An aero-engine life prediction method based on knowledge and data fusion driving

Abstract

The application provides a kind of based on knowledge and data fusion driven aero-engine life prediction method, belong to aero-engine life prediction technical field;The technical problem to be solved is to provide a kind of based on knowledge and data fusion driven aero-engine life prediction method;The technical scheme for solving the technical problem is that the overall architecture of encoder-decoder is used, the encoder extracts high-level feature representation from past observations, and the decoder summarizes past information and continuously optimizes the prediction result;A GRU-SDE module is designed in the encoder, and a stochastic differential equation SDE is used to simulate the engine degradation process, which more accurately simulates the uncertainty and randomness of degradation while extracting timing dependence;A knowledge graph is constructed for the inherent knowledge of the engine to obtain a knowledge-based multi-sensor relationship matrix;The data-based relationship matrix is randomly initialized, and the two are adaptively fused during the training process;The application is applied to aero-engine life prediction.

Description

TECHNICAL FIELD

[0001] The application provides a method for predicting the service life of an aero-engine based on knowledge and data fusion driving, and belongs to the technical field of aero-engine service life prediction. BACKGROUND

[0002] With the rise and promotion of intelligent manufacturing and the concept of Industry 4.0, the importance of Prognostics and Health Management (PHM) and Predictive Maintenance (PdM) is increasingly highlighted, and they are increasingly important in ensuring the reliability, safety and economic benefits of production systems; in this context, Remaining Useful Life (RUL) as an important research direction in industrial applications requires effective estimation of the expected duration from the current time to complete failure of mechanical equipment.

[0003] As the power component of an airplane, the reliability and safety of an aero-engine are directly related to the safety of flight, however, the engine is inevitably affected by various factors during operation, such as high temperature, high pressure, vibration, corrosion, etc., which leads to gradual degradation of its performance and continuous reduction of its remaining useful life; therefore, accurate prediction of the remaining useful life of an aero-engine is of great significance for ensuring flight safety, optimizing maintenance strategies and reducing operating costs.

[0004] Currently, the main methods for predicting the remaining useful life of an aero-engine are: predicting RUL by deeply understanding the physical degradation process and failure mechanism of the engine, however, due to the high complexity of the system operation mechanism, the cost of obtaining an accurate physical model is high; in addition, deep learning technology has become an effective means to handle highly nonlinear and changing sequence data, including fully connected neural networks (FCN), recurrent neural networks (RNN), convolutional neural networks (CNN), attention mechanism-based Transformers, and graph neural networks (GNN), etc., which can use the above advanced network architecture to obtain better results in the RUL prediction task than traditional methods.

[0005] Aero-engine RUL prediction technology based on deep learning can be divided into two categories: time pattern-based methods and spatio-temporal pattern-based methods:

[0006] The first type of strategy mainly focuses on mining the complex temporal variations within the sequence; the classic network architecture has long short-term memory (LSTM) and gated recurrent unit (GRU), which relies on the Markov assumption to model the sequence data. For example, Sayah et al. proposed a multi-layer LSTM for aero-engine RUL prediction. Luo et al. combined posterior analysis with a large number of GRU classifiers, effectively solving the multi-path degradation problem in life prediction. These methods have achieved some good results, but their efficiency is affected by the sequential computing paradigm. In contrast, the temporal convolution network (TCN) has unique advantages in the RUL prediction task due to its parallel computing capability and efficient feature extraction mechanism. Wang et al. used TCN to process multi-dimensional and multi-channel sensor data and proposed a new type of time convolution neural network with a soft threshold mechanism for engine RUL prediction. Cao et al. used TCN to capture temporal dependencies and extract advanced feature representations from vibration signals and achieved good prediction results. However, due to the limitations of one-dimensional convolution kernels, they can only model the changes between adjacent time points. In recent years, the Transformer with attention mechanism has been widely used to extract long-term dependencies of time series data. Zhang et al. used the attention mechanism to process the correlation between non-adjacent time points and proposed an integrated multi-head double sparse self-attention network based on improved Transformer to improve the RUL prediction effect.

[0007] The second type of strategy often considers the potential spatial correlation between multiple sensors on the basis of the first type of strategy. Related research usually uses attention mechanisms or graph neural networks to learn the relationship between sensors to enhance the feature contribution value that plays a key role in prediction. For example, Liu et al. proposed a new data-driven framework for aero-engine RUL estimation, which uses a CNN with channel attention to obtain more important key elements. Then, the Transformer model focuses attention on these important features. To reveal the features most closely related to the degradation process, Chen et al. added an attention mechanism to the LSTM, which enabled them to give more weight to important samples and features. Sun et al. constructed multiple sensors into a sensor network and used hierarchical graph representation layers to model spatial dependencies to achieve effective information fusion of sensors.

