Aero-engine cross-domain remaining service life prediction method based on transferable topology perception graph pooling network

Abstract

The invention discloses an aero-engine cross-domain residual service life prediction method based on a migratable topology perception graph pooling network, belongs to the subject crossing field of combination of aero-engine residual service life prediction and artificial intelligence, and particularly relates to an aero-engine cross-domain residual service life prediction method. The invention aims to solve the problem that the residual service life of the aero-engine is difficult to predict under the cross-domain condition. The method comprises the following steps: collecting sensor historical data of an aero-engine working in a service environment, and processing the sensor historical data to obtain processed sensor historical data; building a transferable topology awareness graph pooling network; constructing a target loss function of the transferable topology awareness graph pooling network; obtaining a trained migratable topology perception graph pooling network based on the target loss function; and on the basis of the trained transferable topology perception graph pooling network, carrying out online residual service life prediction on the aero-engine.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of the combination of the remaining useful life prediction of an aero-engine and artificial intelligence, and particularly relates to a cross-field remaining useful life prediction method for an aero-engine. BACKGROUND

[0002] As a core link of the health management of an aero-propulsion system, the remaining useful life (RUL) prediction technology is directly related to flight safety and the economy of the whole life cycle of an engine. For typical failure modes such as crack propagation of a high-pressure turbine blade and creep damage of a hot end part of a combustion chamber, existing remaining useful life prediction methods can be divided into two paradigms of mechanism-driven and data-driven. The mechanism-driven method constructs a low-cycle fatigue life model of a turbine disc based on aerodynamic thermodynamic equations, but it is difficult to cope with the nonlinear degradation process under the coupling of multiple physical fields, because it depends on the accurate stress-strain constitutive relationship assumption. In comparison, without empirical formula and theoretical assumption, the data-driven method analyzes the time series data (such as exhaust temperature, fuel flow, etc.) of the gas path parameters and the vibration spectral features of the blade, constructs an end-to-end degradation representation model, and gradually becomes the mainstream technical path of intelligent operation and maintenance of an aero-engine. In recent years, the deep learning method provides a good opportunity for the application of the data-driven remaining useful life method.

[0003] The existing deep learning methods, such as long short-term memory network and convolutional neural network, have technical bottlenecks. The traditional long short-term memory network can capture the single-channel time series degradation trend, but it is difficult to model the spatial coupling characteristics of the turbine blade and the vibration signal between the compressor stages. On the other hand, the convolutional neural network extracts the spatial features of the sensor data through a local convolution kernel, but its fixed receptive field cannot adaptively capture the dynamic correlation of the degradation mode under different engine conditions (such as take-off, cruising, and high altitude). In addition, the traditional convolutional neural network and long short-term memory network mainly use local pooling and recurrent operations, which can only consider the local information of the input data and are difficult to model in non-Euclidean space. In comparison, the graph neural network (GNN) that considers the spatial hierarchical modeling in the global range and is effective in learning the graph structure data representation provides a good opportunity to solve the above challenges. By regarding the input features as graph nodes and the relationship between the features as the edges of the graph, the spatial dependency hidden in the graph data is mined. In the degradation process of the system, the spatial dependency between the nodes in the graph also changes, which provides an opportunity for the effective prediction of the remaining useful life of the aero-engine.

[0004] The actual service environment of an aero-engine faces the challenge of cross-domain migration: the monitoring data generated by the same model engine in different service environments (such as tropical humid climate and polar low temperature environment) has a significant distribution deviation, and obtaining the remaining useful life label of the whole life cycle requires a large amount of data acquisition cost. However, the supervised data-driven learning method assumes that the offline training data domain and the online test data domain must come from the same distribution, which is difficult to meet in the actual remaining useful life prediction task. How to use the labeled data in one domain for the unlabeled remaining useful life in other domains needs further research. At present, domain adaptation (DA) learning oriented transfer learning is widely used in the remaining useful life prediction task under cross-domain conditions. The remaining useful life prediction under cross-domain conditions uses labeled data in the source domain for training, and uses a small amount of unlabeled samples in the target domain to extract domain-invariant representation between the source domain and the target domain, and then realizes the remaining useful life prediction of the target domain. It is worth mentioning that although there are currently aero-engine remaining useful life prediction methods based on graph neural networks and transfer learning, few reports combine the two together for research. SUMMARY

[0005] The present application is to solve the problem of aero-engine remaining useful life prediction under cross-domain conditions, and proposes an aero-engine cross-domain remaining useful life prediction method based on a transferable topological perception graph pooling network.

[0006] The aero-engine cross-domain remaining useful life prediction method based on the transferable topological perception graph pooling network has the following specific process:

[0007] Step 1, collect sensor historical data of the aero-engine working in a service environment, process the sensor historical data to obtain processed sensor historical data; is a positive integer greater than or equal to 2;

[0008] Step 2, build a transferable topological perception graph pooling network;

[0009] Step 3, construct a loss function of hybrid domain distance measurement;

[0010] Step 4, based on the loss function of hybrid domain distance measurement, construct the target loss function of the transferable topological perception graph pooling network;

[0011] Step 5, based on the processed sensor historical data in step 1, the target loss function of the migratable topological perception graph pooling network in step 4, the migratable topological perception graph pooling network in step 2 is trained until the target loss function of the migratable topological perception graph pooling network converges, and a trained migratable topological perception graph pooling network is obtained;

[0012] Step 6, based on the trained migratable topological perception graph pooling network, an online remaining useful life prediction of an aero-engine is carried out.

[0013] The beneficial effects of the present application are:

[0014] (1) In order to effectively learn the spatial dependence and non-Euclidean characteristics of aero-engine degradation data features, the present application adopts a combination of local voting and global voting, and constructs a topological perception graph pooling network, thereby effectively mining the feature information in the hidden non-Euclidean space.

[0015] (1) In order to effectively construct a loss function to describe the domain difference, the present application considers designing a hybrid domain distance measurement loss function, specifically including a maximum mean difference function and a CORAL loss function, to better learn the domain-invariant representation, thereby improving the cross-domain remaining useful life prediction effect.

