Combustion chamber shape and state quantity reconstruction method based on discrete time series data driving

Abstract

The application discloses a combustion chamber shape and state quantity reconstruction method based on discrete time sequence data driving, relates to the field of aircraft engine combustion chamber design, and obtains full-field node coordinates, temperature and strain data of an outer wall of a combustion chamber through multi-working-condition multi-physical-field numerical simulation; selects limited discrete node temperature and strain time sequence data corresponding to the positions of measuring points as input samples, takes the full-field node coordinates, temperature and strain of the outer wall in the same time window as output samples, constructs a supervised learning data set and trains a data-driven reconstruction model. When reconstruction is performed, the limited discrete time sequence data under a to-be-reconstructed working condition is input into the trained model, and the full-field node coordinates, temperature and strain of the outer wall of the combustion chamber are output, so that the combustion chamber shape and state quantity are jointly reconstructed. The method can effectively reduce the full-field direct measurement and online coupling calculation requirement and improve the reconstruction efficiency.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine combustor design. More specifically, this invention relates to a method for reconstructing combustor shape and state variables based on discrete time-series data, which applies engine combustor state monitoring, structural shape reconstruction, and data-driven inversion to aircraft engine combustor design. Background Technology

[0002] With the development of hypersonic vehicles and air-breathing propulsion systems, the combustion chamber, as a key hot-end component of the engine, directly affects engine performance and structural safety due to changes in its shape and state. When operating under high temperature, high pressure, and complex combustion environments, the outer wall of the combustion chamber is prone to temperature rise, strain changes, and structural deformation. Therefore, it is necessary to monitor and reconstruct the shape and state parameters of the combustion chamber's outer wall.

[0003] Currently, combustion chamber condition monitoring mainly relies on sensors for temperature, strain, and pressure to acquire data from local measurement points. Due to limitations in the combustion chamber's working environment and sensor placement space, only a small number of discrete measurement points' temperature and strain time-series data can typically be obtained, making it difficult to directly acquire the spatial coordinates, temperature, and strain of all nodes on the outer wall.

[0004] Furthermore, although existing physical model-based inversion methods have a certain degree of physical interpretability, they are complex to model and have high computational costs, making it difficult to meet the needs of rapid online reconstruction.

[0005] Furthermore, data-driven methods can be used to establish a mapping relationship between discrete measurement point data and the overall response using simulation or experimental data, and the reconstruction results can be quickly output after model training. However, existing methods mostly focus on predicting single temperature fields, flow fields, or local states, and research on simultaneously reconstructing the spatial coordinates, temperature, and strain of the combustion chamber exterior nodes based on finite discrete time-series data is still insufficient.

[0006] Therefore, a method is needed to quickly reconstruct the shape and state variables of the combustion chamber exterior wall using a small amount of discrete time-series data. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0008] To achieve these objectives and other advantages of the present invention, a method for reconstructing combustion chamber shape and state variables based on discrete time-series data is provided, characterized by comprising: S1. Establish a three-dimensional geometric model of the combustion chamber and conduct multi-condition multi-physics field numerical simulations to obtain response data corresponding to all field nodes of the outer wall of the combustion chamber. S2. Based on the measurement point arrangement scheme in the combustion chamber, select a finite number of discrete nodes corresponding to the measurement point positions from all field nodes, and use the node data corresponding to the finite number of discrete nodes as input samples to characterize the local time-series response under the condition of finite measurement points. S3. Extract the response data within the same time window as the input sample as the output sample to characterize the full-field shape, temperature field and strain field response of the combustion chamber exterior wall, and thus obtain the sample set; S4. Construct a data-driven reconstruction model based on a temporal convolutional residual neural network model, and train and validate the data-driven reconstruction model based on a sample set to build a nonlinear mapping relationship between input samples and output samples. S5. During reconstruction, the finite discrete time series data under the working condition to be reconstructed is used as the input of the data-driven reconstruction model, and the reconstruction response data is output based on the mapping relationship to complete the joint reconstruction of the combustion chamber shape and state variables. In S1, the response data includes: The spatial coordinates corresponding to the full-field nodes of the outer wall of the combustion chamber are used to characterize the shape changes of the outer wall of the combustion chamber. Corresponding to the full-field nodes of the outer wall of the combustion chamber, temperature data and strain time series data are used to characterize the changes in the state variables of the outer wall of the combustion chamber.

[0009] Preferably, in S1, the material properties and boundary conditions are pre-set before the multi-condition multi-physics numerical simulation; The multi-condition multiphysics numerical simulation is performed in a sequential coupling manner, including the following steps: Fluid field simulation was performed to obtain the temperature field I and pressure field at the coupling interface of the combustion chamber through fluid field calculation; Temperature field I is used as the input to the transient thermal analysis module to obtain the transient thermal simulation of temperature field II corresponding to the combustion chamber structure; Pressure field and temperature field II are used as input loads for the transient structural analysis module to obtain transient structural simulation data of the combustion chamber outer wall nodal response data.

