Aero-engine life prediction method fusing dynamic knowledge graph and dynamic attention mechanism

Abstract

The application provides an aero-engine life prediction method fusing a dynamic knowledge graph and a dynamic attention mechanism, and comprises the following steps: obtaining historical monitoring data of an aero-engine and performing preprocessing to obtain sample data; constructing a dynamic knowledge graph by using entity relationships and dynamic coupling relationships among sensors; extracting spatial features among the sensors by using a graph neural network; designing a dynamic attention mechanism with adaptive weights and constructing a Transformer network architecture to extract time series features; establishing and training an aero-engine life prediction model based on the spatial features and the time series features among the sensors; and performing aero-engine life prediction by using the trained aero-engine life prediction model. The aero-engine life prediction method can solve the problems of ignoring sensor entity connections, coupling relationships and long-range dependence information loss in traditional data-driven degradation feature prediction methods, and precise prediction is achieved.

Description

Technical Field

[0001] This invention relates to the field of aero-engine life prediction, and in particular to an aero-engine life prediction method that integrates dynamic knowledge graphs and dynamic attention mechanisms. Background Technology

[0002] Predicting the remaining life of aero-engines in real-world scenarios is a typical example of multivariate time-series data modeling in complex systems. It requires not only accurately capturing the physical relationships and dynamic coupling between sensors but also effectively addressing potential anomalies in the time-series data (such as noise, missing data, or even erroneous data). While traditional data-driven methods can solve multivariate time-series data modeling problems to some extent, they often neglect the spatial correlations and dynamic relationships between sensors. Furthermore, these methods are prone to losing long-range dependency information when dealing with long-term series, leading to inaccurate and unstable prediction results.

[0003] Therefore, providing a method to accurately predict the lifespan of aero engines has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the present invention provides a method for predicting the lifespan of aero-engines that integrates dynamic knowledge graphs and dynamic attention mechanisms, in order to solve the problems existing in traditional aero-engine lifespan prediction methods.

[0005] This invention provides a method for predicting the lifespan of aero-engines that integrates dynamic knowledge graphs and dynamic attention mechanisms, comprising:

[0006] Step 1: Acquire historical monitoring data from 21 sensors installed on the aero-engine and preprocess it to obtain sample data. Then, divide the sample data into training set and test set. The 21 sensors are used to collect data on the following parameters: total fan inlet temperature, total low-pressure compressor outlet temperature, total high-pressure compressor outlet temperature, total low-pressure turbine outlet temperature, fan inlet pressure, total bypass pipe pressure, total high-pressure compressor outlet pressure, fan speed, core shaft speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow rate to Ps30 ratio, fan correction speed, core shaft correction speed, bypass ratio, combustion chamber fuel-air ratio, extraction enthalpy, fan demand speed, fan correction demand speed, high-pressure turbine cooling exhaust volume, and low-pressure turbine cooling exhaust volume.

[0007] Step 2: Construct entity relationships between sensors based on their physical locations and functions to obtain an entity knowledge graph;

[0008] Step 3: Calculate the correlation between each sensor and apply evidence theory to evaluate the dynamic coupling relationship between the sensors, and construct a coupling knowledge graph;

[0009] Step 4: Merge the entity knowledge graph and the coupled knowledge graph into a unified dynamic knowledge graph;

[0010] Step 5: Use a graph neural network to embed the nodes and edges in the dynamic knowledge graph, extract the spatial features between sensors, and generate a spatial feature vector to characterize the sensor associations;

[0011] Step 6: Design a dynamic attention mechanism with adaptive weights to dynamically adjust the attention weights based on the importance of the spatiotemporal features of the sensor input;

[0012] Step 7: Construct a Transformer network architecture based on dynamic attention mechanism to extract long-term sequence features of sensor input;

[0013] Step 8: Based on the spatial features between sensors extracted by the graph neural network and the time series features extracted by the Transformer, a deep neural network model for predicting the remaining life of aero-engines is established, and the model is trained and validated using the training set and the test set to obtain the trained aero-engine life prediction model.

[0014] Step 9: Use the trained aero-engine life prediction model to predict the life of the aero-engine.

[0015] Preferably, in step 1, the preprocessing includes: data normalization, slicing the normalized data, and setting lifetime labels.