[0008] However, although the above methods can achieve aero-engine life prediction, they mainly benefit from the progress in capturing temporal dependencies and spatial correlations from sensor monitoring data, and still have the following limitations and defects:

[0009] (1) In the field of aviation, the operating environment of engine system is complex and changeable, and is affected by many unpredictable factors, such as changes in operating conditions, environmental disturbances and equipment aging, etc. These factors lead to high uncertainty of the degradation process of aero-engine, and the degradation trajectory is covered by noise, which is difficult to accurately model.

[0010] (2) There is a lack of potential space correlation based on data acquisition sensors, and they usually lack in-depth understanding of the physical characteristics and operating mechanism of the engine, which limits the generalization ability and robustness of the model when facing data distribution changes. In fact, domain knowledge has been proven to guide the modeling of deep learning, but how to combine it with data-driven methods is still a big difficulty. SUMMARY

[0011] The present application aims to overcome the deficiencies in the prior art, and the technical problems to be solved are to provide a knowledge and data fusion driven aero-engine life prediction method.

[0012] In order to solve the above technical problems, the technical scheme adopted by the present application is as follows: a knowledge and data fusion driven aero-engine life prediction method, comprising the following prediction steps:

[0013] Step 1: Use K-means clustering algorithm to obtain different condition classifications based on operation settings, and use z-score operation to standardize the data for each condition classification category, and the calculation formula for processing is:

[0014]

[0015] In the formula: S represents the set of sensors, C represents the set of conditions, i and j represent the i-th sensor and the j-th condition classification respectively, u i，j and σ i，j respectively represent the mean and standard deviation of the i-th sensor under the j-th condition classification, x i，j and z i，j respectively represent the data before and after standardization;

[0016] Then use the exponential weighted moving average method to smooth the data, and the calculation formula for processing is:

[0017] z t =α*z t +(1-α)*z t-1 ;

[0018] In the formula: z′ t represents the filtered value of z t , α represents the filter coefficient, and t represents the time step.

[0019] Step two: Constructing the RUL prediction model of the aero-engine based on stochastic differential equation and domain knowledge:

[0020] Step 2-1: Adopting the standard 1x1 convolution, the preprocessed data Z T×N Projecting to the potential high-dimensional space to obtain the embedding representation X of the original data, the expression is: X = Embedding(Z);

[0021] Step 2-2: Extracting the advanced feature representation of the data using the encoder:

[0022] Step 2-2-1: Processing the data through the gated recurrent unit GRU to capture the temporal dependency relationship and obtain the hidden state;

[0023] Step 2-2-2: Modeling the engine degradation process using stochastic differential equation SDE;

[0024] Step 2-2-3: Solving the stochastic differential equation SDE using the gated recurrent unit GRU output to obtain the feature matrix of the original data;

[0025] Step 2-2-4: Constructing the aero-engine knowledge graph, specifically summarizing the composition relationship, connection relationship, flow relationship, and monitoring relationship between the components of the aero-engine;

[0026] Step 2-2-5: Knowledge transformation to obtain the knowledge-based multi-sensor correlation matrix;

[0027] Step 2-2-6: Knowledge embedding and adaptive fusion;

[0028] Step 2-3: Adopting the linear residual module as the decoder of the model, mapping the encoded hidden representation to the latent space, mapping the hidden layer to the output linear layer, using dropout at the same time, and using layer normalization at the output, the expression is as follows:

[0029] Y = Decoder(O);

[0030] Step 2-4: The decoder output Y will obtain the final aero-engine RUL prediction value through the linear layer, the expression is:

[0031] RUL = Linear(Y);

[0032] Step three: Training and testing the network constructed in step two, selecting the root mean square error RMSE as the loss function, adopting the Adam optimization algorithm with adaptive learning rate, adjusting the learning rate of each parameter dynamically to accelerate the convergence speed of the model and improve the stability of the training.

[0033] The specific method for capturing the time sequence dependency relationship to obtain the hidden state in step 2-2-1 is as follows:

[0034] The data is processed by a gated recurrent unit GRU to capture the time sequence dependency relationship and obtain a hidden state h t , and the expression is as follows:

[0035] r t =σ(x t W xr +h t-1 W hr +b r );

[0036] u t =σ(x t W xu +h t-1 W hu +b u );

[0037] c t =RELU(x t W xc +r t ☉W hc +b c );

[0038] h t =(1-u t )☉h t-1 +u t ⊙c t ;

[0039] In the formula, ⊙ is an element-wise product, σ is a sigmoid activation function, x t represents the input of the layer at time t, W xr , W hr , W xu , W hu , W xc , W hc , b r , b u and b u are all model parameters.

