An engine vibration prediction method, system, device and medium

Abstract

This invention discloses an engine vibration prediction method, system, device, and medium, relating to the field of engine vibration prediction technology. The method includes: acquiring engine assembly data of the target domain to be predicted; using a pre-classifier to predict the engine assembly data of the target domain to obtain pseudo-labels; based on the pseudo-labels, selecting a generator of a corresponding local domain adversarial adaptive network to perform domain transformation on the engine assembly data of the target domain to obtain engine assembly data of the pseudo-source domain to be predicted; and using a source domain classifier to predict the engine assembly data of the pseudo-source domain to obtain the vibration prediction result. This invention can accurately predict engine vibration under unbalanced small sample conditions, improving engine vibration prediction performance.

Description

Technical Field

[0001] This invention relates to the field of engine vibration prediction technology, and in particular to an engine vibration prediction method, system, device and medium. Background Technology

[0002] Heterogeneous transfer learning methods can be mainly divided into two categories: methods based on symmetric feature transformation and methods based on asymmetric feature transformation. Methods based on symmetric feature transformation transform the source and target domains into a common subspace where the data distributions of the source and target domains are more similar than before. For example, Duan et al. proposed Heterogeneous Feature Augmentation (HFA), which uses two transformation matrices to map the source and target domains to the common subspace and merges the two transformation matrices, optimizing the merged matrix by minimizing the structural risk function of the SVM. Methods based on asymmetric feature transformation transform the feature space of the source domain to the feature space of the target domain or vice versa. Sukhija et al. proposed Supervised Heterogeneous Domain Adaptation via Random Forests (SHDA-RF). SHDA-RF uses the shared label distribution of the target and source domains as the core of the feature transformation, defining the relationship between the shared label distribution and features through random forests, thereby obtaining the relationship between the feature spaces of the source and target domains. Feuz et al. proposed a Feature-Space Remapping (FSR) method. FSR defines the meta-features of the target and source domains, constructs a similarity matrix of the target and source domain features, and finally maps the target domain samples to the source domain sample space through feature mapping relationships.

[0003] With the development of deep learning, Yosinski et al. proved that deep neural networks are transferable, and transfer can be achieved by reusing deep learning network layers. Tzeng et al. proposed the DDC (Deep Domain Confusion) method to solve the adaptation problem of deep networks. DDC uses an AlexNet network trained on ImageNet, fixes the first 7 layers of AlexNet, and adds an adaptive metric layer that applies the Maximum Mean Discrepancy (MMD) criterion to the layer before the classifier. Long et al. improved DDC and proposed the DAN (Deep Adaptation Networks) method, adding three adaptive layers and using multi-core MMD instead of single-core MMD, integrating the parameter learning of multi-core MMD into the training of the deep network.

[0004] Since Goodfellow proposed Generative Adversarial Networks (GANs), many scholars have attempted to apply the idea of ​​network adversarialism to transfer learning, proposing various deep adversarial network methods. For example, Ganin et al. proposed Domain-Adversarial Neural Networks (DANNs), which consist of a feature extractor, a class predictor, and a domain classifier. Through adversarial learning between the feature extractor and the domain classifier, target domain samples and source domain samples are mapped to a certain space, and the class predictor constrains the feature extractor to ensure that the mapped samples can be accurately classified. Bousmalis et al. extended DANNs, proposing Domain Separation Networks (DSNs). DSNs divide the source and target domains into a common part with learnable common features and a private part that maintains the independent characteristics of each domain, and further define the loss function accordingly.

[0005] However, when dealing with imbalanced small sample classification prediction problems, existing technologies often suffer from mistransfer of minority class samples during the transfer process, leading to label bias in the trained classification model. Using overly complex neural networks with too many layers can result in overfitting. Therefore, for target engine models with small and imbalanced training data, existing technologies often perform poorly in engine vibration prediction. Summary of the Invention

[0006] The purpose of this invention is to provide an engine vibration prediction method, system, device, and medium to accurately predict engine vibration under unbalanced small sample conditions and improve engine vibration prediction performance.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] An engine vibration prediction method, the method comprising:

[0009] Obtain engine assembly data for the target domain to be predicted;

[0010] A pre-classifier is used to predict the engine assembly data of the target domain to be predicted, and pseudo-labels are obtained;

[0011] Based on the pseudo-labels, the generator of the corresponding local domain adversarial adaptive network is selected to perform domain transformation on the engine assembly data of the target domain to be predicted, so as to obtain the engine assembly data of the pseudo source domain to be predicted.

[0012] A source domain classifier is used to predict the engine assembly data of the pseudo-source domain to be predicted, and vibration prediction results are obtained.