[0010] Preferably, in S2, the finite discrete nodes include: Temperature nodes corresponding to the locations of combustion chamber temperature measuring points; Strain nodes corresponding to the locations of strain measurement points in the combustion chamber.

[0011] Preferably, S3 also includes: preprocessing the sample set through temporal segmentation, input / output sample matching, dataset partitioning, and normalization. The time series segmentation refers to dividing continuous long time series data into multiple short time series samples using a fixed-length sliding window method. The normalization process refers to using the maximum-minimum normalization method to map spatial coordinates, temperature data, and strain data of different dimensions to a unified numerical range.

[0012] Preferably, in S4, the data-driven reconstruction model includes: Input data transformation module used for dimensional adjustment and feature organization of input samples; A time-series feature extraction module used to extract dynamic features from node data; A residual feature enhancement module is used to enhance the ability of data-driven reconstruction models to express complex nonlinear mapping relationships. The full-field response reconstruction module is used to output reconstruction response data.

[0013] Preferably, the input data transformation module operates as follows: The input samples are reshaped to convert 4D data into 3D data. After dimensional transformation, the permute operation is used to swap the dimensional positions of the channel bits and the time steps.

[0014] Preferably, the temporal feature extraction module includes a one-dimensional convolutional layer and a ReLU activation function, and the first... d The hidden passage is in the first j The convolutional output at each time step Characterized by the following formula: in, b To represent the number of samples in a batch, d Indicates the sequence number of the hidden feature channel, and d =1,2,…,256 r Indicates the sequence number of the input channel, and r =1,2,…,8, j Indicates the time step number. Represents the convolution weights. Indicates the bias term. This indicates the relative position of the convolution kernel in the time dimension. This represents the convolution kernel.

[0015] Preferably, the residual feature enhancement module includes residual block I, residual block II, and residual block III connected in series, and the output of each residual block... H m It is characterized by the following formula: In the above formula, R m This represents the output of the convolutional branch in each residual block, and K represents the kernel size. This indicates the output of the previous residual block.

[0016] Preferably, the full-field response reconstruction module operates as follows: The output features of residual block III are permuteed using the following formula, and the dimensional positions of the channel bits and time steps are swapped using the following formula: In the above formula, t The number of time steps contained in a single sample; The 256-dimensional hidden features corresponding to each time step are fully connectedly mapped using the following formula: In the above formula, For the output of the fully connected layer, the data dimension change process of the fully connected mapping is as follows: ; The output of the fully connected layer is rearranged using the following formula to obtain the corresponding reconstructed response data. : In the above formula, This is a rearrangement function.

[0017] The present invention has at least the following beneficial effects: Firstly, this invention can reconstruct the spatial coordinates, temperature, and strain of all nodes on the outer wall of the combustion chamber using time-series data of temperature and strain from a finite number of discrete nodes, thus overcoming the shortcomings of traditional finite-point monitoring in obtaining full-field information and thus having a stronger full-field reconstruction capability.

[0018] Secondly, the output of this invention includes the spatial coordinates of the nodes on the outer wall of the combustion chamber, temperature, and strain. It can characterize the shape changes of the outer wall of the combustion chamber and reflect its temperature field and strain field state, thus achieving synchronous output of shape and state quantities.

[0019] Third, this invention drives the reconstruction model through offline training data, eliminating the need to repeatedly solve complex multiphysics coupling models during the reconstruction stage. This is beneficial for improving the reconstruction efficiency of combustion chamber shape and state variables, and also reduces the complexity of online computation.

[0020] Fourth, this invention uses discrete node time-series data as input, and can utilize the dynamic characteristics of temperature and strain changes over time to enhance the characterization of the evolution of the shape and state variables of the outer wall of the combustion chamber, making it more suitable for processing relevant data of time-series dynamic response.

[0021] Fifth, the spatial coordinates, temperature, and strain of all nodes output by this invention can be used for three-dimensional shape reconstruction, temperature field display, and strain field display of the combustion chamber, providing data support for combustion chamber condition monitoring and structural health assessment, and facilitating engineering applications and result display.

