Intelligent early warning method for surge fault of multi-stage high-pressure-ratio axial flow compressor of aero-engine

Abstract

The invention discloses an intelligent surge fault early warning method for a multi-stage high-pressure-ratio axial flow compressor of an aero-engine, and belongs to the technical field of aero-engine fault prediction. Firstly, pressure signals of the aero-engine in various operation states are collected to obtain original data, and a graph network is constructed; secondly, preprocessing the original data, and dividing the original data into a training set and a verification set; then, constructing a graph attention physical information neural network model composed of a graph attention layer, a probability prediction network and a physical information prediction network; training a graph attention physical information neural network model; then, verifying and finely adjusting the graph attention physical information neural network model; and finally, surge fault prediction is carried out on the multi-stage high-pressure-ratio axial flow compressor, and a prediction result is output to a surge early warning system. According to the method provided by the invention, the generalization ability of the model can be remarkably improved; an early warning signal can be given in advance, and a high-reliability solution is provided for surge prevention and control of the multi-stage high-pressure-ratio axial flow compressor of the aero-engine.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aero-engine fault diagnosis, and relates to an intelligent early warning method for a surge fault of a multistage high-pressure-ratio axial compressor of an aero-engine. BACKGROUND

[0002] An aero-engine is a core power device of a modern aircraft, and is a high-precision complex thermal mechanical system integrating aerodynamic thermodynamics, structural mechanics, material science and control engineering. The running stability of the aero-engine directly affects the propulsion efficiency and economy of the aircraft, and is a decisive factor for flight safety and mission reliability. Among various engine fault modes, the aero-engine compressor surge fault is concerned due to its suddenness and destructiveness, and is considered as one of the most threatening fault modes of an aviation propulsion system. The surge phenomenon is a strong nonlinear and unsteady aerodynamic dynamics problem caused by flow instability in the compressor. When the surge occurs, the flow field in the engine changes in nature. Firstly, a rotating stall is formed in the compressor section, and then a large-scale flow separation occurs in the whole circumferential direction, which causes severe pressure fluctuations and significant reverse flow. This flow instability can cause the engine thrust to drop sharply within a millisecond time scale, and can also induce secondary faults such as high-cycle fatigue and flutter of the compressor blades. More seriously, the surge-induced sudden change of the combustion chamber inlet pressure can cause flameout, and in extreme cases, can even cause abnormal contact of the rotor system in the radial direction or structural damage of the whole engine. Therefore, it is of great significance for the stable and efficient operation of the aircraft to predict the surge before it occurs and take timely measures.

[0003] Due to the strong coupling, multi-scale and high dynamic characteristics of the internal flow field of the aero-engine compressor, and the millisecond-level rapid evolution characteristics of the surge phenomenon, it is difficult for traditional monitoring methods to achieve effective early warning in engineering practice. Specifically, the suddenness and nonlinearity of the surge cause the precursor signals to be weak and easily masked by noise, and the dynamic characteristics of the engine under complex working conditions further increase the difficulty of state recognition. Therefore, an early warning method capable of accurately identifying abnormal operating states at the initial stage of the surge or even before the surge occurs is urgently needed to gain a critical time window for active flow control such as bleed valve adjustment, so as to avoid disastrous consequences.

[0004] The traditional surge fault warning methods used in the field of aeronautical engineering are mainly divided into two categories: threshold-based judgment and signal processing-based analysis methods. However, both methods have significant limitations. The traditional threshold-based method triggers an alarm by setting fixed thresholds for parameters such as compressor outlet pressure and flow rate. Although this method is simple and easy to implement, it relies heavily on expert experience and cannot adapt to the dynamic characteristics of engine operation under varying conditions. It is prone to false positives or false negatives during non-steady-state transition processes. The signal processing-based analysis method includes time-frequency analysis techniques such as fast Fourier transform, wavelet transform, and Kalman filter. It achieves early warning by extracting the frequency domain features of the surge precursor. In document [1] {Liu Y, Li J, Du J. Application of Fast Wavelet Analysis on Early Stall Warning in Axial Compressors [J]. Journal of Thermal Science, 2019, Volume 28, pages 837–849}, the fast wavelet analysis is used to study the early surge fault warning of axial compressors. By performing wavelet transform on dynamic pressure signals, characteristic signals before surge are identified, and early warning of surge failure is achieved. In document [2] {Xia K L, Zhu M M, Feng JD. Numerical Research on Near Stall Characteristics of a Transonic Axial Compressor Based on Wavelet Analysis [J]. Aerospace Systems, 2024, Volume 7, pages 509–523}, numerical simulation and wavelet analysis are used to study the characteristics of a transonic axial compressor near surge. Time-frequency analysis of pressure signals using wavelet transform reveals the flow instability characteristics before surge. Although this method can capture non-stationary signal features to some extent, it has high computational complexity and cannot meet the real-time requirements of onboard systems. It also relies on an accurate engine aerodynamic model to determine the characteristic frequency band, and model errors in actual applications will lead to performance degradation.

[0005] In recent years, with the development of artificial intelligence technology, data-driven surge fault early warning methods have gradually become a research hotspot, mainly including gated recurrent neural networks, convolutional neural networks, and long short-term memory networks. Literature [3] = {Zhang K X, Lin B, Chen J X. Aero-Engine Surge Fault Diagnosis Using Deep Neural Network [J]. Computer Systems Science and Engineering, 2022, 42(1):351-360.} This study proposes a neural network fusion method for aero-engine surge fault detection. This model combines sequence-to-sequence, convolutional neural networks, and long short-term memory networks, and can perform multi-dimensional and multi-step prediction on multi-sensor time series data, thereby achieving effective detection of surge faults. Literature [4] = {Tan R, Liu Z H, Huang B. Aero Engine Instability Prediction and Detection Method Based on Gated Recurrent Neural Networks [C] / / Chinese Intelligent Systems Conference, 2024.} This paper proposes a method for predicting aero-engine surge faults based on gated recurrent neural networks. Compared with traditional wavelet analysis methods, this method can more stably and accurately predict long-term data while preserving local surge signal characteristics, making it suitable for long-term monitoring of engines. This type of method uses machine learning models to mine surge features from historical data, reducing the dependence on physical models to some extent, but requires a large amount of labeled surge experimental data for training. However, aero-engine surge experiments are extremely costly and risky, resulting in a scarcity of available samples. Secondly, purely data-driven models perform poorly in conditions not covered by the training data, and the model decision-making process lacks physical knowledge support. Therefore, such methods have inherent defects such as data dependence, insufficient generalization, and physical unexplainability.