[0013] The pre-classifier is trained by taking the target domain engine assembly data sample as input and the vibration label corresponding to the target domain engine assembly data sample as output.

[0014] The local domain generative adversarial network corresponds one-to-one with the vibration condition label; the local domain adversarial adaptive network includes a generator and a discriminator; the generator is trained with the target domain engine assembly data sample as input and the pseudo source domain engine assembly data sample corresponding to the target domain engine assembly data sample as output; the discriminator is trained with the source domain engine assembly data sample and the pseudo source domain engine assembly data sample as input samples and the degree of similarity between the input sample and the real source domain engine assembly data as output.

[0015] The source domain classifier is trained by taking the pseudo-source domain engine assembly data and the source domain engine assembly data samples as inputs and the vibration labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as outputs.

[0016] Optionally, the method further includes:

[0017] Obtain the target domain dataset and the source domain dataset; the target domain dataset includes target domain engine assembly data samples labeled with vibration conditions; the source domain dataset includes source domain engine assembly data samples labeled with vibration conditions.

[0018] Max-min normalization is used to transform the value of each data in the source domain dataset and the target domain dataset to the range [0,1].

[0019] Based on the vibration condition labels, the target domain dataset is divided into multiple local target domain datasets, and the source domain dataset is divided into multiple local source domain datasets;

[0020] Based on the local target domain dataset and local source domain dataset with the same vibration condition label, train a local domain adversarial adaptation network;

[0021] The generator in the local domain adversarial adaptive network is used to convert the target domain engine assembly data samples into corresponding pseudo source domain engine assembly data samples to obtain the pseudo source domain dataset.

[0022] Train a source domain classifier based on the pseudo source domain dataset and the source domain dataset;

[0023] A pre-classifier is trained based on the target domain dataset.

[0024] Optionally, training a local domain adversarial adaptation network based on a local target domain dataset and a local source domain dataset with the same vibration condition label specifically includes:

[0025] The target domain engine assembly data samples in the local target domain dataset are input into the generator of the corresponding local domain adversarial adaptation network to obtain the corresponding pseudo source domain engine assembly data samples.

[0026] The pseudo-source domain engine assembly data sample is input into the discriminator of the corresponding local domain adversarial adaptation network to determine the degree of similarity between the pseudo-source domain engine assembly data sample and the real source domain engine assembly data, and obtain the first degree of similarity.

[0027] The source domain engine assembly data samples in the local source domain dataset are input into the discriminator of the corresponding local domain adversarial adaptation network to determine the degree of similarity between the source domain engine assembly data samples and the real source domain engine assembly data, and to obtain the second degree of similarity.

[0028] The generator is trained with the goal of increasing the first degree of proximity.

[0029] The discriminator is trained with the goal of reducing the first proximity and increasing the second proximity.

[0030] Optionally, training the discriminator with the goal of reducing the first proximity and increasing the second proximity specifically includes:

[0031] The loss function of the discriminator of the local domain adversarial adaptation network is determined based on the first proximity degree and the second proximity degree;

[0032] The parameters of the discriminator in the local adversarial adaptive network are adjusted according to the discriminator's loss function.

[0033] Optionally, the loss function of the discriminator is:

[0034]

[0035] Among them, Loss D Indicates the training loss, x t x represents the target domain engine assembly data sample in the local target domain dataset. s Let G() represent the source domain engine assembly data sample in the local source domain dataset, D() represent the generator function, E() represent the discriminator function, and ρ represent the expected value. s x represents s The distribution it follows, ρ t x represents t The distribution it follows, G(x) t ) represents x t The corresponding pseudo-source domain engine assembly data sample, D(G(x) t )) represents x t The degree of similarity between the corresponding pseudo-source domain engine assembly data sample and the real source domain engine assembly data, D(x) s ) represents x s The second degree of closeness to the actual source domain engine assembly data, where λ represents the penalty constant and gp represents the regularization term.

[0036] Optionally, training the generator with the goal of increasing the first proximity specifically includes:

[0037] The loss function of the generator of the local domain adversarial adaptation network is determined based on the first degree of proximity.

[0038] The parameters of the generator of the local domain adversarial adaptation network are adjusted according to the generator's loss function.

[0039] Optionally, the loss function of the generator is:

[0040] Loss G =-E[D(G(x) t ))]

[0041] Among them, Loss G Indicates the training loss, x t G(x) represents the target domain engine assembly data sample in the local target domain dataset, G() represents the generator function, D() represents the discriminator function, E() represents the expected value, and G(x) represents the generator function. t ) represents x t The corresponding pseudo-source domain engine assembly data sample, D(G(x) t )) represents x t The degree of similarity between the corresponding pseudo-source domain engine assembly data sample and the real source domain engine assembly data.

