Aero-engine residual useful life prediction method based on spatiotemporal knowledge graph and sdcnn

Abstract

The application provides an aero-engine residual service life prediction method based on a space-time knowledge graph and an SDCNN, and belongs to the field of aero-engine health management.The application adopts a space-time knowledge graph and an SDCNN neural network architecture.The application innovatively designs the construction of a space-time knowledge graph for an aero-engine, adopts a BERT model for data type conversion, and adopts a multi-head graph attention network and a pooling graph attention network to complete feature extraction and feature fusion to obtain fused features;finally, the fused features are input into a stacked dilated convolutional neural network to perform regression learning on the feature data, and then the residual life of the aero-engine is predicted.The application can effectively predict the residual service life of the aero-engine under limited data by modeling and predicting complex space-time feature data, provides data support for the formulation of a maintenance strategy of the aero-engine, and provides a new idea for the residual life prediction of other industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine operation and maintenance, and in particular relates to a method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN. Background Technology

[0002] Remaining service life (RSW) prediction technology is a crucial research area in aero-engine maintenance. It can predict aero-engine failures, identifying potential malfunctions early so maintenance personnel can address them promptly. This reduces the likelihood of in-flight failures, ensuring flight safety. RSW technology also optimizes overall aircraft maintenance management, reducing downtime and repair costs, thereby improving operational efficiency and economic benefits. This research method differs from traditional maintenance strategies, which rely on technician experience or fixed maintenance cycles. RSW technology provides guidance for maintenance based on the aircraft's current remaining service life.

[0003] Currently, in the field of remaining service life prediction for aero-engines, these prediction technologies can be broadly categorized into two types based on the prediction steps: model-driven and data-driven. Regarding data-driven remaining service life prediction, Chen et al. published a paper titled "Dynamic Predictive Maintenance Scheduling Using Deep Learning Ensemble for System Health Prognostics" in the IEEE Sensors Journal in 2021. This paper proposed using an ensemble algorithm of deep autoencoder (DAE) and bidirectional LSTM (Bi-LSTM) to construct two degradation prediction models: a health status assessment model and a remaining service life prediction model. These models successfully assessed the health status and predicted the remaining service life of aero-engines. In 2025, Cai et al. published a paper titled "Knowledge embedded spatial–temporal graph convolutional networks for remaining useful life prediction" in the journal *Reliability Engineering & System Safety*. This paper proposed a knowledge-embedded spatiotemporal graph convolutional network (KBST-GCN) method for remaining useful life prediction. Knowledge embedding establishes relational triples based on system structure knowledge and sensor location information. These triples are then converted into low-dimensional vector embeddings using an energy-based knowledge embedding algorithm, generating a graph dataset. A GCN layer with multi-head attention, an LSTM layer, and a fully connected layer are used to extract spatiotemporal degradation patterns and obtain the Remaining Useful Life (RUL) prediction results. Regarding model-driven remaining useful life prediction, Si et al. published a paper titled "Remaining useful life estimation – A review on the statistical data driven approaches" in the journal *European Journal of Operational Research*, systematically reviewing current mainstream statistical-data-driven methods.The authors categorize existing models into two main types: first, direct observation state process models, including regression, Wiener process, Gamma process, and Markov chain, which take directly detectable degradation indicators as input and extrapolate to the failure threshold to obtain the RUL; second, indirect observation state process models, including stochastic filtering, proportional hazards model (PHM), hidden Markov model (HMM) and its semi-Markov extension (HSMM), which are characterized by inferring hidden health states from monitoring signals and then estimating the RUL.

[0004] Based on the above analysis, this invention proposes a method to predict the remaining life of aero-engines by constructing a spatiotemporal knowledge graph of the device, first converting the data type, then extracting and fusing features. These processed feature data can be used for regression learning in the SDCNN model. Summary of the Invention

[0005] This invention proposes a method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN. Sensor data is constructed into a device-oriented spatiotemporal knowledge graph according to components, functional relationships, and spatiotemporal dimensions. Features are extracted from the spatiotemporal knowledge graph using a BERT model, a multi-head graph attention network, and a pooled graph attention network. Then, the SDCNN model is used to learn the features, and the prediction function is validated using a test set to obtain a complete model for predicting the remaining service life of aero-engines. This addresses the lack of explanation in existing technologies regarding how component and functional relationships affect the predicted remaining service life.

[0006] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0007] A method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN, the method comprising the following steps:

[0008] Step S1: The sensors collect data from the aero-engine components to obtain the collected dataset, add the remaining service life label to obtain the aero-engine degradation dataset; normalize the sensor-collected data in the aero-engine degradation dataset, and construct a spatiotemporal knowledge graph from the normalized data.

[0009] Step S2: Input the spatiotemporal knowledge graph into the BERT model and convert it into node embedding vectors and relation embedding vectors in tensor form that can be read by the neural network. Then, use a multi-head graph attention network and a pooling graph attention network to extract and fuse features from the node embedding vectors and relation embedding vectors to obtain a fused feature vector.