[0022] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process of a combustion chamber shape and state quantity reconstruction method based on discrete time series data driven by the present invention. Figure 2 This is a schematic diagram of the multi-condition numerical simulation and full-field response data extraction process of the combustion chamber in this invention; Figure 3 This is a schematic diagram illustrating the construction of discrete temporal input samples and full-field node output samples in this invention; Figure 4 This is a schematic diagram of the data-driven reconstruction model in this invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0025] This invention provides a method for reconstructing the shape and state variables of a combustion chamber based on discrete time-series data. It aims to address the problems of difficulty in directly measuring the overall shape and state variables of the combustion chamber exterior wall, the computational complexity of traditional multiphysics inversion methods, and the inability of finite measurement point data to characterize the overall response. This method, based on multi-condition numerical simulation results, selects finite discrete node temperature and strain time-series data corresponding to sensor measurement point locations as input samples, and uses the spatial coordinates, temperature, and strain of the entire combustion chamber exterior wall nodes as output samples. A supervised learning dataset is constructed and a data-driven reconstruction model is trained to achieve rapid reconstruction of the overall shape and state variables of the combustion chamber exterior wall from finite discrete node time-series data. Figure 1 As shown, specifically, it includes three parts: the dataset construction stage (including steps 1-6), the model construction stage (including steps 7-8), and the application reconstruction stage (including steps 9-10), which specifically includes the following steps: Step 1: Establish a three-dimensional geometric model of the combustion chamber In this step, a three-dimensional geometric model is established based on the actual structure of the combustion chamber, retaining the main flow channel, cavity structure, inlet end, outlet end and outer wall structure of the combustion chamber, and setting material properties and boundary conditions, wherein the boundary conditions include: inlet boundary conditions, outlet boundary conditions, wall boundary conditions and structural constraint conditions.

[0026] Step 2: Perform multiphysics coupled numerical simulation of the combustion chamber. The multi-condition multiphysics numerical simulation in this step is carried out in a sequential coupling manner, which includes steady-state calculation of fluid field, transient thermal analysis and transient structural analysis in sequence. The sequential coupling means that the fluid field and the structural field are coupled in the order of fluid field and structural field. In this step, firstly, the temperature and pressure fields of the combustion chamber coupling interface are obtained through fluid field calculation; then, the temperature field is input into the transient thermal analysis module to obtain the temperature field of the combustion chamber structure; finally, the pressure field and structural temperature field are input as loads into the transient structural analysis module to obtain the spatial coordinates (i.e., x-coordinate, y-coordinate, z-coordinate), temperature and strain time-series response data of the nodes on the outer wall of the combustion chamber. The spatial coordinates are used to characterize the shape change of the outer wall of the combustion chamber, and the temperature and strain are used to characterize the state variables change of the outer wall of the combustion chamber.

[0027] Step 3: Extract the response data corresponding to all nodes on the outer wall. This step extracts data from the full-field nodes of the combustion chamber exterior wall at various times from the numerical simulation results. The spatial coordinates, temperature, and strain of each node are used as feature channels to construct the response data of the full-field nodes of the combustion chamber exterior wall. The spatial coordinates are used to characterize the shape changes of the combustion chamber exterior wall, and the temperature and strain are used to characterize the state variables changes of the combustion chamber exterior wall.

[0028] Step 4: Construct input samples for discrete time series. Based on the combustion chamber measuring point layout scheme, this step selects a finite number of discrete nodes corresponding to the measuring point locations from the full-field nodes on the outer wall of the combustion chamber in the numerical simulation, and extracts the temperature and strain time series data of the finite number of discrete nodes to construct the discrete time series input sample. It should be noted that the input sample in this step is not directly derived from the actual sensor data, but from the outer wall node data corresponding to the measuring point locations in the numerical simulation results, which is used to simulate the local time series response information that can be obtained under the condition of finite measuring points. The finite discrete nodes include: temperature nodes for temperature measurement points and strain nodes for strain measurement points. The temperature nodes correspond to the positions of temperature measurement points in the combustion chamber, and the strain nodes correspond to the positions of strain measurement points in the combustion chamber. The discrete time-series input samples during the training phase are derived from the node data corresponding to the measurement point positions in the full-field response data of the outer wall of the combustion chamber.

[0029] Step 5: Construct output samples for all nodes. In the full-field nodes of the combustion chamber exterior wall in numerical simulation, the spatial coordinates, temperature and strain data of the full-field nodes of the combustion chamber exterior wall that are in the same time window as the input samples are extracted, and the full-field node output samples are constructed so that each discrete time-series input sample corresponds to a full-field shape and state quantity output sample.

[0030] This step mainly uses discrete time-series input samples to characterize the local time-series response under finite measurement point conditions, and uses the output samples of the full-field nodes to characterize the full-field shape, temperature field, and strain field response of the combustion chamber exterior wall. During operation, the input samples and the output samples of the full-field nodes are matched according to the same time window to ensure that there is a time correspondence between the temperature and strain time-series data corresponding to the finite discrete nodes and the spatial coordinates, temperature, and strain data of the full-field nodes of the combustion chamber exterior wall.

[0031] Step 6: Data preprocessing and creation of supervised learning dataset This step mainly involves preprocessing the input and output samples obtained from multi-condition simulation data to form a supervised learning dataset for model training, validation, and testing. The preprocessing includes temporal segmentation, input and output sample matching, dataset partitioning, and normalization. The time series segmentation uses a fixed-length sliding window method to divide continuous long time series data into multiple short time series samples.

[0032] The normalization process employs a max-min normalization method to map coordinate, temperature, and strain data to a unified numerical range, thereby reducing the impact of differences in feature dimensions and numerical ranges on model training.

