Compressor surge precursor identification method fusing image feature migration and domain self-adaption

Abstract

The invention discloses a gas compressor surge precursor identification method fusing image feature migration and domain self-adaption, and belongs to the field of aero-engine modeling and fault diagnosis. The method comprises the steps that firstly, blade tip pressure signal data in the process from steady-state throttling to stall of the gas compressor are preprocessed to obtain a plurality of samples, and each sample is converted into a recurrence plot; secondly, constructing a ConvLSTM network to extract spatial and temporal features of the recurrence plot, and extracting graph topology features from the output of the ConvLSTM network through a graph convolutional network; thirdly, a ConvLSTM-GCN model is constructed, and graph topological features including space-time features are extracted; carrying out graph topology feature alignment between the source domain data and the target domain data; the ConvLSTM-GCN model is trained to obtain a prediction model; and finally, performing real-time prediction on the test data set to obtain the surge prediction probability. According to the invention, the problem of low prediction precision of the existing RNN, CNN and variant deep learning models thereof can be solved; in combination with a transfer learning method based on domain self-adaption, the problem that the feature distribution difference between a target domain and a source domain is large is solved, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The invention belongs to the field of aero-engine modeling and fault diagnosis, and relates to a compressor surge precursor recognition method integrating graph feature migration and domain adaptation. Background Art

[0002] The aircraft engine is the "heart" of modern aircraft and the only power source of the aircraft. The technological development of aircraft has put forward more stringent performance indicators and usage standards for engines, and the engine technology needs to be improved synchronously to meet the more complex flight mission requirements in the future. Specifically speaking, as one of the three core components of the aircraft engine, the stable operation of the compressor of the aircraft engine is crucial to the performance and safety of the engine, and directly affects the overall performance. The main function of the compressor is to compress the air entering the engine, increase the air flow pressure, and create favorable conditions for subsequent combustion. The specific working process is briefly described as follows: the air entering the compressor is first accelerated by the rotor blades, and then decelerated in the stator blade channel, converting the kinetic energy of the rotor into static pressure energy. This compression process is repeated between multiple stages of blades until the required total pressure ratio is reached.

[0003] However, when the gas flows axially into the compressor from the air inlet at a certain speed, as the flow rate decreases and the total pressure ratio increases, it is easy to cause gas flow separation, causing aerodynamic instability. The two typical forms of aerodynamic instability of the compressor are rotating stall and surge. Therefore, designers need to fully consider the flow characteristics of the compressor under various operating conditions, take necessary measures to predict instability and prevent the occurrence of aerodynamic instability, and ensure the safe and reliable operation of the engine.

[0004] Surge is a serious airflow disturbance phenomenon related to the engine system. It will cause severe mechanical vibration of engine components and overheating of the hot end in a very short time, and it must be avoided during operation. Surge usually occurs when the compressor transitions from design conditions to non-design conditions. It initially manifests as a non-axisymmetric local disturbance, and after full development, it will evolve into an axisymmetric airflow pulsation propagating along the axial direction. The root cause of surge lies in the instability of the entire engine system. When the system parameters exceed the safe range, this serious aerodynamic instability phenomenon is easily caused. In the process of engine design and operation, it is necessary to attach great importance to the surge problem, and take various effective measures to prevent and suppress this aerodynamic instability to ensure that the engine can operate reliably and safely. Therefore, effective methods must be taken to predict surge to leave enough time margin for the surge control system to work.

[0005] So far, researchers' exploration of the compressor surge prediction process can be divided into two aspects: prediction methods based on traditional time domain methods, frequency domain methods, time-frequency domain methods, and deep learning-based prediction methods:

[0006] Traditional methods mainly rely on signal processing techniques, such as Fourier analysis and wavelet transform, to detect surge precursors. Researchers have successively proposed methods such as spectrum feature analysis, frequency domain feature hump discrimination factor, adjacent cycle rotor correlation coefficient, etc., which can predict the occurrence of surge a certain time in advance. For example, the invention patent with publication number CN117332214A combines wavelet transform with frequency domain coherence function to calculate surge precursors, realizing dual-dimensional prediction in time and frequency domain. However, this method has limited performance in complex working environments, and the subjectivity of threshold selection and false alarm problems are also urgent to be solved.