[0010] Step S3: Construct a stacked dilated convolutional neural network and input the fused feature vector into the stacked dilated convolutional neural network to predict degradation trend features.

[0011] Step S4: Divide 70% of the fused feature vectors obtained from feature extraction and feature fusion into a training set, and the remaining 30% into a test set. Use the remaining service life cycle label as the training label for the training set data, and input it into a stacked dilated convolutional neural network for training to obtain a trained aircraft engine remaining service life prediction model.

[0012] Step S5: Use the trained aircraft engine remaining service life prediction model to predict the remaining service life of the aircraft engine.

[0013] Further, step S1 includes:

[0014] Step S11: The sensor collects data from the aero-engine components to obtain a data set. The remaining service life label is added to each data point in the data set to obtain the aero-engine degradation data set.

[0015] Step S12: Normalize the sensor data collected in the aero-engine degradation dataset to obtain the normalized dataset.

[0016] Step S13: Construct a spatiotemporal knowledge graph from the normalized dataset in the form of triples of "entity-relationship-attribute / entity" from three aspects: structural hierarchy relationship, dynamic transmission relationship, and operation and monitoring relationship.

[0017] Furthermore, the normalization method in step S12 is as follows:

[0018]

[0019] in, To collect data for the dataset; The data is after normalization; It is the minimum value in the collected dataset; It is the maximum value in the collected dataset.

[0020] Further, step S13 includes the following steps:

[0021] Step S131: Perform hierarchical modeling of the overall structure of the aero-engine, express the subordinate relationship between components through the triple form of "component-containment-component", and construct a knowledge graph from the structural hierarchy relationship.

[0022] Step S132: To reflect the aerodynamic and mechanical coupling relationship between components, a knowledge graph is constructed from the power transmission relationship using a triplet form of "component-drive / power supply / transmission-component".

[0023] Step S133: Based on the overall operating status relationship of the engine, integrate the monitoring indicators of each component, unify the association between sensor monitoring data and structural knowledge, and construct a knowledge graph from the operating and monitoring relationship through the triple form of "component - operation / monitoring relationship - monitoring value".

[0024] Step S134: Repeat steps S131 to S133 to complete the construction of the knowledge graph for all time nodes of the normalized dataset, and obtain the spatiotemporal knowledge graph of the complete operation of the aero-engine.

[0025] Furthermore, the spatiotemporal knowledge graph constructed in step S134 is represented as follows:

[0026]

[0027] in, express Knowledge graph of time nodes; This represents an entity-relationship-attribute / entity triple. Represents an entity, Indicates a relationship. Represents an attribute or entity; This represents the set of all entities; the engine components and all components together constitute the set of all entities. A collection representing numerical attributes; Indicates the hierarchical relationship of the structure; Indicates the relationship of power transmission, This indicates the relationship between operation and monitoring.

[0028] Further, step S2 includes the following steps:

[0029] Step S21: Transform the spatiotemporal knowledge graph into a data format using the BERT model, converting the spatiotemporal knowledge graph composed of natural language into tensor-form node embedding vectors and relation embedding vectors, so that the multi-head attention network can extract features.

[0030] Step S22: Input the node embedding vector and relation embedding vector converted into tensor form in step S21 into the multi-head graph attention network for feature extraction to obtain the concatenated feature vector.

[0031] Step S23: Input the spatiotemporal knowledge graph, relation embedding vector, and concatenated feature vector into the pooling graph attention network to perform feature fusion and obtain the fused feature vector.

[0032] Further, step S3 includes the following steps:

[0033] Step S31: Construct a stacked dilated convolutional neural network, which is composed of multiple dilated convolutional layers stacked in a cascaded manner.

[0034] Step S32: Input the fused feature vector into the first dilated convolutional layer, perform weighted summation and activation through the dilated convolutional kernel, and output the first convolutional feature. Input the first convolutional feature into the second dilated convolutional layer and output the second convolutional feature, and so on, until the last dilated convolutional layer outputs the last convolutional feature, which is the degradation trend feature.

[0035] Step S33: Calculate the receptive field of the multi-layer stack and verify the degradation trend characteristics. Does it cover a sufficiently long input history?

[0036] Furthermore, the first The calculation method for each dilated convolutional layer is as follows:

[0037]

[0038]

[0039] in, For the first Each dilated convolutional layer at time step The convolutional features, Indicates the total length of the time series; Indicates the first The convolutional features output by each dilated convolutional layer; Represents a nonlinear activation function; Indicates the kernel size. This represents the index variable inside the convolution kernel; Indicates the first In the dilated convolutional layer, the convolution kernel is at the _th ... Trainable weight matrix at each position; Indicates a time step. Indicates the first The expansion rate of the layer, Indicates the sampling index; Indicates the first The convolutional features output by each dilated convolutional layer Sampling characteristics; Indicates the first The bias of each dilated convolutional layer.

