A Method and System for Predicting Engine Vibration Response Based on Conditional Frequency Domain Diffusion Model

Abstract

This invention discloses a method and system for predicting engine vibration response based on a conditional frequency domain diffusion model. The method acquires the vibration signal of a source condition and the operating parameters of a target condition from a hot-fire test of a liquid rocket engine. It then performs frequency domain transformation on the vibration signal to obtain the PSD of the source condition. The operating parameters of the target condition are input into a pre-trained RMS prediction network to obtain the RMS value of the target condition PSD. Finally, the PSD of the source condition, the operating parameters of the target condition, and the RMS value are input into a pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target condition.

Description

Technical Field

[0001] This invention relates to the field of vibration response prediction technology for liquid rocket engines, specifically to a method and system for predicting engine vibration response based on a conditional frequency domain diffusion model. Background Technology

[0002] Liquid rocket engines typically undergo dynamic environment testing during the development phase to verify their functionality and adaptability under complex dynamic conditions. In such tests, component-level vibration excitation is often based on the vibration response obtained from engine-level hot-fire tests. However, hot-fire tests only cover a portion of the operating conditions; the vibration response under untested conditions remains unknown, leading to difficulties in defining excitation. This not only weakens the effectiveness of dynamic environment testing but also restricts engine design optimization and risk management. Therefore, it is necessary to predict the vibration response under untested conditions based on a limited number of tested conditions, providing a reliable basis for excitation design in component-level tests.

[0003] Currently, rocket engine vibration response prediction technologies are mainly divided into two categories: model-driven and data-driven. Model-driven methods rely on high-precision finite element models and can obtain the response at any location of the structure through mesh generation. However, as the upper limit of the analysis frequency increases, the increased mesh density and higher requirements for structural detail lead to a significant increase in computational cost and a decrease in analysis accuracy. Furthermore, it is difficult to consider the deviations caused by different vibration transmission paths due to the coupling of multiple components in the target structure during modeling. In contrast, data-driven methods do not require building a finite element model of the target; instead, they directly start from experimental data and learn the mapping relationship between operating parameters and response characteristics to predict the structural vibration response, thus becoming another important development direction in this field.

[0004] Existing data-driven vibration response prediction methods primarily focus on predicting the root mean square (RMS) value of the vibration signal. However, RMS compresses the energy across the entire frequency band into a single scalar, making it difficult to reflect key frequency domain characteristics such as resonance peaks, and thus failing to identify and mitigate potential resonance risks in a timely manner. With the rapid iteration of liquid rocket engine models, the continuous expansion of their operating range, and the increased system coupling, prediction based solely on RMS is no longer sufficient to meet engineering requirements. Therefore, it is necessary to shift towards refined modeling in the frequency domain, accurately reconstructing the energy distribution and peak structure at the power spectral density (PSD) level, thereby providing a more discriminative and operable basis for the excitation spectrum design of new engine models. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting engine vibration response based on a conditional frequency domain diffusion model, so as to solve the technical problem that the existing technology is limited to predicting the total energy of a single vibration response and is difficult to avoid potential resonance risks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The method for predicting engine vibration response based on a conditional frequency domain diffusion model includes the following steps: Step 1: Obtain the vibration signal of the source condition and the operating parameters of the target condition of the liquid rocket engine hot test, and perform frequency domain transformation calculation on the vibration signal to obtain the PSD of the source condition; Step 2: Input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the target working condition PSD; Step 3: Input the PSD of the source condition, the condition parameters of the target condition, and the RMS value obtained in Step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target condition.

[0007] Furthermore, in step 1, the Welch method is used to perform frequency domain transformation calculation on the vibration signal of the source condition. The Welch method obtains the PSD by segmenting the vibration signal, windowing it, calculating the discrete Fourier transform of each segment, and then averaging the results. The calculation formula is as follows:

[0008] In the formula, For the first The Fourier transform result of the segment signal indicates its frequency. Spectral amplitude at that location For the length of each segment, This represents the number of segments.