[0006] The present application proposes a new fault early warning method for multi-stage high-pressure ratio axial compressor surge faults in aero-engines. This method uses a graph attention physics-informed neural networks (GAPINN) model that can fuse physical information and data labels, allowing accurate prediction of aero-engine multi-stage high-pressure ratio axial compressor surge faults. SUMMARY

[0007] To address the shortcomings of existing technologies, this invention provides a method for early warning of surge faults in multi-stage high-pressure ratio axial compressors of aero-engines using a graph attention physical information neural network model. This method can determine whether an aero-engine will enter a surge state based on pressure sensor data.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for intelligent early warning of surge faults in a multi-stage high-pressure ratio axial compressor of an aero-engine includes the following steps: First, pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine are used to collect pressure signals under various operating conditions of the aero-engine to obtain raw data, and a graph network is constructed based on the Pearson correlation coefficient; Second, the collected raw data is preprocessed, including data labeling and normalization based on pressure signal change characteristics analysis, to obtain preprocessed data, which is then divided into training and validation sets; Then, a graph attention physical information neural network model is constructed, consisting of a graph attention layer, a probability prediction network, and a physical information prediction network; The graph attention physical information neural network model is trained using the data from the training set; Then, the graph attention physical information neural network model is validated and fine-tuned using the validation set; Finally, the trained graph attention physical information neural network model is used to predict surge faults in the multi-stage high-pressure ratio axial compressor of the aero-engine, and the prediction results are output to the aero-engine surge early warning system. Details are as follows:

[0010] Step S1: Using pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine, pressure signals under various operating conditions of the aero-engine are collected to obtain raw data, and a graph network is constructed based on the Pearson correlation coefficient; specifically:

[0011] First, multiple pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine are used to collect pressure signals under various operating conditions of the aero-engine. The signal data collected by each pressure sensor is defined as a one-dimensional continuous pressure variable. Second, the obtained dataset is analyzed, and a dataset containing the normal operating conditions and abnormal operating conditions of the aero-engine is selected as a sample set for fault prediction. Abnormal operating conditions include rotating stall and surge. Third, each continuous pressure variable is regarded as a node. The Pearson correlation coefficient is used to measure the correlation between the continuous pressure variables, and a correlation threshold is set. If the correlation between two continuous pressure variables is higher than the set correlation threshold, the nodes corresponding to the two continuous pressure variables are considered to be connected, and an edge structure is set for them in the graph network. Otherwise, no edge structure is set, resulting in a graph network structure. Nodes connected by the edge structure are neighbor nodes.

[0012] The Pearson correlation coefficient measures continuous variables. and continuous variables The degree of correlation is shown in equation (1);

[0013] (1)

[0014] in, The sample size for continuous variables; Pearson correlation coefficient; Representing continuous variables No. One variable value; Representing continuous variables No. One variable value; Representing continuous variables The mean; Representing continuous variables The mean;

[0015] Step S2: Preprocess the collected raw data to transform it into a format suitable for the graph attention physical information neural network model. Preprocessing includes data labeling and normalization based on pressure signal change feature analysis, resulting in preprocessed data that is then divided into training and validation sets. Specifically:

[0016] First, signal analysis is performed on the continuous pressure signal variables of the raw data obtained in step S1. Based on their changes and traditional experience, the engine operating state is divided into normal operating state and abnormal operating state. Sample points corresponding to the engine in normal operating state are labeled with a true label "0", and sample points corresponding to the engine in abnormal operating state are labeled with a true label "1". Second, a time sequence is added to the raw data. Then, the raw data is normalized through linear transformation, and the normalized time series is denoted as... Finally, the dataset is divided into training datasets. and validation dataset Both datasets are shuffled, and the splitting process must ensure that both datasets contain data with different labels.

[0017] Step S3: Construct a graph attention-based physical information neural network model, which consists of a graph attention layer, a probability prediction network, and a physical information prediction network; the specific construction process is as follows:

[0018] Step S3.1: Construct the graph attention layer;

[0019] First, based on the graph network structure obtained in step S1, the nodes corresponding to each continuous pressure variable are obtained. The neighboring nodes are identified, and a feature vector with initial features is selected for each node. Select a shared weight matrix For each feature vector A linear transformation is performed to map the original node features to a new shared feature space, as shown in Equation (2);

[0020] (2)

[0021] In equation (2), The weight matrix is ​​a learnable matrix; It is a node The representation of in the shared feature space;

[0022] Then, the node The transformed features from neighboring nodes are vertically concatenated along the feature dimension, and the concatenated feature vector is multiplied by a learnable attention vector before being processed. Activation function obtains nodes The original attention coefficients of each neighbor node are shown in Equation (3);

[0023] (3)

[0024] In equation (3), For nodes with neighboring nodes The original attention coefficients between nodes represent the nodes' attention coefficients. right The importance of; This is a learnable attention vector; This represents the vertical concatenation operation of vectors; Represents a node Features after linear transformation; Represents a node neighboring nodes Features after linear transformation; Indicates use Activation function;

[0025] Secondly, the softmax function is used to normalize all the original attention coefficients to obtain the standardized attention weights, as shown in Equation (4);

[0026] (4)

[0027] In equation (4), Represents a node The set of all neighboring nodes; Represents a set All neighboring nodes; Represents a node Its neighboring nodes The original attention coefficient; Represents a node with neighboring nodes Standardized attention weights between them;

[0028] Finally, for each node Weighted feature aggregation is performed, the features of all its neighboring nodes after linear transformation are multiplied by the corresponding standardized attention weights, and then added together. Nonlinearity is introduced through the ELU activation function. The graph attention layer adopts a multi-head graph attention mechanism, which uses multiple independent attention heads to process nodes and averages the outputs of each attention head to obtain the output feature vector, as shown in Equation (5).