[0016] Further optimization involves dividing the entity relationships into five entity knowledge graphs in step 2: temperature management chain, pressure management chain, speed control chain, fuel-to-air ratio chain, and cooling and heat dissipation chain.

[0017] Further optimization involves using the Pearson correlation coefficient, Spearman correlation coefficient, and mutual information between sensors in step 3 to measure the correlation, and then constructing a dynamic sensor coupling relationship analysis model based on the evidence theory framework.

[0018] Further preferably, in step 6, the dynamic attention mechanism is formally expressed as follows:

[0019] Dynamic_Q=X t W Dynamic_Q ;

[0020] Dynamic_K = X t W Dynamic_K ;

[0021] Dynamic_V = X t W Dynamic_V ;

[0022] Q = Dynamic_Q·W Q ;

[0023] K = Dynamic_K·W K ;

[0024] V = Dynamic_V·W V ;

[0025]

[0026] in, These represent the query matrix, key matrix, and value matrix after weight allocation, respectively, which will undergo three different linear transformations. Acting on the labeled embedding sequence X t To obtain; These represent the query matrix, key matrix, and value matrix, respectively, which will be transformed through three different linear transformations. Acting on embedded sequences To obtain.

[0027] Further optimization, in step 7, the feedforward network in the Transformer is used to map the features extracted by multi-head attention to more abstract features, and combines each different attention head and uses linear mapping to obtain multi-head dynamic attention values;

[0028] The feedforward network consists of two linear transformations and one non-linear activation layer. In the Transformer layer, residual connections and layer normalization are applied to the output of the multi-head attention mechanism and the feedforward neural network. The entire structure can be regarded as being composed of multiple sub-layers, including the multi-head attention module and the feedforward network. The output of each sub-layer is added to the input through residual connections and then subjected to layer normalization.

[0029] Further optimization involves step 8, where the training is completed by determining whether the model has reached the minimum performance metric or the maximum number of training iterations. If so, the model training is complete.

[0030] Further optimization involves evaluating the performance metrics of the aero-engine life prediction model in step 8, specifically RMSE and Score, as follows:

[0031]

[0032] in, y represents the predicted remaining lifetime. i This represents the actual remaining lifespan.

[0033] This invention provides a method for predicting the lifespan of aero-engines by integrating dynamic knowledge graphs and dynamic attention mechanisms. It constructs a dynamic knowledge graph to systematically describe the entity relationships between sensors and their dynamic coupling relationships over time. A graph neural network is used to extract spatial features between sensors to comprehensively characterize sensor associations. Simultaneously, an adaptive weighted dynamic attention mechanism is designed to optimize the extraction of key features in long-term series modeling. Based on this, an improved Transformer framework is used to fully mine the spatiotemporal information in sensor data, avoiding the loss of long-range dependency information. Furthermore, this method introduces a collaborative mechanism between the dynamic knowledge graph and a deep learning model, performing relationship modeling and feature extraction on sensor data one by one, and dynamically adjusting the contribution weights of each sensor through a modular network structure. Finally, the model achieves accurate prediction of aero-engine lifespan by efficiently fusing spatial and sequential features embedded in the knowledge graph. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of an embodiment of the aero-engine life prediction method that integrates dynamic knowledge graph and dynamic attention mechanism according to the present invention.

[0037] Figure 2 Diagram of the sliding window method;

[0038] Figure 3 This is a schematic diagram of a turbofan engine.

[0039] Figure 4 For a knowledge graph of the temperature management chain;

[0040] Figure 5 For stress management chain knowledge graph;

[0041] Figure 6 For velocity control chain knowledge graph;

[0042] Figure 7 A knowledge graph of the fuel-to-air ratio;

[0043] Figure 8 A knowledge graph of cooling and heat dissipation chains;

[0044] Figure 9 A knowledge graph of entity relationships and coupling relationships for a single input;

[0045] Figure 10 Here is a diagram of the GCN model structure;

[0046] Figure 11 The diagrams show a comparison of the structures of dynamic attention mechanisms and ordinary attention mechanisms, where (a) is the structure of ordinary attention mechanisms and (b) is the structure of dynamic attention mechanisms.

