Aero-engine life prediction method based on knowledge embedded graph neural network

Abstract

The invention belongs to the technical field of aero-engine monitoring, and discloses an aero-engine service life prediction method based on a knowledge embedded graph neural network, which comprises the following steps: firstly, constructing an aero-engine sensor knowledge graph, then converting the knowledge graph into an embedded vector, and then constructing a time-space degradation data set of knowledge embedding; and finally predicting the service life of the aero-engine through the space-time diagram convolutional neural network model. According to the method, time correlation, spatial correlation and sensor global information integration are considered, and the service life of the aero-engine can be effectively and accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft engine monitoring, relates to aircraft engine life prediction technology, and in particular to an aircraft engine life prediction method based on knowledge embedded graph neural network. Background Art

[0002] In recent years, the aviation industry has developed rapidly. As the core power system of aircraft, the performance and health status of aircraft engines are directly related to the economy and safety of aircraft flight. During their service, aircraft engines are affected by a variety of stresses and environmental factors, such as high temperature, high pressure, vibration, and corrosion. These factors may cause engine parts to gradually fatigue, wear, and age, which in turn affects the reliability and service life of the entire engine. Therefore, in order to ensure the safe operation of aircraft and reduce operation and maintenance costs, it is particularly important to accurately predict the remaining life of aircraft engines.

[0003] At present, with the advancement of sensor and communication technology, a large amount of aircraft engine health monitoring data is collected and stored, which has promoted the development of many data-driven remaining life prediction methods. This type of method analyzes and learns a large amount of historical data, automatically mines the degradation pattern of the equipment, and thus achieves efficient prediction of engine life. However, in the prediction of aircraft engine life, the relationship between engine components is complex and has significant mutual influence. Traditional methods mainly focus on the temporal correlation of sensor data, ignore the exploration of its spatial correlation, and cannot effectively process complex graph structure data, so it is difficult to capture these high-order dependencies. Secondly, traditional methods usually extract features in a local range, and it is difficult to fully integrate the global information of each component. In addition, traditional methods usually convert raw data into a tensor of fixed shape. For example, flattening or serializing the relationship data of engine components and inputting them into the network will lose the original sensor topological structure information, affecting the performance of the model. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention aims to provide an aircraft engine life prediction method based on knowledge embedded graph neural network, which solves the problems of limited ability of prior art to handle complex relationships, difficulty in global information integration, and insufficient adaptability and scalability of the model by embedding prior knowledge and utilizing graph neural network.

[0005] Firstly, the structural information of aircraft engines and the location information of sensors are regarded as prior spatial knowledge to establish the knowledge graph of aircraft engine sensors. Secondly, the knowledge graph is converted into a low-dimensional embedding vector using an energy-based knowledge embedding algorithm. Then, an aircraft engine spatiotemporal degradation dataset is constructed, and the graph edges and adjacency matrices are constructed using the above embedding vectors. Finally, a spatiotemporal graph convolutional neural network model is established, and the spatiotemporal degradation pattern of aircraft engines is extracted using the spatial module, the temporal module and the output module to obtain the remaining life prediction result.

[0006] Specifically, the method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network provided by the present invention comprises the following steps:

[0007] Step 1: construct an aircraft engine sensor knowledge graph; the aircraft engine sensor knowledge graph includes a number of engine components, a number of sensors for monitoring each engine component, connection relationships between different engine components, and monitoring relationships between engine components and sensors;

[0008] Step 2: Convert the knowledge graph into an embedding vector; this step includes the following sub-steps,

[0009] Step 2.1: construct a mathematical representation of the knowledge graph using triples; each triple includes a head entity, a tail entity, and a relationship between the head entity and the tail entity; the head entity and the tail entity are selected from a set of transmitter components and sensors; all triples constitute a triple set;

[0010] Step 2.2: Construct a set of replacement triples;

[0011] Step 2.3: Iteratively update each triple in the triple set to obtain the optimal embedding vector for each entity and relationship in the triple;