[0029] (5)

[0030] In equation (5), Represents a node Feature vectors after processing by the graph attention layer; The number of attention heads; Indicates the first One point of attention; Represents a node The set of all neighboring nodes; Represents a set All neighboring nodes; Indicates the first The node obtained by the attention head calculation with neighboring nodes Standardized attention weights between them; Indicates the first Neighbor nodes obtained by attention head calculation Features after linear transformation;

[0031] Step S3.2: Construct a probabilistic prediction network;

[0032] The probabilistic prediction network is a multilayer perceptron used to receive the feature vector output by the graph attention layer and predict the unnormalized probability value of the engine being in an abnormal operating state at the corresponding sample point at the current time. First, all nodes... The feature vector obtained after graph attention layer processing With normalized time series The data is horizontally stitched together; then, each row of the stitched result is input into the multilayer perceptron as the input feature at each time step; finally, the hierarchical nonlinear transformation of the multilayer perceptron maps the input features onto the prediction result.

[0033] The core function of the probabilistic prediction network can be summarized as performing a composite function mapping as shown in equation (6);

[0034] (6)

[0035] In equation (6), This represents the unnormalized probability value predicted by the probability prediction network as being in an abnormal operating state at the current moment; This represents the total number of layers in the multilayer perceptron; Indicates the first The learnable weight matrix of the layer; Indicates the first Learnable bias vectors of the layer; Indicates the first The non-linear activation function used in the layer is the Sin function; Representing the eigenvector With normalized time series Each row of the matrix obtained by horizontal concatenation represents the input features of the probability prediction network at each time step.

[0036] Step S3.3: Construct a physical information prediction network;

[0037] The physical information prediction network is a multilayer perceptron, which receives the output value of the probability prediction network. , Partial derivative of pressure signal, Using the partial derivative of the time signal and the preprocessed data obtained in step S2, a physical information value is predicted. This value is used to approximate the relationship between the unnormalized true probability value of the engine being in an abnormal operating state at the current moment and the time, and the relationship satisfies the partial differential equation shown in equation (7).

[0038] (7)

[0039] In equation (7), This represents the unnormalized true probability value of the engine being in an abnormal operating state at the current moment. This represents the change of the unnormalized true probability value of the engine being in an abnormal operating state at the current moment over time. This represents the true probability that the engine is currently in an abnormal operating state, and the pattern that follows during engine operation. This represents the normalized time series; The number of pressure signal features; Indicates the first The normalized pressure signal features; the physical information prediction network uses a more generalized function. To approximate the function That is, it satisfies the change relationship shown in equation (8);

[0040] (8)

[0041] In equation (8), This represents the relationship between the unnormalized predicted probability value of the engine being in an abnormal operating state at the current moment and time, which is also the predicted value of the physical information prediction network. Indicates the function used for approximation. A more generalized function; The number of pressure signal features; Indicates the first The characteristics of the normalized pressure signal; This represents the normalized time series; To predict the output value of the network using probability; express For the first Partial derivative of a pressure signal; express Partial derivatives with respect to the time series;

[0042] First, the output value of the probability prediction network is... , Partial derivative of pressure signal, The partial derivative of the time signal and the horizontal splicing of the preprocessed data obtained in step S2 are shown in equation (9).

[0043] (9)

[0044] In equation (9), This indicates the result of horizontal stitching; The number of pressure signal features; Indicates the first The characteristics of the normalized pressure signal; This represents the normalized time series; To predict the output value of the network using probability; express For the first Partial derivative of a pressure signal; express Partial derivatives with respect to the time series;

[0045] Then, each row of data in the horizontal data stitching result is used as the input of the physical information prediction network at each moment. The input is mapped to the prediction result of physical information through the hierarchical nonlinear transformation of the multilayer perceptron. This physical information is used to approximate the relationship between the unnormalized prediction probability value of the engine being in an abnormal operating state at the current moment and the change over time.

[0046] The core function of the physical information prediction network can be summarized as executing the functional relationship shown in equation (10);

[0047] (10)

[0048] In equation (10), This represents the relationship between the unnormalized predicted probability value of the engine being in an abnormal operating state at the current moment and time, which is also the predicted value of the physical information prediction network. This represents the input to the physical information prediction network at each time step, which is each row of data in the horizontally stitched result; This represents the total number of layers in the multilayer perceptron in the physical information prediction network; Indicates the first The learnable weight matrix of the layer; Indicates the first Learnable bias vectors of the layer; Indicates the first The non-linear activation function used in the layer is the Sin function;

[0049] Step S4: Train the graph attention physical information neural network model;

[0050] The training objective of the graph attention physical information neural network model is to minimize the overall error. The overall error includes data error terms. and physical information error term Its expression is shown in equations (11), (12) and (13);

[0051] (11)

[0052] (12)

[0053] (13)

[0054] In equation (11), This is the data error term, in the form of cross-entropy loss; The actual label set in step S2; The output value of the probability prediction network; in equation (12), This is the physical information error term; express Time series The partial derivative; This represents the predicted value of the physical information prediction network; in equation (13), It is a comprehensive error; It is the physical information constraint coefficient, which is adjusted. The physical information error term can be changed. The degree of constraint imposed on the model during model training;

[0055] The training process is as follows: First, based on the training dataset in step S2... The size of the parameters is used to select appropriate hyperparameters for model training, including training batch size, number of iterations, learning rate, and initial weights. Then, the Adam optimization algorithm based on the principle of stochastic gradient descent is used to optimize the graph attention physical information neural network model, that is, to minimize the overall error. Secondly, input the training dataset. Model training is then performed; finally, the model training is complete, and the optimal model parameters are obtained, including the graph attention layer weight matrix. Learnable attention vectors Weight matrices and bias vectors in probabilistic prediction networks and weight matrices and bias vectors in physical information prediction networks;