[0016] (1) Considering that existing deep learning methods are mostly black box models, the present application attempts to use the integrated gradient method for the explainability analysis of the proposed migratable topological perception graph pooling network, thereby improving the explainability of the deep neural network. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the workflow diagram of the present application;

[0018] Figure 2 is the local voting schematic diagram of the present application, , , , , is a node;

[0019] Figure 3 is the structure schematic diagram of the migratable topological perception graph pooling network of the present application;

[0020] Figure 4a is the experimental result diagram of the explainability analysis of the feature value of the input sample 1 of the topological perception graph pooling network based on integrated gradient;

[0021] Figure 4b is the experimental result diagram of the explainability analysis of the gradient value of the input sample 1 of the topological perception graph pooling network based on integrated gradient;

[0022] Figure 4c is an experimental result graph of the explainability analysis of the feature values of the input sample 2 of the topology-aware graph pooling network based on integrated gradients;

[0023] Figure 4d is an experimental result graph of the explainability analysis of the gradient values of the input sample 2 of the topology-aware graph pooling network based on integrated gradients;

[0024] Figure 5a is a remaining useful life prediction result graph of the task OC1-OC2;

[0025] Figure 5b is a remaining useful life prediction result graph of the task OC2-OC1. DETAILED DESCRIPTION

[0026] Specific implementation one: the specific process of the method for predicting the cross-domain remaining useful life of an aero-engine based on the migratable topology-aware graph pooling network is as follows:

[0027] In order to effectively learn the non-Euclidean features implied in the aero-engine degradation data and achieve high-performance cross-domain remaining useful life prediction, the present application proposes a method for predicting the remaining useful life based on a migratable topology-aware graph pooling network. Specifically, the labeled data of the source domain and the unlabeled data of the target domain are used to train the topology-aware graph pooling network. At the same time, the domain loss function combining the maximum mean difference function and CORAL is used to reduce the domain difference, and finally the cross-domain prediction of the remaining useful life is realized without the condition of the target domain labeled data. The present application also designs a method based on integrated gradients for the explainability analysis of the proposed migratable topology-aware graph pooling network, which improves the explainability of deep neural networks. The workflow of the present application is shown in Figure 1 ;

[0028] Step 1, collect the sensor historical data of an aero-engine working in a service environment, process the sensor historical data to obtain processed sensor historical data; is a positive integer greater than or equal to 2;

[0029] Step 2, build a migratable topology-aware graph pooling network;

[0030] Step 3, construct a loss function of hybrid domain distance measurement;

[0031] Step 4, based on the loss function of hybrid domain distance measurement, construct a target loss function of the migratable topology-aware graph pooling network;

[0032] Step 5, based on the processed sensor historical data in step 1, the target loss function of the migratable topological perception graph pooling network in step 4, the migratable topological perception graph pooling network in step 2 is trained until the target loss function of the migratable topological perception graph pooling network converges, and a trained migratable topological perception graph pooling network is obtained.

[0033] Step 6, based on the trained migratable topological perception graph pooling network, an online residual useful life prediction of an aero-engine is performed.

[0034] Specific implementation method two: the difference between this implementation method and the specific implementation method one is that in the step 1, the sensor historical data of the aero-engine working in a service environment is collected, the sensor historical data is processed, and the processed sensor historical data is obtained; the specific process is as follows:

[0035] The sensor historical data of the aero-engine working in multiple service environments is collected as source domain data and target domain data; constant features with constant values in the sensor historical data are removed, time windows are used for data combination, and the data is converted into a graph structure data form.

[0036] Step 11, the sensor historical data of the aero-engine working in a service environment (high pressure, high temperature, etc.);

[0037] Step 12, constant features with constant values in the sensor historical data are removed, feature screening is completed, and the sensor historical data after feature screening is obtained, so as to prevent the negative influence of redundant data on the residual useful life prediction and realize data dimension reduction;

[0038] Step 13, the sensor degradation data and the residual useful life label of the aero-engine working in a service environment after feature screening are combined in the form of a time window (a period of time), and the combined time window data is a source domain data set (the source domain data set is a labeled data set);

[0039] Step 14, the sensor degradation data of the aero-engine working in another service environment after feature screening is combined in the form of a time window (a period of time), and the combined time window data is a target domain data set (the target domain data set is an unlabeled data set);

[0040] Step 15, the source domain data set is converted into a graph structure , ;

[0041] the target domain data set is converted into a graph structure , ;

[0042] wherein, is a set of nodes in a graph structure and , is a total number of nodes; is a set of edges connecting nodes of a graph structure , is an adjacency matrix between nodes, , is a set of real numbers; if two nodes numbered and numbered are connected to each other in , otherwise ; is an adjacency matrix between nodes numbered and numbered in ; denotes data in a source domain data set , , is a number of channels of nodes;

[0043] is a set of nodes in a graph structure and , is a total number of nodes; is a set of edges connecting nodes of a graph structure , is an adjacency matrix between nodes, , is a set of real numbers; if two nodes numbered and numbered are connected to each other in , otherwise ; is an adjacency matrix between nodes numbered and numbered in ; denotes data in a target domain data set , , is a number of channels of nodes.

[0044] Other steps and parameters are the same as in the first embodiment.

[0045] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that step 2 involves constructing a transferable topology-aware graph pooling network; the specific process is as follows:

[0046] In graph neural networks (GNNs) for data prediction, a crucial operation is graph pooling, which reduces the computational scale of the graph, decreases the number of network parameters, and thus improves computational efficiency. It's worth noting that some traditional graph pooling methods, such as top-k pooling, do not explicitly consider the graph's topological structure when calculating feature importance scores, potentially reducing the accuracy of remaining lifetime prediction. To address this issue, this invention designs a remaining lifetime prediction method based on a topology-aware graph pooling network. This method analyzes the local and global importance of each node through a two-stage voting process. Specifically, local voting calculates importance based on the similarity between each node and its neighbors; global voting evaluates the node's global importance within the entire graph. By combining these two methods, all nodes in the original graph are sorted, and a new graph is constructed using a top-k approach.