[0040] Compared with the prior art, the beneficial technical effects of the present invention are:

[0041] 1) This invention can efficiently extract degradation features from aero-engine operating data across multiple time scales, fully capture short-term fluctuations and long-term trend changes, thereby achieving high-precision prediction of the remaining service life of the engine.

[0042] 2) This invention exhibits stronger feature learning and robustness under complex operating conditions, significantly improving the ability to intelligently assess the health status of aero engines and judge degradation trends. It provides maintenance personnel with a more scientific basis for maintenance cycle planning and risk warning, thereby realizing intelligent and predictive management of aero engine maintenance and improving the operational safety and economy of the equipment.

[0043] 3) By modeling and predicting complex spatiotemporal characteristic data, this invention can effectively predict the remaining service life of aero engines with limited data, providing data support for the formulation of aero engine maintenance strategies, and also providing new ideas for predicting the remaining service life of other industrial equipment. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN, as proposed in this invention.

[0046] Figure 2 This is a schematic diagram illustrating the change in the dilation rate of each layer in the Stacked Dilated Convolutional Neural Network (SDCNN) of this invention.

[0047] Figure 3 This is a schematic diagram of the computational logic architecture of the Stacked Dilated Convolutional Neural Network (SDCNN) of this invention.

[0048] Figure 4 This is the health status structure of the spatiotemporal knowledge graph of this invention.

[0049] Figure 5 This is the spatiotemporal knowledge graph data that has been fully constructed based on the C-MAPSS dataset example of this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] like Figure 1As shown, this invention proposes a method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN. The method includes the following steps:

[0052] Step S1: The sensors collect data from the aero-engine components to obtain the collected dataset, add the remaining service life label to obtain the aero-engine degradation dataset; normalize the sensor-collected data in the aero-engine degradation dataset, and construct a spatiotemporal knowledge graph from the normalized data.

[0053] Step S11: The sensor collects data from the aero-engine components to obtain a data set. The remaining service life label is added to each data point in the data set to obtain the aero-engine degradation data set.

[0054] In one optional implementation, the engine components acquired by the sensor include the core engine (CE), engine (E), fan, low-pressure compressor (LPC), low-pressure turbine (LPT), combustion chamber (CC), high-pressure compressor (HPC), and high-pressure turbine (HPT).

[0055] In one optional implementation, the data collected by the sensor includes the core machine physical speed, high-pressure compressor outlet total pressure, high-pressure compressor outlet static pressure, fuel flow rate to P30 ratio, corrected core speed, bypass ratio, bleed air enthalpy, low-pressure compressor temperature, corrected fan speed, physical fan speed, low-pressure turbine temperature, low-pressure turbine coolant discharge, high-pressure turbine coolant discharge, and high-pressure compressor temperature.

[0056] The sensors acquire the physical rotational speed of the core engine; the total pressure, static pressure, and temperature at the high-pressure compressor outlet; the fuel flow rate to P30 ratio in the combustion chamber; the calibrated core rotational speed of the core engine; the bypass ratio of the engine; the temperature and bleed air enthalpy of the low-pressure compressor; the calibrated and physical fan speeds; the temperature and coolant emissions from the low-pressure turbine; and the coolant emissions from the high-pressure turbine.

[0057] Each data point in the sensor-collected dataset is labeled with its remaining service life, forming an aero-engine degradation dataset. The remaining service life label is set to the maximum operating cycle minus the current operating cycle.

[0058] Step S12: Normalize the sensor data collected in the aero-engine degradation dataset to obtain the normalized dataset. The normalization method is as follows:

[0059]

[0060] in, To collect data for the dataset; The data is after normalization; It is the minimum value in the collected dataset; It is the maximum value in the collected dataset.

[0061] Step S13: Construct a spatio-temporal knowledge graph (STKG) for the normalized dataset in the form of triples of "entity-relationship-attribute / entity" from three aspects: structural hierarchy relationship, dynamic transmission relationship, and operation and monitoring relationship.

[0062] The knowledge graph at a specific point in time is constructed using a triplet of "entity-relationship-attribute / entity" to uniformly model the engine's structural composition, functional coupling, and condition monitoring information. The spatiotemporal knowledge graph data includes three parts: structural hierarchical relationships, power transmission relationships, and operational and monitoring relationships.

[0063] This invention constructs a spatiotemporal knowledge graph from three aspects: structural hierarchy, power transmission, and operation and monitoring.

[0064] The construction of the knowledge graph for a time node of the normalized dataset includes steps S131 to S132:

[0065] Step S131: Perform hierarchical modeling of the overall structure of the aero-engine, express the subordinate relationship between components through the triple form of "component-containment-component", and construct a knowledge graph from the structural hierarchy relationship.