[0042] An engine vibration prediction system, the system comprising:

[0043] The data acquisition module is used to acquire engine assembly data for the target domain to be predicted.

[0044] The pre-classification module is used to predict the engine assembly data of the target domain to be predicted using a pre-classifier to obtain pseudo-labels;

[0045] The domain transformation module is used to select the generator of the corresponding local domain adversarial adaptive network according to the pseudo-label to perform domain transformation on the engine assembly data of the target domain to be predicted, so as to obtain the engine assembly data of the pseudo source domain to be predicted.

[0046] The vibration prediction module is used to predict the engine assembly data of the pseudo-source domain to be predicted using a source domain classifier to obtain the vibration prediction result.

[0047] The pre-classifier is trained by taking the target domain engine assembly data sample as input and the vibration label corresponding to the target domain engine assembly data sample as output.

[0048] The local domain generative adversarial network corresponds one-to-one with the vibration condition label; the local domain adversarial adaptive network includes a generator and a discriminator; the generator is trained with the target domain engine assembly data sample as input and the pseudo source domain engine assembly data sample corresponding to the target domain engine assembly data sample as output; the discriminator is trained with the source domain engine assembly data sample and the pseudo source domain engine assembly data sample as input samples and the degree of similarity between the input sample and the real source domain engine assembly data as output.

[0049] The source domain classifier is trained by taking the pseudo-source domain engine assembly data and the source domain engine assembly data samples as inputs and the vibration labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as outputs.

[0050] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor running the computer program to enable the electronic device to perform the engine vibration prediction method described above.

[0051] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described engine vibration prediction method.

[0052] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0053] The engine vibration prediction method provided by this invention divides the target domain dataset and the source domain dataset into multiple local domains according to vibration condition labels. Different local domain adversarial adaptation networks are established for each local domain, thereby mapping samples from each local target domain to the corresponding local source domain, ensuring that minority class samples can also achieve reasonable transfer. During vibration prediction, this invention first uses a pre-classifier to determine pseudo-labels for the engine assembly data in the target domain, thus using these pseudo-labels to solve the mapping function selection problem of the adversarial adaptation network. Finally, the source domain classifier provides a reliable final prediction result, correcting any errors that may arise from the pseudo-labels. Therefore, this invention can accurately predict the vibration of the target engine model (i.e., the source domain) even with a small training sample size for the target engine model (i.e., the target domain), by transferring data from other engine models (i.e., the source domain), thereby improving engine vibration prediction performance. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart of the engine vibration prediction method provided by the present invention;

[0056] Figure 2 This is a flowchart illustrating the engine vibration prediction method provided in an embodiment of the present invention.

[0057] Figure 3 A comparison chart of Wilcoxon signed-rank test for AUC statistical analysis of EVP-PDAA and other transfer learning methods provided in the embodiments of the present invention;

[0058] Figure 4 Cohen's d effect size verification comparison chart for AUC statistical analysis of EVP-PDAA and other transfer learning methods provided in the embodiments of the present invention;

[0059] Figure 5 A comparison chart of Wilcoxon signed-rank test results for F1 statistical analysis of EVP-PDAA and other transfer learning methods provided in the embodiments of the present invention;

[0060] Figure 6 A comparison chart of Cohen's d effect size validation for F1 statistical analysis of EVP-PDAA and other transfer learning methods provided in the embodiments of the present invention;

[0061] Figure 7A block diagram of the engine vibration prediction system provided by the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The purpose of this invention is to provide an engine vibration prediction method, system, device, and medium to accurately predict engine vibration under unbalanced small sample conditions and improve engine vibration prediction performance.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Example 1

[0066] This embodiment provides a method for predicting engine vibration. For example... Figure 1 As shown, the method includes:

[0067] Step 101: Obtain engine assembly data for the target domain to be predicted.

[0068] Step 102: A pre-classifier is used to predict the engine assembly data of the target domain to obtain pseudo-labels. The pre-classifier is trained using engine assembly data samples of the target domain as input and vibration labels corresponding to the engine assembly data samples of the target domain as output.

