculating the error and the simplified formula for weighted reverse update are as follows:

[0077]

[0078]

[0079] To avoid overfitting in the network, this paper uses the "early stopping" method for training, such as... Figure 8 As shown in the diagram, this training method creates a separate validation set within the dataset. If the training set error continues to decrease while the validation set error continues to increase, training is stopped early, and the network data is saved. This allows the network to learn the features of the training data while ensuring its generalization ability.

[0080] To ensure the universality of the constructed flow field prediction method, the input parameters are obtained by directly sampling the scatter plot of the inlet flow field. To guarantee the closure and uniqueness of the solution, an additional operating condition definition parameter is required. In this paper, the operating condition definition parameter used in the prediction model is the rotor physical flow rate.

[0081] Furthermore, after decomposing the data using POD, the flow field features are stored in each spatial mode. To achieve high reconstruction accuracy, multiple modes need to be used for prediction. Simultaneously predicting the operating condition coefficients corresponding to multiple modes means that the neural network needs to perform multi-task regression learning. Since the functional relationship between the operating condition coefficients and the operating conditions differs for different spatial modes, a large number of neurons in the neural network are needed to store the function information for each. If there are many prediction targets, a deep neural network and a large number of neurons are required to meet the fitting needs of each function. This results in an extremely large sample requirement for directly using the neural network for fitting.

[0082] From the mathematical principles and expressions of POD, it can be seen that the operating condition coefficients for different orders of spatial modes are... With working conditions The relationships between the parameters are independent. Considering this characteristic of the coefficient arrays for different modal conditions, this paper further decomposes the fitting task of the neural network. This scheme borrows the idea of ​​model ensemble, training a small-scale neural network for each spatial mode to perform single-task learning and predict the corresponding coefficients for each mode independently. Finally, the output parameters are ensembled to obtain the coefficient array of the spatial mode matrix under the corresponding conditions. Multiplying each order of the modal vector with the corresponding element in the array and linearly superimposing them allows for the reconstruction of the flow field. This strategy not only better adapts to the unique functional characteristics of the coefficients for each modal condition but also improves the prediction accuracy and model robustness.

[0083] The plan is as follows Figure 9 As shown, under this architecture, each mode will independently use a subnetwork to predict the response coefficient of that mode under different operating conditions. The next step will be to determine the hyperparameters of each subnetwork.

[0084] Furthermore, within each sub-network, due to the thorough analysis and dimensionality reduction of the prediction target and training data, the network size is small, and the number of learnable parameters is relatively small. The output parameter structure and the sampling method for the input parameters have been preliminarily determined. To determine the specific network structure, activation function, loss function, and other hyperparameters, it is also necessary to confirm the input parameter dimension, i.e., the number of sampling points, in order to determine the number of neurons in the input layer.

[0085] Regarding the selection of the number of sampling points, this method uses the accuracy of the training results of the coefficient array of the first-order mode as a reference, employs mean squared error (MSE) as the loss function, and sigmoid as the activation function. The network has three hidden layers, each with 20, 10, and 5 neurons. The number of neurons in the input layer is the number of sampling points plus one parameter (physical flow rate), and the output layer has only one neuron, which is the response coefficient of the corresponding mode. The training results of the first-order mode are as follows: Figure 10 As shown.

[0086] S104: Input the sampled data of the inlet flow field of the target device into the trained BP neural network to generate the predicted data of the outlet flow field of the target device and the corresponding component performance parameters; perform device flow field reconstruction based on the predicted data and spatial mode matrix.

[0087] Here, the sampling data of the inlet flow field of the target device is input into the trained BP neural network. The fundamental purpose is to directly predict the flow field distribution and overall performance of the target device under the current operating condition using existing training results, without having to perform costly numerical simulations or experiments. Specifically, when the target device is in the operating condition to be evaluated, the sampling data of the inlet flow field of the target device is first obtained under this condition according to the aforementioned matrix region division and sampling method. The sampling data is consistent with the input data used to construct the training set in terms of physical quantity type, sampling location arrangement, and data organization. It is also processed according to the same preprocessing method used in the training stage to ensure that the input features match the parameter system of the trained BP neural network, so that the network can make a stable and reliable response to the inlet flow field data of the new operating condition.