[0066] A hierarchical model of the overall structure of the aero-engine was constructed. The hierarchical relationship between the engine and key components was expressed using a "component-container-component" triplet format, and a knowledge graph was built based on this hierarchical relationship. This indicates an inclusion relationship.

[0067] In one optional implementation, the knowledge graph constructed from the structural hierarchy is represented as follows: an engine includes a core engine, a fan, a low-pressure compressor, and a low-pressure turbine; the core engine further includes a high-pressure compressor, a high-pressure turbine, and a combustion chamber.

[0068] Step S132: To reflect the aerodynamic and mechanical coupling relationship between components, a knowledge graph is constructed from the power transmission relationship using a triplet form of "component-drive / power supply / transmission-component".

[0069] To reflect the aerodynamic and mechanical coupling relationships between components and support knowledge-based reasoning for analyzing unseen performance degradation, a knowledge graph is constructed from the power transmission relationships through the "drive / power supply / transmission" framework. These are driving relationships, propels, and deliveries.

[0070] In one alternative implementation, the knowledge graph representation based on the power transmission relationship is as follows: the core engine drives the engine and fan; the fan propels the engine; the low-pressure turbine provides power to the low-pressure compressor and fan; the high-pressure turbine drives the high-pressure compressor and provides energy to the low-pressure turbine; the combustion chamber provides thermal energy to drive the high-pressure turbine; and the low-pressure compressor transfers gas to the high-pressure compressor.

[0071] Step S133: Based on the overall operating status relationship of the engine, integrate the monitoring indicators of each component, unify the association between sensor monitoring data and structural knowledge, and construct a knowledge graph from the operating and monitoring relationship through the triple form of "component - operation / monitoring relationship - monitoring value".

[0072] To address the overall operational status relationships of the engine, the engine is incorporated into a knowledge graph. This knowledge graph is constructed from operational relationships using a triplet structure of "component—operational relationship—monitoring value." Monitoring values ​​provide crucial operating condition information to the knowledge graph, enabling inference and prediction to dynamically adjust based on changes in engine operation. The operational and monitoring relationships described here... These are the parameters for three operating modes: flight altitude (operating setting 1), Mach number (operating setting 2), and throttle lever solver angle (operating setting 3), and their corresponding entities. For engine.

[0073] Furthermore, by integrating the monitoring indicators of each component and unifying the association between sensor monitoring data and structural knowledge, a knowledge graph is constructed from the monitoring relationships using a "component-monitoring relationship-monitoring value" triple format. This relates to the operational and monitoring relationships. For the first Monitoring indicators ), indicating the first Data collected by a sensor.

[0074] Data collected by sensors from each component—low-pressure compressor, high-pressure compressor, low-pressure turbine, high-pressure turbine, fan, combustion chamber, core engine, and engine—forms a monitoring indicator. Each indicator is represented by a normalized value obtained in step S1, indicating the component's state or health at the current sampling moment.

[0075] The spatiotemporal knowledge graph structure information table constructed through steps S131-S132 is shown in Table 1:

[0076] Table 1. Spatiotemporal Knowledge Graph Structure Information Table

[0077]

[0078] Step S134: Repeat steps S131 to S133 to complete the construction of the knowledge graph for all time nodes of the normalized dataset, and obtain the spatiotemporal knowledge graph of the complete operation of the aero-engine.

[0079] The constructed spatiotemporal knowledge graph is represented as follows:

[0080]

[0081] in, express Knowledge graph of time nodes; This represents an entity-relationship-attribute / entity triple. Represents an entity, Indicates a relationship. Represents an attribute or entity; This represents the set of all entities; the engine components and all components together constitute the set of all entities. A collection representing numerical attributes; Indicates the hierarchical relationship of the structure; Indicates the relationship of power transmission, This indicates the relationship between operation and monitoring.

[0082] Step S2: Input the spatiotemporal knowledge graph into the BERT model and convert it into node embedding vectors and relation embedding vectors in tensor form that can be read by the neural network. Then, use a multi-head graph attention network and a pooling graph attention network to extract and fuse features from the node embedding vectors and relation embedding vectors to obtain a fused feature vector.

[0083] Step S21: Transform the spatiotemporal knowledge graph into a data format using the BERT model, converting the spatiotemporal knowledge graph composed of natural language into tensor-form node embedding vectors and relation embedding vectors, so that the multi-head attention network can extract features.

[0084] The BERT model maps nodes of a spatiotemporal knowledge graph described in natural language and the relationships between nodes to a 768 (BERT_dim) dimensional vector.

[0085]

[0086]

[0087] in, Represents a node Node embedding vectors; Represents a node Node embedding vectors; Represents a node With nodes The relationship between the embedding vectors; Represents the BERT model; This represents the vector dimension, which is 768 dimensions. Node The entity corresponding to the triple of "entity-relationship-attribute / entity" ,node Corresponding to attributes or entities .

[0088] Step S22: Input the node embedding vector and relation embedding vector converted into tensor form in step S21 into the multi-head graph attention network for feature extraction to obtain the concatenated feature vector.