[0007] In recent years, methods combined with deep learning have been widely used in the field of compressor surge prediction. Researchers have successively proposed RNN variants LR-WaveNet and LSTM stall prediction. Some scholars have also applied TCN networks combined with random forest (RF) models and deterministic learning theory to surge prediction. For example, the invention patent with publication number CN112001128A uses an expanded causal convolution structure to fuse statistical features, which can effectively extract surge precursor features to a certain extent, but usually has limitations. The limitation is that this method requires enough fault data with phase distribution and feature space to support the training of neural networks, and the generalization performance of the model needs to be improved urgently. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a compressor surge precursor identification method that integrates graph feature transfer and domain adaptation. The compressor surge precursor identification method integrates the graph topology features of the data to improve the prediction accuracy of the ConvLSTM-GCN model, and at the same time improves the generalization ability of the ConvLSTM-GCN model by combining the idea of ​​transfer learning, wherein transfer learning is defined as given a source domain and a corresponding learning task, a target domain and a corresponding learning task, and using the knowledge in the source domain and the source task to improve the learning performance of the target domain and the target task. The graph topology features can deeply characterize the complex nonlinear structure of the surge precursor dynamics, capture the global time dependency, and enhance the robustness and generalization ability of the ConvLSTM-GCN model. These features, combined with deep neural networks, make it perform better in surge prediction and improve the prediction accuracy.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is:

[0010] A compressor surge precursor recognition method integrating graph feature migration and domain adaptation comprises the following steps:

[0011] S1, preprocessing the tip pressure signal data of different types of compressors of aircraft engines during the process from steady-state throttling to stall. The steps include:

[0012] S1.1, dividing the blade tip pressure signal data into source domain data and target domain data;

[0013] S1.2, divide the source domain data into 4:1 ratio, 80% as training data set and 20% as test data set; similarly, divide the target domain data into 80% as training data set and 20% as test data set according to the above steps;

[0014] S1.3, using a low-pass filter to filter all the data in the training data set obtained in S1.2;

[0015] S1.4, downsample the data after filtering in S1.3; because the surge frequency is between 3 and 166 Hz, in order to meet the requirements of Nyquist sampling theorem, 10 times downsampling is performed;

[0016] S1.5, split the data according to the time step size, and label each sample after the split; set the time step to 256, set a time window of length 256 to slide on the data processed by S1.3, and split the data falling in the time window as a sample. According to whether there is surge or not, assign a label of 1 or 0 to each sample, where the label with surge is 1, and the label without surge is 0.

[0017] S2, convert each sample processed by S1.5 into a recurrence plot (RP) to reveal the evolution of its state over time and the repetitive patterns between states; the specific steps of converting into a recurrence plot are as follows:

[0018] S2.1, for the sample obtained in S1.5 , select the time delay τ and embedding dimension m, and construct the embedding vector ; as shown below:

[0019] (1)

[0020] in, is the embedding vector generated by embedding dimension and time delay; is the sample obtained by S1.5, with a length of N; τ is the time delay; m is the embedding dimension;

[0021] S2.2, use the recursive matrix V to represent the distance between embedded vectors; the elements of the recursive matrix V are The definition is as follows:

[0022] (2)

[0023] in, , , , represents the Euclidean distance; then the recursive matrix is ​​expressed as:

[0024] (3)

[0025] S2.3, define a threshold , binarize the elements in the recursive matrix V to obtain the final recursive graph RP;

[0026] (4)

[0027] in, Represents the element in the i-th row and j-th column of the recursive matrix V; is the set threshold; is the element in the i-th row and j-th column of the recursive graph RP.

[0028] Furthermore, in S2.3, the threshold The range is 0.1 to 0.5.

[0029] S3, building a ConvLSTM network to extract the spatiotemporal features of the recursive graph RP; including the following steps:

[0030] S3.1, taking the recursive graph RP obtained in S2 as input, constructing a ConvLSTM network including two ConvLSTM layers; wherein the two ConvLSTM layers have different numbers and sizes of convolution kernels, and can provide spatiotemporal features of different scales;

[0031] S3.2, the output of the ConvLSTM network is obtained through a fully connected layer and a relu activation function to describe the spatiotemporal features.

[0032] S4, the output of the ConvLSTM network is used to extract graph topology features through a graph convolutional network (GCN), which includes the following steps:

[0033] S4.1, the output of the ConvLSTM network obtains the graph structure through the graph generation layer algorithm (GGL) to satisfy the calculation of the graph convolution network; finally, the graph structure is obtained , where A is the adjacency matrix and T is the node feature; the specific expression of the graph generation layer algorithm is:

[0034] (5)

[0035] (6)

[0036] Among them, T is the output of the ConvLSTM network, which is also used as a node feature; express The transpose of represents the normalization function; is the original adjacency matrix; further, return The indices of the first k largest values ​​of , which makes the adjacency matrix sparse and reduces the computational burden;