[0056] Step S5: Validation and fine-tuning of the graph attention physical information neural network model;

[0057] Step S5.1, firstly, the validation dataset from step S2... Input the graph attention physical information neural network model trained in step S4, and use the sigmoid activation function to adjust the output of the probability prediction network. Normalization is performed to obtain the predicted probability value of the engine being in an abnormal operating state at the current moment;

[0058] Step S5.2, then, set a probability threshold, the probability threshold being in the range of 0.45 to 0.55;

[0059] When the predicted probability value obtained in step S5.1 is greater than the probability threshold, the engine is defined as being in an abnormal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "1"; when the predicted probability value is less than the probability threshold, the engine is defined as being in a normal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "0".

[0060] Compare all the predicted labels obtained with all the real labels set in step S2, and calculate the prediction accuracy of the graph attention physical information neural network model.

[0061] Step S5.3: Adjust the parameters of the graph attention physical information neural network model based on the prediction accuracy.

[0062] Step S5.4 Finally, set a prediction accuracy improvement threshold for the model, and continue to repeat the model validation and fine-tuning. When the improvement of the model's prediction accuracy is lower than the improvement threshold, terminate this process to obtain the trained graph attention physical information neural network model.

[0063] Step S6: Early warning of surge fault in multi-stage high-pressure ratio axial compressor of aero-engine based on trained graph attention physical information neural network model;

[0064] The trained graph attention physical information neural network model is loaded into the early warning equipment of the aircraft engine. The corresponding sensor data is used as the input of the equipment. The early warning equipment performs real-time calculations to predict the probability that the aircraft engine is currently in an abnormal operating state. If the predicted probability value determines that the engine is in an abnormal operating state, a surge warning alarm is issued, and the crew takes countermeasures based on the alarm.

[0065] The beneficial effects of this invention are:

[0066] (1) Compared with the existing early warning technology for compressor surge of aero-engines, the method proposed in this invention uses graph attention method to integrate the correlation between different sensor signals of the collected data into the graph attention physical information neural network model. In addition, physical information and data labels are integrated for model training, and the physical information error term is embedded as a constraint term of the neural network into the loss function, so that the graph attention physical information neural network model can simultaneously satisfy the accuracy of data fitting and the consistency of physical laws, and significantly improve the generalization ability of the model.

[0067] (2) The present invention can predict the probability of the engine being in an abnormal operating state based on the pressure sensor data of the multi-stage high pressure ratio axial compressor of the aero-engine, and then determine whether the engine will enter the surge state, and give an early warning signal in advance, providing a highly reliable solution for surge prevention and control of multi-stage high pressure ratio axial compressor of aero-engine. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method for predicting surge faults in multi-stage high-pressure ratio axial compressors of aero-engines using a graph attention physical information neural network model, provided by the present invention, for early warning of surge faults in multi-stage high-pressure ratio axial compressors of aero-engines;

[0069] Figure 2 This is a data graph of the pressure sensor installed at the inlet of the multi-stage high-pressure ratio axial compressor of the aero-engine in dataset A. The pressure data has been declassified by the research institute.

[0070] Figure 3 A schematic diagram showing the edges between nodes in a graph network structure, based on the correlation between nodes corresponding to continuous pressure variables;

[0071] Figure 4This is a waveform diagram of the 500,000th to 570,000th sampling point segment extracted from the data of the pressure sensor installed at the inlet of the multi-stage high-pressure ratio axial compressor of the aero-engine in dataset A after low-pass filtering.

[0072] Figure 5 This is a graph attention physical information neural network model architecture diagram used for early warning of surge faults in multi-stage high-pressure ratio axial compressors of aero-engines;

[0073] Figure 6 The training set selected from dataset A After training the graph attention physical information neural network model, input the complete dataset A into the trained graph attention physical information neural network model in chronological order without shuffling to obtain the prediction result graph.

[0074] Figure 7 It is Figure 6 After converting the sampling points to time, the region from 12.6s to 13.6s was magnified, and the result diagram was obtained by adding a dashed line indicating the time point when the predicted probability first exceeded the probability threshold.

[0075] Figure 8 The training set selected from dataset A After the graph attention physical information neural network model is trained, the complete dataset B is input into the trained graph attention physical information neural network model in chronological order without shuffling to obtain the prediction result graph.

[0076] Figure 9 It is Figure 8 After converting the sampling points to time, the region from 12.6s to 13.6s was magnified, and the result diagram was obtained by adding a dashed line indicating the time point when the predicted probability first exceeded the probability threshold. Detailed Implementation

[0077] To provide a more detailed description of the technical solution and advantages proposed in this invention, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings and experimental data.

[0078] This implementation presents a method for predicting surge faults in multi-stage high-pressure ratio axial compressors of aero-engines using a graph attention physical information neural network model. Figure 1 The method described is used for early warning of surge faults in multi-stage high-pressure ratio axial compressors of aero-engines, and includes the following steps.

[0079] Step S1: Using pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine, pressure signals under various operating conditions of the aero-engine are collected to obtain raw data, and a graph network is constructed based on the Pearson correlation coefficient; specifically:

[0080] This invention is based on the aero-engine breathing test data provided by Dalian University of Technology in conjunction with an aero-engine research institute, and details the execution process of the method.

[0081] First, a forced-breathing test was conducted on a certain type of aero-engine. Multiple pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine were used to collect pressure signals under various operating conditions of the aero-engine. The signal data collected by each sensor was defined as a one-dimensional continuous pressure variable. A total of 11 sets of experimental data were collected. Each set of experimental data contained data from 38 pressure sensors. The pressure data has been declassified by the research institute.