[0040] The specific method for modeling the engine degradation process in step 2-2-2 is as follows:

[0041] The expression of the stochastic differential equation SDE used is as follows:

[0042] dX t =u(t,X t )dt+σ(t,X t )dW t , X0=x0;

[0043] where X t represents the state of the system at time t; u(t, X t ) and σ(t, X t ) are deterministic functions representing the drift and diffusion characteristics of the system, respectively; dW t is the differential of a Wiener process or Brownian motion, representing the random fluctuations or noise experienced by the system;

[0044] The Euler-Amaru numerical method is used for solving, the solution of SDE is approached step by step by discretizing the time step, and the evolution trajectory of the system state with time is obtained, and the calculation formula is:

[0045]

[0046] where Δt is the length of the unit time period; u(t, X t ) and σ(t, X t ) represent the drift term and diffusion term, respectively; ∈ t represents the noise term;

[0047] According to the above formula, the state X t of each step is obtained by iterative solution.

[0048] The specific method for obtaining the feature matrix of the original data in step 2-2-3 is:

[0049] The hidden state h t extracted by the gated recurrent unit GRU is input into the stochastic differential equation SDE, and the drift term u t (t, X t ), the diffusion term σ t (t, X t ), and the noise term ∈ t (t, X t ) are calculated respectively, and h t is projected into the latent space by using three different linear layers to obtain the unknown term representation.

[0050] For ∈ t , the sigmoid function is used to adaptively adjust the strength of the randomly generated noise z, and the expression is:

[0051] u t (t, x t ) = Linear u (h t );

[0052] σ t (t, x t ) = Linear σ (h t);

[0053] ∈ t (t, x t )=sigmoid(Linea ∈ (h t ))*z t ;

[0054] where Linear u , Linear σ , Linear ∈ represent three different linear layers, sigmoid represents the activation function, z t is the noise generated randomly at time t and obeys the normal distribution;

[0055] Then the drift term u t (t, X t ), the diffusion term σ t (t, X t ), and the noise term ∈ t (t, X t ) are brought into the following formula:

[0056]

[0057] The state X t at each step is obtained by iteration;

[0058] The model is constructed by calculating the difference between adjacent time steps, and the expression is:

[0059]

[0060]

[0061] The difference sequence ΔX={Δx1, Δx2...Δx t ,... Δx T} can reflect the trend and dynamic characteristics of time series data, and the feature matrix X fea of the original data is obtained through the above steps.

[0062] The specific method for constructing the aero-engine knowledge graph in step 2-2-4 is:

[0063] Inducting the composition relationship: describing each component that constitutes the overall system of the turbofan engine;

[0064] Inducting the connection relationship: describing the physical connection mode between each component in the turbofan engine;

[0065] Inducting the flow relationship: describing the flow path of gas and liquid in the turbofan engine;

[0066] Induction of monitoring relationships: describe which sensors monitor the key parameters of each component in the turbofan engine;

[0067] Based on the above induction information, the turbofan engine knowledge graph is constructed.

[0068] The specific method of knowledge transformation in step 2-2-5 is:

[0069] For a triple (u, v, w) in the knowledge graph, the TransE model learns the vector representation of entities and relationships by optimizing the following objective function:

[0070]

[0071] In the formula: u, v, w represent the vector representation of the head entity, the relationship and the tail entity respectively, and θ represents the set of all triples in the knowledge graph;

[0072] By optimizing the above objective function, the embedding e of each entity is obtained, and then the Pearson correlation coefficient is used to calculate the similarity between entities to obtain the knowledge-based multi-sensor correlation matrix A1.

[0073] The specific method of knowledge embedding and adaptive fusion in step 2-2-6 is:

[0074] Randomly initialize the data-based multi-sensor correlation matrix A2, fuse the knowledge-based correlation matrix A1 and the data-based correlation matrix A2 to generate a comprehensive correlation matrix A, and use an adaptive weight distribution method to dynamically adjust the weights of the two matrices:

[0075] A1∈R D×D and A2∈R D×D Expand the dimension to A1∈R 1×D×D and A2∈R 1×D×D Then use 2D convolution to obtain the adaptive weight parameter α, the expression is:

[0076]

[0077] A=αA1+(1-α)A2;

[0078] In the formula: σ is the sigmoid activation function, g(·) represents the 2D convolution layer, represents matrix splicing; α is the adaptive weight parameter, and the value range is between [0, 1];

[0079] By optimizing the value of α, the effective fusion of A1 and A2 is realized;

[0080] Feature matrix X feaThe encoder output will be represented as o, and the expression is:

[0081] O=AX fea .