[0009] Furthermore, the method for constructing the RMS predictive network described in step 2 includes the following steps: Step 2-1: Perform energy integration on the PSD of the source operating condition to obtain the RMS value, and divide the RMS value and its corresponding operating condition parameters into training set and test set according to the ratio; Step 2-2: Calculate the correlation coefficient between the RMS value and the operating condition parameter in the training set described in Step 2-1, filter out the operating condition parameters with a correlation greater than 0.6 with RMS, and replace the corresponding parameters in the original training set and test set with the filtered operating condition parameters to obtain the updated training set and test set. Step 2-3: Construct a multi-scale convolutional network. Train the multi-scale convolutional network using the updated training set described in Step 2-2, and verify the prediction accuracy of the multi-scale convolutional network using the updated test set. Once the preset error limit is reached, the construction of the RMS prediction network is complete.

[0010] Furthermore, in step 2-1, the formula for calculating the energy integral is:

[0011] In the formula, The PSD of the vibration signal For frequency variables, This is the lowest frequency for PSD. This is the highest frequency of the PSD. In step 2-1, according to the proportion p The source working condition sample data is divided into training set and test set. Each sample data includes the working condition parameters of a certain working condition and the RMS value of the PSD of that working condition. In step 2-2, the correlation coefficient is the Pearson correlation coefficient, and the calculation formula is:

[0012]

[0013] In the formula, This represents the total number of samples in the training set. Let be a certain operating condition parameter in the i-th sample. This represents the average value of the parameters for this operating condition across all n samples. Let RMS be the value of the i-th sample. This is the average RMS value across all n samples.

[0014] Furthermore, in steps 2-3, the RMS predictive network adopts a multi-scale convolutional structure, and its construction process includes the following steps: Step 2-3-1: Using operating parameters as input, perform feature encoding through a multilayer perceptron to obtain embedded features; Step 2-3-2: Perform multi-scale convolution processing on the embedded features, use parallel convolution with different kernel sizes to extract local and global information, and fuse them to obtain the basic features; Step 2-3-3: Perform multi-scale convolution on the basic features again to extract higher-level feature information and generate enhanced features; Steps 2-3-4: Input the enhanced features into the multilayer perceptron for dimensionality reduction mapping, and output the predicted RMS value of the target working condition; Steps 2-3-5: During training, the mean squared error is used as the loss function to minimize the deviation between the predicted RMS value and the true RMS value; the specific formula is:

[0015]

[0016]

[0017]

[0018]

[0019] In the formula, These are the operating parameters. The embedded features are obtained by feature encoding of the operating parameters using a multilayer perceptron. Parallel convolutions with different kernel sizes, Basic features, To enhance features, For the predicted RMS value, This is the true RMS value. Let the mean squared error loss function be . This represents the number of samples.

[0020] Furthermore, the method for constructing the conditional frequency domain diffusion model described in step 3 includes the following steps: Step 3-1: Select the PSD of a certain working condition from the source working condition as the reference working condition PSD, and select the PSD of another working condition as the predicted working condition PSD. Combine the working condition parameters and RMS value corresponding to the predicted working condition to form a sample. Then divide the sample into training set and test set according to the ratio p. Step 3-2: Construct a conditional frequency domain diffusion network. Train the conditional frequency domain diffusion network using the training set described in Step 3-1, and verify the prediction accuracy of the conditional frequency domain diffusion network using the test set. Once the preset error limit is reached, the construction of the conditional frequency domain diffusion model is complete.

[0021] Further, step 3-2, training the conditional frequency domain spread network using the training set, includes the following steps: Step 3-2-1: Inject Gaussian noise into the PSD of the predicted operating condition according to the time step t of the diffusion process, and obtain... And record the corresponding real Gaussian noise; Step 3-2-2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The PSDs of the reference working conditions are spliced ​​together along the channel dimension to form network input features, and the network input features are then input into the denoising U-Net network.

[0022] Step 3-2-3: In the attention block of the denoising U-Net network, the working condition parameters and RMS values ​​of the predicted working condition are introduced as conditional information through the cross-attention mechanism; Step 3-2-4: The predicted noise is output from the denoised U-Net network, and the mean square error loss between the predicted noise and the real noise is calculated for training.