[0089] Multi-head image attention network is composed of Unlike single-layer graph attention networks, which consist of parallel computations of single-layer graph attention layers, multi-head graph attention networks maintain an independent set of linear transformation weights for each attention head. These weights not only calculate the normalized attention coefficients from that specific viewpoint but also independently integrate edge features into the features of neighboring nodes for weighted aggregation, resulting in a feature vector. Ultimately, this... The feature vectors output by each single-layer graph attention layer are concatenated to obtain a concatenated feature vector, which captures the rich semantic and structural information of the spatiotemporal knowledge graph in different feature subspaces.

[0090] No. The attention coefficient of a single-layer graph is calculated as follows:

[0091]

[0092] in, Indicates the first Attention coefficients of a single-layer graph, representing the nodes Features of nodes The attention coefficient represents the number of nodes. From the perspective of nodes Information and the relationship between the two nodes How much does it contribute to its current task (such as classification, regression, etc.)? This represents a nonlinear activation function with a negative slope of 0.2; Represents the attention weight vector; Indicates transpose; Indicates a connection operation; Indicates the first Nodes in a single-layer graph attention network The linear transformation weight matrix; Indicates the first Nodes in a single-layer graph attention network The linear transformation weight matrix; Indicates the first Nodes in a single-layer graph attention network With nodes The linear transformation weight matrix of the relationship between them; Represents a node Node embedding vectors; Represents a node Node embedding vectors; Represents a node With nodes The relationship between the embedding vectors.

[0093] To make the attention coefficient easier to calculate and compare, a softmax function is introduced to normalize the attention coefficient. The normalization formula is as follows:

[0094]

[0095] in, This represents the normalized attention coefficient; Show the softmax function; Representing nodes in a spatiotemporal knowledge graph The neighboring nodes; Indicates at node Neighbor nodes The temporary index in the value is used for summation operations; This indicates the power operation with base e.

[0096] After obtaining the normalized attention coefficient After that, regarding the current number Nodes in a single-layer graph attention network Merge neighbor nodes Features and Relationship Nodes Features, thereby updating nodes The feature representation is calculated using the following formula:

[0097]

[0098] in, Indicates the updated node eigenvectors; This represents a non-linear activation function.

[0099] Multi-head graph attention networks utilize multiple independent single-layer graph attention networks to process nodes simultaneously. Each single-layer graph attention network has an independent weight matrix, enabling it to focus on different aspects of node features.

[0100] Will The feature vectors output by each single-layer graph attention layer are concatenated to fuse feature information from different subspaces, resulting in a concatenated feature vector, as shown below:

[0101]

[0102] in, Represents a node The concatenated feature vector output by the multi-head graph attention network.

[0103] Step S23: Input the spatiotemporal knowledge graph, relation embedding vector, and concatenated feature vector into the pooling graph attention network to perform feature fusion and obtain the fused feature vector.

[0104] Step S231: The pooling graph attention network constructs an adjacency matrix based on the spatiotemporal knowledge graph and relation embedding vectors.

[0105] Pooling graph attention networks integrate the local structural information of nodes in a spatiotemporal knowledge graph while identifying and assigning high weights to more important nodes in the spatiotemporal knowledge graph. First, the pooling graph attention network maps the semantic relationship information between entity nodes to an adjacency matrix. Adjacency matrix elements Mapped in the following way:

[0106]

[0107] in, Representing the adjacency matrix The Middle Line number The elements of a column represent nodes. With nodes The connection relationship between them is represented by 1 if there is a connection and 0 otherwise. Indicates the number of nodes.

[0108] Step S232: The pooling graph attention network fuses the adjacency matrix and the concatenated feature vector through graph pooling operation to obtain the fused feature vector.

[0109] The pooled graph attention network utilizes a multi-source fusion mechanism to deeply integrate semantic relationship features from the spatiotemporal knowledge graph with sensor time-series data, constructing a unified graph representation structure. The graph pooling layer fuses the adjacency matrix and concatenated feature vectors to calculate the global graph feature representation of the initial spatiotemporal knowledge graph, obtaining the fused feature vector. The calculation formula is as follows:

[0110]

[0111] in, This represents the fused feature vector, which is a graph-level feature vector. This represents graph pooling operations; Indicates global max pooling; Represents the node feature matrix, This indicates the first [head] after being extracted by the multi-head graph attention network. The concatenated feature vector of each node; This represents a multi-head attention operation; the spatiotemporal knowledge graph processed by the pooling graph attention mechanism achieves efficient fusion of semantic relations and sensor data.

[0112] Step S3: Construct a Stacked Dilated Convolutional Neural Network (SDCNN) and input the fused feature vectors into the Stacked Dilated Convolutional Neural Network to predict degradation trend features.

[0113] Step S31: Construct a stacked dilated convolutional neural network, which is composed of multiple dilated convolutional layers stacked in a cascaded manner.