[0082] The present application has the beneficial effects relative to the prior art: the present application proposes an aero-engine life prediction method based on knowledge and data fusion driving, relates to aero-engine fault diagnosis and life prediction, and specifically aims at the problems that the current aero-engine degradation process has high uncertainty, is difficult to accurately model, and the traditional deep learning method has high data dependence and poor generalization, and cannot effectively utilize domain knowledge, designs a GRU-SDE module, and innovatively simulates the degradation process of the aero-engine by using a random differential equation, more accurately describes the uncertainty and randomness of the degradation process; the present application constructs a knowledge graph for the aero-engine domain knowledge, and combines it with the data-driven learning process through a self-adaptive fusion mode, effectively embeds the domain knowledge into the deep neural network, and improves the interpretability and generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0083] The present application will be further described below in combination with the drawings:

[0084] Figure 1 The step flow chart of the present application based on the random differential equation and the domain knowledge of the aero-engine RUL prediction is shown in the figure;

[0085] Figure 2 The structure diagram of the aero-engine RUL prediction model of the present application is shown in the figure;

[0086] Figure 3 The schematic diagram for obtaining the SDE unknown term expression by using the GRU of the present application is shown in the figure;

[0087] Figure 4 The structure schematic diagram of the turbofan engine of the present application is shown in the figure;

[0088] Figure 5 The knowledge graph schematic diagram of the turbofan engine of the present application is shown in the figure;

[0089] Figure 6 The decoder structure diagram used in the present application is shown in the figure. DETAILED DESCRIPTION

[0090] The application innovatively combines deep learning with domain knowledge, develops a high-precision aero-engine life prediction model, not only promotes the technical progress in the field of fault diagnosis and life prediction, but also provides a new idea for the health management of other complex industrial systems, with the popularization and application of the method, the technology upgrading of related industries can be promoted, the intelligent level of equipment is improved, and the development of industrial internet and intelligent manufacturing is promoted; The technology proposed in the application is rigorously and comprehensively evaluated on the C-MAPSS turbofan engine data set, and the precision is improved by 32% compared with the traditional method, which is an effective aero-engine life prediction method; The prediction method mainly comprises: adopting an overall architecture of an encoder-decoder, the encoder extracts high-level feature representation from past observations, and the decoder summarizes past information and continuously optimizes the prediction result; A GRU-SDE module is designed in the encoder, a random differential equation (SDE) is used to simulate the engine degradation process, the temporal dependence is extracted, and the uncertainty and randomness of the degradation are more accurately simulated; A knowledge graph is constructed according to the inherent knowledge of the engine, and a multi-sensor relationship matrix based on knowledge is obtained; meanwhile, the data-based relationship matrix is randomly initialized, and the two are adaptively fused in the training process; the decoder takes a residual multilayer perceptron (MLP) as the core architecture, and the encoded hidden representation is simply and efficiently mapped into a high-dimensional space; finally, a linear layer is adopted to obtain the final aero-engine RUL prediction.

[0091] In order to more clearly set forth and explain the purpose, measures and key points of the application, the method proposed by the application is described in detail as follows in combination with the drawings:

[0092] The main purpose of the application is to realize the remaining useful life prediction of an aero-engine, as shown in the accompanying Figure 1 The accompanying Figure 2 is a model framework diagram, the accompanying Figure 3 , the accompanying Figure 4 , the accompanying Figure 5 , the accompanying Figure 6 are respectively a graph for obtaining SDE unknown term representation by using GRU, a turbofan engine structure and sensor position diagram, a turbofan engine knowledge graph and a decoder structure diagram; Table 1 is 21 measurement sensors in the turbofan engine and their meanings, and table 2 is the meaning of the constructed triplets; the problems studied by the application and the overall algorithm process are introduced first, and then the design details of each module are described in detail, and the specific implementation mode is as follows:

[0093] I. Problem definition:

[0094] In the field of reliability engineering, the RUL prediction of an aero-engine is a regression problem, and the input is a multi-source time series X T×NWhere T represents the length of the historical time step, and N represents the number of sensors; the task target of the application is to predict the engine RUL in real time by establishing a mapping relationship between the RUL and the historical sensor data, and to construct a mapping function f, expressed as RUL=f(X).

