High-speed combustion unsteady state prediction method based on lightweight interpretable deep learning

Abstract

The invention discloses a high-speed combustion unsteady state prediction method based on lightweight interpretable deep learning, and relates to the technical field of high-speed combustion intelligent diagnos.The high-speed combustion unsteady state prediction method comprises the steps that a ground wind tunnel test and CFD numerical simulation software are utilized to construct combustion chamber configuration flow field data sets under different incoming flow conditions; performing data preprocessing operation on the flow field data set to obtain a preprocessed flow field data set; disrupting the preprocessed flow field data set and segmenting the flow field data set into a training set and a test set according to a proportion; training the lightweight high-speed combustion unsteady state phenomenon rapid prediction model by using the training set to obtain a trained model and a training weight; and embedding the test set, the trained model and the weight into an edge computing platform to obtain an unsteady state phenomenon prediction result. According to the method, high-precision and high-efficiency prediction of the extreme combustion unsteady-state multi-physical field under the high-speed flight condition can be realized, and an important technical support is provided for evaluation of the operation state of the air-breathing ramjet engine and optimization of a combustion organization.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed combustion intelligent diagnosis, and in particular to a high-speed combustion non-steady-state prediction method based on lightweight and interpretable deep learning. Background Art

[0002] High-speed aircraft powered by air-breathing ramjets are a crucial component of future global rapid transportation. Under high-speed flight conditions, the extreme high-temperature, high-pressure combustion environment within the combustion chamber of an air-breathing ramjets creates complex vortex motion, shock wave / boundary layer interference, and unsteady heat release, leading to combustion instability. This can cause engine flameout, reduced combustion efficiency, increased local heat loads, and even structural damage. Rapid analysis and diagnosis of combustion instabilities under these extreme conditions is crucial for the design and development of the next generation of air-breathing ramjets.

[0003] However, the current development of air-breathing ramjet engines relies primarily on numerical simulation. Traditional numerical simulation methods are limited by the enormous computational complexity of the fuel combustion chemical reaction kinetics, resulting in low computational efficiency. High-resolution and rapid simulation of unsteady flows, such as flashover and dynamic processes, is currently difficult, hindering the development of advanced engines. In recent years, a wide range of computational fluid dynamics (CFD) algorithms have been proposed, with significant breakthroughs achieved in numerically solving incompressible and compressible Navier-Stokes (NS) equations using finite element methods, finite difference methods, and meshless methods. However, the development of these algorithms, as well as the collection, organization, and proper utilization of data, rely heavily on accumulated experience and skills. Furthermore, meshing is a time-consuming task for large-scale projects with complex dynamic hysteresis. Various fluid modeling techniques have been developed to reduce the computational cost of DNS, including Reynolds-averaged Navier-Stokes (RANS) modeling, large eddy simulation (LES), and hybrid RANS-LES models. However, this also reduces the fidelity of the numerical simulation results.

[0004] Thanks to their powerful feature extraction and nonlinear fitting capabilities, deep learning models are increasingly being used to solve classic flow control and applied mathematics problems in the field of fluid mechanics. Numerous studies have utilized neural network models such as residual networks, convolutional networks, long short-term memory networks, and pyramid network structures to achieve high-precision reconstruction of high-speed flow fields. However, purely data-driven neural networks require a large amount of high-quality data and computing power as support. To this end, physical information neural networks, by deeply exploring the fluid dynamics and thermochemical reaction kinetics mechanisms hidden in combustion flow data, cleverly adopt a "data + mechanism" dual-drive mode to achieve soft constraints on neural network models, giving neural network models the prior knowledge and interpretability that they lack, and providing reliable data for combustion flow analysis inside combustion chambers. Summary of the Invention

[0005] In response to the above-mentioned deficiencies in the prior art, the present invention provides a high-speed combustion non-steady-state prediction method based on lightweight and interpretable deep learning, which can achieve high-precision and high-efficiency prediction of extreme combustion non-steady-state multi-physical fields under high-speed flight conditions, providing important technical support for the evaluation of the operating status of air-breathing ramjet engines and the optimization of combustion organization.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a high-speed combustion non-steady-state prediction method based on lightweight and interpretable deep learning, comprising the following steps: S1: Using ground wind tunnel tests and CFD numerical simulation software, a flow field dataset of the combustion chamber configuration under different incoming flow conditions is constructed; S2: performing data preprocessing on the combustion chamber configuration flow field data set to obtain a preprocessed flow field data set; S3: Shuffle the preprocessed flow field dataset and divide it into training set and test set in proportion; S4: Using the training set, the fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion is trained to obtain a trained fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion and training weights; S5: Embed the test set, the trained lightweight high-speed combustion non-steady-state phenomenon rapid prediction model and the training weights into the edge computing platform to obtain the prediction results of the non-steady-state phenomenon.