[0082] Secondly, analyzing the obtained dataset, dataset A, which includes both normal and abnormal operating states of the aero-engine, was selected as the sample set. Dataset A consists of data obtained from a forced-surge experiment conducted on the aero-engine at maximum speed. Dataset A contains 614,400 sampling points at a sampling rate of 40,960 Hz. The data from the pressure sensor installed at the inlet of the multi-stage high-pressure ratio axial compressor of the aero-engine in dataset A is shown in the attached figure. Figure 2 As shown;

[0083] Next, each continuous pressure variable is treated as a node. The Pearson correlation coefficient is used to measure the correlation between the continuous pressure variables in dataset A, and a correlation threshold of 0.5 is set. If the correlation between two continuous pressure variables is higher than 0.5, the nodes corresponding to these two continuous pressure variables are considered connected, and an edge structure is set for them in the graph network; otherwise, no edge structure is set, resulting in the graph network structure. The edge settings in the graph network structure are shown in the attached figure. Figure 3 As shown, the red squares represent nodes whose corresponding horizontal and vertical coordinates have edge structures, while the blue squares represent nodes whose corresponding horizontal and vertical coordinates do not have edge structures.

[0084] Step S2: Preprocess the collected raw data to transform it into a format suitable for the graph attention physical information neural network model. Preprocessing includes data labeling and normalization based on pressure signal change feature analysis, resulting in preprocessed data that is then divided into training and validation sets. Specifically:

[0085] First, a sixth-order Butterworth low-pass filter is designed to perform low-pass filtering on the continuous pressure signal variables of the raw data obtained in step S1, with a cutoff frequency of 200Hz. The results of filtering the data from the pressure sensor installed at the inlet of the multi-stage high-pressure ratio axial compressor of the aero-engine in dataset A are shown in the attached figure. Figure 4As shown, the 554717th sampling point, which is the 13.54s point where the pressure signal intensity significantly increases, is selected as the starting point of the surge. The sample points corresponding to the data from sampling point 528835 to 614400 are defined as abnormal operating states that include rotational stall state and surge state and are marked with the real label "1". The sample points corresponding to the data from sampling point 0 to 528835 are defined as normal operating states and are marked with the real label "0".

[0086] Secondly, the sequence from 0 to 614400 is added to the original data as a sequence representing the time order;

[0087] Then, the data is linearly transformed using the maximum-minimum normalization method. Within the range;

[0088] Finally, 80% of the data from the normalized dataset was randomly selected as the training dataset. The remaining 20% ​​of the data was used as a validation dataset. Both datasets were shuffled to ensure that both datasets contained data on normal and abnormal engine operation states during the partitioning process.

[0089] Step S3: Construct a graph attention-based physical information neural network model, which consists of a graph attention layer, a probability prediction network, and a physical information prediction network; the specific construction process is as follows:

[0090] The graph attention physical information neural network model architecture proposed in step S3 of the invention is attached. Figure 5 As shown, in specific implementation, a graph attention layer, a probability prediction network, and a physical information prediction network are constructed sequentially.

[0091] Step S3.1: Construct the graph attention layer;

[0092] First, based on the graph network structure obtained in step S1, the nodes corresponding to each continuous pressure variable are obtained. The neighboring nodes;

[0093] Then, for each continuous pressure variable, it is selected as its own initial eigenvector, and a shared weight matrix is ​​initialized using the Kaiming initialization method. Perform a linear transformation on each initial eigenvector;

[0094] Secondly, for each node All nodes The transformed features of the neighbor nodes are vertically concatenated along the feature dimension, and the concatenation result is multiplied by a learnable attention vector initialized using the Xavier initialization method. and after Activation function obtains nodes The original attention coefficients of each neighbor node are normalized using the softmax function to obtain standardized attention weights.

[0095] Finally, for each node Weighted feature aggregation is performed by multiplying the transformed features of all its neighboring nodes by their corresponding standardized attention weights, then summing them. A non-linearity is introduced through the ELU activation function. The graph attention layer employs a multi-head graph attention mechanism, using three independent attention heads to process nodes, and averaging the outputs of each attention head to obtain 38 output feature vectors. ;

[0096] Step S3.2: Construct a probabilistic prediction network;

[0097] First, all nodes The feature vector obtained after graph attention layer processing With normalized time series Horizontal splicing;

[0098] Then, each row of data from the splicing result is input into the multilayer perceptron as the input feature at each time step, with a feature dimension of 39.

[0099] Finally, the input features are mapped to the prediction results through the hierarchical nonlinear transformation of the multilayer perceptron. The specific network structure is a multilayer perceptron with five hidden layers. The input layer contains 39 nodes to receive input features, each of the five hidden layers contains 256 neurons, and the output layer contains 1 neuron. The activation function is the Sin function, and the Dropout mechanism is added to prevent overfitting. The probability prediction network predicts the unnormalized probability value of the engine being in an abnormal operating state at the current moment.

[0100] Step S3.3: Construct a physical information prediction network;

[0101] First, the output value of the probability prediction network is... , Partial derivative of pressure signal, The partial derivative of the time signal and the preprocessed data obtained in step S2 are horizontally stitched together. Each row of data in the horizontal stitching result is used as the input of the physical information prediction network. The dimension of the input signal features is 79.

[0102] Then, each row of data from the horizontally concatenated data is used as the input to the physical information prediction network at each time step. The input undergoes a hierarchical nonlinear transformation through a multilayer perceptron to map the input features to the predicted physical information. The specific network structure is a multilayer perceptron with four hidden layers. The input layer contains 79 nodes to receive input features, each of the four hidden layers contains 256 neurons, and the output layer contains 1 neuron. The activation function is the Sin function, and a Dropout mechanism is added to prevent overfitting. This predicted physical information value is used to approximate the relationship between the unnormalized predicted probability value of the engine being in an abnormal operating state at the current time and the change over time.