[0069] Step 103: Based on the pseudo-labels, select the generator of the corresponding local domain adversarial adaptation network to perform domain transformation on the engine assembly data of the target domain to be predicted, and obtain the engine assembly data of the pseudo source domain to be predicted. The local domain generative adversarial network corresponds one-to-one with the vibration condition labels; the local domain adversarial adaptation network includes a generator and a discriminator; the generator is trained with the engine assembly data sample of the target domain as input and the pseudo source domain engine assembly data sample corresponding to the engine assembly data sample of the target domain as output; the discriminator is trained with the engine assembly data sample of the source domain and the pseudo source domain engine assembly data sample as input samples and the degree of similarity between the input samples and the real source domain engine assembly data as output.

[0070] Step 104: Use a source domain classifier to predict the pseudo-source domain engine assembly data to obtain vibration prediction results. The source domain classifier is trained using the pseudo-source domain engine assembly data and the source domain engine assembly data samples as inputs, and the vibration labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as outputs.

[0071] Furthermore, the method also includes:

[0072] Step 001: Obtain the target domain dataset and the source domain dataset; the target domain dataset includes target domain engine assembly data samples labeled with vibration conditions; the source domain dataset includes source domain engine assembly data samples labeled with vibration conditions. Specifically, the target domain engine assembly data samples are assembly data of the target engine model for which vibration prediction performance needs to be improved; the source domain engine assembly data samples are assembly data of other engine models that are related to the vibration characteristics of the target engine model and have excellent vibration prediction performance.

[0073] Step 002: Use maximum-minimum normalization to transform the value of each data in the source domain dataset and the target domain dataset to the range [0,1].

[0074] Step 003: Based on the vibration condition labels, divide the target domain dataset into multiple local target domain datasets, and divide the source domain dataset into multiple local source domain datasets.

[0075] Step 004: Train a local domain adversarial adaptation network based on the local target domain dataset and the local source domain dataset with the same vibration condition label.

[0076] Step 005: Using the generator in the local domain adversarial adaptation network, the target domain engine assembly data samples are converted into corresponding pseudo source domain engine assembly data samples to obtain the pseudo source domain dataset.

[0077] Step 006: Train the source domain classifier based on the pseudo source domain dataset and the source domain dataset.

[0078] Step 007: Train a pre-classifier based on the target domain dataset.

[0079] Step 004 specifically includes:

[0080] (1) Input the target domain engine assembly data sample in the local target domain dataset into the generator of the corresponding local domain adversarial adaptation network to obtain the corresponding pseudo source domain engine assembly data sample.

[0081] (2) Input the pseudo-source domain engine assembly data sample into the discriminator of the corresponding local domain adversarial adaptation network to determine the degree of similarity between the pseudo-source domain engine assembly data sample and the real source domain engine assembly data, and obtain the first degree of similarity.

[0082] (3) Input the source domain engine assembly data sample in the local source domain dataset into the discriminator of the corresponding local domain adversarial adaptation network to determine the degree of similarity between the source domain engine assembly data sample and the real source domain engine assembly data, and obtain the second degree of similarity.

[0083] (4) Train the generator with the goal of increasing the first proximity. That is: determine the loss function of the generator of the local domain adversarial adaptation network based on the first proximity; adjust the parameters of the generator of the local domain adversarial adaptation network based on the loss function of the generator.

[0084] (5) The discriminator is trained with the goal of reducing the first proximity and increasing the second proximity. That is, the loss function of the discriminator of the local domain adversarial adaptation network is determined based on the first proximity and the second proximity; the parameters of the discriminator of the local domain adversarial adaptation network are adjusted based on the loss function of the discriminator.

[0085] Furthermore, step 004 also includes:

[0086] (6) The local adversarial adaptation network that meets the set conditions is used as the trained local adversarial adaptation network. The set conditions include the number of training rounds reaching a set value, or the first proximity degree being within a set range within a set number of training rounds, or the expected value of the first proximity degree being within a set range within a set number of training rounds, but these are not restrictions and can be adjusted according to the actual situation.

[0087] like Figure 2 As shown, in practical applications, the specific process of the engine vibration prediction method provided by this invention includes:

[0088] Step 1) Data Acquisition: Labeled (training) assembly data of the target engine model whose vibration prediction performance needs to be improved is used as the target domain dataset D. t The assembly data of other engine models whose vibration characteristics are related to the target domain engine and whose vibration prediction performance is excellent on their data are used as the source domain dataset D. s .

[0089] Step 2) Data Preprocessing: Max-min normalization is used to transform the values ​​of each attribute in the source and target domain data to the range [0,1], eliminating the influence of units; the source and target domain data are then divided into multiple local source and local target domains according to labels. The local target domain is the set of source domain samples labeled γ. The set of target domain samples labeled γ.

[0090] Step 3) Establishment of local adversarial adaptation networks: Establish multiple local adversarial adaptation networks to map each local target domain sample to the corresponding local source domain.