[0033] Step 7: Build a data-driven refactoring model In this step, a data-driven reconstruction model is constructed using a temporal convolutional residual neural network model (TCN-ResNet). This model is used to establish the mapping relationship between finite discrete node time-series data and the coordinates, temperature, and strain of all nodes on the outer wall of the combustion chamber. Specifically, it includes: Input data transformation module for adjusting the dimensions and organizing the features of discrete time series input samples; A temporal feature extraction module for extracting dynamic features from time-series temperature and strain data with finite discrete nodes is provided. The temporal feature extraction module adopts a temporal convolutional network structure, and the residual feature enhancement module adopts a residual network structure. The temporal convolutional network structure is used to extract time-related features from the time-series temperature and strain data with finite discrete nodes, and the residual network structure is used to improve the deep feature representation ability and model training stability. Residual feature enhancement module used to enhance the model's ability to express complex nonlinear mapping relationships; The full-field response reconstruction module is used to output the spatial coordinates, temperature, and strain of all nodes on the outer wall of the combustion chamber.

[0034] Step 8: Model Training and Performance Evaluation In this step, the data-driven reconstruction model is trained using a supervised learning dataset to learn the nonlinear mapping relationship between discrete node temperature and strain time-series data and the full-field node response of the combustion chamber exterior wall. The reconstruction performance of the model is evaluated using a validation set and a test set. The trained data-driven reconstruction model outputs the spatial coordinates, temperature, and strain data corresponding to the full-field nodes of the combustion chamber exterior wall as reconstruction response data. The spatial coordinates in the reconstruction response data are used to recover the three-dimensional shape of the combustion chamber exterior wall. The temperature and strain data in the reconstruction response data are used to characterize the temperature and strain field distribution of the combustion chamber exterior wall and to visualize the shape changes and state changes of the combustion chamber exterior wall.

[0035] Step 9: Model Reconstruction Application During the reconstruction phase, the finite discrete time-series data corresponding to the measurement point locations under the working conditions to be reconstructed are input into the trained data-driven reconstruction model, which outputs the coordinates of all nodes on the outer wall of the combustion chamber, temperature, and strain, thereby achieving joint reconstruction of the combustion chamber shape and state variables.

[0036] Step 10: Visual Representation The three-dimensional shape of the outer wall of the combustion chamber is generated based on the output spatial coordinates of all nodes, and the temperature and strain are mapped to the corresponding nodes, realizing the visualization and characterization of the three-dimensional shape, temperature field, and strain field of the outer wall of the combustion chamber. Therefore, the method of the present invention can utilize the time series data of temperature and strain from finite discrete nodes to establish a mapping relationship between the local discrete response and the coordinates of all nodes, temperature, and strain of the outer wall of the combustion chamber, and can reconstruct the shape and state variables of the outer wall of the combustion chamber without directly measuring the full-field data.

[0037] Example: like Figure 1 As shown, the overall process of this invention consists of a dataset construction stage, a model construction stage, and a reconstruction and application stage. Specifically, the dataset construction stage is used to obtain full-field response data for multiple operating conditions in the combustion chamber; the model training stage is used to establish the mapping relationship between finite discrete node time-series data and full-field node responses; and the reconstruction and application stage is used to output the spatial coordinates, temperature, and strain of the full-field nodes on the outer wall of the combustion chamber.

[0038] In step 1, the three-dimensional geometric model of the combustion chamber is established based on the actual combustion chamber structure. As one embodiment, the combustion chamber is a rectangular cross-section long channel structure, with a total length of approximately 963 mm, a total width of approximately 130 mm, a main body height of approximately 70 mm, and a wall thickness of approximately 10 mm. A concave cavity structure is set in the middle of the model to characterize the internal flow and combustion organization features of the combustion chamber. To improve computational efficiency, details such as threads and local chamfers, which have little impact on the overall temperature field, strain field, and shape changes, are simplified. As another embodiment, in practical applications, multiple operating parameters need to be set, including the incoming Mach number, equivalence ratio, incoming total temperature, pressure boundary, and fuel injection parameters. The incoming Mach number is set to Ma = 4.0, and the equivalence ratios are set to 1.2, 1.1, 1.0, 0.9, and 0.8, respectively.