[0023] Further, in step 3-2-1, the time step t is a preset discrete diffusion step number used to control the noise injection intensity of the current sample. The forward diffusion process progresses from t=0 to t=T, where... This represents the maximum time step in the diffusion process, during which the original spectrum gradually approaches the noise pattern of a standard normal distribution; The Gaussian noise in step 3-2-1 is standard normal noise with zero mean and unit covariance. Noise injection employs parametric Markov forward diffusion, with variance scheduling coefficients set. Where t represents the time step of the diffusion process, and the signal preservation coefficient at the t-th time step is defined. and their cumulative product ,in This represents the signal hold coefficient at the s-th time step, given the predicted operating condition PSD as follows: That is, generating noisy samples at any time step. ; The denoising U-Net network in step 3-2-3 consists of a five-layer encoder, a bottleneck layer, and a five-layer decoder. Each layer of the encoder consists of a downsampling unit, a residual block, and an attention block. The bottleneck layer consists of a residual block and an attention block. Each layer of the decoder consists of an upsampling unit, a residual block, and an attention block. The downsampling unit is a one-dimensional convolution with a stride of 2, the upsampling unit is a deconvolution with a stride of 2, the residual block consists of a convolution unit and residual connections, used to change the number of channels and introduce diffusion time step information to modulate the feature distribution. The attention block includes self-attention and cross-attention. Self-attention calculates the features of this layer, and cross-attention calculates the features of this layer and the conditional information of the prediction condition. The formula for calculating cross attention in step 3-2-3 is as follows:

[0024] In the formula, For the current layer features, The conditional feature vector is obtained by feature encoding from the predicted working condition parameters and the predicted working condition RMS value. , , These are the query vector, key vector, and value vector, respectively. , , It is a learnable linear projection matrix. is the feature dimension of the key vector.

[0025] Furthermore, in step 3, when generating the PSD vibration response of the target working condition, the conditional frequency domain diffusion model gradually reconstructs the target PSD from pure noise. At each time step, it predicts and denoises the noise component based on the current sample and condition information, while adding proportionally scaled random noise to maintain generation diversity. After multiple iterations, the noise component is gradually removed, making the generated PSD gradually approach the true spectrum. The specific formula is as follows:

[0026] In the formula, The sample after denoising. For the current noisy sample, For the noise predicted by the model, For random noise that follows a standard normal distribution, The variance scheduling coefficient is... , .

[0027] An engine vibration response prediction system based on a conditional frequency domain diffusion model includes: The first module is used to acquire the vibration signal of the source condition and the operating parameters of the target condition of the hot test of the liquid rocket engine, and to perform frequency domain transformation calculation on the vibration signal to obtain the power spectral density (PSD) of the source condition. The second module is used to input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the PSD of the target working condition. The third module is used to input the PSD of the source working condition, the working parameters of the target working condition, and the RMS value obtained in step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target working condition.

[0028] Compared with the prior art, the present invention has the following beneficial technical effects: 1) This invention proposes a PSD prediction method based on RMS prior constraints. The RMS of the target working condition PSD is obtained in advance through an independent RMS prediction network and injected into the diffusion model as a prior condition. Energy constraints are applied to the generated full-band PSD, which effectively reduces energy leakage and overall error, and makes the predicted spectrum highly consistent with the true spectrum in terms of total energy.

[0029] 2) The conditional frequency domain diffusion network constructed in this invention introduces a conditional diffusion model in the frequency domain, uses operating parameters and RMS values ​​as conditional information, and uses a stepwise denoising mechanism to characterize the sparse and high-amplitude spectral peak distribution in PSD. Compared with the direct regression method, it significantly alleviates the overfitting problem in the peak region and the underfitting problem in other frequency bands, and improves the accuracy and reliability of spectral peak position and amplitude prediction.

[0030] 3) This invention proposes an engine vibration response prediction method based on a conditional frequency domain diffusion model. It can predict the vibration PSD of untested conditions by relying only on the vibration response of a limited number of source conditions and the condition parameters of the target condition. This provides a basis for defining the excitation of component-level dynamic tests, reduces the number of hot tests and vibration tests while ensuring safety margins, and lowers test costs and development risks.