[0007] Furthermore, the step S2 includes performing data preprocessing operations on the flow field data set generated by the numerical simulation and performing data preprocessing operations on the flow field data set generated by the ground wind tunnel test; The data preprocessing operation on the flow field data set generated by the numerical simulation includes the following steps: A1: Obtain the time coordinates and space coordinates of the flow field boundary grid points; A2: Use the Latin hypercube sampling method to obtain the time coordinates and spatial coordinates of the residual points at the flow field boundary; A3: Obtain the time points and spatial coordinates of the flow field boundary and the flow field within the boundary area; A4: Align the time points and spatial coordinates of each pair of sample data to obtain aligned sample data. Each pair of sample data includes the boundary grid points, residual points, and the flow field within the boundary and boundary region at the same time. The boundary grid points and residual points at the same time are used as input, and the flow field within the boundary and boundary region is used as output. A5: Standardize each aligned sample data to obtain the pre-processed flow field data set generated by the numerical simulation; The data preprocessing operation for the flow field data set generated by the ground wind tunnel test includes the following steps: B1: Convert the input image into a single-channel grayscale space through the weighted fusion method of RGB three channels to achieve grayscale preprocessing; B2: Dynamic range expansion based on cumulative histogram. By constructing a 256-level grayscale histogram statistical distribution, calculating the cumulative distribution function to determine the nonlinear truncation threshold, and performing a piecewise linear transformation, the truncation interval [lower, upper] is mapped to the full dynamic range [0, 255], and the grayscale levels at both ends are saturated to zero. B3: Use a Gaussian low-pass filter to extract low-frequency components, separate high-frequency details through weighted differences between the original image and the low-frequency components, and apply a linear gain to the high-frequency signal; B4: Perform adaptive frequency domain hybrid enhancement, linearly superimpose the high-frequency component after gain with the original image with a weight of 0.7, and construct an improved homomorphic filtering model.

[0008] Furthermore, the Latin hypercube sampling method is: C1: Determine the number of samples N; C2: Divide the data area into M sub-areas; C3: Randomly sample N / M data in each sub-region; C4: Combine the data obtained by random sampling in different sub-areas to obtain the overall sampling result.

[0009] Furthermore, the fast prediction model for the lightweight high-speed combustion non-steady-state phenomenon in S4 includes: The signal expansion module uses a fully convolutional architecture to achieve intelligent mapping of one-dimensional signals to two-dimensional feature maps. It uses two-stage 1×1 convolutions with bilinear upsampling to construct spatial dimensions. The first-layer 1×1 convolution expands the input dimension to 256 dimensions for feature enhancement, and the second-layer 1×1 convolution compresses the channels to the target dimension. 3×3 convolutions are used to refine spatial features, avoiding the parameter explosion problem caused by traditional fully connected layers. The adaptive receptive field module integrates dilated convolution and axial attention mechanisms to construct an adaptive feature extraction unit. It uses 3×3 convolution to expand the effective receptive field to a 7×7 area and cooperates with the axial attention module to achieve feature selection. The axial attention module uses channel reduction technology to compress the intermediate dimension to 1 / 4 of the original channel and uses row-column separated convolution kernels for spatial attention calculation. The multi-scale fusion module consists of parallel multi-branch depthwise separable convolutions. It uses square convolution kernels of three different scales (3, 5, and 7) to capture multi-granularity features. Each branch shares parameters through grouped convolution. Each group processes only a single channel feature. After feature concatenation, cross-channel information fusion is achieved through 1×1 convolution. The progressive upsampling module uses a two-stage learnable upsampling architecture, employs bilinear interpolation with 3×3 convolution for feature refinement, and uses 1×1 convolution to restore the target dimension; The deformable reconstruction module uses dynamic offset prediction combined with deformable convolution to achieve high-precision reconstruction. The 18-channel offset prediction layer learns the deformation parameters of the 3×3 convolution kernel, enabling the receptive field to adaptively adjust based on feature content. Finally, single-channel deformable convolution is used to map features to outputs, completing end-to-end signal reconstruction. The non-stationary phenomenon discrimination module realizes non-stationary feature detection through multi-scale causal convolution and dynamic weighting mechanism. The input time series signal is processed by parallel three-way causal convolution, using 3×1 convolution kernels with expansion rates of 1, 2 and 4 respectively to capture short-, medium- and long-range time-dependent features; the output of each branch is compressed by global average pooling in the spatiotemporal dimension, and then sent to a two-layer fully connected network to generate dynamic weight coefficients, and finally the discrimination result is obtained through weighted fusion.