[0012] Step 3: Based on the embedding vectors of each sensor obtained in step 2, a spatiotemporal degradation dataset for knowledge embedding is constructed;

[0013] Step 4: Substitute the sensor measurement history data into the spatiotemporal degradation dataset to obtain the offline spatiotemporal degradation dataset; and use the offline spatiotemporal degradation dataset to train the spatiotemporal graph convolutional neural network model;

[0014] Step 5: Substitute the online data measured by the sensor into the spatiotemporal degradation dataset to obtain the online spatiotemporal degradation dataset; input the online spatiotemporal degradation dataset into the trained spatiotemporal graph convolutional neural network model to obtain the aircraft engine life prediction result.

[0015] Further, step 1 includes the following steps,

[0016] Step 1.1: Establish the connection relationship between different engine components according to the structure of the aircraft engine and the connection type of each engine component;

[0017] Step 1.2: Establish the monitoring relationship between the engine components and the sensors according to the location layout of the sensors and the monitoring objects;

[0018] Step 1.3: Construct an aviation engine sensor knowledge graph based on the connection relationship between different engine components and the monitoring relationship between engine components and sensors.

[0019] Furthermore, the connection relationship includes mechanical connection and airflow connection.

[0020] Furthermore, the specific method of step 2.1 is:

[0021] The triple (e h ,r,e t ) form to construct the mathematical representation of the knowledge graph, where e h ,e t ∈E is an entity, which takes values ​​from the set E of sensors and engine components; r represents a relationship; all triples constitute a set S; all triples are initialized, e h 、r、e t From uniform distribution Random sampling is performed in, where U represents uniform distribution and M is the dimension of the embedding vector; and r is normalized. Here, ||·|| represents the 2-norm.

[0022] Furthermore, the specific method of step 2.2 is:

[0023] Using the formula For each original triple in the set S, e h or e t Randomly replace with other entities in set E and construct a replacement triple set S′.

[0024] Further, step 2.3 includes the following sub-steps,

[0025] Step 2.3.1: Use the formula The entity is normalized; where ||·|| represents the 2-norm;

[0026] Step 2.3.2: Randomly extract a batch of triplets S of size b from the set S batch ; For S batch For each triple in , a replacement triple is extracted from the set S ′ to form a triple pair set T batch ;

[0027] Step 2.3.3: Construct a loss function using the margin-based ranking criterion

[0028]

[0029] Where d represents the 1-norm, [·] + Indicates taking the positive part;

[0030] Step 2.3.4: Based on minimization Update the embedding vectors of entities and relations in triples in batches;

[0031] Repeat steps 2.3.2-2.3.4 until all triples in set S are updated, and this round of iteration is completed; then repeat step 2.3 until the iteration requirements are met.

[0032] Further, step 3 includes the following sub-steps,

[0033] Step 3.1: Extract the embedding vector of each sensor from the triplet updated in step 2;

[0034] Step 3.2: Use the formula Calculate the edge d between the i-th sensor and the j-th sensor ij ; Among them, s i,m 、s j,m Represent the mth value of the i-th sensor and the j-th sensor respectively;

[0035] Step 3.3: Construct the adjacency matrix W of each sensor;

[0036] Step 3.4: Construct a knowledge-embedded aerospace engine spatiotemporal degradation dataset Specifically:

[0037]

[0038] in, represents the sensor set, s i represents the i-th sensor, E represents the set of edges between sensors, (s i ,s j ) represents the edge between the i-th sensor and the j-th sensor, represents the set of sensor measurements, x i Represents the measurement value of the i-th sensor.

[0039] Furthermore, the spatiotemporal graph convolutional neural network model includes a spatial module, a temporal module and an output module arranged in sequence; the spatial module is used to extract features containing spatial information from the spatiotemporal degradation data set to obtain a first feature; the temporal module is used to extract features containing time series information from the first feature to obtain a second feature; and the output module is used to obtain an aircraft engine life prediction result based on the second feature.