[0031] 4) The vibration response prediction method constructed in this invention achieves an effective balance between overall energy consistency and fine spectrum reconstruction. The RMS prior guides the denoising direction of the diffusion network through global energy constraints, suppressing numerical divergence. Meanwhile, the diffusion network's characterization of local features refines the energy constraints into specific spectral peak distributions. The two are coupled with each other, improving the accuracy of broadband prediction results for untested operating conditions under limited samples. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram of the conditional frequency domain diffusion model. Figure 3 This is a schematic diagram illustrating a specific implementation process of an embodiment of the present invention; Figure 4 This is a predictive result of the PSD response of various engine components under a certain operating condition. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] Example 1 This invention provides a method for predicting engine vibration response based on a conditional frequency domain diffusion model, comprising the following steps: Step 1: Obtain the vibration signal of the source condition and the operating parameters of the target condition of the liquid rocket engine hot test, and perform frequency domain transformation calculation on the vibration signal to obtain the power spectral density (PSD) of the source condition. Step 2: Input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the target working condition PSD; Step 3: Input the PSD of the source condition, the condition parameters of the target condition, and the RMS value obtained in Step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target condition.

[0037] This invention can predict the vibration PSD of untested conditions by relying only on the vibration response of a limited number of source conditions and the condition parameters of the target condition. It provides a basis for defining the excitation of component-level dynamic tests, reduces the number of hot runs and vibration tests while ensuring safety margins, and lowers test costs and development risks.

[0038] Example 2 This invention provides a method for predicting engine vibration response based on a conditional frequency domain diffusion model. (See also...) Figure 1 This includes the following steps: Step 1: Obtain the vibration signal of the source condition and the operating parameters of the target condition of the liquid rocket engine hot test, and perform frequency domain transformation calculation on the vibration signal to obtain the power spectral density (PSD) of the source condition. Step 2: Input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the target working condition PSD; Step 3: Input the PSD of the source condition, the condition parameters of the target condition, and the RMS value obtained in Step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target condition.

[0039] As a preferred embodiment of the present invention, the process of frequency domain transformation calculation of the vibration signal in step 1 is as follows: The Welch method is used to perform frequency domain transformation calculation on the source working condition vibration signal. The Welch method obtains the PSD by segmenting the signal, windowing it, calculating the discrete Fourier transform of each segment, and then averaging the results. The calculation formula is as follows:

[0040] In the formula, For the first The Fourier transform result of the segment signal indicates its frequency. Spectral amplitude at that location For the length of each segment, This represents the number of segments.

[0041] In a preferred embodiment of the present invention, step 2, the method for constructing the RMS predictive network includes the following steps: Step 2-1: Perform energy integration on the PSD of the source operating condition to obtain the RMS value, and divide the RMS value and its corresponding operating condition parameters into training set and test set according to the ratio; Step 2-2: Calculate the correlation coefficient between the RMS value and the operating condition parameter in the training set described in Step 2-1, filter out the operating condition parameters with a correlation greater than 0.6 with RMS, and replace the corresponding parameters in the original training set and test set with the filtered operating condition parameters to obtain the updated training set and test set. Step 2-3: Construct a multi-scale convolutional network. Train the multi-scale convolutional network using the updated training set described in Step 2-2, and verify the prediction accuracy of the multi-scale convolutional network using the updated test set. Once the preset error limit is reached, the construction of the RMS prediction network is completed.

[0042] In a preferred embodiment of the present invention, in step 2-1, the formula for calculating the energy integral is:

[0043] In the formula, The PSD of the vibration signal For frequency variables, This is the lowest frequency for PSD. This is the highest frequency of PSD.

[0044] As a preferred embodiment of the present invention, in step 2-1, according to the proportion p The source working condition sample data is divided into training set and test set. Each sample data includes the working condition parameters of a certain working condition and the RMS value of the PSD of that working condition.

[0045] In a preferred embodiment of the present invention, in step 2-2, the correlation coefficient is the Pearson correlation coefficient, and the calculation formula is:

[0046]

[0047] In the formula, This represents the total number of samples in the training set. Let be a certain operating condition parameter in the i-th sample. This represents the average value of the parameters for this operating condition across all n samples. Let RMS be the value of the i-th sample. This is the average RMS value across all n samples.