[0047] Figure 12 Here is a diagram of the Transformer architecture;

[0048] Figure 13 shows the results of remaining lifetime prediction using the method provided by the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0050] To address the problems existing in current aero-engine life prediction methods, such as... Figure 1 As shown, this invention provides a method for predicting the lifespan of aero-engines that integrates dynamic knowledge graphs and dynamic attention mechanisms, comprising the following steps:

[0051] Step 1: Acquire historical monitoring data from 21 sensors installed on the aero-engine and preprocess it to obtain sample data. Then, divide the sample data into training set and test set for training and validation of the prediction model. The 21 sensors are used to collect data on the following parameters: total fan inlet temperature, total low-pressure compressor outlet temperature, total high-pressure compressor outlet temperature, total low-pressure turbine outlet temperature, fan inlet pressure, total bypass pipe pressure, total high-pressure compressor outlet pressure, fan speed, core shaft speed, engine pressure ratio (P50 / P2), high-pressure compressor outlet static pressure, fuel flow rate to Ps30 ratio, fan correction speed, core shaft correction speed, bypass ratio, combustion chamber fuel-air ratio, extraction enthalpy, fan demand speed, fan correction demand speed, high-pressure turbine cooling exhaust volume, and low-pressure turbine cooling exhaust volume.

[0052] The preprocessing includes: data normalization, slicing the normalized data, and setting lifetime labels.

[0053] The formula for data normalization is:

[0054]

[0055] In the formula, x k This is the kth input data.

[0056] The method for slicing the normalized data is as follows: Figure 2 As shown, a time window method is used for data slicing. The window setting process uses a fixed-size time window to segment normalized data samples. Each window contains sensor data from multiple time steps. The sliding process involves moving the window along the time direction in steps of 1, gradually generating new samples until all time series data are covered. Assuming the time length in the sample is T, the slice length is P, the slice sliding distance is 1, and the sliced ​​samples are (1, P), (2, P+1), ..., (TP, T).

[0057] In the early stages, the performance of aero-engine components is good, and degradation is negligible. Therefore, it is unnecessary to predict the remaining service life (RUL) at this stage, and the RUL can be set to a constant value. The RUL tag threshold for the early stage of the aero-engine is set to 125, and after a period of operation, it transitions to a linear decreasing mode. The RUL tag is calculated by selecting the maximum value from each aero-engine operating cycle, decreasing it sequentially until it reaches 0, marking the end of one operating cycle. The RUL tag formula is:

[0058]

[0059] In the formula: RUL is the remaining lifetime value.

[0060] The training set contains data from various sensors and runtimes of the engine from its initial state until a fault occurs; the test set contains data from various sensors and runtimes of the engine from its initial state to a certain point in time.

[0061] Step 2: Construct entity relationships between sensors based on their physical locations and functions to obtain an entity knowledge graph;

[0062] The entity relationships are divided into five entity knowledge graphs: temperature management chain, pressure management chain, speed control chain, fuel-to-air ratio chain, and cooling and heat dissipation chain.

[0063] The temperature management chain involves several temperature parameters. The total temperature at the fan inlet is raised to the total temperature at the low-pressure compressor outlet after compression, indicating a boundary relationship between these two temperatures. The low-pressure compressor outlet temperature is further raised to the total temperature at the high-pressure compressor outlet after high-pressure compression, again indicating a boundary relationship between these two temperatures. Finally, the high-temperature compressed gas, after combustion and expansion, is cooled by the low-pressure turbine to the temperature recorded by the low-pressure turbine outlet temperature sensor, indicating a boundary relationship between the high-pressure compressor outlet temperature and the low-pressure turbine outlet temperature.

[0064] In the pressure management chain, the fan inlet pressure is distributed to the bypass pipe during bypass flow, meaning there is a boundary relationship between the fan inlet pressure and the total bypass pipe pressure. Some gas enters the compressor and reaches the high-pressure compressor outlet pressure at the high-pressure compressor outlet total pressure, meaning there is a boundary relationship between the bypass pipe total pressure and the high-pressure compressor outlet total pressure. The pressure ratio between the high-pressure compressor outlet and inlet is calculated using an engine pressure ratio (P50 / P2) sensor, meaning there is a boundary relationship between the high-pressure compressor outlet total pressure and the engine pressure ratio (P50 / P2). The high-pressure compressor outlet total pressure affects the high-pressure compressor outlet static pressure, used to monitor compression efficiency. There is a boundary relationship between the high-pressure compressor outlet total pressure and the high-pressure compressor outlet static pressure.