[0040] Furthermore, the spatial module uses a GCN layer with a multi-head attention mechanism, with a spatiotemporal degradation dataset as input, and the convolution operation formula is:

[0041]

[0042] Among them, K represents the number of attention heads, and Φ (k) denote the adjacency matrix elements and weight matrix of the k-th attention head, σ denotes the sigmoid function; h = {h1,h2,...,h n}, represents the output feature of the GCN layer (i.e., the first feature), n represents the number of sensors, represents a set of real numbers, and F represents the dimension of each sensor output feature; represents the first-order neighborhood of the ith sensor (i.e., the sensors directly adjacent to the ith sensor); the hyperparameter Φ (k) Automatically updated by the neural network.

[0043] The time module uses an LSTM network, and the formula is:

[0044]

[0045] Among them, x t represents the input vector of the time module at time t (i.e., the first feature), L t-1 represents the output vector of the time module at time t-1, i t 、f t , o t , g t They represent the activation vectors of the input gate, forget gate, output gate, and unit input at time t, respectively. t represents the unit state vector at time t, L t represents the output vector of the time module at time t (i.e., the second feature), W i , W f , W o , W g Represents the input weight matrix of the input gate, forget gate, output gate and unit input, U i , U f , U o, U g Represents the recurrent connection weight matrix of the input gate, forget gate, output gate and unit input, b i 、b f 、b o 、b g They represent the deviation of the input gate, forget gate, output gate and unit input respectively, tanh represents the hyperbolic tangent function, and ⊙ represents the Hadamard product, that is, the multiplication of the corresponding elements of the matrix;

[0046] The output module uses a fully connected layer, the formula is:

[0047]

[0048] in, represents the output vector of the output module at time t, W FC represents the weight matrix of the fully connected layer, b FC Denotes the bias of the fully connected layer.

[0049] Furthermore, in step 4, the sensor measurement historical data is substituted into the spatiotemporal degradation dataset to obtain an offline spatiotemporal degradation dataset; the spatiotemporal graph convolutional neural network model is trained using the offline spatiotemporal degradation dataset; the mean square error is used as the loss function, and the model parameters are iteratively optimized using the gradient descent optimization algorithm. When the number of training iterations reaches a set threshold, the iteration is stopped to obtain a trained spatiotemporal graph convolutional neural network model.

[0050] Furthermore, in step 4, before training the spatiotemporal graph convolutional neural network model, the adjacency matrix of each sensor is constructed as follows:

[0051] Using the formula Substitute the edge d ij Calculate the corresponding adjacency matrix element w ij , and construct the adjacency matrix W:

[0052]

[0053] During the training of the spatiotemporal graph convolutional neural network model, the attention mechanism is used to update the adjacency matrix. Specifically, the formula e is used. ij =a(Φx i ,Φx j ), calculate the attention coefficient eij; where the hyperparameter a(·,·) is the attention mechanism vector, which is automatically updated by the neural network;

[0054] The attention coefficient e ij Regularization processing obtains the updated adjacency matrix element w ij And the updated adjacency matrix W is calculated as:

[0055]

[0056] in, represents the first-order neighborhood of the ith sensor (i.e., the sensors directly adjacent to the ith sensor).

[0057] Compared with the prior art, the aircraft engine life prediction method based on knowledge embedded graph neural network provided by the present invention has the following technical effects:

[0058] 1) The present invention constructs a knowledge graph of aircraft engine sensors and converts it into an embedded vector to construct a knowledge-embedded aircraft engine spatiotemporal degradation dataset, and then predicts the remaining life of the aircraft engine through a spatiotemporal graph convolutional neural network model; the present invention supplements the correlation between the sensors in the aircraft engine based on prior knowledge, so as to effectively and accurately predict the remaining life of the aircraft engine;