[0047] The input to a transferable topology-aware graph pooling network is the source domain dataset. and target domain dataset ;

[0048] The output of a transferable topology-aware graph pooling network is the source domain dataset. The fusion voting score vector corresponding to the data in the middle and target domain dataset The fusion voting score vector corresponding to the data in the middle .

[0049] Other steps and parameters are the same as in specific implementation method one or two.

[0050] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the working principle of the transferable topology-aware graph pooling network is as follows:

[0051] Step 21: Calculate the source domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Local voting score vector The specific process is as follows:

[0052] Local voting essentially measures the importance of nodes within a local region, reflecting local information about the graph topology. Specifically, it calculates a similarity score between each node and its neighbors. Each node's score is the average of its similarity scores with its neighbors. A high local voting score indicates that the node effectively represents the local graph structure. Mathematically, this is expressed as an adjacency matrix... The feature matrix of the node ;

[0053] Source domain dataset The Middle Data Local voting score vector Expressed in the form described by formulas (1)-(3):

[0054] (1)

[0055] (2)

[0056] (3)

[0057] Target domain dataset The Middle Data Local voting score vector Expressed in the form described by formulas (4)-(6):

[0058] (4)

[0059] (5)

[0060] (6)

[0061] Among them, superscript This indicates the transpose; Represents the similarity matrix of the source domain; The similarity matrix represents the target domain; Represents the diagonalizability matrix of the source domain; The diagonalization degree matrix represents the target domain;

[0062] After adding 3D unit array The source domain adjacency matrix between nodes, , for A dimensional identity matrix;

[0063] After adding 3D unit array The target domain adjacency matrix between nodes, , for A dimensional identity matrix;

[0064] medium elements Represented as , for Chinese elements This indicates the node number after the identity matrix has been added. and number The source domain adjacency matrix between them;

[0065] medium elements Represented as , for Chinese elements This indicates the node number after the identity matrix has been added. and number The target domain adjacency matrix between them;

[0066] This represents the Hadamard product operation;

[0067] This represents the normalized similarity matrix of the source domain. Each row vector in the vector represents the similarity score between a node and its neighboring nodes;

[0068] This represents the similarity matrix of the target domain after normalization. Each row vector in the vector represents the similarity score between a node and its neighboring nodes;

[0069] This represents the softmax function, which is used to normalize similarity scores.

[0070] In summary, local voting essentially encodes local information through inner product operations and incorporates graph topological information, thereby achieving a goal similar to max pooling. A schematic diagram of local voting is shown below. Figure 2 As shown.

[0071] Step 22: Calculate the source domain dataset The Middle Data Global voting score vector and target domain dataset The Middle Data a global voting score vector of the i-th data in the source domain data set ; the specific process is:

[0072] The local voting mainly focuses on the local part of the graph data. Considering that there is an overlap between the local regions of the graph, it is not enough to rely on the local features to realize the graph pooling. Therefore, the application also adopts the global voting method to evaluate the importance degree of each graph node in the global aspect, that is, the importance degree in the whole graph;

[0073] source domain data set th data in the source domain data set th data in the source domain data set a global voting score vector of the i-th data in the source domain data set is expressed in the form described in formulas (7)-(8):

[0074] (7)

[0075] (8)

[0076] target domain data set th data in the target domain data set th data in the target domain data set a global voting score vector of the i-th data in the target domain data set is expressed in the form described in formulas (9)-(10):

[0077] (9)

[0078] (10)

[0079] wherein, represents a source domain trainable projection vector, ; represents a source domain structure aggregated feature matrix;

[0080] represents a target domain trainable projection vector, ; represents a target domain structure aggregated feature matrix;

[0081] Through the global voting method, the adjacency information of each node can be aggregated, and the role information of the node in the global range can be mined.

[0082] Step 23, based on the local voting score vector of the i-th data in the source domain data set th data in the source domain data set th data in the source domain data set th data in the source domain data set th data in the source domain data set th data in the source domain data set th data in the source domain data set Global voting score vector Build source domain dataset The Middle The fusion voting score vector of the data ;

[0083] Based on target domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Global voting score vector Build the target domain dataset The Middle The fusion voting score vector of the data ;

[0084] The specific process is as follows:

[0085] The local voting in step 21 measures the local similarity of each node in the graph among its neighbors, while the global voting in step 22 reveals the importance of a node across the entire graph. Combining these two methods effectively measures the importance of different nodes, thus explicitly considering topological information.

[0086] Source domain dataset The Middle The fusion voting score vector of the data Represented as:

[0087] (11)

[0088] Target domain dataset The Middle The fusion voting score vector of the data Represented as:

[0089] (12)

[0090] Step 24: Process the source domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling (on the source domain dataset) The Middle The fusion voting score vector of the data middle (Sampling at key nodes) to obtain new source domain data and The corresponding new adjacency matrix ;

[0091] For the target domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling (target domain dataset) The Middle The fusion voting score vector of the data middle (Sampling is performed on key nodes) to obtain new target domain data. and The corresponding new adjacency matrix ;

[0092] The specific process is as follows:

[0093] New source domain data and The corresponding new adjacency matrix The acquisition process is as follows:

[0094] (13)

[0095] New target domain data and The corresponding new adjacency matrix The acquisition process is as follows:

[0096] (14)

[0097] in, This indicates the source domain dataset. The Middle The fusion voting score vector of the data middle An index function that samples key nodes; Is it a return? Indexing functions for important nodes; Indicates the target dataset from the source domain. The Middle The index list of the selected nodes for each data point; Indicates the target dataset of the source domain The Middle The index list of the selected nodes for each data point. Corresponding source domain dataset Chinese data; Indicates the target dataset of the source domain The Middle The index list of the selected nodes for each data point. Corresponding adjacency matrix Middle adjacency matrix;

[0098] This indicates the target domain dataset. The Middle The fusion voting score vector of the data middle An index function that samples key nodes; Indicates targeting the dataset The Middle The index list of the selected nodes for each data point; Indicates targeting the dataset The Middle The index list of the selected nodes for each data point. Corresponding target domain dataset Chinese data; Indicates targeting the dataset The Middle The index list of the selected nodes for each data point. Corresponding adjacency matrix Middle adjacency matrix;

[0099] First cycle In the original step one Find the corresponding index in the middle; second loop. In the previous loop Find the corresponding index in the middle; third loop. In the previous loop Find the corresponding index in the adjacency matrix; same;

[0100] First cycle In the original step one Find the corresponding index in the middle; second loop. In the previous loop Find the corresponding index in the middle; 3rd loop In the previous loop Find the corresponding index in the adjacency matrix; same;

[0101] Step 25: Transfer the new source domain data from Step 24. and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the source domain data.