[0010] Furthermore, the edge computing platform in S5 is an NPU module and its matching LED display module.

[0011] The beneficial effects of the present invention are as follows: the high-speed combustion non-steady-state prediction method based on lightweight interpretable deep learning proposed in the present invention can achieve high-precision and high-efficiency prediction of extreme combustion non-steady-state multi-physical fields under high-speed flight conditions, and provide important technical support for the evaluation of the operating status of air-breathing ramjet engines and the optimization of combustion organization. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of the high-speed combustion non-steady-state prediction method based on lightweight interpretable deep learning of the present invention.

[0013] Figure 2 Schematic diagram of the scramjet engine model structure.

[0014] Figure 3 Schematic diagram of the structure of a fast prediction model for non-steady-state phenomena in high-speed combustion based on lightweight and interpretable deep learning.

[0015] Figure 4 Run the flow chart for the field deployment of the algorithm. DETAILED DESCRIPTION

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown in FIG, a high-speed combustion non-steady-state prediction method based on lightweight interpretable deep learning includes the following steps: S1: Using ground wind tunnel tests and CFD numerical simulation software, a flow field dataset of the combustion chamber configuration under different incoming flow conditions is constructed; S2: performing data preprocessing on the combustion chamber configuration flow field data set to obtain a preprocessed flow field data set; S3: Shuffle the preprocessed flow field dataset and divide it into training set and test set in proportion; S4: Using the training set, the fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion is trained to obtain a trained fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion and training weights; S5: Embed the test set, the trained lightweight high-speed combustion non-steady-state phenomenon rapid prediction model and the training weights into the edge computing platform to obtain the prediction results of the non-steady-state phenomenon.

[0018] Step S1 is specifically as follows: The flow and operating conditions required for CFD numerical calculations and ground wind tunnel tests are determined based on the conditions for the generation of supersonic unsteady phenomena and the operating conditions of air-breathing ramjet engines. The two-dimensional numerical simulation verification work is based on the finite volume method and is performed in an internal and external flow coupling numerical simulation application software. This software is a three-dimensional large-scale parallel computing software developed based on unstructured, hybrid grids and can simulate the compressible turbulence of the diffusion or combustion of complete gases and multi-component mixed gases. The three-dimensional NS equations in the rectangular coordinate system are solved in the following form:

[0019] In the formula 0 , represents the inviscid flux, represents the viscous flux, is the source term, for Directional speed, Indicates the gas density and mass fraction of the components, the total internal energy of the gas ,in represents the thermodynamic internal energy.

[0020] The inlet cross-section of the isolation section of the combustion chamber configuration provided by the present invention is 30×150 mm², and the length of the isolation section is 0.3m, which includes two parts: a straight section with a length of 0.22m and an expansion section with a length of 0.08m and an upper wall expansion angle of 1.4°. The combustion section length ranges from 0.30m to 1.07m, including a groove with a length-to-depth ratio of 11 and an expansion section with four expansion sections. The groove is 11mm deep, and the expansion angle of the groove bottom is 1.4°. The four expansion sections are a first expansion section with a length of 105mm and a 1.4° expansion angle, a second expansion section with a length of 155mm and a 2° expansion angle, a third expansion section with a length of 142mm and an 8° expansion angle, and a fourth expansion section with a length of 103mm and a 15° expansion angle. The schematic diagram of the model is as follows Figure 2 .