[0065] The speed control chain, specifically the actual fan speed and the corrected speed, forms the basis for real-time monitoring and adjustment of the fan's operating status; that is, there is a direct relationship between the fan speed and the corrected speed. The core shaft speed is linked to the correction value to ensure that the output thrust meets requirements; that is, there is a direct relationship between the core shaft speed and the corrected speed. The required fan speed serves as a reference for the fan's workload, and the corrected required fan speed takes into account environmental influences; that is, there is a direct relationship between the required fan speed and the corrected required fan speed.

[0066] Among these factors, the fuel-to-air ratio, the fuel flow rate to Ps30 ratio, and the combustion chamber fuel-air ratio directly affect combustion efficiency; that is, the fuel flow rate to Ps30 ratio and the combustion chamber fuel-air ratio have a lateral relationship. The bypass ratio is closely related to the combustion chamber fuel-air ratio, and it determines the fuel-air mixing ratio; that is, the fuel flow rate to Ps30 ratio has a lateral relationship.

[0067] Among them, the cooling and heat dissipation chain, the extraction enthalpy affects the cooling demand, and the cooling exhaust volume of the high-pressure turbine and the low-pressure turbine are measured by the high-pressure turbine cooling exhaust volume and the low-pressure turbine cooling exhaust volume sensors, respectively. That is, there is a boundary relationship between the extraction enthalpy and the high-pressure turbine cooling exhaust volume, the extraction enthalpy and the low-pressure turbine cooling exhaust volume, and the high-pressure turbine cooling exhaust volume and the low-pressure turbine cooling exhaust volume.

[0068] Each sensor is treated as a node, and the edge relationships between sensors are treated as edge relationships, thus constructing an entity knowledge graph;

[0069] Step 3: Calculate the correlation between each sensor and apply evidence theory to evaluate the dynamic coupling relationship between sensors, constructing a coupling knowledge graph, such as... Figures 4-8 As shown;

[0070] The correlation between sensors is assessed using multiple correlation measures, including the Pearson correlation coefficient, Spearman Rank correlation coefficient, and mutual information. A dynamic sensor coupling relationship analysis model is constructed using an evidence theory framework. The correlation index of each input data is calculated sequentially to evaluate the correlation strength between sensors in real time. Based on evidence theory, the results of multiple correlation measures are integrated to achieve dynamic updates of sensor coupling relationships.

[0071] The Pearson correlation coefficient is used to measure the linear correlation between sensor output data, and it is defined as follows:

[0072]

[0073] In the formula, and denoted as the means of variables X and Y, respectively, and n is the sample size.

[0074] The Spearman correlation coefficient reflects the nonlinear relationship by comparing the order of variables, and its calculation formula is as follows:

[0075]

[0076] In the formula, d i =rank(x) i )-rank(y i ), where n is the number of samples.

[0077] To capture the statistical dependencies between variables, mutual information quantifies the degree of information sharing by comparing the joint probability distribution and the marginal probability distribution. Its mathematical expression is as follows:

[0078]

[0079] In the formula, p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions of X and Y, respectively.

[0080] Subsequently, evidence theory is introduced to fuse the results of the various relevance measures mentioned above. By defining the Basic Probability Assignment (BPA), the relevance results obtained by different methods are transformed into independent evidence representations, where BPA satisfies:

[0081] m(A)∈[0,1];

[0082]

[0083] Where Ω represents the proposition space.

[0084] Subsequently, the Dempster combination rule was used to fuse the multi-source evidence, the mathematical expression of which is:

[0085]

[0086] Where m1(A) and m2(B) are the BPAs of different correlation methods, respectively. c (C) represents the merged BPA, with the denominator used to exclude conflicting evidence.

[0087] Each sensor is considered a node. When the calculated deterministic value in the evidence theory is greater than 0.8, it is assumed that there is a coupling relationship between two sensors, i.e., there is an edge relationship, thus constructing a coupled knowledge graph.

[0088] Step 4: Merge the entity knowledge graph and the coupled knowledge graph into a unified dynamic knowledge graph to comprehensively reflect the complex relationships between sensors;

[0089] Specifically, nodes are merged and connected by edge relationships, thereby constructing a dynamic knowledge graph;

[0090] Among them, the dynamic knowledge graph of entity relationships and coupling relationships constructed based on a single input data is as follows: Figure 9 As shown.