[0091] The local adversarial adaptive network consists of a generator G and a discriminator D. The generator learns the mapping relationship from the local target domain to the local source domain sample space, taking local target domain samples as input and outputting samples transformed into the local source domain sample space. The discriminator learns to determine whether data is transformed target domain data or source domain data, taking local source domain samples and the generator's output as input and outputting the degree of proximity between the input sample and the local source domain. Through adversarial training between the generator and the discriminator, the generator can accurately map local target domain samples to the corresponding local source domain.

[0092] To prevent mode collapse and other issues caused by the discriminator getting stuck in local optima during adversarial training, this invention uses Earth Mover Distance (EMD) as the optimization objective of the discriminator. When the function f represented by the discriminator satisfies the Lipschitz constraint, i.e., |f(x1)-f(x2)|≤K|x1-x2|, the objective function of the discriminator is:

[0093]

[0094] Where, ρ s and ρ t These are local source domain samples x s and local target domain sample x t The distribution it follows.

[0095] To ensure that the function represented by the discriminator always satisfies the Lipschitz constraint, a gradient penalty strategy is introduced, which involves adding a regularization term to the discriminator's loss function. The calculation formula is as follows:

[0096]

[0097] in, The samples are random. By constraining the L2 norm of the gradient to around 1, the parameters of the discriminator are kept within a certain constant, ensuring that the discriminator satisfies the Lipschitz constraint.

[0098] random samples The calculation method is as follows:

[0099]

[0100] Where ε is a random number between 0 and 1.

[0101] Therefore, the loss function of discriminator D is:

[0102]

[0103] Among them, Loss D Indicates the training loss, x t x represents the target domain engine assembly data sample in the local target domain dataset. s Let G() represent the source domain engine assembly data sample in the local source domain dataset, D() represent the generator function, E() represent the discriminator function, and ρ represent the expected value. s x represents s The distribution it follows, ρ t x represents t The distribution it follows, G(x) t ) represents x t The corresponding pseudo-source domain engine assembly data sample, D(G(x) t )) represents x t The degree of similarity between the corresponding pseudo-source domain engine assembly data sample and the real source domain engine assembly data, D(x) s ) represents x s The second degree of closeness to the actual source domain engine assembly data, where λ represents the penalty constant and gp represents the regularization term.

[0104] The loss function of generator G is:

[0105] Loss G =-E[D(G(x) t ))]

[0106] Among them, Loss G This represents the training loss generated.

[0107] Specifically, both the generator and discriminator consist of an input layer, hidden layers, and an output layer. Due to the small amount of data, the hidden layer network should not be designed to be too complex to avoid overfitting. Furthermore, since batch normalization (BN) normalizes a batch of samples, adding a BN layer to the discriminator network would cause errors in the gradient calculation for each sample; therefore, no BN layer is added to the discriminator.

[0108] Step 4) Train the source domain classifier and pre-classifier: Select the appropriate generator from multiple local domain adversarial adaptation networks, and map each local target domain sample to the corresponding local source domain sample space to obtain the transformed data, i.e., the pseudo source domain dataset D. t_trans Using Dt_trans and D s The source domain classifier is trained using two sets of data, and D is used. t The pre-classifier and classification algorithm both use random forest.

[0109] Step 4) Domain transformation based on pseudo-labels: Use a pre-classifier to perform domain transformation on the unlabeled target domain engine assembly data U to be predicted. t Make predictions and provide pseudo-labels for the samples; based on the pseudo-labels, select the generator of the corresponding local domain adversarial adaptive network to apply the algorithm to U. t Perform a domain transformation to obtain the transformed data U. t_trans .

[0110] Step 5) Vibration prediction: Use the source domain classifier to predict the vibration of U. t_trans Make predictions and provide the final results of vibration prediction.

[0111] Example 2

[0112] This invention uses a heterogeneous transfer learning method based on adversarial networks to improve engine vibration prediction performance. A specific embodiment is provided below to illustrate the above technical effects in detail.

[0113] Since the engine vibration prediction method provided in this invention is an engine vibration prediction model based on Partial Domain Adversarial Adaptation Network (EVP-PDAA), this method is referred to as EVP-PDAA in this embodiment. The core idea of ​​this method is to divide the domain into multiple local domains according to labels, establish multiple local domain adversarial adaptation networks to map each local target domain sample to the corresponding local source domain, use pseudo-labels to solve the mapping function selection problem during vibration prediction, and use a source domain classifier to provide reliable prediction results, correcting errors that may occur with pseudo-labels. Multiple experiments show that EVP-PDAA significantly improves upon other methods in terms of Area Under the Curve (AUC) and F1 (F1-Score) evaluation metrics. The performance of this invention on a real-world dataset is described in detail below, and comparisons are made with other methods.