[0048] In a preferred embodiment of the present invention, in steps 2-3, the RMS predictive network adopts a multi-scale convolutional structure, and its construction process includes the following steps: Step 2-3-1: Using operating parameters as input, perform feature encoding through a multilayer perceptron to obtain embedded features; Step 2-3-2: Perform multi-scale convolution processing on the embedded features, use parallel convolution with different kernel sizes to extract local and global information, and fuse them to obtain the basic features; Step 2-3-3: Perform multi-scale convolution on the basic features again to extract higher-level feature information and generate enhanced features; Steps 2-3-4: Input the enhanced features into the multilayer perceptron for dimensionality reduction mapping, and output the predicted RMS value of the target working condition; Steps 2-3-5: During training, the mean squared error is used as the loss function to minimize the deviation between the predicted RMS and the true RMS; the specific formula is:

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, These are the operating parameters. The embedded features are obtained by feature encoding of the operating parameters using a multilayer perceptron. Parallel convolutions with different kernel sizes, Basic features, To enhance features, For the predicted RMS value, This is the true RMS value. Let the mean squared error loss function be . This represents the number of samples.

[0054] In a preferred embodiment of the present invention, step 3, the method for constructing the conditional frequency domain diffusion model includes the following steps: Step 3-1: Select the PSD of a certain working condition from the source working condition as the reference working condition PSD, and select the PSD of another working condition as the predicted working condition PSD. Combine the working condition parameters and RMS value corresponding to the predicted working condition to form a sample. Then divide the sample into training set and test set according to the ratio p. Step 3-2: Construct a conditional frequency domain diffusion network. Train the network using the training set described in Step 3-1, and verify the prediction accuracy of the network using the test set. Once the preset error limit is reached, the construction of the conditional frequency domain diffusion model is complete.

[0055] As a preferred embodiment of the present invention, the training set in step 3-1 contains the same working conditions as the training set in step 2-1, and the test set is also consistent.

[0056] As a preferred embodiment of the present invention, training the conditional frequency domain diffusion network described in step 3-2 includes the following steps: Step 3-2-1: Inject Gaussian noise into the PSD of the predicted operating condition according to the time step t of the diffusion process, and obtain... And record the corresponding real Gaussian noise; Step 3-2-2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The reference working condition PSD is spliced ​​along the channel dimension to form the network input feature, and the network input feature is then input into the denoising U-Net network.

[0057] Step 3-2-3: In the attention block of the denoising U-Net network, the working condition parameters and RMS values ​​of the predicted working condition are introduced as conditional information through the cross-attention mechanism; Step 3-2-4: The predicted noise is output from the denoised U-Net network, and the mean square error loss between the predicted noise and the real noise is calculated for training.

[0058] In a preferred embodiment of the present invention, in step 3-2-1, the time step t is a preset discrete diffusion step number used to control the noise injection intensity of the current sample. The forward diffusion process progresses from t=0 to t=T, where... This represents the maximum time step in the diffusion process, during which the original spectrum gradually approaches a standard normally distributed noise pattern. Gaussian noise is standard normally distributed noise with zero mean and unit covariance. Noise injection employs parametric Markov forward diffusion, with variance scheduling coefficients set. Where t represents the time step of the diffusion process, and the signal preservation coefficient at the t-th time step is defined. and their cumulative product ,in This represents the signal hold coefficient at the s-th time step, given the predicted operating condition PSD as follows: That is, generating noisy samples at any time step. .

[0059] In a preferred embodiment of the present invention, in step 3-2-3, the denoising U-Net network consists of a five-layer encoder, a bottleneck layer, and a five-layer decoder. Each layer of the encoder consists of a downsampling unit, a residual block, and an attention block; the bottleneck layer consists of a residual block and an attention block; and each layer of the decoder consists of an upsampling unit, a residual block, and an attention block. The downsampling unit is a one-dimensional convolution with a stride of 2, the upsampling unit is a deconvolution with a stride of 2, the residual block consists of a convolutional unit and residual connections, used to change the number of channels and introduce diffusion time step information to modulate the feature distribution; the attention block includes self-attention and cross-attention, the self-attention calculates the features of the layer, and the cross-attention calculates the features of the layer and the conditional information of the prediction condition.