[0095] II. Model framework

[0096] The aviation engine life prediction method based on knowledge and data fusion driving provided by the application relates to a degradation process modeling method based on stochastic differential equations and domain knowledge, and is used for aviation engine RUL prediction, and a model framework is established as shown in Figure 2 The model adopts a standard encoder-decoder structure, wherein the encoder extracts information from past observations, and the decoder aggregates past information and improves the prediction from simple initialization; specifically, a GRU-SDE module is designed in the encoder, a stochastic differential equation (SDE) is used to simulate the system degradation process, the timing dependence is extracted, and the uncertainty and randomness of the degradation are more accurately simulated; in addition, a knowledge graph is constructed based on the inherent knowledge of the aviation engine, a multi-sensor relationship matrix is obtained, and data is adaptively fused to find the balance between knowledge and data in the training process; the decoder used in the application utilizes a residual linear layer, which takes a multi-layer perceptron (MLP) as a core architecture, and simply and efficiently maps the encoded hidden representation to a high-dimensional space; finally, a linear layer is used to obtain the final RUL prediction.

[0097] The detailed modeling and prediction method process is as follows:

[0098] Step one: data preprocessing

[0099] Step 1-1: In order to effectively cope with the substantial influence of the changing working conditions of the aviation engine on the sensor data, the K-means clustering algorithm is used to obtain different condition classifications based on operation settings, and the z-score is used to standardize the data for each category, and the calculation formula for processing is:

[0100]

[0101] In the formula, s represents a set of sensors, c represents a set of conditions, i and j represent the i th sensor and the j th condition classification respectively, u i，j and σ i，j respectively represent the mean and standard deviation of the i th sensor under the j th condition classification, x i，j and z i，j respectively represent the data before and after standardization.

[0102] Step 1-2: The data is smoothed using an exponentially weighted moving average method to reduce the impact of local oscillations in the sensor data and obtain a more stable and reliable representation of the sensor data. The calculation formula used for this process is as follows:

[0103] z t =α*z t +(1-))*z t-1 ;

[0104] In the formula: z′ t Indicate z t The filtered value, where α represents the filter coefficient and t represents the time step.

[0105] Step 2: Constructing a RUL prediction model for aero-engines based on stochastic differential equations and domain knowledge:

[0106] Step 2-1: Data embedding:

[0107] Using standard 1×1 convolution, the preprocessed data Z∈R T×N Projecting onto a latent high-dimensional space, we obtain the embedded representation X of the original data, as shown in the formula:

[0108] X = Embedding(Z).

[0109] Step 2-2: Extract high-level feature representations of the data using the encoder:

[0110] Step 2-2-1: Process the data through a Gated Recurrent Unit (GRU) to capture temporal dependencies and obtain the hidden state h. t The expression is:

[0111] r t =σ(x t W xr +h t-1 W hr +b r );

[0112] u t =σ(x t W xu +h t-1 W hu +b u );

[0113] c t =RELU(x t W xc +r t ☉W hc +b c );

[0114] ht = (1 - u t )☉h t-1 + u t ☉c t ;

[0115] where ⊙ is element-wise multiplication, σ is the sigmoid activation function, x t denotes the input to the layer at time t, w xr , W hr , W xu , W hu , W xc , W hc , b r , b u , and b u are model parameters.

[0116] Step 2-2-2: Model the engine degradation process using stochastic differential equations. Stochastic Differential Equations (SDE) is a mathematical tool used to describe the evolution of dynamic systems affected by random forces. The general form of SDE can be represented as:

[0117] dX t = u(t, X t )dt + σ(t, X t )dW t , X0= x0;

[0118] where X t denotes the state of the system at time t; u(t, Z t ) and σ(t, X t ) are deterministic functions representing the drift and diffusion characteristics of the system, respectively; dW t is the differential of a Wiener process or Brownian motion, representing the random fluctuations or noise experienced by the system;

[0119] The solution of SDE is usually difficult to obtain directly, so numerical methods are needed to solve it. Euler-Amaru is one of the simplest and most intuitive numerical methods, which approximates the solution of SDE by discretizing the time step and gradually approaching the evolution trajectory of the system state over time. The Euler-Amaru method gives the following equation:

[0120]

[0121] where Δt is the length of the unit time period; u(t, X t ) and σ(t, X t ) represent the drift and diffusion terms, respectively; ∈ tThis represents the noise term; using the above formula, the state X at each step can be obtained through iterative solution. t .