[0114] This invention conducted experiments on three source domain datasets (DR, EP-1, EP-2) and three target domain datasets (SR-1, SR-2, SR-3). The actual assembly data of the single-shaft engine contained 248 assembly parameters. Three rounds of feature selection were performed using different methods, resulting in the selection of 9, 12, and 23 key assembly parameters, respectively, for SR-1, SR-2, and SR-3. DR was collected from the actual assembly process of a dual-shaft engine, which shares common key vibration characteristics with the single-shaft engine. EP-1 and EP-2 were collected from an experimental platform for the single-shaft engine, a simplified simulation of the single-shaft engine. Statistical information for each dataset is shown in Table 1.

[0115] Table 1. Statistical information of the dataset

[0116]

[0117] This invention selects five comparison methods: 1) RF-T: classifier is built directly using target domain data without using transfer learning; 2) TCA (Transfer Component Analysis): resolves the data distribution difference between the source and target domains by minimizing the distance between their marginal probability distributions; 3) CORAL (CORrelation Alignment): resolves the data distribution difference by aligning the covariance of the source and target domains; 4) FSR (Feature-Space Remapping): establishes a feature mapping relationship between the target and source domains using meta-features and a similarity matrix; 5) SHDA-RF (Supervised Heterogeneous Domain Adaptation via Random Forests): obtains a feature space mapping relationship between the two domains using shared labels and random forests. Since CORAL and TCA require the target and source domains to have the same feature space, UMR is used to unify the feature spaces of the target and source domains before transfer learning. UMR consists of three parts: source domain-specific features, target domain-specific features, and common features between the source and target domains. The UMR of the source domain is achieved by setting the features specific to the target domain to 0, and the UMR of the target domain is achieved by setting the features specific to the source domain to 0. All methods use the Random Forest (RF) algorithm as the classifier.

[0118] This invention uses AUC and F1 to evaluate the performance of vibration prediction models established by each method. F1 measures the model's prediction performance for minority classes, while AUC evaluates the overall model performance. To avoid the influence of random factors on the experiments, each experiment was repeated 20 times. The Wilcoxon signed-rank test was used to determine whether the experimental results of the two methods were statistically different, with a confidence level set at 95%. Furthermore, this invention uses Cohen's d effect size to quantify the difference between the two methods. The correspondence between Cohen's d value and effect level is shown in Table 2.

[0119] Table 2 Cohen's d Effect Size Ranking Table

[0120]

[0121] As shown in Tables 3 and 4, EVP-PDAA significantly improves AUC and F1 scores compared to RF-T when facing different source and target domains. This indicates that EVP-PDAA fully utilizes the knowledge from the source domain, performs effective transfer, and has strong generalization ability. In most transfer learning scenarios, the vibration prediction model built by CORAL shows some improvement in AUC and F1 scores compared to RF-T. However, in some scenarios, CORAL exhibits negative transfer, with prediction performance decreasing instead of improving, indicating poor generalization of CORAL. Other transfer learning methods generally perform worse than RF-T in most cases, showing severe negative transfer, indicating that these methods cannot perform reasonable transfer when faced with imbalanced small samples.

[0122] Table 3 Comparison of AUC indices for different methods

[0123]

[0124] Table 4 Comparison of F1 Scores for Each Method

[0125]

[0126]

[0127] For statistical analysis of the AUC and F1 scores of the EVP-PDAA method provided in this invention compared with other transfer learning methods, please refer to the following figures. Figure 3 , Figure 4 , Figure 5 and Figure 6It can be seen that the prediction models established by EVP-PDAA and other transfer learning methods are statistically significantly different in performance. The effect size levels of AUC and F1 are almost all L, and the effect size values ​​are much larger than the threshold of L, indicating that EVP-PDAA can more effectively transfer knowledge of other engine models and establish a more powerful vibration prediction model than other transfer learning methods.

[0128] Example 3

[0129] To implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, an engine vibration prediction system is provided below. For example... Figure 7 As shown, the system includes:

[0130] The data acquisition module 701 is used to acquire engine assembly data for the target domain to be predicted.

[0131] The pre-classification module 702 is used to predict the engine assembly data of the target domain to be predicted using a pre-classifier, and obtain pseudo-labels. The pre-classifier is trained by taking the engine assembly data samples of the target domain as input and the vibration label corresponding to the engine assembly data samples of the target domain as output.