[0059] 2) When constructing the embedding vector of the knowledge graph, the present invention simultaneously utilizes the engine components, the connection relationship between the engine components, and the monitoring relationship between the engine components and the sensors, thereby comprehensively integrating the global information of each component to improve the accuracy and stability of the aircraft engine life prediction;

[0060] 3) The present invention combines the sensor set, the set of sensor edges, the adjacency matrix of each sensor, and the set of sensor measurement values ​​to construct a knowledge-embedded aero-engine spatiotemporal degradation data set, taking into account the topological structure of the sensor, and further improving the accuracy of aero-engine life prediction;

[0061] 4) The spatiotemporal graph convolutional neural network model provided by the present invention extracts features containing spatial information through the spatial module and extracts features containing time series information through the temporal module, while taking into account the spatial correlation and temporal correlation of sensor data, thereby ensuring the effectiveness of aircraft engine life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the aircraft engine life prediction method based on knowledge embedding graph neural network;

[0063] Figure 2 The knowledge graph of aviation engine sensors constructed;

[0064] Figure 3 A schematic diagram of the process of converting a knowledge graph into an embedding vector;

[0065] Figure 4 Schematic diagram of the process of constructing spatiotemporal degradation datasets for knowledge embedding;

[0066] Figure 5 This is a diagram showing the life prediction results of the embodiment. DETAILED DESCRIPTION

[0067] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0068] The aircraft engine life prediction method based on knowledge embedded graph neural network provided in this embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0069] Step 1: Build the knowledge graph of aircraft engine sensors;

[0070] Step 2: Convert the knowledge graph into an embedding vector;

[0071] Step 3: Based on the embedding vectors of each sensor obtained in step 2, a spatiotemporal degradation dataset for knowledge embedding is constructed;

[0072] Step 4: Substitute the sensor measurement history data into the spatiotemporal degradation dataset to obtain the offline spatiotemporal degradation dataset; and use the offline spatiotemporal degradation dataset to train the spatiotemporal graph convolutional neural network model;

[0073] Step 5: Substitute the online data measured by the sensor into the spatiotemporal degradation dataset to obtain the online spatiotemporal degradation dataset; input the online spatiotemporal degradation dataset into the trained spatiotemporal graph convolutional neural network model to obtain the aircraft engine life prediction result.

[0074] The specific implementation of the above steps 1 to 5 is explained below.

[0075] In the above step 1, the structural information of the aircraft engine and the location information of the sensor are combined to construct the aircraft engine sensor knowledge graph, which is regarded as prior spatial knowledge.

[0076] The connection relationship between different engine components is established according to the structure of the aircraft engine and the connection type of each engine component (that is, the various parts that make up the engine). The connection relationship includes mechanical connection and airflow connection: mechanical connection means that two components are connected by a mechanical structure, such as a fan fixed to the main shaft through a coupling; airflow connection means that two components are connected by airflow, such as compressed air flowing from a low-pressure chamber to a high-pressure chamber. Then, the monitoring relationship between the engine components and the sensors is established according to the location layout of the sensors and the monitoring objects; the monitoring objects are set according to the factors affecting the performance of the engine components; for example, the factors that affect the performance calculation of the fan include the inlet temperature, pressure and fan speed. The corresponding English numbers of aircraft engine components and sensors are shown in Table 1. According to the connection relationship between different engine components and the monitoring relationship between engine components and sensors, the knowledge graph of aircraft engine sensors is constructed as follows: Figure 2 shown.

[0077] Therefore, the constructed aviation engine sensor knowledge graph includes several engine components, several sensors for monitoring each engine component, the connection relationship between different engine components, and the monitoring relationship between engine components and sensors.

[0078] Table 1 Aircraft engine components and sensor codes and corresponding monitoring objects of sensors

[0079]

[0080] In the above step 2, the knowledge graph is converted into an embedding vector using an energy-based knowledge embedding algorithm.

[0081] In the specific implementation, Figure 3 As shown, this step includes the following sub-steps:

[0082] Step 2.1: Construct a mathematical representation of the knowledge graph using triples; each triple includes a head entity, a tail entity, and a relationship between the head entity and the tail entity; the head entity and the tail entity are selected from a set of transmitter components and sensors; all triples constitute a triple set.