[0021] To accurately obtain flow field image information, a variety of non-contact optical measurement methods are used, including schlieren and high-speed photography, to obtain relevant flow information in different states of cold and hot flows, further assisting with wall pressure measurement information and analyzing flame propagation and stability.

[0022] Step S2 is specifically as follows: Optionally, the data preprocessing operations performed on the flow field data set generated by the numerical simulation include: A1. Obtain the time coordinates and space coordinates of the flow field boundary grid points; A2. Obtaining the time coordinates and spatial coordinates of residual points at the flow field boundary using Latin hypercube sampling; wherein the number of the residual points and the number of the flow field boundary grid points are 1:1; In actual operation, the ratio of the number of residual points to the number of flow field boundary grid points is approximately 1:1. In a specific embodiment of the present invention, the number of residual points is 1000, and the number of flow field boundary grid points is 1023.

[0023] Furthermore, the Latin hypercube sampling method is: (1) First determine the sample size N; (2) Divide the data area into M sub-areas; (3) Randomly sample N / M data in each sub-region; (4) Combine the data obtained by random sampling in different sub-areas to obtain the overall sampling results.

[0024] A3. Obtain the time points and spatial coordinates of the flow field boundary and the flow field within the boundary area; A4. Align the time points and spatial coordinates of each pair of sample data to obtain aligned sample data. Each pair of sample data includes boundary grid points, residual points, and the flow field within the boundary and boundary region at the same time. The boundary grid points and residual points at the same time serve as input, and the flow field within the boundary and boundary region serves as output. A5. Perform normalization on each aligned sample data set to obtain a preprocessed flow field dataset. Normalization reduces the magnitude difference between the data. For each sample data set, subtract its mean and divide it by its variance to normalize all data values ​​to between -1 and 1.

[0025] Optionally, data preprocessing operations for the flow field dataset generated by the ground wind tunnel test include: For the data generated by ground wind tunnel tests, since the directly collected wall pressure and image data contain noise, it is necessary to perform denoising and image enhancement operations on the original data. At the same time, frequency change technology and modal extraction methods are used to determine the criteria for distinguishing non-steady-state phenomena, and the data set is annotated according to this standard. Among them, the image enhancement algorithm adopts a cascade processing architecture, which prioritizes high-frequency enhancement and then implements dynamic range expansion. This sequential design has significant advantages: ① Avoiding the amplification of quantization noise in the contrast stretching stage; ② Ensuring that frequency domain operations are performed in an undistorted state of the original grayscale distribution, maintaining the physical consistency of detail enhancement. The specific process is as follows: B1: Grayscale preprocessing. During the initialization phase of the comprehensive enhancement process, the input image is first converted to a single-channel grayscale space. This operation extracts luminance features through weighted fusion of the three RGB channels, eliminating the interference of chrominance information on subsequent frequency domain processing and contrast adjustment, conforming to the typical preprocessing paradigm for grayscale image enhancement.

[0026] B2: Dynamic range expansion based on the cumulative histogram. By constructing a 256-step grayscale histogram statistical distribution, the cumulative distribution function is calculated to determine the nonlinear truncation threshold: ① When setting the lower threshold, the low-frequency end of the histogram is iteratively accumulated until it covers a default 2% of pixels. ② The upper threshold is truncated to a default 90% of pixels using a reverse accumulation method. A piecewise linear transformation is then applied: the truncation interval [lower, upper] is mapped to the full dynamic range [0, 255], with the grayscale levels at both ends saturated to zero. This method effectively eliminates the impact of extreme outliers on contrast and improves overall grayscale uniformity.

[0027] B3: Multi-scale frequency domain decomposition and high-frequency gain: A Gaussian low-pass filter is used to extract low-frequency components. High-frequency details are separated by weighted differences between the original image and the low-frequency components. A linear gain is applied to the high-frequency signal to achieve edge sharpening and texture enhancement.

[0028] B4: Adaptive frequency-domain hybrid enhancement: This strategy linearly superimposes the amplified high-frequency components with the original image at a weight of 0.7, constructing an improved homomorphic filtering model. This strategy significantly improves detail visibility while maintaining the stability of the main structure. Its weight parameters are empirically tuned to balance noise amplification and sharpening intensity.

[0029] In step S3, the present invention selects 80% of the data as a training set; and 20% of the data as a test set.