[0037] S4.2, build a multi-receptive field graph convolutional network, and transform the graph structure obtained in S4.1 As the input of the multi-receptive field graph convolutional network to extract graph topology features; the standard graph convolution GCN can only aggregate information in a fixed-size receptive field. In order to solve this problem, the present invention adopts a multi-receptive field graph convolutional network (MRF-GCN, Multi-receptiveField Graph Convolutional Networks) method to obtain powerful feature representation. The multi-receptive field graph convolutional network consists of three graph convolutional layers with different receptive field sizes, and the three receptive fields of different sizes are respectively denoted as K1, K2, and K3;

[0038] Graph Structure The features learned by the graph convolutional layers with receptive field sizes of K1, K2, and K3 are expressed as follows:

[0039] (7)

[0040] (8)

[0041] (9)

[0042] in, They represent the graph convolutional layers with receptive field sizes of K1, K2, and K3 respectively. and Represent the feature representations learned by the three graph convolutional layers respectively; , They represent the trainable weight matrices of the three graph convolutional layers respectively; A is the adjacency matrix obtained by S4.1, and T is the node feature obtained by S4.1;

[0043] S4.3, the feature representation learned by the three graph convolutional layers and Splicing them together, we get the output of the final graph convolutional network, which is the graph topology feature.

[0044] S5, through the ConvLSTM network constructed in S3 and the multi-receptive field graph convolution network constructed in S4, a ConvLSTM-GCN model is formed to extract graph topology features containing spatiotemporal characteristics, referred to as graph topology features; and combined with transfer learning and domain adaptation methods, the JS divergence joint domain adversarial network is used to align the graph topology features between the source domain data and the target domain data, so that the ConvLSTM-GCN model can adapt to a variety of data distributions, including the following steps:

[0045] S5.1: The source domain data in the training data set is extracted through the graph topological features of the source domain data from S2 to S5. Represents: The target domain data in the training data set is extracted through S2 to S5. The graph topology features of the target domain data are expressed as Representation; and use the graph topological features of the source domain data and the target domain data to represent the distribution of the source domain data and the target domain data, and finally use JS divergence to perform preliminary feature alignment;

[0046] Furthermore, the basic steps for preliminary feature alignment using JS divergence are as follows:

[0047] (10)

[0048] (11)

[0049] in, Represents the JS divergence of the graph topology features of the source domain data and the graph topology features of the target domain data; and yes With M and The KL divergence between and M; M is and average;

[0050] S5.2, based on the preliminary alignment of features by JS divergence in S5.1, the idea of ​​domain adversarial learning is introduced to further align the graph topological features of the source domain data and the target domain data, that is, through the domain adversarial learning method, adversarial training is used to reduce the distribution difference between the source domain data and the target domain data. It is necessary to first define a discriminator model to distinguish the source domain data from the target domain data. The structure of the discriminator model includes two layers of fully connected networks, and a sigmoid activation function is set in the last fully connected layer so that the output value is between 0 and 1, indicating the probability that the input data belongs to the source domain data;

[0051] S5.3, define the label of source domain data as 1, and the label of target domain data as 0. For the graph topological features of source domain data and target domain data mentioned in S5.1 and , respectively through the discriminator model The calculation represents the probability that the input data is judged to be source domain data. The domain adversarial loss is calculated by the binary cross entropy loss function. as follows:

[0052] (12)

[0053] in, and are the number of source domain data and target domain data respectively; and They are graph topological features of source domain data and target domain data respectively; is the discriminator model;

[0054] Domain Adversarial Loss The purpose is to make the classification ability of the discriminator model for source domain data and target domain data as weak as possible to reduce the distribution difference between source domain data and target domain data, so that the ConvLSTM-GCN model can obtain better generalization ability.

[0055] S6, combining the loss function of the ConvLSTM-GCN model in S5 for the source domain data and the target domain data, the JS divergence of the source domain data and the target domain data, and the domain adversarial loss of the source domain data and the target domain data as the final joint loss function, and training the ConvLSTM-GCN model, including the following steps:

[0056] S6.1, the loss function of the ConvLSTM-GCN model for both source domain data and target domain data adopts the binary cross entropy loss function, the JS divergence is calculated according to S5.1, and the domain adversarial loss is calculated according to S5.3. The Adam optimizer is used to update the weights of the ConvLSTM-GCN model on the training data set; the ConvLSTM-GCN model parameters are updated through the back propagation algorithm. The final joint loss function expression is:

[0057] (13)

[0058] in, and is the loss function of the ConvLSTM-GCN model for the source domain data and the target domain data respectively; It is the JS divergence mentioned in S5.1; is the domain adversarial loss mentioned in S5.3; λ, is a hyperparameter used to balance and ;

[0059] S6.2, save the trained ConvLSTM-GCN model and test it on the test data set, adjust the hyperparameters of the ConvLSTM-GCN model according to the evaluation index, and use the F2 index as the evaluation index. The ConvLSTM-GCN model with the best evaluation index is used as the final prediction model;

[0060] The F2 index is: (14);

[0061] Among them, P is the precision, which means the proportion of samples classified as positive that are actually positive: , TP is the number of true positive examples, FP is the number of false positive examples; R is the recall rate, which indicates the proportion of positive classes in the sample that are correctly predicted: , TP is the number of true positive cases, FN is the number of false negative cases; is an important weight parameter used to adjust the relative importance of precision and recall. When >1, more emphasis is placed on recall. When <1, more emphasis is placed on accuracy, and the value is usually 2; F2-score is an indicator that comprehensively evaluates the performance of the classification model. It places more emphasis on recall rate and is suitable for scenarios where it is necessary to prioritize reducing missed reports.