[0083] In this embodiment, the triple (e h ,r,e t ) form to construct the mathematical representation of the knowledge graph, where e h ,e t ∈E is an entity, taking values ​​from the set E of sensors and engine components, which includes Figure 1 There are 21 sensors and 9 components shown; r represents the relationship, including 3 types of connections, namely mechanical connection, airflow connection, and monitoring connection; for example, Figure 2 In the example, (C1, mechanical connection, C8) is one of the triples. All triples constitute a set S; initialize all triples, e h 、r、e t From uniform distribution Random sampling is performed in, where U represents uniform distribution and M is the dimension of the embedding vector; and r is normalized. Here, ||·|| represents the 2-norm.

[0084] Step 2.2: Construct a set of replacement triples.

[0085] In this embodiment, e in each original triple in the set S is h or e t Randomly replace with other entities in the set E (that is, any other entity other than itself) to construct a replacement triple set S′, as shown in the following formula:

[0086]

[0087] Among them, e h ′ and e t ' indicates the replaced entity.

[0088] Step 2.3: Iteratively update each triple in the triple set to obtain the optimal embedding vector for each entity and relation in the triple.

[0089] In each iteration, follow these steps:

[0090] Step 2.3.1: Use the formula The entities are normalized; where ||·|| represents the 2-norm, which is the square root of the sum of the squares of the elements.

[0091] Step 2.3.2: Randomly extract a batch of triplets S of size b from the set S batch ; For S batch For each triple in , a replacement triple is extracted from the set S ′ to form a triple pair set T batch .

[0092] For each triple It should be The nearest neighbor of therefore, and The distance between and On this basis, the margin-based sorting criterion is used as the loss function

[0093] Step 2.3.3: Construct a loss function using the margin-based ranking criterion

[0094]

[0095] Where d represents the 1-norm, which is the sum of the absolute values ​​of each element; [·] + Indicates taking the positive part;

[0096] Step 2.3.4: Based on minimization The embedding vectors of entities and relations in triples are updated in batches.

[0097] Using the gradient descent algorithm, all entities and relations in each triple in the current batch The embedding vector is updated until the loss function is 0 or the number of iterations reaches 1000.

[0098] Based on the above analysis, repeat steps 2.3.2-2.3.4 until all triples in set S are updated, and then this round of iteration is completed. Repeat step 2.3 again until the iteration requirements are met (for example, the number of iterations reaches 20 times), and the optimal embedding vectors of each entity and relationship in the triple are obtained.

[0099] In the above step 3, the embedding vector of each sensor is extracted from the triplet updated in step 2, and then the spatiotemporal degradation dataset of knowledge embedding is constructed.

[0100] In the specific implementation, Figure 4 As shown, this step includes the following sub-steps:

[0101] Step 3.1: Extract each sensor embedding vector from the updated triplet in step 2.

[0102] Through step 2, each sensor is assigned an M-dimensional vector to represent its position in the low-dimensional space. For example, the embedding vector of the i-th sensor is s i =(s i,1 ,s i,2 ,…,s i,M ).

[0103] Step 3.2: Use the formula Calculate the edge d between the i-th sensor and the j-th sensor ij ; Among them, s i,m 、s j,m represent the mth value of the i-th sensor and the j-th sensor respectively.

[0104] Step 3.3: Construct the adjacency matrix W of each sensor.

[0105] Initially, before training the spatiotemporal graph convolutional neural network model, the adjacency matrix of each sensor is constructed as:

[0106] Using the formula Substitute the edge d ij Calculate the corresponding adjacency matrix element w ij , and construct the adjacency matrix W:

[0107]

[0108] Here, exp(·) represents an exponential function.