[0039] In step 2, as Figure 2 As shown, a sequential coupling approach is used to conduct multiphysics numerical simulations. First, after modeling the 3D solid, the data is imported into Fluent for steady-state fluid field calculations to extract the coupling interface data, thereby obtaining the temperature and pressure fields at the fluid-structure interaction interface. Then, the coupled interface temperature field and convective heat transfer data are imported into the transient thermal analysis module for transient thermal analysis, and boundary condition I is input to perform solid transient thermal analysis and solve the temperature response of the combustion chamber structure over time to obtain the transient temperature field distribution, and extract the corresponding temperature field from the transient temperature field distribution. Finally, the coupled interface pressure field obtained from the steady-state fluid field calculation and the transient temperature field distribution obtained from the transient thermal analysis module are combined as loads and imported into the transient structural analysis module for transient structural analysis. Boundary condition II is input, and through solid transient structural analysis and solution calculation, the transient displacement field distribution and transient strain field distribution are obtained respectively. The corresponding displacement field and strain field are extracted from the transient displacement field distribution and transient strain field distribution, respectively. Combined with the temperature field obtained from the transient thermal analysis module, the spatial coordinates, temperature, and strain response data of the outer wall nodes of the combustion chamber are obtained. As an example, the total transient calculation time is 10 seconds, and the number of time points is 2000.

[0040] In step 3, combined Figure 2The data extraction process shown extracts data from all nodes of the combustion chamber exterior wall in batches. Each node contains five feature channels at each time point: x-coordinate, y-coordinate, z-coordinate, temperature, and strain. The x-coordinate, y-coordinate, and z-coordinate describe the shape changes of the combustion chamber exterior wall, while temperature and strain describe the state changes. In one embodiment, 37,864 combustion chamber exterior wall nodes are extracted, forming a data structure of 2000×37864×5. Here, 2000 represents the number of time points, 37864 represents the total number of nodes on the combustion chamber exterior wall, and 5 represents the number of feature channels.

[0041] like Figure 3 As shown, steps 4 and 5 complete the construction of the model input and output samples. The input does not use all node data from the outer wall of the combustion chamber, but rather selects a finite number of discrete nodes corresponding to the measurement point locations from the numerical simulation outer wall nodes according to the measurement point layout scheme, serving as discrete time-series input samples. It should be noted that the input samples in this embodiment during the training phase do not come directly from actual sensor data, but rather from the node data corresponding to the measurement point locations in the numerical simulation results. Furthermore, the node data consists of temperature time-series data obtained from the nodes corresponding to temperature measurement points, and strain time-series data obtained from the nodes corresponding to strain measurement points. As one embodiment, the input end selects eight discrete nodes as input nodes from the nodes corresponding to temperature measurement points and strain measurement points. Among them, four nodes correspond to the temperature measurement point locations (used to obtain finite discrete node temperature time-series data from temperature time-series data), and four nodes correspond to the strain measurement point locations (used to obtain finite discrete node strain time-series data from strain time-series data), thus obtaining discrete time-series input samples. The output end uses the x-coordinate, y-coordinate, z-coordinate, temperature, and strain data corresponding to the full-field node data of the combustion chamber exterior wall matched in the same time window as the full-field node output samples.

[0042] In step 6, for Figure 3 The input and output data shown are used for time series segmentation and sample matching. Let the total length of the long time series data be T, the window length be W, and the sliding step size be S. Then the number of short time series samples N that can be generated is: This step involves constructing time-series samples using a fixed-length sliding window for long-term multiphysics simulation data of the combustion chamber. This adapts to the dynamic time-series input from the sensors, facilitating the later construction of a supervised learning dataset. In practice, the 10-second time-series data can be divided into 1-second time windows, with each window containing 200 time points and a sliding step size of 0.2 seconds (40 time points). After segmentation, the discrete time-series input samples are matched with the output samples of all nodes within the same time window. A total of 230 samples are generated for five operating conditions: 160 for the training set, 35 for the validation set, and 35 for the test set. Due to the different dimensions of spatial coordinates, temperature, and strain, a maximum-minimum normalization method is used for normalization. The calculation method is as follows: in, x This is the original data. For the normalized data, x min and x max These represent the minimum and maximum values ​​of the corresponding features. For large-scale full-field node data, HDF5 format can be used for block storage and on-demand retrieval. This method can simultaneously handle three types of dimensional data with huge differences: spatial coordinates, temperature, and strain. It eliminates the interference of the differences in the numerical span of multiple physical features on the training of the TCN-ResNet model, and thus adapts to the multi-channel output after the joint reconstruction of shape and state variables.

[0043] In step 7, as an example, the data-driven reconstruction model adopts a temporal convolutional residual network (TCN-ResNet) structure to establish a nonlinear mapping relationship between the temporal data of finite discrete measurement points and the full-field node response of the outer wall of the combustion chamber. Figure 4 The overall structure and data flow of the data-driven reconstruction model are shown. The data-driven reconstruction model takes the finite discrete time-series samples input above as input, and after feature extraction, enhancement, and mapping by intermediate functional modules, outputs the full-field nodal response results of the combustion chamber exterior wall below. The output results include the x-coordinate, y-coordinate, z-coordinate, temperature, and strain of the full-field nodal nodes on the combustion chamber exterior wall.

[0044] Specifically, such as Figure 4 As shown, the input is a finite discrete-time input sample. The finite discrete-time input sample consists of the response data of several temperature and strain measurement points over continuous time steps. In one specific implementation, the finite discrete-time input sample can be 8-channel time-series data, including 4 temperature measurement points and 4 strain measurement points. The horizontal axis represents the time step, and the vertical axis represents the response amplitude of the corresponding measurement point, reflecting the dynamic changes of the discrete measurement points in the time domain.