[0122] Step 2-2-3: Solve the SDE using the GRU output. The hidden state h extracted by the GRU... t Input into SDE, and calculate the drift term u respectively. t (t,X t )Diffusion term σ t (t,X t ) and noise terms ∈ t These terms can be estimated using neural networks or other function approximators. In this invention, three different linear layers are used to estimate h. t Projected onto the latent space to obtain a representation of the unknowns (e.g.) Figure 3 To effectively accommodate noise in the data, for ∈ t The sigmoid function is used to adaptively adjust the strength of the randomly generated noise z;

[0123] u t (t,x t Linear u (h t );

[0124] σ t (t,x t Linear σ (h t );

[0125] ∈ t (t,x t = sigmoid(Linear) ∈ (h t ))*z t ;

[0126] In the formula: Linear u Linear σ Linear ∈ This represents three different linear layers, where sigmoid represents the activation function, and z... t The noise is randomly generated at time t and follows a normal distribution;

[0127] Then the drift term u t (t,X t ), diffusion term σ t (t,X t ) and noise terms ∈ t (t,X t Substitution The state X at each step can be obtained through iterative solution. tThe initial state X0 is crucial for the modeling of the entire time series. However, in practical problems, X0 is often unknown, and random initialization can have a significant impact on the performance of the model. In addition, the complexity of iterative calculation can significantly increase the computational cost, affecting the efficiency and performance of the model. Inspired by the idea of difference, the present invention constructs a model by calculating the difference between adjacent time steps, which ingeniously alleviates the above problems, as follows:

[0128]

[0129]

[0130] The difference sequence ΔX = {Δx1, Δx2...Δx t ,...Δx T} can reflect the trend and dynamic characteristics of time series data, so through the above steps, the feature matrix X fea of the original data can be obtained. fea The time series dependence in the data is extracted, and the random dynamics and disturbances in the degradation process are learned.

[0131] Step 2-2-4: Constructing an aero-engine knowledge graph:

[0132] The performance of an aero-engine depends on its individual components. By analyzing the relationships between aero-engine components, we can better utilize the relationships between measurement points and the inherent knowledge of the aero-engine. Figure 4 The structure of a turbofan engine used in the C-MAPSS simulation data set provided by NASA is shown in Table 1, which gives the 21 measurement sensors placed inside the turbofan engine and their meanings;

[0133]

[0134] Table 1: 21 measurement sensors inside the turbofan engine and their meanings

[0135] The composition relationship, connection relationship, flow direction relationship, and monitoring relationship between engine components are summarized as follows:

[0136] Composition relationship: refers to the individual components that make up the overall system of the turbofan engine. For example, (engine, composed of, fan) indicates that the engine is composed of a fan;

[0137] Connection relationship: describes the physical connection methods between individual components in the turbofan engine. These connections ensure that individual components can cooperate to achieve the overall function of the engine. For example, (low-pressure rotor, connected to, high-pressure compressor) indicates that the low-pressure rotor is connected to the high-pressure compressor;

[0138] Flow direction relationship: describes the flow path of gas and liquid in a turbofan engine. These flow direction relationships are crucial for the proper functioning of the engine, determining the efficiency and stability of energy transfer. For example, (fan, flow direction, low pressure compressor) indicates that air flows from the fan to the low pressure compressor.

[0139] Monitoring relationship: describes which sensors monitor the key parameters of each component in a turbofan engine. For example, (fan, monitoring, S1) indicates that the fan is monitored by S1.

[0140] Based on the above, for the construction of the turbofan engine knowledge graph, 29 entities, 4 relationships, and a total of 42 triples are summarized, as shown in Table 1 and Table 2. Figure 5

[0141]

[0142] Table 2: Meaning of constructed triples

[0143] Step 2-2-5: Knowledge transformation:

[0144] Depending on the algorithm in TransE, it learns a low-dimensional vector for each entity and relationship. Specifically, for a triple (u, v, w) in the knowledge graph, the TransE model learns the vector representation of entities and relationships by optimizing the following objective function:

[0145]

[0146] In the formula: u, v, w represent the vector representation of the head entity, the relationship and the tail entity respectively, and θ represents the set of all triples in the knowledge graph.

[0147] By optimizing the above objective function, the embedding e of each entity can be obtained. Then, the similarity between entities is calculated using the Pearson correlation coefficient, and the knowledge-based multi-sensor correlation matrix A1 is obtained.

[0148] Step 2-2-6: Knowledge embedding and adaptive fusion:

[0149] The data-based multi-sensor correlation matrix A2 is randomly initialized and learns channel dependencies during training. Adaptive fusion aims to fuse the knowledge-based correlation matrix A1 and the data-based correlation matrix A2 to generate a comprehensive correlation matrix A; the present application adopts an adaptive weight allocation method to dynamically adjust the weights of the two matrices.

[0150] First, A1∈R D×D and A2∈R D×D are extended to A1∈R 1×D×D and A2∈R 1×D×D ​In order to facilitate matrix splicing. Then, by using 2D convolution, the adaptive weight parameter a is obtained;

[0151]

[0152] A = aA1 + (1-a)A2;

[0153] In the formula: sigma is a sigmoid activation function, g(·) represents a 2D convolution layer, Indicates matrix splicing; a is an adaptive weight parameter, and the value range is between [0, 1].