[0091] Step 5: Use a graph neural network (GCN) to embed the nodes and edges in the dynamic knowledge graph, extract the spatial features between sensors, and generate a spatial feature vector to characterize the sensor associations;

[0092] Among them, the knowledge graph, i.e., the graph structure G, can be expressed as G = (V G E G V G Let E represent the main body that constructs the network of relationships. G Let edges represent the relationships between entities. The mathematical expression for a Graph Neural Network (GCN) is:

[0093]

[0094] Among them, H (l) This represents the feature matrix of the nodes in the l-th layer, with dimensions N×F. (l) N is the number of nodes, F (l) It is the feature dimension. The adjacency matrix A of the graph plus the identity matrix I N This indicates that each node includes its own adjacency information. It is an adjacency matrix The degree matrix, where the diagonal elements are the degrees of the nodes, W (l) It is the trainable weight matrix of the l-th layer, with dimension F. (l) ×F (l+1) σ(·) is an activation function, such as ReLU;

[0095] The graph neural network (GCN) structure diagram is as follows: Figure 10 As shown, the graph neural network (GCN) is set to 2 layers;

[0096] Step 6: Design a dynamic attention mechanism with adaptive weights to dynamically adjust the attention weights based on the importance of the spatiotemporal features of the sensor input, including the allocation of query (Q), key (K), and value (V), to optimize the capture capability of important features and suppress the influence of redundant features.

[0097] The dynamic attention mechanism allows the model to flexibly adjust attention weights, using various Q, K, and V values ​​for attention calculation. This ensures that each position in the information extracted by the attention mechanism draws useful information from other positions, making it more effective for modeling long-term series dependencies. Compared to static attention mechanisms, dynamic attention mechanisms can flexibly adjust attention weights based on different input data, focusing on the most relevant parts of the current task. This allows the model to better handle input data that is variable and diverse. Formulaically, this can be expressed as:

[0098] Dynamic_Q=X t W Dynamic_Q ;

[0099] Dynamic_K = X t W Dynamic_K ;

[0100] Dynamic_V = X t W Dynamic_V ;

[0101] Q = Dynamic_Q·W Q ;

[0102] K = Dynamic_K·W K ;

[0103] V = Dynamic_V·W V ;

[0104]

[0105] in, These represent the query matrix, key matrix, and value matrix after weight allocation, respectively, which will undergo three different linear transformations. Acting on the labeled embedding sequence X t To obtain. These represent the query matrix, key matrix, and value matrix, respectively, which will be transformed through three different linear transformations. Acting on embedded sequences To obtain;

[0106] The comparison diagram between dynamic attention mechanisms and traditional attention mechanisms is shown below. Figure 11 As shown.

[0107] Step 7: Construct a Transformer network architecture based on dynamic attention mechanism to extract long-term series features from sensor input; through multi-head attention mechanism and residual connection design, fully explore long-range dependencies in the data and improve the model's ability to model multivariate time series data;

[0108] The Transformer structure diagram is as follows: Figure 12 As shown, the feedforward network in the Transformer is used to map the features extracted by multi-head attention to more abstract features. Specifically, the attention mechanism described emphasizes the differences in features only in one representation space, while the multi-head dynamic attention mechanism can perform weighted operations on features at different positions in multiple different representation subspaces, thus having stronger feature representation capabilities. Specifically, each different attention head is combined, and a linear mapping is used to obtain the multi-head dynamic attention value, which is formulaically expressed as:

[0109] MH(Q,K,V)=Concat(H1,...,H h )·W o ;

[0110]

[0111] in, This represents the weight matrix of Q, K, and V in the i-th attention head. This represents the weight matrix for multi-head attention, where Concat represents the number of connections, h represents the number of attention heads, and d represents the weight matrix. v The dimension representing the attention value.

[0112] The feedforward network consists of two linear transformations and one non-linear activation layer. To improve convergence, the traditional ReLU activation function is used. In the Transformer layers, residual connections and layer normalization are applied to the multi-head attention mechanism and the output of the feedforward neural network. Specifically, the entire structure can be viewed as composed of multiple sublayers, including the multi-head attention module and the feedforward network. The output of each sublayer is added to the input via residual connections, followed by layer normalization. This structure not only stabilizes gradient flow but also effectively prevents gradient vanishing, thereby enhancing the model's learning ability. Their outputs will be expressed as:

[0113] o=LayerNorm(x+Sublayer(x));

[0114] In this way, the model can better capture long-term dependency features, improving accuracy and generalization ability. Therefore, multivariate long-term series data, through stacked Transformers of 6 layers, can effectively extract hidden temporal dependency features and ultimately be used for remaining lifespan prediction tasks.