[0062] S7, real-time prediction on the test dataset, includes the following steps:

[0063] S7.1, preprocess the test data set according to S1.3 to S1.5, and convert it into a recursive graph according to S2;

[0064] S7.2, using the prediction model saved in S6.2 to give the surge prediction probability of the data in each test data set.

[0065] Beneficial effects of the present invention:

[0066] (1) Compared with the previous methods, the method provided by the present invention is used to predict the surge of aircraft engines. The method of converting the pressure signal into a recursive graph is adopted to reveal the evolution of the state of the aircraft engine compressor tip pressure signal data over time and the repetitive patterns between the states;

[0067] (1) At the same time, compared with previous methods, the ConvLSTM network constructed by S3 and the graph convolutional network constructed by S4 are combined to construct a ConvLSTM-GCN model to extract graph topology features, revealing the essential characteristics of source domain data and target domain data, and improving prediction accuracy;

[0068] (2) By combining transfer learning and domain adaptation methods, the JS divergence joint domain adversarial loss method proposed by S5 can significantly enhance the generalization ability when the feature distributions of source domain data and target domain data are different. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The structural flow chart of the compressor surge precursor identification method integrating graph feature transfer and domain adaptation;

[0070] Figure 2 This is the data preprocessing flow chart;

[0071] Figure 3 A set of representative prediction results of the ConvLSTM-GCN model in the source domain data of the test dataset (surge warning time 7.49 seconds); Figure 3 (a) is the graph showing the dynamic pressure at the first-stage stator tip changing with time. Figure 3 (b) is the graph showing the change of surge prediction probability over time given by the model. Figure 3 (c) in the figure is the surge warning signal given by the model;

[0072] Figure 4 A set of representative prediction results of the ConvLSTM-GCN model in the target domain data of the test dataset (surge warning time 8.95 seconds); Figure 4 (a) is the time-varying diagram of the dynamic pressure at the tip of the first-stage stator. Figure 4 (b) is the graph showing the change of surge prediction probability over time given by the model. Figure 4 (c) in the figure is the surge warning signal given by the model. DETAILED DESCRIPTION

[0073] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This embodiment verifies the effect of the present invention through experimental data of compressor surge of two types of aircraft engines A and B. The structure of the compressor surge precursor recognition method integrating graph feature migration and domain adaptation is as follows: Figure 1 shown.

[0074] S1, preprocessing of blade tip pressure signal data of two types of aircraft engine compressors during steady-state throttling to stall. Figure 2 The data preprocessing flow chart includes the following steps:

[0075] S1.1, the tip pressure signal data is divided into source domain data and target domain data; the source domain experimental data comes from one of the models, with a total of 10 groups, each group of experiments contains 10 measurement points measuring the dynamic pressure values ​​from normal to surge for a total of 10s, the sensor measurement frequency is 6kHz, and the 10 measurement points are located at: the tip of the inlet guide vane stator, the tip of the zero-stage stator, the tip of the first-stage stator (three circumferentially), the tip of the second-stage stator, the tip of the third-stage stator, the tip of the fourth-stage stator, the tip of the fifth-stage stator, and the outlet wall; the target domain data comes from another model, and the rest are the same as the source domain data;

[0076] S1.2, divide the source domain data into 4:1 ratio, 80% as training data set and 20% as test data set; similarly, divide the target domain data into the same way;

[0077] S1.3, using Butterworth low-pass filter to filter all data in the training data set;

[0078] S1.4, downsample the filtered data; because the surge frequency is between 3 and 166 Hz, in order to meet the requirements of the Nyquist sampling theorem, downsample by 10 times;

[0079] S1.5, split the data according to the time step size, and label each sample after the split; to ensure that each time step covers the entire range of surge precursors and maximize the model operation efficiency, the time step is set to 256. Set a time window of length 256 to slide on the data processed by S1.3, and split the data falling within the time window as a sample. Assign a label of 1 or 0 to each sample based on whether there is surge or not, where the label with surge is 1 and the label without surge is 0.