[0114] Stacked dilated convolutional neural networks are composed of The dilated convolutional layers are stacked in a cascaded manner. This indicates the number of dilated convolutional layers stacked. To effectively capture multi-scale contextual information, the dilation rate of each layer is typically set to increase with the number of dilated convolutional layers (e.g., exponential growth), such as... Figure 2 As shown, this is an optional implementation, with the number of dilated convolutional layers stacked... The expansion rate is 4, and the expansion rate of each layer is 4. The growth rate increases exponentially with the number of dilated convolutional layers. In the processing of stacked dilated convolutional neural networks, the fused feature vector is used as input and flows sequentially through... Each of the dilated convolutional layers utilizes an expanded convolutional kernel to aggregate broader contextual information, ultimately outputting deep features rich in semantics. Dilated convolutional neural networks can exponentially increase the receptive field by changing the dilation rate, without completely abandoning pooling layers. This allows them to capture extensive long-range dependencies in the fused vector while maintaining high-resolution feature information, achieving a transformation from shallow to deep features.

[0115] The formula for calculating the receptive field of a single dilated convolutional layer is:

[0116]

[0117] in, The dilation rate represents the interval span of the convolutional kernel across the input features; Indicates the kernel size; This indicates the receptive field size of the dilated convolutional layer, which is the range of length of the input sequence that a single dilated convolutional layer can sense. This allows for better capture of local and long-term dependency information and more accurate modeling of the degradation trend of long sequences.

[0118] Step S32: Input the fused feature vector into the first dilated convolutional layer, perform weighted summation and activation through the dilated convolution kernel, and output the first convolutional feature. The first convolutional feature is input into the second dilated convolutional layer, and outputs the second convolutional feature, and so on, until the last dilated convolutional layer outputs the last convolutional feature, which is the degradation trend feature, such as... Figure 3 As shown.

[0119] No. The calculation method for each dilated convolutional layer is as follows:

[0120]

[0121]

[0122] in, For the first Each dilated convolutional layer at time step The convolutional features, Indicates the total length of the time series; Indicates the first The convolutional features output by each dilated convolutional layer; Represents a nonlinear activation function; Indicates the kernel size. This represents the index variable inside the convolution kernel; Indicates the first In the dilated convolutional layer, the convolution kernel is at the _th ... Trainable weight matrix at each position; Indicates a time step. Indicates the first The expansion rate of the layer, Indicates the sampling index; Indicates the first The convolutional features output by each dilated convolutional layer The sampling features define how feature data is acquired in an expanded manner; Indicates the first The bias of each dilated convolutional layer.

[0123] The output of the last dilated convolutional layer is represented as follows:

[0124]

[0125] in, This represents the last convolutional feature output by the last dilated convolutional layer, i.e., the degradation trend feature. This represents a stacked dilated convolutional neural network; This indicates that the dilated convolutional layers are nested. This indicates that the fused feature vectors are input into the first dilated convolutional layer for processing. This indicates that the first convolutional feature is input into the second dilated convolutional layer for processing; This indicates the processing of the last dilated convolutional layer. The output of each dilated convolutional layer.

[0126] Step S33: Calculate the receptive field of the multi-layer stack and verify the degradation trend characteristics. Does it cover a sufficiently long input history? The calculation method for the receptive field of multi-layer stacked receptive fields is as follows:

[0127]

[0128] in, This represents the receptive field of the entire Stacked Dilated Convolutional Neural Network (SDCNN), used to ensure that the network can cover a sufficiently long input sequence, but will not directly participate in local convolution calculations; Indicates the number of stacked dilated convolutional layers; Indicates the first The expansion rate of the layer; This indicates the kernel size.

[0129] Furthermore, the receptive field of the entire stacked dilated convolutional neural network is increased. With fused feature vectors The sequence lengths are compared, if This indicates that the stacked dilated convolutional neural network is stacked deep enough to completely cover the fused feature vectors, capturing all causal relationships from the earliest time to the current time without missing important historical information; conversely, it indicates that the network's "field of vision" is too narrow and it has not captured all causal relationships.

[0130] Step S4: Divide 70% of the fused feature vectors obtained from feature extraction and feature fusion into a training set, and the remaining 30% into a test set. Use the remaining service life cycle label from step S11 as the training label for the training set data, and input it into the stacked dilated convolutional neural network for training to obtain the trained aircraft engine remaining service life prediction model.

[0131] During training, the input training data is processed layer by layer through multiple dilated convolutional layers. Each convolutional layer can automatically extract local features and gradually expand the receptive field by stacking. After extracting the sequence features, the degradation trend feature vector output by the stacked dilated convolutional neural network is passed to the fully connected layer for prediction. The fully connected layer maps the features extracted by the stacked convolutional layers to the prediction space to generate the remaining lifespan prediction value corresponding to each input sample.