[0154] By optimizing the value of a, the effective fusion of A1 and A2 can be realized. Finally, the feature matrix X fea Will be multiplied by the fused sensor relationship matrix A to realize the effective fusion of multi-source data, and the encoder output will be represented as O, and the expression is as follows:

[0155] O = AX fea .

[0156] Step 2-3: Adopting a linear residual module as the decoder of the model Figure 6 , which maps the encoded hidden representation to the latent space. Specifically, the linear residual block is an MLP with a hidden layer with ReLU activation. In addition, a residual connection is used to alleviate the vanishing gradient. Finally, dropout is used on the linear layer that maps the hidden layer to the output, and layer normalization is used at the output, and the expression is as follows:

[0157] Y = Decoder(O).

[0158] Step 2-4: The decoder output Y will obtain the final RUL prediction value through a linear layer, as follows:

[0159] RUL = Linear(Y).

[0160] Step three: train and test the network constructed in step two:

[0161] In order to optimize the performance of the model, the root mean square error (RMSE) is selected as the loss function; the Adam optimization algorithm with adaptive learning rate is adopted, which adjusts the learning rate of each parameter dynamically, accelerates the convergence speed of the model, and improves the stability of the training.

[0162] The aviation engine remaining useful life prediction method provided by the application adopts an overall framework design of an encoder-decoder, innovatively combines a stochastic differential equation with a recurrent neural network, more accurately simulates uncertainty and randomness in the degradation process, and simultaneously allows adjustment of noise intensity to adapt to random degradation dynamics; meanwhile, a domain knowledge graph is efficiently constructed, a correlation matrix is calculated, and the weight thereof is dynamically adjusted in the learning process; by incorporating knowledge in the network, the model can accurately understand and reflect complex multi-sensor relationships; and a large number of experiments on a C-MAPSS public data set show that the method achieves the most advanced level in the RUL prediction task. Compared with the current prediction method, the application achieves significant improvement in the root mean square error (RMSE) and the scoring function (score).

[0163] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application.

Claims

1. An aero-engine life prediction method based on knowledge and data fusion driving, characterized in that: The prediction steps include the following: Step 1: Obtain different condition classifications based on operation settings using the K-means clustering algorithm, and standardize the data for each condition classification using the z-score operation. The calculation formula is: where S denotes a set of sensors, C denotes a set of conditions, i and j denote the ith sensor and the jth condition class, respectively, u i，j and σ i，j denote the mean and standard deviation of the ith sensor under the jth condition class, respectively, x i，j and z i,j denote the data before and after normalization, respectively; Then, use the exponential weighted moving average method to smooth the data, and the calculation formula is: z' t = a * z t + (1 - a) * z' t-1 ; where: z' t represents z t filtered values, a represents a filter coefficient, and t represents a time step. Step 2: Construct an aero-engine RUL prediction model based on stochastic differential equations and domain knowledge: Step 2-1: Apply a standard 1x1 convolution to the pre-processed data Z ∈ R T×N Project to a potentially high-dimensional space to obtain an embedding representation X of the original data, expressed as: X = Embedding(Z); Step 2-2: Use the encoder to extract high-level feature representations of the data: Step 2-2-1: Process the data through the gated recurrent unit (GRU) to capture temporal dependencies and obtain hidden states. Step 2-2-2: Model the engine degradation process using stochastic differential equations (SDE). Step 2-2-3: Use the gated recurrent unit (GRU) output to solve the stochastic differential equations (SDE) and obtain the feature matrix of the original data. Step 2-2-4: Construct an aero-engine knowledge graph, which includes the composition relationship, connection relationship, flow relationship, and monitoring relationship between aero-engine components. The specific method is: Induce the composition relationship: describe the components that make up the overall system of a turbofan engine. Induce the connection relationship: describe the physical connection methods between components in a turbofan engine. Induce the flow relationship: describe the flow paths of gases and liquids in a turbofan engine. Induce the monitoring relationship: describe which sensors monitor the key parameters of each component in a turbofan engine. Construct a turbofan engine knowledge graph based on the above induction information. Step 2-2-5: Perform knowledge transformation to obtain a knowledge-based multi-sensor correlation matrix. The specific method is: For a triple (u, v, w) in the knowledge graph, use the TransE model to learn the vector representations of entities and relationships by optimizing the following objective function: Where: u, v, w represent the vector representations of the head entity, relationship, and tail entity, respectively, and θ represents the set of all triples in the knowledge graph. By optimizing the above objective function, obtain the embedding e of each entity, then calculate the similarity between entities using the Pearson correlation coefficient to obtain a knowledge-based multi-sensor correlation matrix A1. Step 2-2-6: Perform knowledge embedding and adaptive fusion. The specific method is: Randomly initialize the data-based multi-sensor correlation matrix A2, and fuse the knowledge-based correlation matrix A1 and the data-based correlation matrix A2 to generate a comprehensive correlation matrix A. Use the adaptive weight allocation method to dynamically adjust the weights of the two matrices: A1∈R D×D and A2∈R D×D A1∈R 1×D×D and A2∈R 1×D×D and then use 2D convolution to obtain adaptive weight parameters α, the expression is: A = αA1 + (1 - α)A2 where σ is a sigmoid activation function, g(·) represents a 2D convolution layer, denotes matrix concatenation; and a is an adaptive weight parameter with a value range of [0, 1]. Optimize the value of α to achieve effective fusion of A1 and A2. Feature matrix X fea The encoder output will be denoted as O and expressed as: O = AX fea ; Step 2-3: Use a linear residual module as the decoder of the model to map the encoded hidden representation to the latent space. Map the hidden layer to the output linear layer while using dropout and layer normalization at the output. The expression is as follows: Y = Decoder(O) Step 2-4: The decoder output Y will obtain the final aero-engine RUL prediction value through a linear layer, and the expression is: RUL = Linear(Y); Step three: train and test the network constructed in step two, select the root mean square error RMSE as the loss function, use the Adam optimization algorithm with adaptive learning rate, adjust the learning rate of each parameter dynamically, accelerate the convergence speed of the model, and improve the stability of the training.