[0080] S2, converting the samples processed by S1.5 into a recurrence plot (RP) to reveal the evolution of its state over time and the repetitive patterns between states; comprising the following steps:

[0081] S2.1, for each sample obtained in S1.5 , select the time delay τ and embedding dimension m adapted to this embodiment, and construct the embedding vector ; as shown below:

[0082] (1)

[0083] in, is the embedding vector generated by embedding dimension and time delay; is the sample obtained by S1.5, with a length of N; τ is the time delay, which is 2 in this embodiment; m is the embedding dimension, which is 3 in this embodiment;

[0084] S2.2, use the recursive matrix V to represent the distance between embedded vectors; the elements of the recursive matrix V are The definition is as follows:

[0085] (2)

[0086] in , , , represents the Euclidean distance; then the recursive matrix is ​​expressed as:

[0087] (3)

[0088] S2.3, define a threshold , binarize the elements in V to obtain the final recursive graph RP;

[0089] (4)

[0090] in, Represents the element in the i-th row and j-th column of the recursive matrix V; is the set threshold, and the value of this embodiment is 0.3; is the element in the i-th row and j-th column of the recursive graph RP.

[0091] S3, building a ConvLSTM network to extract the spatiotemporal features of the recursive graph RP; including the following steps:

[0092] S3.1, taking the recursive graph RP obtained in S2 as input, construct a ConvLSTM network containing two ConvLSTM layers;

[0093] Furthermore, the ConvLSTM network is a network architecture that combines a convolutional neural network (CNN) and a long short-term memory (LSTM) network, and is specifically used to process spatiotemporal sequence data. The ConvLSTM network extracts spatial features from the recursive graph through convolution operations, and captures the dynamic information of these features changing over time through the temporal dependency of LSTM. Compared with traditional CNN, the ConvLSTM network can better handle temporal dependencies and provide powerful spatiotemporal feature modeling capabilities;

[0094] S3.2, the two ConvLSTM layers have different numbers and sizes of convolution kernels; the first ConvLSTM layer has 32 convolution kernels and a convolution kernel size of , the number of convolution kernels in the second ConvLSTM layer is 64, and the convolution kernel size is ; Therefore, spatiotemporal features of different scales can be extracted; Finally, the output of the ConvLSTM network is obtained by passing a fully connected layer and activating it with the relu activation function, which can describe the spatiotemporal features and has a dimension of (6,1).

[0095] S4, extracts graph topology features from the output of the ConvLSTM network in S3 through a graph convolutional network (GCN), including the following steps:

[0096] S4.1, the output of the ConvLSTM network obtains the graph structure through the graph generation layer algorithm (GGL) to satisfy the calculation of the graph convolution network. The specific expression of the graph generation layer algorithm is:

[0097] (5)

[0098] (6)

[0099] Among them, T is the output of the ConvLSTM network, which is also used as a node feature; express The transpose of represents the normalization function; is the original adjacency matrix; further, Indicates return The indices of the first k largest values ​​of , which makes the adjacency matrix sparse and reduces the computational burden;

[0100] Finally, we get the graph structure , where A is the adjacency matrix and T is the node feature;

[0101] S4.2, build a multi-receptive field graph convolutional network, and transform the graph structure obtained in S4.1 As input to extract graph topology features; standard graph convolution GCN can only aggregate information in a fixed-size receptive field. To solve this problem, the present invention adopts a multi-receptive field graph convolutional network (MRF-GCN, Multi-receptive Field Graph Convolutional Networks) method to obtain powerful feature representation. The multi-receptive field graph convolutional network consists of three graph convolution layers with different receptive field sizes, and the three receptive fields of different sizes are denoted as K1, K2, and K3 respectively;

[0102] Graph Structure The features learned by the graph convolutional layers with receptive field sizes of K1, K2, and K3 are expressed as follows:

[0103] (7)

[0104] (8)

[0105] (9)

[0106] in, They represent the graph convolutional layers with receptive field sizes of K1, K2, and K3 respectively. and Represent the feature representations learned by the three graph convolutional layers respectively; , They represent the trainable weight matrices of the three graph convolutional layers respectively; A is the adjacency matrix obtained by S4.1, and T is the node feature obtained by S4.1;

[0107] S4.3, the feature representation learned by the three graph convolutional layers and Splicing them together, we get the output of the final multi-receptive field graph convolutional network, that is, the graph topology feature.