[0045] like Figure 4As shown, the data-driven reconstruction model includes an input data transformation module, a temporal feature extraction module, a residual feature enhancement module, and a full-field response reconstruction module, and these modules are connected in sequence.

[0046] For ease of explanation Figure 4 The data transfer relationships between the various modules are defined as follows: Let the discrete-time input sample be X, and the model output be the full-field nodal response of the combustion chamber exterior wall. The overall mapping relationship of the data-driven reconstruction model can be expressed as: in, This indicates the data-driven reconstruction model. θ It represents the model parameters and establishes a nonlinear mapping between a small number of discrete measurement point time series inputs X and the joint output Y of spatial coordinates, temperature, and strain of all nodes on the outer wall of the combustion chamber through an overall mapping relationship. This is different from the traditional single temperature / flow field prediction model and simultaneously realizes the reconstruction of geometry and multiple physical state variables.

[0047] The discrete-time input sample X can be represented as: in, b Indicates the number of samples in the batch. t This indicates the number of time steps contained in a single sample. p Represents a finite number of discrete nodes. c This represents the input feature dimension corresponding to each finite discrete node. This expression mainly enables the four-dimensional structure of batch, time step, number of measurement points, and single measurement point features to adapt to parallel time-series input of mixed temperature and strain measurement points, and supports unified encoding of multi-sensor mixed data.

[0048] In this embodiment, the input sample includes nodes corresponding to 4 temperature measurement points and 4 nodes corresponding to 4 strain measurement points, constituting a total of 8-channel input. p =8, c =1, satisfying: The model ultimately outputs the full-field nodal response of the outer wall of the combustion chamber, and the output data format is as follows: Wherein, 37864 represents the total number of nodes on the outer wall of the combustion chamber, and 5 represents the five types of response features corresponding to each node, namely x-coordinate, y-coordinate, z-coordinate, temperature and strain. The four-dimensional output tensor simultaneously carries the five types of features of the three-dimensional spatial coordinates (shape) + temperature + strain (state) of all nodes in the field, realizing the synchronous output of geometric shape and multi-physics field distribution of a single model, which is used to distinguish it from the existing single-field reconstruction methods.

[0049] The input data transformation module is used to compress and rearrange the dimensions of the original finite discrete time-series input samples so that the input data meets the requirements of subsequent one-dimensional convolution calculations.

[0050] Specifically, the input tensor is first reshaped to transform the original data from a four-dimensional form to a three-dimensional form: In this embodiment, due to Therefore, the reshape result is: Subsequently, a permute operation is performed on the reshaped input data, so that the finite discrete node channels are located in the second dimension and the time step is located in the third dimension, resulting in: It should be noted that the above formula is specifically designed for the dimension transformation logic of mixed discrete measurement points of temperature and strain, adjusting the time series of multiple measurement points to the standard input format of temporal convolutional network (TCN), solving the engineering problem that mixed time series of multiple sensors cannot be directly convolved in one dimension. Furthermore, through the above data transformation, the time series data of the corresponding nodes of the four temperature measurement points and the corresponding nodes of the four strain measurement points are organized into eight input channels and input along the time axis into the subsequent time series feature extraction module.

[0051] The temporal feature extraction module is used to perform convolutional encoding on 8-channel finite discrete time-series data along the time dimension to extract dynamic change features in temperature and strain time-series responses.

[0052] Specifically, the temporal feature extraction module includes a one-dimensional convolutional layer and a ReLU activation function, the calculation process of which can be expressed as follows: in, k =3 indicates that the kernel size is 3. p =1 indicates that the padding length is 1.

[0053] The data dimension change process of the time-series feature extraction module is as follows: That is, the module maps the 8-channel temperature and strain time series inputs of a finite number of discrete nodes to 256-channel hidden time series features, while maintaining the number of time steps. t constant.

[0054] Furthermore, the first d The hidden passage is in the... j The convolution output at each time step can be represented as: in,d Indicates the hidden feature channel number. d =1,2,…,256; r Indicates the input channel number. r =1,2,…,8; j Indicates the time step number. Represents the convolution weights. Indicates the bias term. This indicates the relative position of the convolution kernel in the time dimension. The temporal feature extraction module uses one-dimensional temporal convolution to extract the dynamic temporal evolution features of the measurement points. It is specifically adapted to the transient combustion dynamic temporal data of the combustion chamber, and mines the coupling law of temperature and strain changes with time, providing temporal feature support for dynamic deformation reconstruction.

[0055] The residual feature enhancement module is used to further enhance the 256-channel hidden features output by the temporal feature extraction module. This module includes three residual blocks connected in series: residual block 1, residual block 2, and residual block 3.