[0102] The new target domain data from step 24 and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the target domain data.

[0103] Step 26: Repeat steps 21 to 25 until the source domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the source domain dataset. ;

[0104] Repeat steps 21 to 25 until the target domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the target domain dataset. ;

[0105] Source domain dataset and target domain dataset The number of data points is the same.

[0106] By repeating the calculation process of formulas (1)-(14), the layer stacking of the topology-aware graph pooling convolutional neural network that integrates local and global information can be realized, thereby mining the non-Euclidean features hidden in the graph. Through the designed topology-aware graph pooling network, the non-Euclidean features hidden in the sensor data collected in the aero-engine can be mined, and then a high-dimensional feature vector can be constructed for subsequent transfer learning tasks;

[0107] In implementation, this invention designs an embedding layer based on a graph convolutional network to generate a low-dimensional embedded representation of nodes in the graph. Then, through local voting in formulas (1)-(6) in step 21 and global voting in formulas (7)-(10) in step 22, the local similarity of each node in the graph among its neighborhoods is measured to explore the importance of nodes at the entire graph level. Information focusing is achieved through pooling layers that fuse local and global information using formulas (11)-(12); by repeating the calculation process of formulas (1)-(14), the number of neural network layers that fuse local and global information can be stacked to construct a high-dimensional feature vector. This will be used for subsequent transfer learning tasks.

[0108] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0109] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the loss function for constructing the hybrid domain distance metric is described in step 3; the specific process is as follows:

[0110] The loss functions for mixed-domain distance metrics include the maximum mean difference loss function and the CORAL loss function;

[0111] This invention considers designing a loss function for hybrid domain distance measurement, specifically including maximum mean difference and CORAL loss function, to facilitate the learning of domain-invariant representations, which will be described in detail below. Maximum mean difference is an important method for measuring the difference between the source domain data distribution p and the target domain data distribution q. Its core idea is to map the samples of the two distributions to a high-dimensional space through the kernel function φ(.), and compare the difference between the sample means of the two probability distributions to determine their similarity. The formula for calculating maximum mean difference is shown in formula (15):

[0112] (15)

[0113] in, and Representing the source domain dataset respectively and target domain dataset The data in Indicates by significant kernel Defined regenerative kernel Hilbert space, The mapping relationship from the original data to the reproducible kernel Hilbert space is defined, and the kernel function relationship can be defined as follows: , This means inner product operation.

[0114] Assuming the source domain dataset This can be formally expressed as: ;

[0115] in, and They represent the first Source domain data The corresponding fusion voting score vector and the corresponding remaining lifetime label;

[0116] Target domain dataset This can be formally expressed as: ;

[0117] in, Indicates the first The feature variables represented by each target domain sample.

[0118] The maximum mean difference expression in formula (15) can be further written in the form described in formula (16);

[0119] Step 31: Construct the maximum mean difference loss function, expressed as:

[0120] (16)

[0121] in, and These represent the data from the source domain dataset in step 13. and the target domain dataset in step 14 The number of samples drawn from the sample; ;

[0122] express The kernel function; express The kernel function;

[0123] Indicates the first Source domain data The corresponding fusion voting score vector; Indicates the first Data from each target domain The corresponding fusion voting score vector;

[0124] This represents the source domain dataset from step 13. This represents the target domain dataset from step 14.

[0125] Indicates by significant kernel Defined regenerable kernel Hilbert space; Represents the L2 norm; This represents the loss due to the maximum mean difference.

[0126] Step 32: As a widely used domain adaptation method, while the maximum mean difference can achieve first-order feature alignment between two domains, it may still lack a comprehensive metric description. Therefore, the CORAL loss function is designed to achieve second-order feature alignment between source and target domain data.

[0127] Construct the CORAL loss function, as shown in formula (17):

[0128] (17)

[0129] in, Indicates CORAL loss; Denotes the Frobenius norm; and They represent dimensions as follows: The feature covariance matrices of the source domain data and the target domain data are shown in formulas (18)-(19):

[0130] (18)

[0131] (19)

[0132] in, Represents a column vector where all elements are 1;

[0133] Step 33: To more comprehensively reflect domain differences, this invention considers incorporating the maximum mean difference loss. and CORAL loss By combining these methods, a loss function for a hybrid domain distance metric is constructed.

[0134] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0135] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that step 33 incorporates the maximum mean difference loss. and CORAL loss By combining these methods, a loss function for the hybrid domain distance metric is constructed, as shown in Equation (20):

[0136] (20)

[0137] in, The loss is used for the distance metric in the mixed domain.

[0138] This is the weighted trade-off coefficient between the maximum mean difference loss function and the CORAL loss;

[0139] Formula (20) can be further written in the form described by formula (21):

[0140] (twenty one)

[0141] In the formula, Indicates the first Source domain data The corresponding fusion voting score vector;

[0142] Indicates the first Data from each target domain The corresponding fusion voting score vector.

[0143] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0144] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that, in step 4, a target loss function for a transferable topology-aware graph pooling network is constructed based on a loss function using a hybrid domain distance metric; the specific process is as follows:

[0145] In addition to measuring the domain differences between source domain data and target domain data, it is also necessary to construct a mapping relationship between the source domain input data and its remaining lifetime.

[0146] Step 41: Construct the final fusion voting score vector corresponding to the source domain data obtained in Step 25. The mapping relationship between the remaining useful life and the predicted value is shown in Equation (22):

[0147] (twenty two)

[0148] in, For the first Predicted remaining lifetime values ​​for each source domain sample;

[0149] and Here are the weight matrix and bias vector of the fully connected layer. Indicates the activation function;

[0150] Step 42: Combining formulas (21) and (22), the target loss function of the transferable topology-aware graph pooling network is shown in formula (23):

[0151] (twenty three)

[0152] in, The mean squared error loss for a transferable topology-aware graph pooling network;

[0153] The mean squared error loss for transferable topology-aware graph pooling networks Loss with distance metric from mixed domain The weighted trade-off coefficient between them;

[0154] The target loss for a transferable topology-aware graph pooling network;

[0155] For the first The true remaining lifetime value of each source domain sample.