[0115] Step 8: Based on the spatial features between sensors extracted by the graph neural network (GCN) and the time series features extracted by the Transformer, a deep neural network model for predicting the remaining life of aero-engines is established, and the model is trained and validated using the training set and the test set to obtain the trained aero-engine life prediction model.

[0116] The training loop proceeds, and the completion of training is determined by whether the model has reached the minimum performance metric or the maximum number of training iterations. If so, the model training is complete.

[0117] The performance metrics for evaluating aero-engine life prediction models are RMSE and Score, as detailed below:

[0118]

[0119] in, y represents the predicted remaining lifetime. i This represents the actual remaining lifespan.

[0120] Step 9: Use the trained aero-engine life prediction model to predict the life of the aero-engine.

[0121] At each prediction time, dynamically updated sensor data is input, and the model's structure and parameters are dynamically adjusted through joint processing of knowledge graphs and deep learning models to output the remaining life prediction results of the aero-engine in real time; combined with the prediction results, support is provided for subsequent continuous decision-making.

[0122] This invention provides a method for predicting the lifespan of aero-engines by integrating dynamic knowledge graphs and dynamic attention mechanisms. This addresses the problems of traditional data-driven degradation feature prediction methods that neglect sensor entity relationships and coupling, as well as the loss of long-range dependency information. The method first converts the entity relationships and dynamic coupling between sensors into a knowledge graph representation, extracting the embedding features of nodes and edges in the graph through a graph neural network. Then, it introduces an adaptive weighted dynamic attention mechanism, enabling the Transformer-based deep learning model to dynamically allocate the weights of Q, K, and V, thereby enhancing the model's ability to capture key features. This significantly reduces the risk of losing long-range dependency information, especially when processing long-term series data from complex systems.

[0123] This invention provides a method for predicting the lifespan of aero-engines by integrating dynamic knowledge graphs and dynamic attention mechanisms. It uses different sensors within the aero-engine as entities, establishing a relational network between entities based on their physical location and function, generating a structured entity relationship knowledge graph. Coupling relationships exist between the various sensors; these relationships are constructed based on evidence theory and fused with the entity relationships to form the knowledge graph. The knowledge graph, built based on sensor location and function, can reflect the entity connections and dynamic coupling relationships between sensors in a structured manner, making the model more consistent with actual physical meaning when utilizing input data. Graph Convolutional Networks (GCNs) are used to extract relational features between sensors and latent patterns in time-series data, embedding these features into the knowledge graph and representing it as long-term temporal features. This method can fully exploit the spatial dependencies between sensors, enhance the physical interpretability of input features, and effectively reduce the interference of redundant information on model predictions. Through this representation, the model can more accurately capture the complex relationships between sensors, providing more discriminative input features for lifespan prediction. The knowledge graph features embedded in GCN can capture complex spatial relationships between sensors, providing high-quality, low-redundancy input features for deep learning models and significantly improving the model's ability to perceive multivariate data. Simultaneously, this invention proposes a Transformer neural network based on multivariate long-term series, retaining the multi-head attention mechanism and residual connection design of the traditional Transformer, effectively capturing long-term dependency features relevant to prediction. Unlike the classic Transformer, this invention replaces the decoder module with an efficient fully connected network, simplifying the model structure and significantly improving training efficiency. Furthermore, the proposed dynamic attention mechanism optimizes the weight allocation of Q, K, and V values ​​according to the importance of the input data, focusing more on temporal features crucial for lifetime prediction. Replacing the decoder module with a fully connected network reduces network complexity, lowers computational costs, and accelerates the model training process. The dynamic attention mechanism adaptively allocates attention weights, enabling the model to capture key features more accurately and reducing the interference of irrelevant features on prediction performance. By retaining the multi-head attention mechanism, the model can efficiently capture dependencies in long-term series, thereby improving the accuracy and robustness of RUL prediction for aero-engines. When dealing with complex, multivariate, long-term series tasks, the model exhibits higher generalization ability and noise resistance, resulting in smoother and more reliable predictions. By constructing a dynamic knowledge graph and optimizing the Transformer network structure, this invention effectively integrates spatial and temporal information from various sensors, significantly improving the accuracy, efficiency, and interpretability of aero-engine life prediction, and providing strong technical support for the health management of complex systems.