[0132] The domain transformation module 703 is used to select the generator of the corresponding local domain adversarial adaptation network based on the pseudo-labels to perform domain transformation on the engine assembly data of the target domain to be predicted, thereby obtaining the pseudo-source domain engine assembly data to be predicted. The local domain generative adversarial network corresponds one-to-one with the vibration condition label; the local domain adversarial adaptation network includes a generator and a discriminator; the generator is trained with the target domain engine assembly data sample as input and the pseudo-source domain engine assembly data sample corresponding to the target domain engine assembly data sample as output; the discriminator is trained with the source domain engine assembly data sample and the pseudo-source domain engine assembly data sample as input samples and the degree of similarity between the input samples and the real source domain engine assembly data as output.

[0133] The vibration prediction module 704 is used to predict the vibration of the pseudo-source domain engine assembly data using a source domain classifier, thereby obtaining a vibration prediction result. The source domain classifier is trained using the pseudo-source domain engine assembly data and the source domain engine assembly data samples as inputs, and the vibration labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as outputs.

[0134] Example 4

[0135] This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the engine vibration prediction method of Embodiment 1. The electronic device may be a server.

[0136] In addition, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the engine vibration prediction method in Embodiment 1.

[0137] This invention provides a method, system, device, and medium for predicting engine vibration. By dividing the domain into multiple local domains according to labels, multiple local domain adversarial adaptation networks are established to map target domain samples to the corresponding labeled local source domains, ensuring reasonable transfer even for minority class samples. Due to the small sample size, to ensure the stability of the local domain adversarial adaptation network training and the correctness of domain transformation, bulldozer distance is used as the network optimization objective, and a gradient penalty strategy is employed when updating network parameters. During vibration prediction, pseudo-labels are used to address the local domain adversarial adaptation network selection problem for the target samples to be predicted, and a label correction mechanism is designed to correct potential errors in pseudo-labels. Experimental results show that the method provided by this invention outperforms other transfer learning methods when facing imbalanced small samples, achieving the transfer of data from other engine models to improve the vibration prediction effect of the target engine.

[0138] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0139] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An engine vibration prediction method characterized by, The method comprises: acquiring target domain engine assembly data to be predicted; using a pre-classifier to predict the target domain engine assembly data to be predicted, to obtain pseudo-labels; according to the pseudo-labels, selecting a corresponding generator of a local domain adversarial adaptation network to perform domain conversion on the target domain engine assembly data to be predicted, to obtain pseudo-source domain engine assembly data to be predicted; using a source domain classifier to predict the pseudo-source domain engine assembly data to be predicted, to obtain a vibration prediction result; the pre-classifier is trained with target domain engine assembly data samples as input and vibration condition labels corresponding to the target domain engine assembly data samples as output; the local domain adversarial adaptation network corresponds one-to-one to the vibration condition labels; the local domain adversarial adaptation network comprises a generator and a discriminator; the generator is trained with the target domain engine assembly data samples as input and pseudo-source domain engine assembly data samples corresponding to the target domain engine assembly data samples as output; the discriminator is trained with source domain engine assembly data samples and the pseudo-source domain engine assembly data samples as input samples and the closeness of the input samples to real source domain engine assembly data as output; the source domain classifier is trained with the pseudo-source domain engine assembly data and the source domain engine assembly data samples as input and vibration condition labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as output.

2. The engine vibration prediction method according to claim 1, characterized by, The method further comprises: acquiring a target domain data set and a source domain data set; the target domain data set comprises target domain engine assembly data samples with vibration condition labels; the source domain data set comprises source domain engine assembly data samples with vibration condition labels; using maximum-minimum value normalization to transform the value of each data in the source domain data set and the target domain data set to [0, 1]; according to the vibration condition labels, dividing the target domain data set into multiple local target domain data sets and dividing the source domain data set into multiple local source domain data sets; training a local domain adversarial adaptation network according to local target domain data sets and local source domain data sets with the same vibration condition labels; using a generator in the local domain adversarial adaptation network to convert the target domain engine assembly data samples into corresponding pseudo-source domain engine assembly data samples respectively, to obtain a pseudo-source domain data set; training a source domain classifier according to the pseudo-source domain data set and the source domain data set; training a pre-classifier according to the target domain data set.