[0156] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0157] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that, in step 5, the transferable topology-aware graph pooling network of step 2 is trained based on the sensor historical data processed in step 1 and the target loss function of the transferable topology-aware graph pooling network in step 4, until the target loss function of the transferable topology-aware graph pooling network converges, thus obtaining the trained transferable topology-aware graph pooling network; the specific process is as follows:

[0158] Step 51: Take the source domain dataset obtained in Step 1. Chinese data and target domain dataset The input data is a transferable topology-aware graph pooling network, and the output data is a fusion voting score vector for each data point in the source domain dataset and a fusion voting score vector for each data point in the target domain dataset.

[0159] Step 52: Input the fusion voting score vector corresponding to each data point in the source domain dataset into the fully connected layer and the activation function layer in sequence. The activation function layer outputs the remaining service life of the aero-engine.

[0160] Step 53: Repeat steps 51 and 52 until the target loss of the transferable topology-aware graph pooling network is reached. Convergence yields a well-trained, transferable topology-aware graph pooling network.

[0161] The entire transferable topology-aware graph pooling network structure is as follows: Figure 3 As shown.

[0162] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0163] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that step 6 involves predicting the online remaining service life of aero-engines based on a trained transferable topology-aware graph pooling network; the specific process is as follows:

[0164] Sensor data of aero-engines is acquired online in the form of time windows (a period of time). Constant features with constant values ​​in the sensor data are removed to obtain sensor data of aero-engines after feature filtering, so as to prevent the negative impact of redundant data on the prediction of remaining service life and achieve data dimensionality reduction; the sensor data of aero-engines after feature filtering is transformed into a graph structure.

[0165] The graph structure is input into the transferable topology-aware graph pooling network trained in step 5. The trained transferable topology-aware graph pooling network outputs a fused voting score vector. The fused voting score vector is then input into the fully connected layer and the activation function layer in sequence. The activation function layer outputs the predicted value of the remaining service life of the aero-engine.

[0166] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0167] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that it further includes step 7, performing interpretability analysis of the trained transferable topology-aware graph pooling network based on ensemble gradients; the specific process is as follows:

[0168] This invention designs an interpretability analysis method for topology-aware graph pooling networks based on integrated gradients. The essence of this interpretability analysis method is an attribution analysis method that measures the degree of influence of input on the result. Based on this, an attribution-based metric is designed to measure the contribution of features to the model output.

[0169] Specifically, assuming the mapping relationship of the trained transferable topology-aware graph pooling network is F, the input of the trained transferable topology-aware graph pooling network is... Considering relative to Baseline input , for The zero-dimensional vector is used to calculate the gradient at all points along the path (from 0 to 1 of the integral variable); the integral gradient is obtained by accumulating the gradient, specifically expressed as formula (24):

[0170] (twenty four)

[0171] in, For integrating gradients; yes In the Gradient of dimension express Input a trained topology-aware graph pooling network; For integration variables; express The Middle Dimensional data; express The Middle Dimensional data, It is the set of real numbers.

[0172] In this sense, the influence of the input on the output can be measured by the path gradient integral from the input to the output of a transferable topology-aware graph pooling network, and the more influential part of the input can be determined by comparing the path gradient integral.

[0173] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0174] Evaluation of the remaining useful life prediction effect: The root mean square error (RMSE) and the two other indicators are used to evaluate the prediction effect of the proposed adaptive remaining useful life. The expressions for the absolute mean error and the root mean square error are shown in Equation (25):

[0175] (25)

[0176] in, The number of test data samples. This is the sample number. and The first The predicted and actual values ​​of remaining useful life for each sample. The smaller the values ​​of absolute mean error and root mean square error, the better the online prediction effect of the adaptive remaining useful life proposed in this invention.

[0177] The beneficial effects of the present invention are verified through the following embodiments:

[0178] Example 1: This invention uses a dataset generated by the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) platform to validate the proposed method. This dataset reproduces the performance degradation behavior of aero-engines through multi-dimensional operating condition coupling simulation technology. The experimental data includes full-lifecycle degradation data of turbofan engines under four different operating modes (covering diverse combinations of thrust load and environmental parameters) (as shown in Table 1). Each mode is divided into training and validation sets according to standards.

[0179] The training set records complete time-series monitoring data of the engine from its initial healthy state to its final failure, while the validation set contains degradation segments from the early stages of engine operation, with the remaining lifetime at the end of each segment serving as the basis for model performance evaluation. Data samples for each operating mode consist of a 26-dimensional multivariate time series, specifically including: engine identifier, number of operating cycles, three types of operating parameters (flight altitude, Mach number, and sea level temperature), and measurements from 21 sensor channels (such as exhaust temperature, compressor outlet pressure, engine pressure ratio, and engine speed).

[0180] The four operating condition subsets of the dataset exhibit significant heterogeneity in data distribution due to differentiated failure mechanisms and combinations of thrust environment parameters. Therefore, achieving cross-domain degradation knowledge transfer under unsupervised conditions presents a severe challenge. To address this issue, this invention selects labeled data for a single operating condition as the source domain, trains a transferable topology-aware graph pooling network, bridges the cross-domain distribution differences, and ultimately achieves prediction of the remaining service life of aero-engines under the target domain operating condition. This invention designs multiple sets of cross-domain experiments (source domain → target domain), covering asymmetric transfer scenarios, with specific configurations shown in Table 1.

[0181]

[0182] Step 1: Collect aircraft engine data. Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; the specific process is as follows:

[0183] Historical sensor data from aircraft engines operating under various service environments were collected as source and target domain data. Feature filtering was performed on the aircraft turbofan engine dataset provided by NASA Ames Research Center, selecting 14 types of sensor data numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21. Next, the feature-filtered data were combined using a time window approach, and the data was transformed into a graph-structured data format.