[0132] By monitoring the training process of the stacked dilated convolutional neural network, it can be determined whether the model has converged to the optimal prediction model or completed the training within the predetermined number of training iterations. If so, the model training has been completed, and a well-trained prediction model for the remaining service life of the aero-engine is obtained.

[0133] Step S5: Use the trained aircraft engine remaining service life prediction model to predict the remaining service life of the aircraft engine.

[0134] This invention establishes evaluation indicators for the prediction model of the remaining service life of this aero-engine, namely the root mean square error (RMSE) and the coefficient of determination (R²), and the calculation formulas for the evaluation indicators are as follows:

[0135]

[0136]

[0137] in, Indicates the root mean square error; Indicates the coefficient of determination; The number of samples; Represents the actual value; Indicates the model's predicted value; The mean of the true values; This represents the sum of squared residuals, which measures the error between the model's predicted values ​​and the actual values. It represents the total sum of squares, which measures the variance of the target variable itself.

[0138] To better understand and implement this invention, specific embodiments are provided below using simulation data and accompanying drawings to illustrate the method used in this invention in detail.

[0139] This embodiment uses the problem of predicting the remaining service life of aero-engines using the C-MAPSS dataset as an example to illustrate a method for predicting the remaining service life of aero-engines based on spatiotemporal knowledge graphs and SDCNN.

[0140] Step S1: Obtain data from 21 sensors in the C-MAPSS dataset, and generate a health status of the knowledge graph at a specific time point, as shown below. Figure 4 As shown, a remaining lifespan label is added to each data point in the collected dataset; the sensor data in the dataset is normalized, and a spatiotemporal knowledge graph is constructed. The constructed spatiotemporal knowledge graph data is as follows. Figure 5 As shown.

[0141] Step S2: Input the spatiotemporal knowledge graph into the BERT model and convert it into node embedding vectors and relation embedding vectors in tensor form that can be read by the neural network. Then, use a multi-head graph attention network and a pooling graph attention network to extract and fuse features from the node embedding vectors and relation embedding vectors to obtain a fused feature vector.

[0142] The hyperparameters of the BERT model, multi-head graph attention, and pooling graph attention network are set as shown in Table 2.

[0143] Table 2. Hyperparameter settings for BERT model, multi-head graph attention, and pooling graph attention networks.

[0144]

[0145] Step S3: Construct a stacked dilated convolutional neural network and input the fused feature vector into the stacked dilated convolutional neural network to predict degradation trend features.

[0146] Step S4: Divide 70% of the fused feature vectors obtained from feature extraction and feature fusion into a training set, and the remaining 30% into a test set. Use the remaining service life cycle label from step S11 as the training label for the training set data, and input it into the stacked dilated convolutional neural network for training to obtain the trained aircraft engine remaining service life prediction model.

[0147] For the Stacked Dilated Convolutional Neural Network (SDCNN), this invention performs hyperparameter tuning and optimization to achieve optimal performance in specific task scenarios. Hyperparameters include, but are not limited to, key parameters such as learning rate, batch size, number of hidden units, and number of layers. The hyperparameter settings for the Stacked Dilated Convolutional Neural Network are shown in Table 3.

[0148] Table 3. Hyperparameter settings for stacked dilated convolutional neural networks

[0149]

[0150] The training set is input into the model for training, and a validation set is introduced during the training process to verify the current learning status of the model. The loss curve is monitored using the Wandb library. If the validation set index does not improve for several consecutive rounds, the training is stopped and the optimal weights are saved. The trained aircraft engine remaining service life prediction model is obtained.

[0151] Step S7: Input the test set data into the trained aircraft engine remaining service life prediction model to predict the remaining service life of the aircraft engine.

[0152] The prediction performance of the aero-engine remaining service life prediction model on the test set is evaluated as shown in Table 4. The results demonstrate that the method of this invention exhibits excellent prediction accuracy and robustness, better copes with uncertainties in complex systems, significantly improves system reliability, provides important technical support and theoretical basis for practical applications, and offers valuable reference and new research directions for the technological development of aero-engine remaining service life prediction.

[0153] Table 4. Prediction Results of Remaining Service Life of Aero Engines

[0154]