2. The method of claim 1, wherein: The specific method for capturing the time-dependent relationship in step 2-2-1 to obtain the hidden state is: The data is processed by a gated recurrent unit GRU to capture the temporal dependency and obtain a hidden state h t The expression is: r t = σ(x t W xr +h t-1 W hr +b r ); u t = σ(x t W xu + h t-1 W hu + b u ); c t = RELU(x t W xc + r t ⊙ W hc + b c ); h t = (1 - u t ) * h t-1 + u t * c t ; where: is element-wise multiplication, σ is the sigmoid activation function, x t represents the input to the layer at time t, W xr ,W hr ,W xu ,W hu ,W xc ,W hc ,b r ,b u and b u are model parameters.

3. The method of claim 2, wherein: The specific method for modeling the engine degradation process in step 2-2-2 is: The expression of the stochastic differential equation SDE used is: dX t = u(t, X t )dt + σ(t, X t )dW t , X0= x0; where: X t represents the state of the system at time t; u(t, X t ) and σ(t, X t ) are deterministic functions representing the drift and diffusion characteristics of the system, respectively; dW t is the differential of a Wiener process or Brownian motion representing the random fluctuations or noise to which the system is subjected. The Euler-Amaru numerical method is used for solving, the solution of SDE is approximated step by step by discretizing the time step, and the evolution trajectory of the system state with time is obtained, and the calculation formula is: where: Δt is the length of the unit time period; u(t, X t ) and σ(t, X t ) represent the drift term and the diffusion term, respectively; ∈ t represents the noise term; According to the above equation, the state X of each step is obtained by iterative solution t .

4. The method of claim 3, wherein: The specific method for obtaining the feature matrix of the original data in step 2-2-3 is: The hidden state h extracted by the gated recurrent unit GRU t is input into a stochastic differential equation SDE to calculate a drift term u t t , a diffusion term σ t t , and a noise term ∈ t t respectively, and h t is projected into a latent space by using three different linear layers to obtain an unknown term representation;​​​ For ∈ t , the sigmoid function is used to adaptively adjust the strength of the randomly generated noise z, expressed as: u t (t, x t ) = Linear u (h t ) ; σ t (t, x t ) = Linear σ (h t ); ∈ t (t, x t ) = sigmoid(Linear ∈ (h t ))*z t ; In the formula: Linear u Linear σ Linear ∈ This represents three different linear layers, where sigmoid represents the activation function, and z... t The noise is randomly generated at time t and follows a normal distribution; The drift term u t (t, X t ) is then brought into the equation t (t, X t ) is the diffusion term t (t, X t ) is the noise term Solve for each step state X iteratively t ; The model is constructed by calculating the difference between adjacent time steps, and the expression is: The differential sequence ΔX = {Δx1, Δx2...Δx t ,...Δx T} can reflect the change trend and dynamic characteristics of the time series data, and the feature matrix X fea of the original data is obtained through the above steps.

Images