[0184] Step 2: Construct a transferable topology-aware graph pooling network; the specific process is as follows:

[0185] The input to a transferable topology-aware graph pooling network is the source domain dataset. and target domain dataset ;

[0186] The output of a transferable topology-aware graph pooling network is the source domain dataset. China Data Corresponding fusion voting score vector and target domain dataset China Data Corresponding fusion voting score vector ;

[0187] The working principle of transferable topology-aware graph pooling networks is as follows:

[0188] Step 21: Calculate the source domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Local voting score vector ;

[0189] The similarity matrix is ​​calculated based on the self-multiplication of node features, and then normalized. Each row vector represents the similarity score between a node and its neighboring nodes. Finally, the similarity score is normalized using the softmax function. Local voting essentially encodes local information through inner product operations and incorporates the graph's topological structure information, thereby achieving a goal similar to max pooling.

[0190] Step 22: Calculate the source domain dataset The Middle Data Global voting score vector and target domain dataset The Middle Data Global voting score vector ;

[0191] Step 23: Based on the source domain dataset The Middle Data Local voting score vector and source domain dataset The Middle Data Global voting score vector Build source domain dataset The Middle The fusion voting score vector of the data ;

[0192] Based on target domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Global voting score vector Build the target domain dataset The Middle The fusion voting score vector of the data ;

[0193] Step 24: Process the source domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling (source domain dataset) The Middle The fusion voting score vector of the data middle (Sampling at key nodes) to obtain new source domain data and The corresponding new adjacency matrix ;

[0194] For the target domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling (on the target domain dataset) The Middle The fusion voting score vector of the data middle (Sampling is performed on key nodes) to obtain new target domain data. and The corresponding new adjacency matrix ;

[0195] Step 25: Transfer the new source domain data from Step 24. and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the source domain data.

[0196] The new target domain data from step 24 and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the target domain data.

[0197] Step 26: Repeat steps 21 to 25 until the source domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the source domain dataset. ;

[0198] Repeat steps 21 to 25 until the target domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the target domain dataset. ;

[0199] Source domain dataset and target domain dataset The number of data points is the same.

[0200] Step 3: Construct the loss function for the hybrid domain distance metric; the specific process is as follows:

[0201] Step 31: Construct the maximum mean difference loss function;

[0202] Step 32: Construct the CORAL loss function;

[0203] Step 33: To more comprehensively reflect domain differences, this invention considers incorporating the maximum mean difference loss. and CORAL loss By combining these methods, a loss function for a hybrid domain distance metric is constructed.

[0204] Step 4: Construct the target loss function for the transferable topology-aware graph pooling network based on the loss function of the hybrid domain distance metric; the specific process is as follows:

[0205] In addition to measuring the domain differences between source domain data and target domain data, it is also necessary to construct a mapping relationship between the source domain input data and its remaining lifetime.

[0206] Step 41: Construct the final fusion voting score vector corresponding to the source domain data obtained in Step 25. The mapping relationship between the remaining useful life and the predicted value;

[0207] Step 42: Construct the objective loss function for a transferable topology-aware graph pooling network;

[0208] Step 5: Based on the historical sensor data processed in Step 1 and the target loss function of the transferable topology-aware graph pooling network in Step 4, train the transferable topology-aware graph pooling network from Step 2 until the target loss function of the transferable topology-aware graph pooling network converges, thus obtaining the trained transferable topology-aware graph pooling network; the specific process is as follows:

[0209] Step 51: Take the source domain dataset obtained in Step 1. and target domain dataset Input a transferable topology-aware graph pooling network, output a fused voting score vector corresponding to the source domain data. The fused voting score vector corresponding to the target domain data ;

[0210] Step 52: Combine the fusion voting score vector corresponding to the source domain data. The fully connected layer and the activation function layer are input sequentially. The activation function layer outputs the remaining service life of the aero-engine.

[0211] Step 53: Repeat steps 51 and 52 until the target loss of the transferable topology-aware graph pooling network is reached. Convergence yields a well-trained, transferable topology-aware graph pooling network.

[0212] The optimization algorithm is the Adam optimization algorithm, the learning rate is 0.001, and the network training process is carried out in a hardware environment with one GPU (GTX 4060 graphics card).

[0213] Step 6: Predict the online remaining service life of aero-engines based on the trained transferable topology-aware graph pooling network; the specific process is as follows:

[0214] Sensor data of aero-engines is acquired online in the form of time windows (a period of time). Constant features with constant values ​​in the sensor data are removed to obtain sensor data of aero-engines after feature filtering, so as to prevent the negative impact of redundant data on the prediction of remaining service life and achieve data dimensionality reduction; the sensor data of aero-engines after feature filtering is transformed into a graph structure.

[0215] The graph structure is input into the transferable topology-aware graph pooling network trained in step 5. The trained transferable topology-aware graph pooling network outputs a fused voting score vector. The fused voting score vector is then input into the fully connected layer and the activation function layer in sequence. The activation function layer outputs the predicted value of the remaining service life of the aero-engine.

[0216] Step 7: Perform interpretability analysis on the trained transferable topology-aware graph pooling network based on ensemble gradients. The essence of this interpretability analysis method is an attribution analysis method that measures the degree of influence of the input on the result. Based on this, an attribution-based metric is designed to measure the contribution of features to the model output. Figure 4a , 4b , Figure 4c , 4d Experimental results are presented for an interpretability analysis method for topology-aware graph pooling networks based on integrated gradients.

[0217] Step 8: Evaluate the remaining useful life prediction effect; use root mean square error (RMSE) and other indicators to evaluate the online remaining useful life prediction effect of the method proposed in this invention.

[0218] Figure 5a , 5b Table 1 shows the predicted and actual values ​​of the proposed method for remaining service life on the aero-engine test dataset under the OC1-OC2 tasks. Table 2 compares the results of this invention with those of various existing transfer learning methods in various cross-domain combinations for remaining service life prediction tasks. Figure 4a , 4b , Figure 4c , 4d As can be seen from Table 2, the technical solution proposed in this invention achieves high-precision prediction of the remaining service life of aero engines in multiple scenarios through a cross-domain degradation feature transfer mechanism, which can provide real-time health status assessment and maintenance decision optimization support for operation and maintenance personnel, thereby systematically improving the intelligent operation and maintenance level of aero engines.