[0108] S5, the ConvLSTM network constructed by S3 and the multi-receptive field graph convolution network constructed by S4 form a ConvLSTM-GCN model, which can extract graph topology features containing spatiotemporal characteristics, referred to as graph topology features; and combined with transfer learning and domain adaptation methods, the JS divergence joint domain adversarial network is used to align the graph topology features between the source domain data and the target domain data in the training data set, so that the ConvLSTM-GCN model can adapt to a variety of data distributions, including the following steps:

[0109] S5.1: The source domain data in the training data set is extracted through the graph topological features of the source domain data from S2 to S5. The target domain data in the training data set is represented by the graph topology features of the target domain data extracted by S2 to S5 above. Representation, and use the graph topology features of the source domain data and the target domain data to represent the distribution of the source domain data and the target domain data; the basic steps for preliminary feature alignment using JS divergence are as follows:

[0110] (10)

[0111] (11)

[0112] in, Represents the JS divergence between the graph topology features of the source domain data and the graph topology features of the target domain data; and yes With M and The KL divergence between and M; M is and average;

[0113] S5.2, based on JS divergence, the idea of ​​domain adversarial learning is introduced to further align the graph topological features of the source domain data and the target domain data, that is, through the domain adversarial learning method, adversarial training is used to reduce the distribution difference between the source domain data and the target domain data. A discriminator model needs to be defined first. The purpose of the discriminator model is to distinguish whether the input data is from the source domain data or the target domain data. The discriminator model structure contains two layers of fully connected networks, and the sigmoid activation function is set in the last fully connected layer so that the output value is between 0 and 1, indicating the probability that the input data belongs to the source domain data;

[0114] S5.3, define the label of source domain data as 1; define the label of target domain data as 0. For the graph topological features of source domain data and target domain data mentioned in S5.1 and , respectively through the discriminator model The calculation represents the probability that the input data is judged to be source domain data. The domain adversarial loss is calculated by the binary cross entropy loss function. as follows:

[0115] (12)

[0116] in, and are the number of source domain data and target domain data respectively; and They are graph topological features of source domain data and target domain data respectively; is the discriminator model;

[0117] Domain Adversarial Loss The purpose is to make the discriminator's classification ability of source domain data and target domain data as weak as possible to reduce the distribution difference between source domain data and target domain data, so that the ConvLSTM-GCN model can obtain better generalization ability.

[0118] S6, the loss function of the source domain data and the target domain data, the JS divergence of the source domain data and the target domain data, and the domain adversarial loss of the source domain data and the target domain data of the ConvLSTM-GCN model introduced in step S5 are used as the final joint loss function to train the ConvLSTM-GCN model, including the following steps:

[0119] S6.1, the loss function of the ConvLSTM-GCN model for both the source domain data and the target domain data adopts the binary cross entropy loss function, the JS divergence is calculated according to step S5.1, and the domain adversarial loss is calculated according to step S5.3. The Adam optimizer is used to update the weights of the ConvLSTM-GCN model on the training data set; the model parameters are updated through the back propagation algorithm. The final joint loss function expression is:

[0120] (13)

[0121] in, and is the loss function of the ConvLSTM-GCN model for the source domain data and the target domain data respectively, is the JS divergence, is the domain adversarial loss mentioned in S5.3; λ, is a hyperparameter used to balance and ; In this embodiment, λ is 0.3, is 0.4;

[0122] S6.2, save the trained ConvLSTM-GCN model and test it on the test data set, adjust the model hyperparameters according to the evaluation index, and use the F2 index as the evaluation index. The ConvLSTM-GCN model with the best evaluation index is used as the final prediction model;

[0123] The F2 index is: (14);

[0124] Among them, P is the precision, which means the proportion of samples classified as positive that are actually positive: , TP is the number of true positive examples, FP is the number of false positive examples; R is the recall rate, which indicates the proportion of positive classes in the sample that are correctly predicted: , TP is the number of true positive cases, and FN is the number of false negative cases.

[0125] S7, real-time prediction on the test dataset, includes the following steps:

[0126] S7.1, preprocess the test data set according to S1.3 to S1.5, and convert it into a recursive graph according to S2;

[0127] S7.2, using the prediction model saved in S6.2 to give the surge prediction probability of the data in each test data set. In the test data set, a group of representative data from the source domain and the target domain are selected respectively as follows Figure 3 and Figure 4 As shown. Figure 3(a) is the time-varying graph of the dynamic pressure of the first-stage stator tip in the source domain data of the test data set. Figure 3 (b) is the graph showing the change of surge prediction probability over time given by the prediction model. Figure 3 (c) in the figure is the surge warning signal given by the prediction model. Figure 4 (a) is the time-varying diagram of the dynamic pressure at the first-stage stator tip in the target domain of the test data set. Figure 4 (b) is the graph showing the change of surge prediction probability over time given by the prediction model. Figure 4 (c) in the figure is the surge warning signal given by the prediction model.