[0056] Specifically, let the first m The input features of each residual block are Hm−1, where m If the values ​​are 1, 2, and 3, then the output of the convolution branch in the residual block can be expressed as: Subsequently, the output of the convolution branch is added element-wise to the input of the shortcut residual connection, and ReLU activation is performed to obtain the... m Output of each residual block: The feature dimensions of each residual block remain unchanged before and after, that is: By performing feature enhancement sequentially through three residual blocks, the model's ability to express the complex nonlinear relationship between finite discrete node time-series data and the full-field node response of the combustion chamber exterior wall is improved while keeping the number of hidden feature channels and time steps constant.

[0057] The full-field response reconstruction module is used to map the 256-channel hidden temporal features enhanced by residual features to five types of response features of the full-field nodes of the combustion chamber exterior wall.

[0058] Specifically, the features output by residual block 3 are first permuteed to convert the data from channel-priority format to time-step-priority format: The purpose of this step is to strengthen the nonlinear correlation between the measurement point timing and the full-field multiphysics field for the cascaded multi-layer one-dimensional residual blocks without changing the timing and channel dimensions, solve the gradient vanishing problem in deep temporal networks, and adapt to large-scale full-field node high-dimensional output training.

[0059] Subsequently, a fully connected mapping is performed on the 256-dimensional hidden features corresponding to each time step: The data dimension change process of the fully connected mapping is as follows: The fully connected layer maps the 256-dimensional hidden features at each time step to five types of response features for 37,864 combustion chamber outer wall nodes.

[0060] Finally, the output results of the fully connected layer are rearranged to obtain the full-field nodal response of the outer wall of the combustion chamber: Furthermore, such as Figure 4 As shown, the overall nodal response result can be further decomposed into five output components: x-coordinate, y-coordinate, z-coordinate, temperature, and strain. The x-coordinate, y-coordinate, and z-coordinate together characterize the spatial positional changes of the nodes on the outer wall of the combustion chamber, reflecting the shape of the outer wall; temperature and strain characterize the distribution of state variables on the outer wall of the combustion chamber.

[0061] Through the above-mentioned input data transformation, time series feature extraction, residual feature enhancement and full-field response reconstruction process, the data-driven reconstruction model realizes the joint mapping from finite discrete node temperature and strain time series data to the full-field node spatial coordinates, temperature and strain of the combustion chamber outer wall, thereby realizing the joint reconstruction of the shape and state variables of the combustion chamber outer wall.

[0062] In step 8, combined Figure 4 The model structure shown utilizes a training set to train a data-driven reconstruction model. During training, discrete-time input samples are fed into the model, with the corresponding spatial coordinates, temperature, and strain of the full-field nodes on the outer wall of the combustion chamber serving as supervision labels. The error between the model's predictions and the actual output samples is calculated, and the model parameters are updated using an optimization algorithm. After training, the model is evaluated using a validation set and a test set. The evaluation includes reconstruction errors in different feature channels such as x-coordinate, y-coordinate, z-coordinate, temperature, and strain. As one embodiment, the Mean Absolute Error (MAE) can be used to evaluate the reconstruction results, and its calculation method is as follows: in, This represents the model's predicted value. The true value is represented by n, which represents the total number of valid data points participating in the error statistics. This step mainly involves reconstructing the error by unifying and quantifying the five output channels of spatial coordinates, temperature, and strain, thereby achieving an integrated quantitative evaluation of the combustion chamber shape deformation accuracy and the temperature and strain field state accuracy, which is suitable for the performance verification requirements of the multi-objective joint reconstruction of this invention.

[0063] In step 9, the finite discrete time-series data of the corresponding measurement points under the working condition to be reconstructed are input into the trained reconstruction model. After forward inference, the model outputs the x, y, and z coordinates, temperature, and strain of all nodes on the outer wall of the combustion chamber. The shape of the outer wall of the combustion chamber can be recovered from the output node spatial coordinates, and the distribution of state variables of the outer wall of the combustion chamber can be obtained from the output temperature and strain.

[0064] In step 10, the spatial coordinates of all nodes output in step 9 are used to generate the three-dimensional shape of the outer wall of the combustion chamber. Temperature and strain values ​​are mapped to the corresponding nodes, forming temperature and strain field display results, respectively. This step allows for a visual representation of the shape changes, temperature distribution, and strain distribution of the outer wall of the combustion chamber, providing support for combustion chamber condition monitoring and structural health assessment.

[0065] In summary, the method in this embodiment uses multi-condition numerical simulation results as its data foundation. By employing finite discrete node temperature and strain time-series data corresponding to the measurement point locations, it learns the mapping relationship between these data and the spatial coordinates, temperature, and strain of the entire combustion chamber exterior wall nodes. This method can achieve joint reconstruction of the combustion chamber exterior wall shape, temperature field, and strain field without directly acquiring the full-field measurement data, reducing the computational cost of traditional multiphysics inversion and improving the efficiency of combustion chamber shape reconstruction and state characterization.