[0088] After completing the above sampling and preprocessing, the modal coefficient matrix from the sampled data of the target equipment's inlet flow field is fed as input features into the already trained BP neural network model. Since the BP neural network has already learned the mapping relationship between the inlet flow field features, the outlet flow field modal coefficients, and the component performance parameters during the training phase, there is no need to update the weights in this step. Instead, the above input features are directly input into the trained BP neural network, and the predicted data corresponding to the inlet flow field can be output through a single forward propagation calculation. This predicted data includes the predicted modal coefficient vector of the target equipment's outlet flow field, as well as the predicted results of the component performance parameters corresponding to the modal coefficient vectors of the inlet and outlet flow fields under this operating condition, including total pressure ratio, efficiency, etc., realizing a rapid mapping from inlet flow field sampling to outlet flow field features and performance indicators.

[0089] After obtaining the predicted modal coefficients of the target equipment under its current operating condition, the flow field of the target equipment is reconstructed by combining the spatial modal matrix extracted and fixed in the previous steps. Specifically, the spatial modes are weighted and superimposed according to their corresponding predicted modal coefficients, and then superimposed with the reference flow field obtained during the aforementioned mean-reduction processing. This allows the predicted flow field distribution of the target equipment under that operating condition to be recovered at the selected spatial location. This reconstructed flow field is physically equivalent to the numerical simulation results of the outlet flow field or the section of interest of the target equipment. However, the reconstruction process relies only on the trained model and the calculation of finite-dimensional modal coefficients, resulting in a computational cost far lower than that of traditional three-dimensional numerical simulation. Through this flow field reconstruction based on modal superposition, this invention achieves low-dimensional representation and rapid reconstruction of complex three-dimensional flow fields.

[0090] Specifically, in the model's prediction process, a new unknown snapshot (the field under a certain operating condition) is denoted as... The modal coefficient vector under the known modes needs to be obtained through projection operations. .

[0091]

[0092] Among them, here The mean should be removed first, consistent with the base used in POD.

[0093] This projection process is essential for ensuring the consistency of the feature space; only through... right Orthogonal projection is required to transform the new data into a format similar to the input during the training phase (the original coefficient matrix). Isomorphic low-dimensional representations ensure compatibility with already trained neural networks. and (Dimensional matching and consistent physical meaning of features). If the projection is skipped and the original sampled data is used directly, the prediction will fail due to feature space mismatch.

[0094] Specifically, the process of obtaining the coefficient vector of the test object in the known spatial modes involves discretizing the space into... Construct a matrix using sampling points: for The modal matrix, the first Listed in Mode at each sampling point . k The selected number of modes; for A vector, the elements of which are the values ​​of the new snapshot at each sampling point; The spatial mean vector used during training ( ),Right now .

[0095] To remove the mean:

[0096] If the modes are orthogonally normalized under the discrete inner product (i.e.) (This is guaranteed during the decomposition process), coefficients:

[0097]

[0098] Furthermore, after the prediction model has been trained, the network can now output a set of coefficients for each mode for any given inlet conditions of the predicted object. Based on the POD decomposition equation, the reconstruction equation can be obtained, and the spatial modes can be decomposed. Based on prediction coefficients The export forecast results can be reconstructed by weighting.

[0099]

[0100] Specifically, the predicted data includes the modal coefficient vector of the outlet flow field; and the steps of performing device flow field reconstruction based on the predicted data and sampled data to obtain the spatial modal matrix include: Based on the sampled data of the inlet flow field and the spatial mode matrix, the target mode coefficient vector of the inlet flow field is determined. The target mode coefficient vector is then input into the trained BP neural network to obtain the predicted mode coefficient vector of the outlet flow field of the target device and the corresponding component performance parameters. The flow field of the target device is reconstructed by predicting the modal coefficient vector and the spatial mode matrix.

[0101] Regarding the specific description above, after obtaining the inlet flow field sampling data under the current operating condition of the target equipment, the sampling data is first arranged in the same sampling point order as in the training phase, and then projected onto the spatial mode obtained by the POD method beforehand. The weights of the inlet flow field on each spatial mode are calculated, thereby obtaining the target modal coefficient vector used to characterize the inlet operating condition. This target modal coefficient vector is essentially a low-dimensional feature representation of the high-dimensional inlet flow field information, which retains the key features related to the outlet flow field and component performance parameters, while compressing the original data into the modal coefficient space consistent with the training phase. Subsequently, the target modal coefficient vector is fed as input into the trained BP neural network, which can output the predicted modal coefficient vector of the outlet flow field corresponding to the current inlet operating condition and the corresponding predicted values ​​of component performance parameters, realizing the mapping of directly inferring the modal features and performance indicators of the outlet flow field from the modal features of the inlet flow field.