[0060] In a preferred embodiment of the present invention, in step 3-2-3, the formula for calculating cross-attention is:

[0061] In the formula, For the current layer features, The conditional feature vector is obtained by feature encoding from the predicted working condition parameters and the predicted working condition RMS value. , , These are the query vector, key vector, and value vector, respectively. , , It is a learnable linear projection matrix. is the feature dimension of the key vector.

[0062] In a preferred embodiment of the present invention, when generating the PSD vibration response of the target working condition in step 3, the conditional frequency domain diffusion model gradually reconstructs the target PSD from pure noise. At each time step, the noise component is predicted and denoised based on the current sample and condition information. At the same time, proportionally scaled random noise is added to maintain the diversity of generation. After multiple iterations, the noise component is gradually removed, so that the generated PSD gradually approaches the true spectrum. The specific formula is as follows:

[0063] In the formula, The sample after denoising. For the current noisy sample, For the noise predicted by the model, For random noise that follows a standard normal distribution, The variance scheduling coefficient is... , .

[0064] Example 3 To better illustrate the technical effects of the present invention, a specific embodiment is used to experimentally verify the invention. See also... Figure 3 This embodiment is based on measured vibration data from a hot-run test of a certain type of engine. The data used involves vibration response signals of three key engine components—the thrust chamber, gas generator, and kerosene pump—under various typical operating conditions. Each component is equipped with sensors in three directions: axial, radial, and tangential, with a sampling frequency of 25600 Hz.

[0065] This embodiment compares the present invention with four other state-of-the-art sequence prediction models. The details of the comparison are as follows: 1) CNN+MLP: CNN is used to extract local features of PSD sequences, and MLP is combined to model global nonlinear relationships, so as to realize multi-scale feature representation while ensuring computational efficiency.

[0066] 2) CWGAN-GP: It adopts conditional Wasserstein GAN to introduce adversarial learning, which improves the realism of the PSD generation distribution; the gradient penalty mechanism enhances training stability and improves convergence performance.

[0067] 3) cVAE: It realizes the uncertainty expression of PSD prediction through probabilistic modeling of the latent space, and improves the continuity and controllability of the generated results by reparameterization techniques.

[0068] 4) Transformer: Based on the self-attention mechanism, it models the global dependencies of PSD sequences. Multi-head attention improves the ability to model long sequence features and the accuracy of prediction.

[0069] Under the same training conditions, the average RMS error of each model is shown in Table 1.

[0070] Table 1. Predicted Vibration Response Errors of Various Engine Components

[0071] As can be seen from the data in the table, the method of this invention achieves the lowest RMS error in all directions for prediction of the three types of components. In the propulsion chamber, the axial, radial, and tangential errors are 0.487 dB, 0.291 dB, and 0.168 dB, respectively, significantly lower than the other methods (all greater than 0.7 dB). In the gas generator, the three-dimensional errors of the method of this invention are 0.749 dB, 0.513 dB, and 0.253 dB, while the comparative methods are generally in the range of 1.7–3.8 dB. In the kerosene pump, the method of this invention has axial and tangential errors of 0.379 dB and 0.280 dB, respectively, and radial error of 0.808 dB, still outperforming the comparative methods overall. The results show that the method significantly reduces the PSD prediction error in different components and directions.

[0072] Example 4 The present invention also provides an engine vibration response prediction system based on a conditional frequency domain diffusion model, comprising: The first module is used to acquire the vibration signal of the source condition and the operating parameters of the target condition of the hot test of the liquid rocket engine, and to perform frequency domain transformation calculation on the vibration signal to obtain the power spectral density (PSD) of the source condition. The second module is used to input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the PSD of the target working condition. The third module is used to input the PSD of the source working condition, the working parameters of the target working condition, and the RMS value obtained in step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target working condition.