[0124] Example

[0125] This embodiment uses the aircraft engine life prediction problem on the C-MAPSS dataset as an example to illustrate an aircraft engine life prediction method that integrates dynamic knowledge graphs and dynamic attention mechanisms.

[0126] Step 1, according to Figure 3 The location structure of the various sensors of the turbofan engine shown is equipped with 21 sensors, and the functions of the 21 sensors are shown in Table 1.

[0127] Table 1 Sensor Monitoring Information

[0128]

[0129] Subsequently, the sensor monitoring data was normalized using the Max-Min Normalization method. In predicting the remaining life of aero-engines, data from multiple time steps can more comprehensively reflect the engine's degradation trend compared to data from a single time step. Therefore, this invention employs a sliding window technique to process the normalized data. Considering the actual situation of the engine's operating cycle, its degradation process can be divided into two stages: the healthy operation stage and the performance degradation stage. Finally, the data is divided into a training set and a validation set.

[0130] Figure 2 The method for dividing the time series data is illustrated. The window setting process involves using a fixed-size time window to segment the normalized data samples, with each window containing sensor data from multiple time steps. The sliding process involves moving the window along the time direction in steps of 1, gradually generating new samples until all time series data are covered.

[0131] Step 2, construct an entity relationship knowledge graph based on location and function (e.g., Figure 4-8 (as shown), Figure 4 A knowledge graph of the temperature management chain is shown; Figure 5 A knowledge graph of the stress management chain is shown; Figure 6 A knowledge graph of the speed control chain is shown; Figure 7 A knowledge graph of the fuel-to-air ratio chain is shown; Figure 8 A knowledge graph of the cooling and heat dissipation chain is shown;

[0132] Step 3: Analyze the dynamic coupling relationship between sensors. Taking sensors 1, 5, 16, and 21 as examples, construct a dynamic sensor coupling relationship analysis model, as shown in Table 2:

[0133] Table 2 Evidence Theory Table

[0134]

[0135] In the first set of input data, according to the deterministic and uncertainty analysis results, there is a coupling relationship between No. 1 and No. 5, and a coupling relationship between No. 5 and No. 21.

[0136] Step 4: Construct a fused knowledge graph that integrates entity relationships and coupling relationships, such as... Figure 9 As shown;

[0137] Step 5: Graph Neural Network (GCN) is used to efficiently embed and represent the knowledge graph, generating spatial feature vectors that characterize the sensor associations;

[0138] Steps 6 and 7 involve designing a dynamic attention mechanism and a Transformer network, which not only effectively extracts key features from multivariate time series data, but also reduces the interference of redundant information through dynamic weight allocation and multi-head attention mechanism, significantly improving the model's ability to model complex data.

[0139] Step 8: Fuse spatial features and temporal features to form a unified deep learning model for accurately predicting engine remaining life.

[0140] Step 9 involves adaptively optimizing the model structure and parameters using dynamically updated sensor data and a joint inference method, ensuring the real-time performance and accuracy of the prediction results. Experimental results are shown in Figure 13(ad). The results demonstrate that, compared to traditional remaining lifetime prediction methods, the method of this invention exhibits superior prediction accuracy and robustness, better copes with uncertainties in complex systems, significantly improves system reliability, and provides important technical support and theoretical basis for practical applications.