[0102] The reconstructed flow field on the inlet side can be obtained by weighting and summing the spatial modes in each column of the spatial mode matrix according to the elements in the target mode coefficient vector. Similarly, the spatial modes of the outlet flow field can be weighted and superimposed according to the elements in the predicted mode coefficients to obtain the reconstructed flow field on the outlet side. If necessary, the mean field obtained through the aforementioned mean-reduction processing can also be superimposed, so that the reconstruction result simultaneously includes the overall background distribution and modal fluctuation characteristics. Through the above flow field reconstruction process, this invention can recover the spatial flow field structure of the target equipment under complex operating conditions, relying only on finite-dimensional modal coefficients and the spatial mode matrix, without the need for high-cost three-dimensional numerical simulation. Combined with the component performance parameters output by the BP neural network, it achieves a unified characterization and visualization analysis of the flow field evolution characteristics and performance response of the target equipment.

[0103] As can be seen from the above, by using the weighted summation of spatial modes based on predicted mode coefficients, the present invention can quickly obtain the reconstructed flow field of the target device at spatial discrete points without performing full three-dimensional numerical simulation, thereby achieving rapid prediction and evaluation of the performance of the target device.

[0104] The data-driven compressor flow field and performance prediction method provided by this invention constructs a snapshot matrix, extracts the spatial mode matrix and modal coefficient matrix, and combines it with a BP neural network for training and prediction. This enables rapid prediction of the outlet flow field and overall performance of turbomachinery under complex inflow conditions while maintaining the complete expression of high-dimensional flow field characteristics. It effectively solves the problems of high-dimensional flow field information being easily lost during the transfer to low-dimensional performance models in existing numerical scaling techniques, difficulty in accurately characterizing distorted flow fields and the response relationship of strongly coupled components, and reliance on high-cost three-dimensional numerical simulation to obtain performance results. Thus, while ensuring prediction accuracy and adaptability to distorted operating conditions, it significantly reduces simulation computation overhead and time costs, and improves the performance evaluation efficiency and engineering application value of turbomachinery under complex inflow conditions.

[0105] A second aspect of the present invention provides a system 2. In some embodiments of the present invention, such as... Figure 13 As shown, system 2 includes: The data acquisition module 201 acquires raw data at low cost and high efficiency based on the volume force model, and preprocesses the raw data to form a flow field snapshot matrix.

[0106] The snapshot matrix decomposition module 202 is used to perform POD method analysis on the snapshot matrix, extract the spatial mode matrix and mode coefficient matrix, and provide the required training set for training set construction through dimensionality reduction and feature extraction of the snapshot matrix.

[0107] The neural network training module 203 is used to input the training set into the BP neural network and continuously update the mapping relationship between the inlet flow field and the outlet flow field in the modal coefficient vector until the neural network completes training.

[0108] The flow field reconstruction and performance prediction module 204 is used to input the sampled data of the inlet flow field of the target device into the trained BP neural network to predict the performance parameters of the target device.

[0109] The system provided by the present invention is used to execute the compressor flow field and performance prediction method in any of the above embodiments, and therefore has all the beneficial effects of the above compressor flow field and performance prediction method, which will not be repeated here.

[0110] An embodiment of the third aspect of the present invention provides an electronic device. In some embodiments of the present invention, such as... Figure 14 As shown, an electronic device is provided, which may include: a desktop computer, a laptop, a handheld computer, and a cloud server, etc. The electronic device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 14 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0111] Processor 301 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0112] The memory 302 can be an internal storage unit of the electronic device 3, specifically, a hard disk or RAM of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, specifically, a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store the computer program 303 and other programs and data required by the electronic device.

[0113] An embodiment of the fourth aspect of the present invention provides a computer-readable storage medium. In some embodiments of the present invention, a computer-readable storage medium is provided that, when executed by processor 301, implements the steps of the above-described method. Therefore, the computer-readable storage medium provided in the fourth aspect of the present invention has all the technical effects of the above-described steps, which will not be repeated here.