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A method for predicting engine vibration response based on a conditional frequency domain diffusion model, characterized in that, Includes the following steps: Step 1: Obtain the vibration signal of the source condition and the operating parameters of the target condition of the liquid rocket engine hot test, and perform frequency domain transformation calculation on the vibration signal to obtain the PSD of the source condition; Step 2: Input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the target working condition PSD; Step 3: Input the PSD of the source condition, the condition parameters of the target condition, and the RMS value obtained in Step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target condition.

2. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 1, characterized in that, In step 1, the Welch method is used to perform frequency domain transformation calculation on the vibration signal of the source condition. The Welch method obtains the PSD by segmenting the vibration signal, windowing it, calculating the discrete Fourier transform of each segment, and then averaging the results. The calculation formula is as follows: In the formula, For the first The Fourier transform result of the segment signal indicates its frequency. Spectral amplitude at that location, For the length of each segment, This represents the number of segments.

3. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 1, characterized in that, The method for constructing the RMS predictive network described in step 2 includes the following steps: Step 2-1: Perform energy integration on the PSD of the source operating condition to obtain the RMS value, and divide the RMS value and its corresponding operating condition parameters into training set and test set according to the ratio; Step 2-2: Calculate the correlation coefficient between the RMS value and the operating condition parameter in the training set described in Step 2-1, filter out the operating condition parameters with a correlation greater than 0.6 with RMS, and replace the corresponding parameters in the original training set and test set with the filtered operating condition parameters to obtain the updated training set and test set. Step 2-3: Construct a multi-scale convolutional network. Train the multi-scale convolutional network using the updated training set described in Step 2-2, and verify the prediction accuracy of the multi-scale convolutional network using the updated test set. Once the preset error limit is reached, the construction of the RMS prediction network is complete.

4. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 3, characterized in that, In step 2-1, the formula for calculating the energy integral is: In the formula, The PSD of the vibration signal For frequency variables, This is the lowest frequency for PSD. This is the highest frequency of the PSD. In step 2-1, according to the proportion p The source working condition sample data is divided into training set and test set. Each sample data includes the working condition parameters of a certain working condition and the RMS value of the PSD of that working condition. In step 2-2, the correlation coefficient is the Pearson correlation coefficient, and the calculation formula is: In the formula, This represents the total number of samples in the training set. Let be a certain operating condition parameter in the i-th sample. This represents the average value of the parameters for this operating condition across all n samples. Let RMS be the value of the i-th sample. This is the average of the RMS values ​​across all n samples.

5. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 3, characterized in that, In steps 2-3, the RMS predictive network adopts a multi-scale convolutional structure, and its construction process includes the following steps: Step 2-3-1: Using operating parameters as input, perform feature encoding through a multilayer perceptron to obtain embedded features; Step 2-3-2: Perform multi-scale convolution processing on the embedded features, use parallel convolution with different kernel sizes to extract local and global information, and fuse them to obtain the basic features; Step 2-3-3: Perform multi-scale convolution on the basic features again to extract higher-level feature information and generate enhanced features; Steps 2-3-4: Input the enhanced features into the multilayer perceptron for dimensionality reduction mapping, and output the predicted RMS value of the target working condition; Steps 2-3-5: During training, the mean squared error is used as the loss function to minimize the deviation between the predicted RMS value and the true RMS value; the specific formula is: In the formula, These are the operating parameters. The embedded features are obtained by feature encoding of the operating parameters using a multilayer perceptron. Parallel convolutions with different kernel sizes, Basic features, To enhance features, For the predicted RMS value, This is the true RMS value. Let the mean squared error loss function be . This represents the number of samples.

6. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 1, characterized in that, The method for constructing the conditional frequency domain diffusion model described in step 3 includes the following steps: Step 3-1: Select the PSD of a certain working condition from the source working condition as the reference working condition PSD, and select the PSD of another working condition as the predicted working condition PSD. Combine the working condition parameters and RMS value corresponding to the predicted working condition to form a sample. Then divide the sample into training set and test set according to the ratio p. Step 3-2: Construct a conditional frequency domain diffusion network. Train the conditional frequency domain diffusion network using the training set described in Step 3-1, and verify the prediction accuracy of the conditional frequency domain diffusion network using the test set. Once the preset error limit is reached, the construction of the conditional frequency domain diffusion model is complete.

7. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 6, characterized in that, Step 3-2, which involves training the conditional frequency domain spread network using the training set, includes the following steps: Step 3-2-1: Inject Gaussian noise into the PSD of the predicted operating condition according to the time step t of the diffusion process, and obtain... And record the corresponding real Gaussian noise; Step 3-2-2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation The PSDs of the reference working conditions are spliced ​​together along the channel dimension to form network input features, and the network input features are then input into the denoising U-Net network. Step 3-2-3: In the attention block of the denoising U-Net network, the working condition parameters and RMS values ​​of the predicted working condition are introduced as conditional information through the cross-attention mechanism; Step 3-2-4: The predicted noise is output from the denoised U-Net network, and the mean square error loss between the predicted noise and the real noise is calculated for training.

8. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 7, characterized in that, In step 3-2-1, the time step t is a preset discrete diffusion step number used to control the noise injection intensity of the current sample. The forward diffusion process progresses from t=0 to t=T, where... This represents the maximum time step in the diffusion process, during which the original spectrum gradually approaches the noise pattern of a standard normal distribution; The Gaussian noise in step 3-2-1 is standard normal noise with zero mean and unit covariance. Noise injection employs parametric Markov forward diffusion, with variance scheduling coefficients set. Where t represents the time step of the diffusion process, and the signal preservation coefficient at the t-th time step is defined. and their cumulative product ,in This represents the signal hold coefficient at the s-th time step, given the predicted operating condition PSD as follows: That is, generating noisy samples at any time step. ; The denoising U-Net network in step 3-2-3 consists of a five-layer encoder, a bottleneck layer, and a five-layer decoder. Each layer of the encoder consists of a downsampling unit, a residual block, and an attention block. The bottleneck layer consists of a residual block and an attention block. Each layer of the decoder consists of an upsampling unit, a residual block, and an attention block. The downsampling unit is a one-dimensional convolution with a stride of 2, the upsampling unit is a deconvolution with a stride of 2, the residual block consists of a convolution unit and residual connections, used to change the number of channels and introduce diffusion time step information to modulate the feature distribution. The attention block includes self-attention and cross-attention. Self-attention calculates the features of this layer, and cross-attention calculates the features of this layer and the conditional information of the prediction condition. The formula for calculating cross attention in step 3-2-3 is as follows: In the formula, For the current layer features, The conditional feature vector is obtained by feature encoding from the predicted working condition parameters and the predicted working condition RMS value. , , These are the query vector, key vector, and value vector, respectively. , , It is a learnable linear projection matrix. is the feature dimension of the key vector.

9. The engine vibration response prediction method based on the conditional frequency domain diffusion model according to claim 8, characterized in that, In step 3, when generating the PSD vibration response of the target working condition, the conditional frequency domain diffusion model gradually reconstructs the target PSD from pure noise. At each time step, the noise component is predicted and denoised based on the current sample and condition information. At the same time, proportionally scaled random noise is added to maintain the diversity of generation. After multiple iterations, the noise component is gradually removed, so that the generated PSD gradually approaches the true spectrum. The specific formula is as follows: In the formula, The sample after denoising. For the current noisy sample, For the noise predicted by the model, For random noise that follows a standard normal distribution, The variance scheduling coefficient is... , .

10. An engine vibration response prediction system based on a conditional frequency domain diffusion model, characterized in that, include: The first module is used to acquire the vibration signal of the source condition and the operating parameters of the target condition of the hot test of the liquid rocket engine, and to perform frequency domain transformation calculation on the vibration signal to obtain the power spectral density (PSD) of the source condition. The second module is used to input the working condition parameters of the target working condition into the pre-trained RMS prediction network to obtain the RMS value of the PSD of the target working condition. The third module is used to input the PSD of the source working condition, the working parameters of the target working condition, and the RMS value obtained in step 2 into the pre-trained conditional frequency domain diffusion model to generate the PSD vibration response of the target working condition.

Images