[0141] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0142] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0143] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0144] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0145] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0146] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting the lifespan of aero-engines that integrates dynamic knowledge graphs and dynamic attention mechanisms, characterized in that, include: Step 1: Acquire historical monitoring data from 21 sensors installed on the aircraft engine and preprocess it to obtain sample data. The preprocessing includes: data normalization, slicing the normalized data, and setting lifetime labels. Each slice contains sensor data at multiple time steps. The sample data is divided into training set and test set. Step 2: Using each sensor as a node, construct entity relationships between sensors based on their physical location and function to obtain an entity knowledge graph. Based on these entity relationships, five entity knowledge graphs are divided: temperature management chain knowledge graph, pressure management chain knowledge graph, speed control chain knowledge graph, fuel-to-air ratio chain knowledge graph, and cooling and heat dissipation chain knowledge graph. Step 3: Calculate the correlation of each sensor and apply evidence theory to evaluate the dynamic coupling relationship between sensors. Construct a coupling knowledge graph based on each sensor node and the coupling relationship. Specifically, this includes: calculating the correlation index of each input data one by one, evaluating the correlation strength between sensors in real time, and integrating multiple correlation measurement results based on evidence theory to dynamically update the coupling relationship of sensor nodes. Step 4: Merge the entity knowledge graph and the coupled knowledge graph into a unified dynamic knowledge graph; Step 5: Use a graph neural network to embed the nodes and edges in the dynamic knowledge graph, extract the relationship features between sensors and the latent patterns in the time series data, and generate a spatial feature vector to characterize the sensor associations; Step 6: Design a dynamic attention mechanism with adaptive weights to dynamically adjust the attention weights based on the importance of the spatiotemporal features of the sensor input; Step 7: Construct a Transformer network architecture based on the dynamic attention mechanism, extract long-term dependent features related to the prediction, and output the predicted value of the remaining lifetime through a linear layer to establish a deep neural network model for predicting the remaining lifetime of aero-engines. Step 8: Train and validate the model using the training and test sets to obtain a well-trained aero-engine life prediction model; Step 9: Use the trained aero-engine life prediction model to predict the life of the aero-engine.

2. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 1, characterized in that: The temperature management chain knowledge graph includes sensor nodes for collecting the total temperature at the fan inlet, the total temperature at the low-pressure compressor outlet, the total temperature at the high-pressure compressor outlet, and the total temperature at the low-pressure turbine outlet. The pressure management chain knowledge graph includes sensor nodes for collecting the fan inlet pressure, the total pressure in the bypass pipeline, the total pressure at the high-pressure compressor outlet, the engine pressure ratio, and the static pressure at the high-pressure compressor outlet. The speed control chain knowledge graph includes sensor nodes for collecting the fan speed, the core shaft speed, the fan correction speed, the core shaft correction speed, the fan demand speed, and the fan correction demand speed. The fuel-to-air ratio chain knowledge graph includes sensor nodes for collecting the fuel flow rate to Ps30 ratio, the bypass ratio, and the combustion chamber fuel-to-air ratio. The cooling and heat dissipation chain knowledge graph includes sensor nodes for collecting the extraction enthalpy, the high-pressure turbine cooling exhaust volume, and the low-pressure turbine cooling exhaust volume.

3. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 1, characterized in that: In step 3, the correlation between sensors is measured using the Pearson correlation coefficient, Spearman correlation coefficient, and mutual information between sensors. Combined with the evidence theory framework, a dynamic sensor coupling relationship analysis model is constructed.

4. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 1, characterized in that: In step 6, the dynamic attention mechanism is formally expressed as follows: Dynamic_Q=X t W Dynamic_Q ; Dynamic_K=X t W Dynamic_K ; Dynamic_V=X t W Dynamic_V ; Q=Dynamic_Q·W Q ; K=Dynamic_K·W K ; V=Dynamic_V·W V ; in, These represent the query matrix, key matrix, and value matrix after weight allocation, respectively, which will undergo three different linear transformations. Acting on the labeled embedding sequence X t To obtain; These represent the query matrix, key matrix, and value matrix, respectively, which will be transformed through three different linear transformations. Acting on embedded sequences To obtain.

5. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 1, characterized in that: In step 7, the feedforward network in the Transformer is used to map the features extracted by multi-head attention to more abstract features. It combines each different attention head and uses a linear mapping to obtain the multi-head dynamic attention value. The feedforward network consists of two linear transformations and one non-linear activation layer. In the Transformer layer, residual connections and layer normalization are applied to the output of the multi-head attention mechanism and the feedforward neural network. The entire structure consists of multiple sub-layers, including the multi-head attention module and the feedforward network. The output of each sub-layer is added to the input through residual connections and then subjected to layer normalization.

6. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 1, characterized in that: In step 8, the training is completed by determining whether the model has reached the minimum performance metric or the maximum number of training iterations. If so, the model training is completed.

7. The aero-engine life prediction method integrating dynamic knowledge graph and dynamic attention mechanism according to claim 6, characterized in that: In step 8, the performance metrics for evaluating the aero-engine life prediction model are RMSE and Score, as detailed below: in, y represents the predicted remaining lifetime. i This represents the actual remaining lifespan.

Images