[0030] Step S4 is specifically as follows: 1. Build a high-performance computing environment based on the NVIDIA GPU cluster, and use the Anaconda virtual environment manager to create an isolated development framework for Python 3.8+ and PyTorch 1.9+. Integrate scientific computing libraries such as NumPy, Pandas, and OpenCV through the conda install command, and load the CUDA 11.1 parallel computing driver and the corresponding version of the cuDNN8.0.5 acceleration library. To ensure hardware compatibility, it is necessary to verify the driver hierarchy relationship: CUDA Toolkit→GPU Driver→PyTorch (CUDA-enabled)→cuDNN dynamic link library, and perform device availability testing through torch.cuda.is_available(). For multi-GPU training scenarios, NCCL (NVIDIA Collective Communications Library) must be additionally configured to achieve cross-GPU gradient synchronization; 2. Set hyperparameters, including the initial learning rate Lr, the piecewise decay strategy learning rate adjustment method, the optimizer SDG, and the batch size of samples input to the model at one time; 3. Read data and input the model to start training; 4. Save the trained weights.

[0031] As a preferred method, refer to Figure 3 As shown, the role of the fast prediction model for lightweight high-speed combustion non-steady-state phenomena in step S4 is to construct a nonlinear mathematical relationship between data sample input and output. The present invention uses a neural network to fit the mathematical relationship between input parameters and output parameters.

[0032] The neural network architecture proposed in this paper is based on multimodal feature fusion and dynamic receptive field control technology, and achieves efficient feature expression and lightweight inference through modular design. The following is a detailed description of the functional modules: (1) Signal expansion module: This module uses a fully convolutional architecture to achieve intelligent mapping from one-dimensional signals to two-dimensional feature maps. Two-stage 1×1 convolutions are used with bilinear upsampling to construct the spatial dimension. The first-layer 1×1 convolution expands the input dimension to 256 dimensions for feature enhancement, and the second-layer 1×1 convolution compresses the channels to the target dimension. Spatial features are refined using 3×3 convolution kernels to avoid the parameter explosion problem caused by traditional fully connected layers.

[0033] (2) Adaptive receptive field module, integrating dilated convolution and axial attention mechanism to construct an adaptive feature extraction unit. 3×3 convolution with dilation rate 2 is used to expand the effective receptive field to 7×7 area, and axial attention is used to achieve feature selection. Axis attention compresses the intermediate dimension to 1 / 4 of the original channel through channel reduction technology, and uses row-column separated convolution kernels for spatial attention calculation.

[0034] (3) The multi-scale fusion module consists of parallel multi-branch depthwise separable convolutions, using square convolution kernels of three different scales (3, 5, and 7) to capture multi-granular features. Each branch shares parameters through grouped convolution, with each group processing only a single channel feature, significantly reducing computational complexity. After feature concatenation, cross-channel information fusion is achieved through 1×1 convolution.

[0035] (4) Progressive upsampling module: A two-stage learnable upsampling architecture is designed, using bilinear interpolation combined with 3×3 convolution for feature refinement. Nonlinear expression capabilities are enhanced through an intermediate channel expansion strategy (number of channels × 4), and finally a 1×1 convolution restores the target dimension. Compared with traditional transposed convolution schemes, this design effectively suppresses the generation of upsampling artifacts.

[0036] (5) The deformable reconstruction module uses dynamic offset prediction combined with deformable convolution to achieve high-precision reconstruction. The deformation parameters of the 3×3 convolution kernel are learned through an 18-channel offset prediction layer, so that the receptive field can be adaptively adjusted according to the feature content. Finally, the feature-to-output mapping is achieved through a single-channel deformable convolution, completing end-to-end signal reconstruction.

[0037] (6) Non-stationary phenomenon discrimination module. This module detects non-stationary features through multi-scale causal convolution and dynamic weighting mechanism. The input time series signal is processed by three parallel causal convolutions, using 3×1 convolution kernels with expansion rates of 1, 2, and 4, respectively, to capture short-, medium-, and long-range time dependency features. After the output of each branch is compressed by global average pooling, it is sent to a two-layer fully connected network to generate dynamic weight coefficients (softmax normalization). Finally, the discrimination result is obtained through weighted fusion. This design innovatively adopts the following strategies: ① Parallel multi-scale convolution realizes cross-cycle detection of non-stationary features; ② Dynamic weight mechanism automatically adjusts branch importance according to signal characteristics; ③ The fully connected layer compresses the intermediate dimension to 64 dimensions to achieve lightweight.