[0103] Step S4: Train the graph attention physical information neural network model;

[0104] First, based on the training dataset in step S2 The training batch size is defined as 256, the number of iterations is 200, and the learning rate scheduling adopts a combination of linear warm-up and cosine decay. The linear warm-up phase is set to linearly increase the learning rate from the initial value 1e-5 to the preset maximum value 1e-4 in the first 50 rounds. The cosine decay phase is set to smoothly decrease the learning rate from the initial value 1e-4 to 1e-5 according to the cosine function curve after the linear warm-up. The weight matrix in the probabilistic prediction network and the weight matrix in the physical information prediction network are initialized using the Kaiming initialization method, and the bias vector in the probabilistic prediction network and the bias vector in the physical information prediction network are initialized using the zero initialization method.

[0105] Then, the Adam optimization algorithm based on the principle of stochastic gradient descent is used to optimize the graph attention physical information neural network model, that is, to minimize the comprehensive error. When training the model, physical information is constrained by the coefficients. Set to 0.8;

[0106] Secondly, input the training dataset. Train a graph attention physical information neural network model;

[0107] Finally, the graph attention physical information neural network model is trained, and the optimal model parameters are obtained, including the graph attention layer weight matrix. Learnable attention vectors Weight matrices and bias vectors in probabilistic prediction networks and weight matrices and bias vectors in physical information prediction networks;

[0108] Step S5: Validation and fine-tuning of the graph attention physical information neural network model;

[0109] Step S5.1, firstly, the validation dataset from step S2... Input the graph attention physical information neural network model trained in step S4, and use the sigmoid activation function to adjust the output of the probability prediction network. Normalization is performed to obtain the predicted probability value of the engine being in an abnormal operating state at the current moment;

[0110] Step S5.2, then, set the probability threshold to 0.5;

[0111] When the predicted probability value obtained in step S5.1 is greater than 0.5, the engine is defined as being in an abnormal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "1"; when the predicted probability value is less than 0.5, the engine is defined as being in a normal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "0".

[0112] Compare all the predicted labels obtained with all the real labels set in step S2, and calculate the prediction accuracy of the graph attention physical information neural network model.

[0113] Step S5.3, based on the verification sample set To improve prediction accuracy, adjust the parameters of the graph attention physical information neural network model;

[0114] Step S5.4 Finally, set the prediction accuracy improvement threshold to 0.01%, and continue to repeat the model verification and fine-tuning. When the improvement of the model's prediction accuracy is lower than the improvement threshold, terminate this process to obtain the trained graph attention physical information neural network model.

[0115] Finally, the trained graph attention physical information neural network model was input into the complete dataset A in chronological order without shuffling, and the prediction results are shown in the appendix. Figure 6 With appendix Figure 7 As shown, attached Figure 6 This indicates that the model's output probability of the engine being in an abnormal state can accurately reflect the operating status of the aero-engine and demonstrates significant effectiveness in overall performance. Figure 7 This indicates that if the time when the predicted probability first exceeds the probability threshold is taken as the warning time, then the warning time of the model is 12.96s, which is 0.58s earlier than the surge outbreak time of 13.54s.

[0116] To verify that the graph attention physical information neural network model has good generalization ability by embedding the error term with physical information as a constraint term of the neural network into the loss function, a dataset B containing the normal and abnormal operating states of the aero-engine was selected from 11 sets of breath-force experiment data. The experimental conditions of the breath-force experiment corresponding to dataset B and dataset A are different. Dataset B is the data obtained by the breath-force experiment of the aero-engine at 0.8 times the maximum speed, containing 614,400 sampling points.

[0117] Following the feature selection, normalization, and data labeling operations performed on dataset A in steps S1 and S2, feature selection, normalization, and data labeling were performed on dataset B. The 553912th sampling point, which represents a significant increase in pressure signal intensity and is located at 13.52 seconds, was selected as the starting point of the surge. Then, the complete dataset B, without shuffling, was input into the trained graph attention physical information neural network model in chronological order, and the prediction results are shown in the attached figure. Figure 8 and attached Figure 9 As shown, attached Figure 8 The results show that the graph attention physical information neural network model of this invention performs well in processing experimental data obtained under different set speeds, and the probability of the engine being in an abnormal state output by the model can reflect the operating status of the aero-engine well. Figure 9 This indicates that the model's warning time point is 12.99s, which is 0.53s earlier than the surge outbreak time point of 13.52s in dataset B;

[0118] To verify the advantage of the graph attention physical information neural network model in capturing key features when the engine enters an abnormal operating state, a comparative experiment was conducted on the physical information neural network model (PINN). The parameters of the multilayer perceptron in the physical information neural network model were set to be consistent with those in the graph attention physical information neural network model. The difference is that the physical information neural network model does not have a graph attention layer. Datasets A and B were used as test sets, and the accuracy and cross-entropy were used as the basis for verifying the prediction results. The prediction results are summarized in Table 1.

[0119] Table 1 Comparison of Predictive Performance of Different Models

[0120]

[0121] Where 80%A represents 80% of the data randomly selected from dataset A;

[0122] Experimental results show that the graph attention physical information neural network model of this invention can accurately predict the probability of an engine being in an abnormal operating state. When processing datasets obtained at different set speeds, it can also predict the probability of an engine being in an abnormal operating state quite well. In contrast, the prediction accuracy of the physical information neural network model is much lower than that of the graph attention physical information neural network model. These phenomena all demonstrate that the graph attention physical information neural network model has significant advantages in fault prediction and generalization capabilities.

[0123] Step S6: Early warning of surge fault in multi-stage high-pressure ratio axial compressor of aero-engine based on trained graph attention physical information neural network model;

[0124] The trained graph attention physical information neural network model is loaded into the early warning equipment of the aircraft engine. The corresponding sensor data is used as the input of the equipment. The early warning equipment performs real-time calculations to predict the probability that the aircraft engine is currently in an abnormal operating state. If the predicted probability value determines that the engine is in an abnormal operating state, a surge warning alarm is issued, and the crew takes countermeasures based on the alarm.

[0125] The above embodiments are only used to illustrate the implementation of the present invention, but should not be construed as limiting the present invention. It should be noted that those skilled in the art can make several modifications and variations to the above embodiments without departing from the principles and concepts of the present invention, within the scope of the present invention.