3. The engine vibration prediction method according to claim 1, characterized by, The training of the local domain adversarial adaptation network according to the local target domain data sets and the local source domain data sets with the same vibration condition labels specifically comprises: inputting target domain engine assembly data samples in the local target domain data set into a corresponding generator of a local domain adversarial adaptation network, to obtain corresponding pseudo-source domain engine assembly data samples; inputting the pseudo-source domain engine assembly data sample into a discriminator of a corresponding local domain adversarial adaptation network, determining the closeness of the pseudo-source domain engine assembly data sample to the real source domain engine assembly data, and obtaining a first closeness; inputting the source domain engine assembly data sample in the local source domain data set into a discriminator of a corresponding local domain adversarial adaptation network, determining the closeness of the source domain engine assembly data sample to the real source domain engine assembly data, and obtaining a second closeness; training the generator with the goal of increasing the first closeness; training the discriminator with the goal of reducing the first closeness and increasing the second closeness.

4. The engine vibration prediction method according to claim 3, characterized by, The training of the discriminator with the goal of reducing the first closeness and increasing the second closeness specifically includes: determining a loss function of the discriminator of the local domain adversarial adaptation network according to the first closeness and the second closeness; adjusting the parameters of the discriminator of the local domain adversarial adaptation network according to the loss function of the discriminator.

5. The engine vibration prediction method according to claim 4, characterized by, The loss function of the discriminator is: Wherein, Loss D The discriminant training loss is represented by L, x t The target domain engine assembly data sample in the local target domain data set is represented by x, x s The source domain engine assembly data sample in the local source domain data set is represented by x, ρ s The distribution is represented by p, x s The distribution is represented by p, ρ t The distribution is represented by p, x t The distribution is represented by p, x t The corresponding pseudo source domain engine assembly data sample is represented by G, x t The corresponding pseudo source domain engine assembly data sample is represented by G, x t The first closeness of the corresponding pseudo source domain engine assembly data sample to the real source domain engine assembly data is represented by D, x t The second closeness to the real source domain engine assembly data is represented by D, x s The second closeness to the real source domain engine assembly data is represented by D, x s The second closeness to the real source domain engine assembly data is represented by D, λ The penalty constant is represented by λ, gp The regularization term is represented by R.

6. The engine vibration prediction method according to claim 3, characterized by, The training of the generator with the goal of increasing the first closeness specifically includes: determining a loss function of the generator of the local domain adversarial adaptation network according to the first closeness; adjusting the parameters of the generator of the local domain adversarial adaptation network according to the loss function of the generator.

7. The engine vibration prediction method according to claim 6, characterized by, The loss function of the generator is: Wherein, Loss G The generated training loss is represented by L, x t The target domain engine assembly data sample in the local target domain data set is represented by x, The generation function of the generator is represented by G(x), The expectation value is represented by E(), The generated target domain engine assembly data sample is represented by G(x), x t The generated target domain engine assembly data sample is represented by G(x), x t The corresponding pseudo source domain engine assembly data sample is represented by D(G(x)), x t The first closeness of the corresponding pseudo source domain engine assembly data sample to the real source domain engine assembly data is represented by D(G(x)), x t The first closeness of the corresponding pseudo source domain engine assembly data sample to the real source domain engine assembly data is represented by D(G(x)).

8. An engine vibration prediction system characterized by, The system includes: a data acquisition module configured to acquire target domain engine assembly data to be predicted; a pre-classifier configured to predict the target domain engine assembly data to be predicted to obtain pseudo-labels; a domain conversion module configured to select a corresponding generator of a local domain adversarial adaptation network to perform domain conversion on the target domain engine assembly data to be predicted according to the pseudo-labels, and obtain pseudo-source domain engine assembly data to be predicted; a vibration prediction module configured to predict the pseudo-source domain engine assembly data to be predicted using a source domain classifier to obtain a vibration prediction result; The pre-classifier is trained with target domain engine assembly data samples as input and vibration condition labels corresponding to the target domain engine assembly data samples as output; The local domain adversarial adaptation network and the vibration condition labels correspond to each other. The local domain adversarial adaptation network includes a generator and a discriminator. The generator is trained with the target domain engine assembly data samples as input and pseudo-source domain engine assembly data samples corresponding to the target domain engine assembly data samples as output. The discriminator is trained with source domain engine assembly data samples and the pseudo-source domain engine assembly data samples as input samples and the closeness of the input samples to real source domain engine assembly data as output; The source domain classifier is trained with the pseudo-source domain engine assembly data and the source domain engine assembly data samples as input and vibration condition labels corresponding to the pseudo-source domain engine assembly data and the source domain engine assembly data samples as output.

9. An electronic device, comprising: An electronic device comprising a memory for storing a computer program and a processor for executing the computer program to cause the electronic device to perform the engine vibration prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A non-transitory computer-readable medium storing a computer program, the computer program being executed by a processor to implement the engine vibration prediction method according to any one of claims 1 to 7.

Images