[0038] Purely data-driven neural network algorithms are highly dependent on data, and the supersonic unsteady combustion flow field has strong shock wave interference characteristics and turbulence coupling effects, which inevitably produce non-physical solutions. Under supersonic conditions, unsteady supersonic combustion is not only affected by changes in the flow state of the airflow, but also significantly affected by the dynamic changes of the chemical reaction. The ignition process of the mixed gas in the combustion chamber is a transition from a state without chemical reaction to a state of continuous chemical reaction. At the same time, with the increase of Mach number, the total enthalpy of the incoming flow increases, causing the high-temperature gas effect to gradually become prominent, accompanied by the thermochemical non-equilibrium effect in the ultra-high-speed and lower-pressure airflow. Therefore, we deeply explore the fluid dynamics, chemical reaction kinetics mechanism and shock wave boundary layer theory of supersonic unsteady combustion, establish a complete physical model, and express it in the form of partial differential equations (PDEs).

[0039] The intelligent prediction model of unsteady combustion flow field based on physical information constraints is a typical multi-task learning algorithm. The embedding of physical information is embedded in the loss function in the form of regularization. Therefore, the loss function is shown in the following formula.

[0040]

[0041] represents the prediction loss of the neural network, represents the prediction loss of physical information constraints, Represents the weight balance parameter between the two losses. Corresponding to multi-task learning problems, The reasonable setting of is the key to preventing malicious adversarial optimization between the two tasks.

[0042] Step S5 is specifically as follows: refer to Figure 4 As shown, the edge computing platform is an NPU module and its supporting LED display module. In addition, since the convolutional neural network belongs to a black box model, the convolution calculation process is essentially a matrix operation. In order to increase the interpretability of the convolutional neural network model, the feature layer of the convolutional neural network is visualized using the matplotlib database and the LED display module to display the calculation process of the convolutional neural network and compare the visualized feature map with the flow field characteristics. The process of the convolutional neural network is analyzed for interpretability from a macroscopic perspective and physical principles. In this way, it is convenient for those skilled in the art to adjust the relevant parameters of the convolutional neural network in a timely manner according to the display and interpretability analysis results to obtain a more optimized supersonic combustion chamber flow field reconstruction result. Finally, the non-steady-state real-time monitoring is achieved through the non-steady-state phenomenon discrimination module, and the monitoring signal is transmitted back to the engine.

[0043] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the invention.

Claims

1. A high-speed combustion non-steady-state prediction method based on lightweight and interpretable deep learning, characterized by: The following steps are involved: S1: Using ground wind tunnel tests and CFD numerical simulation software, a combustion chamber flow field dataset under different incoming flow conditions is constructed; S2: performing data preprocessing on the combustion chamber configuration flow field data set to obtain a preprocessed flow field data set; S3: Shuffle the preprocessed flow field dataset and divide it into training set and test set in proportion; S4: Using the training set, the fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion is trained to obtain a trained fast prediction model for the unsteady-state phenomenon of lightweight high-speed combustion and training weights; S5: Embed the test set, the trained lightweight high-speed combustion non-steady-state phenomenon rapid prediction model and the training weights into the edge computing platform to obtain the prediction results of the non-steady-state phenomenon.