Claims

1. A method for intelligent early warning of surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine, characterized in that, The early warning method includes the following steps: Step S1: Using pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine, pressure signals under various operating conditions of the aero-engine are collected to obtain raw data, and a graph network is constructed based on the Pearson correlation coefficient. Step S2: Preprocess the collected raw data to make it suitable for the graph attention physical information neural network model. The preprocessing includes data labeling and normalization based on the pressure signal change feature analysis to obtain the preprocessed data, and divide it into training set and validation set. Step S3: Construct a graph attention physical information neural network model, which consists of a graph attention layer, a probability prediction network, and a physical information prediction network; Step S4: Train the graph attention physical information neural network model using the data from the training set; Step S5: Validate and fine-tune the graph attention physical information neural network model using a validation set; Step S6: Early warning of surge fault in multi-stage high-pressure ratio axial compressor of aero-engine based on trained graph attention physical information neural network model.

2. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 1, characterized in that, The specific steps of S1 are as follows: First, multiple pressure sensors on the multi-stage high-pressure ratio axial compressor of the aero-engine are used to collect pressure signals under various operating conditions of the aero-engine. The signal data collected by each pressure sensor is defined as a one-dimensional continuous pressure variable. Second, a dataset containing the normal operating conditions and abnormal operating conditions of the aero-engine is selected as a sample set for fault prediction. Abnormal operating conditions include rotating stall and surge. Third, each continuous pressure variable is regarded as a node. The Pearson correlation coefficient is used to measure the correlation between the continuous pressure variables, and a correlation threshold is set. If the correlation between two continuous pressure variables is higher than the set correlation threshold, the nodes corresponding to the two continuous pressure variables are considered to be connected, and an edge structure is set for them in the graph network. Otherwise, no edge structure is set, resulting in a graph network structure. Nodes connected by the edge structure are neighbor nodes. The Pearson correlation coefficient measures continuous variables. and continuous variables The degree of correlation is shown in equation (1); (1) ； in, The sample size for continuous variables; Pearson correlation coefficient; Representing continuous variables No. One variable value; Representing continuous variables No. One variable value; Representing continuous variables The mean; Representing continuous variables The mean.

3. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 2, characterized in that, The specific steps of S2 are as follows: First, signal analysis is performed on the continuous pressure signal variables of the raw data obtained in step S1. Based on their changes and traditional experience, the engine operating state is divided into normal operating state and abnormal operating state. Sample points corresponding to the engine in normal operating state are labeled with a true label "0", and sample points corresponding to the engine in abnormal operating state are labeled with a true label "1". Second, a time sequence is added to the raw data. Then, the raw data is normalized through linear transformation, and the normalized time series is denoted as... Finally, the dataset is divided into training datasets. and validation dataset Both datasets are shuffled, and the splitting process must ensure that both datasets contain data with different labels.

4. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 3, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Construct the graph attention layer; First, based on the graph network structure obtained in step S1, the nodes corresponding to each continuous pressure variable are obtained. The neighboring nodes are identified, and a feature vector with initial features is selected for each node. Select a shared weight matrix For each feature vector A linear transformation is performed to map the original node features to a new shared feature space, as shown in Equation (2); (2) ； In equation (2), The weight matrix is ​​a learnable matrix; It is a node The representation of in the shared feature space; Secondly, the nodes The transformed features from neighboring nodes are vertically concatenated along the feature dimension, and the concatenated feature vector is multiplied by a learnable attention vector before being processed. Activation function obtains nodes The original attention coefficients of each neighbor node are shown in Equation (3); (3) ； In equation (3), For nodes with neighboring nodes The original attention coefficients between nodes represent the original attention coefficients between nodes. right The importance of; This is a learnable attention vector; This represents the vertical concatenation operation of vectors; Represents a node Features after linear transformation; Represents a node neighboring nodes Features after linear transformation; Indicates use Activation function; Then, all the original attention coefficients are normalized to obtain the standardized attention weights, as shown in Equation (4); (4) ； In equation (4), Represents a node The set of all neighboring nodes; Represents a set All neighboring nodes; Represents a node Its neighboring nodes The original attention coefficient; Represents a node with neighboring nodes Standardized attention weights between them; Finally, for each node Weighted feature aggregation is performed by multiplying the linearly transformed features of all its neighboring nodes with the corresponding standardized attention weights, then summing them, and introducing nonlinearity; the graph attention layer adopts a multi-head graph attention mechanism, using multiple independent attention heads to process nodes, and averaging the outputs of each attention head to obtain the output feature vector, as shown in Equation (5); (5) ； In equation (5), Represents a node Feature vectors after processing by the graph attention layer; The number of attention heads; Indicates the first One point of attention; Represents a node The set of all neighboring nodes; Represents a set All neighboring nodes; Indicates the first The node obtained by the attention head calculation with neighboring nodes Standardized attention weights between them; Indicates the first Neighbor nodes obtained by attention head calculation Features after linear transformation; Step S3.2: Construct a probabilistic prediction network; The probabilistic prediction network is a multilayer perceptron used to receive the feature vector output by the graph attention layer and predict the unnormalized probability value of the engine being in an abnormal operating state at the corresponding sample point at the current time; first, all nodes... The feature vector obtained after graph attention layer processing With normalized time series The data is horizontally stitched together; then, each row of the stitched result is input into the multilayer perceptron as the input feature at each time step; finally, the hierarchical nonlinear transformation of the multilayer perceptron maps the input features onto the prediction result. Step S3.3: Construct a physical information prediction network; The physical information prediction network is a multilayer perceptron, which receives the output value of the probability prediction network. , Partial derivative of pressure signal, Using the partial derivative of the time signal and the preprocessed data obtained in step S2, a physical information value is predicted. This is used to approximate the relationship between the unnormalized true probability value of the engine being in an abnormal operating state at the current moment and the time. First, the output value of the probability prediction network is... , Partial derivative of pressure signal, The partial derivative of the time signal and the horizontal splicing of the preprocessed data obtained in step S2 are shown in equation (9). (9) ； In equation (9), This indicates the result of horizontal stitching; The number of pressure signal features; Indicates the first A normalized pressure signal characteristic; This represents the normalized time series; To predict the output value of the network using probability; express For the Partial derivative of a pressure signal; express Partial derivatives with respect to the time series; Then, each row of data from the horizontally stitched data result is used as the input to the physical information prediction network at each moment. The input is mapped to the physical information prediction result through the hierarchical nonlinear transformation of the multilayer perceptron. This physical information is used to approximate the relationship between the unnormalized prediction probability value of the engine being in an abnormal operating state at the current moment and the change over time.

5. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 4, characterized in that, In step S3.2, the core function of the probability prediction network is to perform the composite function mapping as shown in equation (6); (6) ； In equation (6), This represents the unnormalized probability value predicted by the probability prediction network as being in an abnormal operating state at the current moment; This represents the total number of layers in the multilayer perceptron; Indicates the first The learnable weight matrix of the layer; Indicates the first Learnable bias vectors of the layer; Indicates the first The non-linear activation function used in the layer is the Sin function; Representing the eigenvector With normalized time series Each row of the horizontally concatenated matrix represents the input features of the probability prediction network at each time step.

6. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 5, characterized in that, In step S3.3: The changing relationship satisfies the partial differential equation shown in equation (7); (7) ； In equation (7), This represents the unnormalized true probability value of the engine being in an abnormal operating state at the current moment. This represents the change of the unnormalized true probability value of the engine being in an abnormal operating state at the current moment over time. This represents the true probability that the engine is currently in an abnormal operating state, and the pattern that follows during engine operation. This represents the normalized time series; The number of pressure signal features; Indicates the first The normalized pressure signal features; the physical information prediction network uses a more generalized function. To approximate the function That is, it satisfies the change relationship shown in equation (8); (8) ； In equation (8), This represents the relationship between the unnormalized predicted probability value of the engine being in an abnormal operating state at the current moment and time, which is also the predicted value of the physical information prediction network. Indicates the function used for approximation. A more generalized function; The number of pressure signal features; Indicates the first A normalized pressure signal characteristic; This represents the normalized time series; To predict the output value of the network using probability; express For the Partial derivative of a pressure signal; express Partial derivatives with respect to the time series; The core function of the physical information prediction network is to execute the functional relationship shown in equation (10); (10) ； In equation (10), This represents the relationship between the unnormalized predicted probability value of the engine being in an abnormal operating state at the current moment and time, which is also the predicted value of the physical information prediction network. This represents the input to the physical information prediction network at each time step, which is each row of data in the horizontally stitched result; This represents the total number of layers in the multilayer perceptron in the physical information prediction network; Indicates the first The learnable weight matrix of the layer; Indicates the first Learnable bias vectors of the layer; Indicates the first The non-linear activation function used in the layer is the Sin function.

7. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 6, characterized in that, Step S4 specifically involves: The training objective of the graph attention physical information neural network model is to minimize the overall error. The overall error includes data error terms. and physical information error term Its expression is shown in equations (11), (12) and (13); (11) ； (12) ； (13) ； In equation (11), This is the data error term, in the form of cross-entropy loss; The actual label set in step S2; To predict the output value of the network using probability; In equation (12), This is the physical information error term; express Time series The partial derivative; This represents the predicted value of the physical information prediction network; in equation (13), It is a comprehensive error; It is the physical information constraint coefficient, which is adjusted. The physical information error term can be changed. The degree of constraint imposed on the model during model training; The training process is as follows: First, based on the training dataset in step S2... The size of the parameters is used to select appropriate hyperparameters for model training, including training batch size, number of iterations, learning rate, and initial weights. Then, the Adam optimization algorithm based on the principle of stochastic gradient descent is used to optimize the graph attention physical information neural network model, that is, to minimize the overall error. Secondly, input the training dataset. Model training is then performed; finally, the model training is complete, and the optimal model parameters are obtained, including the graph attention layer weight matrix. Learnable attention vectors The weight matrix and bias vector in the probabilistic prediction network and the weight matrix and bias vector in the physical information prediction network.

8. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 7, characterized in that, Step S5 specifically involves: Step S5.1, take the validation dataset from step S2. Input the graph attention physical information neural network model trained in step S4, and use the sigmoid activation function to adjust the output of the probability prediction network. Normalization is performed to obtain the predicted probability value of the engine being in an abnormal operating state at the current moment; Step S5.2: Set the probability threshold; When the predicted probability value obtained in step S5.1 is greater than the probability threshold, the engine is defined as being in an abnormal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "1"; when the predicted probability value is less than the probability threshold, the engine is defined as being in a normal operating state at that moment, and the predicted label of the corresponding sample at that moment is set to "0". Compare all the predicted labels obtained with all the real labels set in step S2, and calculate the prediction accuracy of the graph attention physical information neural network model. Step S5.3: Adjust the parameters of the graph attention physical information neural network model based on the prediction accuracy. Step S5.4: Set a prediction accuracy improvement threshold for the model, and continuously repeat the model verification and fine-tuning. When the improvement of the model's prediction accuracy is lower than the improvement threshold, terminate this process to obtain the trained graph attention physical information neural network model.

9. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial flow compressor of an aero-engine according to claim 8, characterized in that, The probability threshold in S5.2 is in the range of 0.45 to 0.

55.

10. The intelligent early warning method for surge fault in a multi-stage high-pressure ratio axial compressor of an aero-engine according to claim 9, characterized in that, Step S6 specifically involves: The trained graph attention physical information neural network model is loaded into the early warning equipment of the aero-engine. The corresponding sensor data is used as the input of the equipment. The early warning equipment performs real-time calculations to predict the probability that the aero-engine is currently in an abnormal operating state. If the predicted probability value determines that the engine is in an abnormal operating state, a surge warning alarm is issued.

Images