[0155] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An aero-engine residual useful life prediction method based on a space-time knowledge graph and an SDCNN, characterized in that, The method comprises the following steps: Step S1: The sensor collects data of the aero-engine component to obtain a collected data set, adds a remaining use life cycle label to obtain an aero-engine degradation data set, and performs normalization processing on the sensor collected data in the aero-engine degradation data set, and constructs a space-time knowledge graph based on the normalized data; Step S2: inputting the space-time knowledge graph into a BERT model to convert it into node embedding vectors and relationship embedding vectors in the form of tensors readable by a neural network, and then performing feature extraction and feature fusion on the node embedding vectors and the relationship embedding vectors through a multi-head graph attention network and a pooling graph attention network to obtain a fusion feature vector; Step S3: constructing a stacked dilated convolutional neural network, inputting the fusion feature vector into the stacked dilated convolutional neural network to predict a degradation trend feature; Step S4: dividing 70% of the fusion feature vector after feature extraction and feature fusion into a training set, and the remaining 30% into a test set, taking the remaining use life cycle label as a training label for the training set data, inputting the training set data into the stacked dilated convolutional neural network for training, and obtaining a trained aero-engine remaining use life prediction model; Step S5: using the trained aero-engine remaining use life prediction model to predict the remaining use life of the aero-engine; Constructing a space-time knowledge graph based on the normalized data comprises the following steps A to D: Step A: hierarchically modeling the overall structure of the aero-engine, expressing the subordinate relationship between components through a triple form of "component - contains - component", and constructing a knowledge graph from the structural hierarchical relationship; Step B: reflecting the aerodynamic and mechanical coupling relationship between components by a triple form of "component - drive / energy supply / transmission - component", and constructing a knowledge graph from the power transmission relationship; Step C: for the overall operating state relationship of the engine, integrating the monitoring indexes of each component, uniformly associating the sensor monitoring data with the structural knowledge, and constructing a knowledge graph from the operation and monitoring relationship through a triple form of "component - operation / monitoring relationship - monitoring value"; Step D: repeatedly performing steps A to C to complete the knowledge graph construction of all time nodes of the normalized data set, and obtaining a space-time knowledge graph of the complete operation of the aero-engine; Step S2 comprises the following steps: Step S21: converting the space-time knowledge graph into node embedding vectors and relationship embedding vectors in the form of tensors through a BERT model, so as to extract features through a multi-head attention network; Step S22: inputting the node embedding vectors and the relationship embedding vectors converted into tensors in step S21 into a multi-head graph attention network to extract features and obtain a splicing feature vector; Step S23: inputting the space-time knowledge graph, the relationship embedding vectors and the splicing feature vector into a pooling graph attention network to perform feature fusion and obtain a fusion feature vector.

2. The aero-engine residual useful life prediction method based on spatiotemporal knowledge graph and SDCNN according to claim 1, characterized in that, The step S1 comprises: Step S11: The sensor collects data of the aero-engine component to obtain a collected data set, and adds a remaining use life cycle label to each piece of data in the collected data set to obtain an aero-engine degradation data set; Step S12: The sensor collected data in the aero-engine degradation data set is normalized to obtain a normalized data set; Step S13: The normalized data set is constructed into a space-time knowledge graph in the form of a triple of "entity-relation-attribute / entity" from three aspects of a structural hierarchical relationship, a power transmission relationship and an operation and monitoring relationship.

3. The aero-engine remaining useful life prediction method based on spatiotemporal knowledge graph and SDCNN according to claim 2, characterized in that, The normalization method of step S12 is as follows: wherein, is the data of the acquired data set; is the normalized data; is the minimum value in the data of the acquired data set; is the maximum value in the data of the acquired data set.

4. The aero-engine residual useful life prediction method based on spatiotemporal knowledge graph and SDCNN according to claim 1, characterized in that, The space-time knowledge graph constructed in step D is represented as follows: wherein, represents a knowledge graph of time nodes; represents an entity-relation-property / entity triple, represents an entity, represents a relation, represents a property or an entity; represents a set of all entities, all components of the engine assembly constitute the set of all entities; represents a set of numerical properties; represents a structural hierarchy relation; represents a power transmission relation, represents a running and monitoring relation.

5. The aero-engine residual useful life prediction method based on spatiotemporal knowledge graph and SDCNN according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: A stacked dilated convolutional neural network is constructed, and the stacked dilated convolutional neural network is stacked by a plurality of dilated convolutional layers in a cascaded manner; Step S32: The fusion feature vector is input into a first dilated convolutional layer, weighted summation and activation are performed through a dilated convolutional kernel, a first convolutional feature is output, the first convolutional feature is input into a second dilated convolutional layer, a second convolutional feature is output, and the like is sequentially performed until a last dilated convolutional layer outputs a last convolutional feature, i.e., a degradation trend feature. Step S33: Compute multi-layer stack receptive field, verify degenerate trend signature Whether enough long input history information is covered.

6. The aero-engine residual useful life prediction method based on spatiotemporal knowledge graph and SDCNN according to claim 5, characterized in that, No. The calculation method for each dilated convolutional layer is as follows: wherein, is the output of the th dilated convolutional layer at time step is the convolutional feature of the th dilated convolutional layer; is the total length of the time series; is the convolutional feature of the th dilated convolutional layer; is the size of the convolutional kernel, is an index variable within the convolutional kernel; is the matrix of trainable weights of the th dilated convolutional layer at the th position of the convolutional kernel; is the time step, is the dilation rate of the th layer, is the sampling index; is the sampled feature of the convolutional feature of the th dilated convolutional layer; is the bias of the th dilated convolutional layer.

Images