[0219]

[0220] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A cross-domain remaining service life prediction method for aero-engines based on transferable topology-aware graph pooling networks, characterized by: The specific process of the method is as follows: Step 1: Collect aircraft engine data. Historical data of sensors operating in various service environments are collected and processed to obtain processed historical sensor data. It is a positive integer greater than or equal to 2; Step 2: Construct a transferable topology-aware graph pooling network; Step 3: Construct the loss function for the hybrid domain distance metric; Step 4: Construct the target loss function for the transferable topology-aware graph pooling network based on the loss function of the hybrid domain distance metric; Step 5: Based on the sensor historical data processed in Step 1 and the target loss function of the transferable topology-aware graph pooling network in Step 4, train the transferable topology-aware graph pooling network in Step 2 until the target loss function of the transferable topology-aware graph pooling network converges, and obtain the trained transferable topology-aware graph pooling network. Step 6: Predict the online remaining service life of aero-engines based on the trained transferable topology-aware graph pooling network.

2. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 1, characterized in that: In step 1, the aero-engine is collected. The historical data of sensors operating in various service environments is collected and processed to obtain processed historical sensor data. The specific process is as follows: Step 11: Collect aircraft engine data. Historical data of sensors operating in various service environments; Step 12: Remove constant features with constant values ​​from the sensor historical data to complete feature filtering and obtain the sensor historical data after feature filtering. Step 13: Combine the sensor degradation data and remaining service life labels of the aero-engines operating under a certain service environment, which have been filtered by features, using a time window approach. The combined time window data is the source domain dataset. ; Step 14: Combine the sensor degradation data of the aero-engines operating in another service environment, which have been selected by feature filtering, using a time window approach. The combined time window data is the target domain dataset. ; Step 15: Transfer the source domain dataset Transform into a graph structure , ; target domain dataset Transform into a graph structure , ; in, Graph structure The set of nodes in the middle, and , This represents the total number of nodes. Graph structure The set of edges connecting nodes in a graph. This is the adjacency matrix between nodes. , It is the set of real numbers; Represents the source domain dataset Chinese data, , The number of channels in the node; Graph structure The set of nodes in the middle, and , This represents the total number of nodes. Graph structure The set of edges connecting nodes in a graph. This is the adjacency matrix between nodes. , It is the set of real numbers; Represents the target domain dataset Chinese data, , This represents the number of channels in the node.

3. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 2, characterized in that: Step 2 involves constructing a transferable topology-aware graph pooling network; the specific process is as follows: The input to a transferable topology-aware graph pooling network is the source domain dataset. and target domain dataset ; The output of a transferable topology-aware graph pooling network is the source domain dataset. The fusion voting score vector corresponding to the data in the middle and target domain dataset The fusion voting score vector corresponding to the data in the middle .

4. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 2, characterized in that: The working principle of the transferable topology-aware graph pooling network is as follows: Step 21: Calculate the source domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Local voting score vector ; The specific process is as follows: Source domain dataset The Middle Data Local voting score vector Expressed in the form described by formulas (1)-(3): (1) (2) (3) Target domain dataset The Middle Data Local voting score vector Expressed in the form described by formulas (4)-(6): (4) (5) (6) Among them, superscript This indicates the transpose; Represents the similarity matrix of the source domain; The similarity matrix represents the target domain; Represents the diagonalizability matrix of the source domain; Represents the diagonalizability matrix of the target domain; After adding 3D unit array The source domain adjacency matrix between nodes, , for A dimensional identity matrix; After adding 3D unit array The target domain adjacency matrix between nodes, , for A dimensional identity matrix; This represents the Hadamard product operation; This represents the softmax function; This represents the normalized similarity matrix of the source domain. Each row vector in the vector represents the similarity score between a node and its neighboring nodes; This represents the similarity matrix of the target domain after normalization. Each row vector in the vector represents the similarity score between a node and its neighboring nodes; Step 22: Calculate the source domain dataset The Middle Data Global voting score vector and target domain dataset The Middle Data Global voting score vector The specific process is as follows: Source domain dataset The Middle Data Global voting score vector Expressed in the form described by formulas (7)-(8): (7) (8) Target domain dataset The Middle Data Global voting score vector Expressed in the form described by formulas (9)-(10): (9) (10) in, This represents the trainable projection vector from the source domain. ; This represents the aggregated feature matrix of the source domain structure; This represents the trainable projection vector of the target domain. ; This represents the aggregated feature matrix of the target domain structure. Step 23: Based on the source domain dataset The Middle Data Local voting score vector and source domain dataset The Middle Data Global voting score vector Build source domain dataset The Middle The fusion voting score vector of the data ; Based on target domain dataset The Middle Data Local voting score vector and target domain dataset The Middle Data Global voting score vector Build the target domain dataset The Middle The fusion voting score vector of the data ; The specific process is as follows: Source domain dataset The Middle The fusion voting score vector of the data Represented as: (11) Target domain dataset The Middle The fusion voting score vector of the data Represented as: (12) Step 24: Process the source domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling to obtain new source domain data. and The corresponding new adjacency matrix ; For the target domain dataset The Middle The fusion voting score vector of the data Perform top-k sampling to obtain new target domain data. and The corresponding new adjacency matrix ; The specific process is as follows: New source domain data and The corresponding new adjacency matrix The acquisition process is as follows: (13) New target domain data and The corresponding new adjacency matrix The acquisition process is as follows: (14) in, This indicates the source domain dataset. The Middle The fusion voting score vector of the data middle An index function that samples key nodes; Is it a return? Indexing functions for important nodes; Indicates the target dataset from the source domain. The Middle The index list of the selected nodes for each data point; Indicates the target dataset from the source domain. The Middle The index list of the selected nodes for each data point. Corresponding source domain dataset Chinese data; Indicates the target dataset from the source domain. The Middle The index list of the selected nodes for each data point. Corresponding adjacency matrix Middle adjacency matrix; This indicates the target domain dataset. The Middle The fusion voting score vector of the data middle An index function that samples key nodes; Indicates targeting the dataset The Middle The index list of the selected nodes for each data point; Indicates targeting the dataset of the target domain The Middle The index list of the selected nodes for each data point. Corresponding target domain dataset Chinese data; Indicates targeting the dataset of the target domain The Middle The index list of the selected nodes for each data point. Corresponding adjacency matrix Middle adjacency matrix; Step 25: Transfer the new source domain data from Step 24. and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the source domain data. The new target domain data from step 24 and The corresponding new adjacency matrix Substitute steps 21 and 22, and repeat steps 21 to 24 until the maximum number of iterations is reached to obtain the final fusion voting score vector corresponding to the target domain data. Step 26: Repeat steps 21 to 25 until the source domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the source domain dataset. ; Repeat steps 21 to 25 until the target domain dataset has been traversed. From all the data, obtain the final set of fused voting score vectors corresponding to the target domain dataset. ; Source domain dataset and target domain dataset The number of data points is the same.

5. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 4, characterized in that: The loss function for constructing the hybrid domain distance metric in step 3 is as follows: Step 31: Construct the maximum mean difference loss function, expressed as: (16) in, and These represent the data from the source domain dataset in step 13. and the target domain dataset in step 14 The number of samples drawn from the sample; ; express Kernel function; express Kernel function; Indicates the first Source domain data The corresponding fusion voting score vector; Indicates the first Data from each target domain The corresponding fusion voting score vector; This represents the source domain dataset from step 13. This represents the target domain dataset from step 14. Indicates a significant kernel Defined regenerable kernel Hilbert space; Represents the L2 norm; This represents the loss due to the maximum mean difference. Step 32: Construct the CORAL loss function, the expression of which is shown in formula (17): (17) in, Indicates CORAL loss; Denotes the Frobenius norm; and They represent dimensions as follows: The feature covariance matrices of the source domain data and the target domain data are shown in formulas (18)-(19): (18) (19) in, Represents a column vector where all elements are 1; Step 33: Calculate the maximum mean difference loss. and CORAL loss By combining these methods, a loss function for a hybrid domain distance metric is constructed.

6. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 5, characterized in that: In step 33, the maximum mean difference loss is... and CORAL loss By combining these methods, a loss function for the hybrid domain distance metric is constructed, as shown in Equation (20): (20) in, The loss is used for the distance metric in the mixed domain. This is the weighted trade-off coefficient between the maximum mean difference loss function and the CORAL loss; Formula (20) can be further written in the form described by formula (21): (21) In the formula, Indicates the first Source domain data The corresponding fusion voting score vector; Indicates the first Data from each target domain The corresponding fusion voting score vector.

7. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 6, characterized in that: In step 4, a target loss function for a transferable topology-aware graph pooling network is constructed based on a loss function using a hybrid domain distance metric. The specific process is as follows: Step 41: Construct the final fusion voting score vector corresponding to the source domain data obtained in Step 25. The mapping relationship between the remaining useful life and the predicted value is shown in Equation (22): (22) in, For the first Predicted remaining lifetime values ​​for each source domain sample; and Here are the weight matrix and bias vector of the fully connected layer. Indicates the activation function; Step 42: Combining formulas (21) and (22), the target loss function of the transferable topology-aware graph pooling network is shown in formula (23): (23) in, The mean squared error loss for a transferable topology-aware graph pooling network; The mean squared error loss for transferable topology-aware graph pooling networks Loss with distance metric from mixed domain The weighted trade-off factor between them; The target loss for a transferable topology-aware graph pooling network; For the first The true remaining lifetime value of each source domain sample.

8. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 7, characterized in that: In step 5, the transferable topology-aware graph pooling network from step 2 is trained based on the historical sensor data processed in step 1 and the target loss function of the transferable topology-aware graph pooling network from step 4, until the target loss function of the transferable topology-aware graph pooling network converges, thus obtaining the trained transferable topology-aware graph pooling network; the specific process is as follows: Step 51: Take the source domain dataset obtained in Step 1. Chinese data and target domain dataset The input data is a transferable topology-aware graph pooling network, and the output data is a fusion voting score vector for each data point in the source domain dataset and a fusion voting score vector for each data point in the target domain dataset. Step 52: Input the fusion voting score vector corresponding to each data point in the source domain dataset into the fully connected layer and the activation function layer in sequence. The activation function layer outputs the remaining service life of the aero-engine. Step 53: Repeat steps 51 and 52 until the target loss of the transferable topology-aware graph pooling network is reached. Convergence yields a well-trained, transferable topology-aware graph pooling network.

9. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 8, characterized in that: In step 6, the online remaining service life prediction of the aero-engine is performed based on a trained transferable topology-aware graph pooling network; the specific process is as follows: Sensor data of aero-engines is acquired online using a time window approach. Constant features with constant values ​​in the sensor data are removed to obtain sensor data of aero-engines after feature filtering. The sensor data of the aero-engine after feature filtering is transformed into a graph structure; The graph structure is input into the transferable topology-aware graph pooling network trained in step 5. The trained transferable topology-aware graph pooling network outputs a fused voting score vector. The fused voting score vector is then input into the fully connected layer and the activation function layer in sequence. The activation function layer outputs the predicted value of the remaining service life of the aero-engine.

10. The method for predicting the remaining service life of aero-engines across domains based on a transferable topology-aware graph pooling network according to claim 9, characterized in that: The method further includes step 7, which performs interpretability analysis of the trained transferable topology-aware graph pooling network based on ensemble gradients; the specific process is as follows: Assuming the mapping relationship of the trained transferable topology-aware graph pooling network is F, and the input of the trained transferable topology-aware graph pooling network is... Considering relative to Baseline input , for The zero-dimensional vector is used to calculate the gradient at all points along the path; the integral gradient is obtained by accumulating the gradient, specifically expressed as formula (24): (24) in, For integrating gradients; yes In the Gradient of dimension express Input a trained topology-aware graph pooling network; For integration variables; express The Middle Dimensional data; express The Middle Dimensional data, It is the set of real numbers.

Images