[0066] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0067] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for reconstructing the shape and state variables of a combustion chamber based on discrete time-series data, characterized in that, include: S1. Establish a three-dimensional geometric model of the combustion chamber and conduct multi-condition multi-physics field numerical simulations to obtain response data corresponding to all field nodes of the outer wall of the combustion chamber. S2. Based on the measurement point arrangement scheme in the combustion chamber, select a finite number of discrete nodes corresponding to the measurement point positions from all field nodes, and use the node data corresponding to the finite number of discrete nodes as input samples to characterize the local time-series response under the condition of finite measurement points. S3. Extract the response data within the same time window as the input sample as the output sample to characterize the full-field shape, temperature field and strain field response of the combustion chamber exterior wall, and thus obtain the sample set; S4. Construct a data-driven reconstruction model based on a temporal convolutional residual neural network model, and train and validate the data-driven reconstruction model based on a sample set to build a nonlinear mapping relationship between input samples and output samples. S5. During reconstruction, the finite discrete time series data under the working condition to be reconstructed is used as the input of the data-driven reconstruction model, and the reconstruction response data is output based on the mapping relationship to complete the joint reconstruction of the combustion chamber shape and state variables. In S1, the response data includes: The spatial coordinates corresponding to the full-field nodes of the outer wall of the combustion chamber are used to characterize the shape changes of the outer wall of the combustion chamber. Corresponding to the full-field nodes of the outer wall of the combustion chamber, temperature data and strain time series data are used to characterize the changes in the state variables of the outer wall of the combustion chamber.

2. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 1, characterized in that, In S1, material properties and boundary conditions are pre-set before multi-condition multi-physics numerical simulation; The multi-condition multiphysics numerical simulation is performed in a sequential coupling manner, including the following steps: Fluid field simulation was performed to obtain the temperature field I and pressure field at the coupling interface of the combustion chamber through fluid field calculation; Temperature field I is used as the input to the transient thermal analysis module to obtain the transient thermal simulation of temperature field II corresponding to the combustion chamber structure; Pressure field and temperature field II are used as input loads for the transient structural analysis module to obtain transient structural simulation data of the combustion chamber outer wall nodal response data.

3. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 1, characterized in that, In S2, the finite discrete nodes include: Temperature nodes corresponding to the locations of combustion chamber temperature measuring points; Strain nodes corresponding to the locations of strain measurement points in the combustion chamber.

4. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 1, characterized in that, S3 also includes: preprocessing the sample set through time-series segmentation, input-output sample matching, dataset partitioning, and normalization. The time series segmentation refers to dividing continuous long time series data into multiple short time series samples using a fixed-length sliding window method. The normalization process refers to using the maximum-minimum normalization method to map spatial coordinates, temperature data, and strain data of different dimensions to a unified numerical range.

5. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 1, characterized in that, In S4, the data-driven reconstruction model includes: Input data transformation module used for dimensional adjustment and feature organization of input samples; A time-series feature extraction module used to extract dynamic features from node data; Residual feature enhancement module used to enhance the model's ability to express complex nonlinear mapping relationships; The full-field response reconstruction module is used to output reconstruction response data.

6. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 5, characterized in that, The input data transformation module operates as follows: The input samples are reshaped to convert 4D data into 3D data. After dimensional transformation, the permute operation is used to swap the dimensional positions of the channel bits and the time steps.

7. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 6, characterized in that, The temporal feature extraction module includes a one-dimensional convolutional layer and a ReLU activation function, and the first... d The hidden passage is in the... j The convolutional output at each time step Characterized by the following formula: in, b To represent the number of samples in a batch, d Indicates the sequence number of the hidden feature channel, and d =1,2,…,256 r Indicates the sequence number of the input channel, and r =1,2,…,8, j Indicates the time step number. Represents the convolution weights. Indicates the bias term. This indicates the relative position of the convolution kernel in the time dimension. This represents the convolution kernel.

8. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 7, characterized in that, The residual feature enhancement module includes residual block I, residual block II, and residual block III connected in series, and the output of each residual block... H m It is characterized by the following formula: In the above formula, R m This represents the output of the convolutional branch in each residual block, and K represents the kernel size. This indicates the output of the previous residual block.

9. The method for reconstructing combustion chamber shape and state variables based on discrete time-series data as described in claim 8, characterized in that, The full-field response reconstruction module operates as follows: The output features of residual block III are permuteed using the following formula, and the dimensional positions of the channel bits and time steps are swapped using the following formula: In the above formula, t The number of time steps contained in a single sample; The 256-dimensional hidden features corresponding to each time step are fully connectedly mapped using the following formula: In the above formula, For the output of the fully connected layer, the data dimension change process of the fully connected mapping is as follows: ; The output of the fully connected layer is rearranged using the following formula to obtain the corresponding reconstructed response data. : In the above formula, This is a rearrangement function.

Images