[0128] S7.3, Figure 3 (a) and Figure 4 (a) in the figure shows the pressure change diagrams of the representative source domain and target domain in the test dataset. As can be seen from the figure, Figure 3 The dynamic pressure in (a) begins to develop an upward spike at around 7.49s, which is in the initial stage of surge precursor. As the surge precursor develops, it begins to fluctuate violently at 7.82s and completely develops into stall surge. Figure 4 The dynamic pressure in (a) begins to show periodic modal waves at around 8.95s, which is also the initial stage of the precursor. The difference is that there is no characteristic Figure 3 (a) in the figure is clear. With the development of surge precursor, it begins to fully develop into stall surge at 9.57s.

[0129] S7.4, using the ConvLSTM-GCN model to give the surge prediction probability of two representative data, observe the prediction results Figure 3 (b) and Figure 4 In (b), we can see that the surge probability rises rapidly in the initial disturbance stage, and a surge warning signal is given after exceeding the threshold value based on actual engineering experience. After that, the surge probability remains at a very high value, and the probability increases to 100% when surge occurs. Therefore, the ConvLSTM-GCN model can respond to small changes in the surge precursor stage in a timely manner; at the same time, for data with different feature distributions in the two domains, the ConvLSTM-GCN model can give timely surge warnings, indicating that the model has excellent generalization ability.

[0130] The above-described embodiments merely express the implementation methods of the present invention, but they should not be understood as limiting the scope of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A compressor surge precursor recognition method integrating graph feature migration and domain adaptation, characterized in that: The compressor surge precursor identification method comprises the following steps: S1, preprocessing the tip pressure signal data of different types of compressors of aircraft engines during the process from steady-state throttling to stalling to obtain multiple samples; S2, converts each sample obtained in S1 into a recurrence graph to reveal the evolution of its state over time and the repetitive patterns between states; S3, build a ConvLSTM network to extract the spatiotemporal features of the recursive graph RP; S4, the output of the ConvLSTM network is used to extract graph topology features through a graph convolutional network; S5, the ConvLSTM network constructed by S3 and the multi-receptive field graph convolution network constructed by S4 form a ConvLSTM-GCN model to extract graph topology features containing spatiotemporal characteristics, referred to as graph topology features; and combined with transfer learning and domain adaptation methods, the JS divergence joint domain adversarial network is used to align the graph topology features between the source domain data and the target domain data, so that the ConvLSTM-GCN model can adapt to a variety of data distributions; S6, combining the loss function of the ConvLSTM-GCN model in S5 for the source domain data and the target domain data, the JS divergence of the source domain data and the target domain data, and the domain adversarial loss of the source domain data and the target domain data as the final joint loss function, and training the ConvLSTM-GCN model to obtain the final prediction model; S7, real-time prediction on the test dataset: S7.1, preprocess the test data set and convert it into a recursive graph according to S2; S7.2, using the prediction model obtained in S6, gives the surge prediction probability of the data in each test data set.

2. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: The S1 comprises the following steps: S1.1, dividing the blade tip pressure signal data into source domain data and target domain data; S1.2, divide the source domain data into a training data set and a test data set; divide the target domain data into a training data set and a test data set; S1.3, using a low-pass filter to filter all the data in the training data set obtained in S1.2; S1.4, downsampling the data after filtering in S1.3; S1.5, divide the data according to the time step size, and label each sample after the division; assign a label of 1 or 0 to each sample according to whether there is surge or not, where the label with surge is 1 and the label without surge is 0.

3. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: In S2, the specific steps of converting into a recursive graph are as follows: S2.1, for the sample obtained in S1 , select the time delay τ and embedding dimension m, and construct the embedding vector ; as shown below: （1） in, is the embedding vector generated by embedding dimension and time delay; is the sample obtained by S1.5, with a length of N; τ is the time delay; m is the embedding dimension; S2.2, use the recursive matrix V to represent the distance between embedded vectors; the elements of the recursive matrix V are The definition is as follows: （2） in, , , , represents the Euclidean distance; then the recursive matrix is ​​expressed as: （3） S2.3, define a threshold , binarize the elements in the recursive matrix V to obtain the final recursive graph RP; （4） in, Represents the element in the i-th row and j-th column of the recursive matrix V; is the set threshold; is the element in the i-th row and j-th column of the recursive graph RP.

4. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 3 is characterized in that: In S2.3, the threshold The range is 0.1 to 0.

5.

5. The compressor surge precursor identification method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: The S3 comprises the following steps: S3.1, taking the recursive graph RP obtained in S2 as input, constructing a ConvLSTM network including two ConvLSTM layers; wherein the two ConvLSTM layers have different numbers of convolution kernels and sizes of convolution kernels; S3.2, the output of the ConvLSTM network is obtained through a fully connected layer and a relu activation function to describe the spatiotemporal features.

6. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: The S4 comprises the following steps: S4.1, the output of the ConvLSTM network obtains the graph structure through the graph generation layer algorithm to meet the calculation of the graph convolution network; finally, the graph structure is obtained , where A is the adjacency matrix and T is the node feature; the specific expression of the graph generation layer algorithm is: （5） （6） Among them, T is the output of the ConvLSTM network, which is also used as a node feature; express The transpose of represents the normalization function; is the original adjacency matrix; further, return The indices of the first k largest values ​​of , which makes the adjacency matrix sparse and reduces the computational burden; S4.2, build a multi-receptive field graph convolutional network, and transform the graph structure obtained in S4.1 As the input of a multi-receptive field graph convolutional network to extract graph topological features; the standard graph convolutional GCN can only aggregate information in a fixed-size receptive field; the multi-receptive field graph convolutional network consists of three graph convolutional layers with different receptive field sizes, and the three receptive fields of different sizes are denoted as K1, K2, and K3 respectively; Graph Structure The features learned by the graph convolutional layers with receptive field sizes of K1, K2, and K3 are expressed as follows: （7） （8） （9） in, They represent the graph convolutional layers with receptive field sizes of K1, K2, and K3 respectively. and Represent the feature representations learned by the three graph convolutional layers respectively; , They represent the trainable weight matrices of the three graph convolutional layers respectively; A is the adjacency matrix obtained by S4.1, and T is the node feature obtained by S4.1; S4.3, the feature representation learned by the three graph convolutional layers and Splicing them together, we get the output of the final graph convolutional network, which is the graph topology feature.

7. The compressor surge precursor identification method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: The S5 comprises the following steps: S5.1: The source domain data in the training data set is extracted through the graph topological features of the source domain data from S2 to S5. Represents: The target domain data in the training data set is extracted through S2 to S5. The graph topology features of the target domain data are expressed as Representation; and use the graph topological features of the source domain data and the target domain data to represent the distribution of the source domain data and the target domain data, and finally use JS divergence to perform preliminary feature alignment; S5.2, based on the preliminary alignment of features by JS divergence in S5.1, the idea of ​​domain adversarial learning is introduced to further align the graph topological features of source domain data and target domain data. Through the domain adversarial learning method, adversarial training is used to reduce the distribution difference between source domain data and target domain data. First, a discriminator model is defined to distinguish source domain data from target domain data. S5.3, define the label of source domain data as 1, and the label of target domain data as 0; for the graph topological features of source domain data and target domain data mentioned in S5.1 and , respectively through the discriminator model The calculation represents the probability that the input data is judged to be source domain data; the domain adversarial loss is calculated by the binary cross entropy loss function as follows: （12） in, and are the number of source domain data and target domain data respectively; and They are graph topological features of source domain data and target domain data respectively; is the discriminator model.

8. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 1 is characterized in that: In S5.1, the steps of performing preliminary feature alignment using JS divergence are as follows: （10） （11） in, Represents the JS divergence of the graph topology features of the source domain data and the graph topology features of the target domain data; and yes With M and The KL divergence between and M; M is and average; In S5.2, the structure of the discriminator model includes two layers of fully connected networks, and a sigmoid activation function is set in the last fully connected layer so that the output value is between 0 and 1, indicating the probability that the input data belongs to the source domain data.

9. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 7 is characterized in that: The S6 comprises the following steps: S6.1, the loss function of the ConvLSTM-GCN model for both source domain data and target domain data adopts the binary cross entropy loss function, calculates the JS divergence according to S5.1, calculates the domain adversarial loss according to S5.3, and uses the Adam optimizer to update the weights of the ConvLSTM-GCN model on the training data set; updates the ConvLSTM-GCN model parameters through the back propagation algorithm; the final joint loss function expression is: （13） in, and is the loss function of the ConvLSTM-GCN model for the source domain data and the target domain data respectively; It is the JS divergence mentioned in S5.1; is the domain adversarial loss mentioned in S5.3; λ, is a hyperparameter used to balance and ; S6.2, save the trained ConvLSTM-GCN model and test it on the test data set, adjust the hyperparameters of the ConvLSTM-GCN model according to the evaluation index, and use the F2 index as the evaluation index. The ConvLSTM-GCN model with the best evaluation index will be used as the final prediction model.

10. The compressor surge precursor recognition method integrating graph feature migration and domain adaptation according to claim 9 is characterized in that: In S6.2, the F2 index is: ; Among them: P is the precision rate, which means the proportion of samples classified as positive that are actually positive: , TP is the number of true positive cases, FP is the number of false positive cases; R is the recall rate, which indicates the proportion of positive classes in the sample that are correctly predicted: , TP is the number of true positive cases, FN is the number of false negative cases; is the weight parameter.

Images