[0114] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure 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 of the technical features. The above modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A data-driven method for predicting compressor flow field and performance, characterized in that, include: Based on the volume force model, operational data of the inlet and outlet flow fields of the target equipment under different working conditions and the corresponding component performance parameters under each physical flow rate are obtained; a snapshot matrix is ​​formed using the operational data. The snapshot matrix is ​​decomposed using the POD method to extract the spatial mode matrix and mode coefficient matrix of the inlet flow field and the outlet flow field, respectively; a training set is constructed using all mode coefficient matrices, the physical flow rate, and the component performance parameters. The training set is input into the BP neural network to continuously update the mapping relationship between the physical flow rate and the modal coefficient vector of the inlet flow field, the modal coefficient vector of the outlet flow field, and the component performance parameters, until the BP neural network completes training; The sampled flow field data of the target device's inlet is input into the trained BP neural network to generate predicted flow field data of the target device's outlet and corresponding component performance parameters; the device flow field is reconstructed based on the predicted data and the spatial mode matrix.

2. The compressor flow field and performance prediction method according to claim 1, characterized in that, Before collecting the operational data, the compressor flow field and performance prediction method further includes: The inlet flow field and the outlet flow field are divided in space according to their axial positions to determine the sampling positions, and sampling points for collecting the operation data are set according to the sampling positions. Define the form and intensity of the flow field distortion in the inlet flow field, and set the sample patterns in the training set according to the form and intensity.

3. The compressor flow field and performance prediction method according to claim 2, characterized in that, The steps of acquiring operational data of the inlet and outlet flow fields of the target equipment under different operating conditions and forming a snapshot matrix based on the volume force model include: The volume force model is used to obtain operational data containing the flow field state of the inlet and outlet flow fields during operation at low cost and high efficiency, and the operation data is used to form a flow field snapshot; The flow field snapshots are flattened into vectors according to the sampling points in a predetermined order, and a snapshot matrix is ​​constructed based on the vectors of all operating conditions.

4. The compressor flow field and performance prediction method according to claim 2, characterized in that, The example pattern is used to classify the flow field features in the training set; The training steps of the BP neural network include: The modal coefficient vector of the outlet flow field and the component performance are used as labels to calibrate the modal coefficient vector and physical flow rate of the inlet flow field. During training, the BP neural network is trained based on the flow field characteristics corresponding to the labels.

5. The compressor flow field and performance prediction method according to claim 4, characterized in that, The example patterns are generated by combining categories based on the form and intensity of the flow field distortion; and The forms include radial distortion, circumferential distortion, and combined distortion occurring in the matrix region; The intensity is defined based on the ratio between the total air pressure in the distorted region and the atmospheric pressure in the undistorted region of the matrix area.

6. The compressor flow field and performance prediction method according to claim 1, characterized in that, The step of decomposing the snapshot matrix using the POD method to extract the spatial mode matrix and modal coefficient matrix of the inlet flow field and the outlet flow field respectively includes: The snapshot matrix is ​​subjected to dimensionality reduction and feature extraction using the intrinsic orthogonal decomposition method to obtain orthogonal basis vectors that can characterize spatial features. The orthogonal basis vectors are then arranged in columns to generate the spatial mode matrix. For each operating condition, obtain the modal coefficient vector associated with the spatial mode and the snapshot matrix, and generate the modal coefficient matrix by arranging the modal coefficient vectors of all operating conditions in rows.

7. The compressor flow field and performance prediction method according to any one of claims 2-6, characterized in that, The predicted data includes the modal coefficient vector of the outlet flow field; and the step of performing device flow field reconstruction based on the spatial modal matrix obtained from the predicted data and the sampled data includes: The target modal coefficient vector of the inlet flow field is determined based on the flow field image and spatial modal matrix in the sampled data. The target modal coefficient vector and the physical flow rate in the sampled data are input into the trained BP neural network to obtain the predicted modal coefficient vector of the outlet flow field of the target device and the corresponding component performance parameters. The flow field of the target device is reconstructed using the target modal coefficient vector, the predicted modal coefficient vector, and the spatial modal matrix.

8. A system for implementing the compressor flow field and performance prediction method according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect operation data and form a snapshot matrix based on the operation data; The snapshot matrix decomposition module is used to perform POD decomposition on the snapshot matrix, extract the spatial mode matrix and mode coefficient matrix, and provide the required data for training set construction through dimensionality reduction and feature extraction of the snapshot matrix. The neural network training module is used to input the training set into the BP neural network and continuously learn the mapping relationship between the inlet flow field and the outlet flow field in the modal coefficient vector until the neural network completes training. The flow field reconstruction and performance prediction module is used to input the sampled data of the inlet flow field of the target device into the trained BP neural network to predict the flow field and performance parameters of the target device.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images