2. The method for predicting high-speed combustion non-steady-state based on lightweight interpretable deep learning according to claim 1, characterized in that: Said S2 includes performing data preprocessing operations on the flow field data set generated by the numerical simulation and performing data preprocessing operations on the flow field data set generated by the ground wind tunnel test; The data preprocessing operation on the flow field data set generated by the numerical simulation includes the following steps: A1: Obtain the time coordinates and space coordinates of the flow field boundary grid points; A2: Use the Latin hypercube sampling method to obtain the time coordinates and spatial coordinates of the residual points at the flow field boundary; A3: Obtain the time points and spatial coordinates of the flow field boundary and the flow field within the boundary area; A4: Align the time points and spatial coordinates of each pair of sample data to obtain aligned sample data. Each pair of sample data includes the boundary grid points, residual points, and the flow field within the boundary and boundary region at the same time. The boundary grid points and residual points at the same time are used as input, and the flow field within the boundary and boundary region is used as output. A5: Standardize each aligned sample data to obtain the pre-processed flow field data set generated by the numerical simulation; The data preprocessing operation is performed on the flow field data set generated by the ground wind tunnel test, The following steps are included: B1: Convert the input image into a single-channel grayscale space through the weighted fusion method of RGB three channels to achieve grayscale preprocessing; B2: Dynamic range expansion based on cumulative histogram. By constructing a 256-level grayscale histogram statistical distribution, calculating the cumulative distribution function to determine the nonlinear truncation threshold, and performing a piecewise linear transformation, the truncation interval [lower, upper] is mapped to the full dynamic range [0, 255], and the grayscale levels at both ends are saturated to zero. B3: Use a Gaussian low-pass filter to extract low-frequency components, separate high-frequency details through weighted differences between the original image and the low-frequency components, and apply a linear gain to the high-frequency signal; B4: Perform adaptive frequency domain hybrid enhancement, linearly superimpose the high-frequency component after gain with the original image with a weight of 0.7, and construct an improved homomorphic filtering model.

3. The method for predicting high-speed combustion non-steady-state based on lightweight interpretable deep learning according to claim 2, characterized in that: The Latin hypercube sampling method is: C1: Determine the number of samples N; C2: Divide the data area into M sub-areas; C3: Randomly sample N / M data in each sub-region; C4: Combine the data obtained by random sampling in different sub-areas to obtain the overall sampling result.

4. The method for predicting high-speed combustion non-steady-state based on lightweight interpretable deep learning according to claim 1, characterized in that: The fast prediction model for the non-steady-state phenomenon of lightweight high-speed combustion in S4 includes: The signal expansion module uses a fully convolutional architecture to achieve intelligent mapping of one-dimensional signals to two-dimensional feature maps. It uses two-stage 1×1 convolutions with bilinear upsampling to construct spatial dimensions. The first-layer 1×1 convolution expands the input dimension to 256 dimensions for feature enhancement, and the second-layer 1×1 convolution compresses the channels to the target dimension. 3×3 convolutions are used to refine spatial features, avoiding the parameter explosion problem caused by traditional fully connected layers. The adaptive receptive field module integrates dilated convolution and axial attention mechanisms to construct an adaptive feature extraction unit. It uses 3×3 convolution to expand the effective receptive field to a 7×7 area and cooperates with the axial attention module to achieve feature selection. The axial attention module uses channel reduction technology to compress the intermediate dimension to 1 / 4 of the original channel and uses row-column separated convolution kernels for spatial attention calculation. The multi-scale fusion module consists of parallel multi-branch depthwise separable convolutions. It uses square convolution kernels of three different scales (3, 5, and 7) to capture multi-granularity features. Each branch shares parameters through grouped convolution. Each group processes only a single channel feature. After feature concatenation, cross-channel information fusion is achieved through 1×1 convolution. The progressive upsampling module uses a two-stage learnable upsampling architecture, employs bilinear interpolation with 3×3 convolution for feature refinement, and uses 1×1 convolution to restore the target dimension; The deformable reconstruction module uses dynamic offset prediction combined with deformable convolution to achieve high-precision reconstruction. The 18-channel offset prediction layer learns the deformation parameters of the 3×3 convolution kernel, enabling the receptive field to adaptively adjust based on feature content. Finally, single-channel deformable convolution is used to map features to outputs, completing end-to-end signal reconstruction. The non-stationary phenomenon discrimination module realizes non-stationary feature detection through multi-scale causal convolution and dynamic weighting mechanism. The input time series signal is processed by parallel three-way causal convolution, using 3×1 convolution kernels with expansion rates of 1, 2 and 4 respectively to capture short-, medium- and long-range time-dependent features; the output of each branch is compressed by global average pooling in the spatiotemporal dimension, and then sent to a two-layer fully connected network to generate dynamic weight coefficients, and finally the discrimination result is obtained through weighted fusion.

5. The method for predicting high-speed combustion non-steady-state based on lightweight interpretable deep learning according to claim 1, characterized in that: The edge computing platform in the S5 is the NPU module and its matching LED display module.

Images