[0109] Step 3.4: Construct a knowledge-embedded aerospace engine spatiotemporal degradation dataset Specifically:

[0110]

[0111] in, represents the sensor set, s i represents the i-th sensor, E represents the set of edges between sensors, (s i ,s j ) represents the edge between the i-th sensor and the j-th sensor, represents the set of sensor measurements, x i Represents the measurement value of the i-th sensor.

[0112] In the above step 4, the spatiotemporal graph convolutional neural network model includes a spatial module, a temporal module and an output module which are arranged in sequence; the spatial module is used to extract features containing spatial information from the spatiotemporal degradation data set to obtain a first feature; the temporal module is used to extract features containing time series information from the first feature to obtain a second feature; and the output module is used to obtain an aircraft engine life prediction result based on the second feature.

[0113] In this embodiment, the spatial module uses the GCN layer of the multi-head attention mechanism, takes the spatiotemporal degradation dataset as input, and the convolution operation formula is:

[0114]

[0115] Among them, K represents the number of attention heads, and Φ (k) denote the adjacency matrix elements and weight matrix of the k-th attention head, σ denotes the sigmoid function; h = {h1,h2,...,h n}, represents the output feature of the GCN layer (i.e., the first feature), n represents the number of sensors, represents a set of real numbers, F represents the dimension of each sensor output feature, and the hyperparameter Φ (k) is the weight matrix, which is automatically updated by the neural network.

[0116] The time module uses the LSTM network, the formula is:

[0117]

[0118] Among them, x t represents the input vector of the time module at time t (i.e., the first feature), L t-1 represents the output vector of the time module at time t-1, i t 、f t , o t , g t They represent the activation vectors of the input gate, forget gate, output gate, and unit input at time t, respectively. t represents the unit state vector at time t, L t represents the output vector of the time module at time t (i.e., the second feature), Wi , W f , W o , W g Represents the input weight matrix of the input gate, forget gate, output gate and unit input, U i , U f , U o , U g Represents the recurrent connection weight matrix of the input gate, forget gate, output gate and unit input, b i , b f , b o , b g They represent the deviation of the input gate, forget gate, output gate and unit input respectively, tanh represents the hyperbolic tangent function, and ⊙ represents the Hadamard product, that is, the multiplication of the corresponding elements of the matrix.

[0119] The output module uses a fully connected layer, the formula is:

[0120]

[0121] in, represents the output vector of the output module at time t, W FC represents the weight matrix of the fully connected layer, b FC Denotes the bias of the fully connected layer.

[0122] The sensor measurement historical data is substituted into the spatiotemporal degradation dataset to obtain the offline spatiotemporal degradation dataset; the spatiotemporal graph convolutional neural network model is trained using the offline spatiotemporal degradation dataset; the mean square error is used as the loss function, and the model parameters are iteratively optimized using the gradient descent optimization algorithm. When the number of training iterations reaches the set threshold, the iteration is stopped to obtain the trained spatiotemporal graph convolutional neural network model.

[0123] During the training iteration of the spatiotemporal graph convolutional neural network model, the attention mechanism is used to update the adjacency matrix. Specifically, the formula e is used. ij =a(Φx i ,Φx j ), calculate the attention coefficient e ij ; Among them, the hyperparameter a(·,·) is the attention mechanism vector, which is automatically updated by the neural network;

[0124] The attention coefficient e ij Regularization processing obtains the updated adjacency matrix element w ij And the updated adjacency matrix W is calculated as:

[0125]

[0126] in, represents the first-order neighborhood of the ith sensor (i.e., the sensors directly adjacent to the ith sensor).

[0127] In the above step 5, the online real-time measurement value of each sensor is substituted into the spatiotemporal degradation dataset to obtain the online spatiotemporal degradation dataset; the online spatiotemporal degradation dataset is input into the trained spatiotemporal graph convolutional neural network model to obtain the aircraft engine life prediction result.

[0128] In this embodiment, the proposed method of the present invention is verified based on the aircraft engine simulation data set published by NASA of the United States.

[0129] From the aviation engine simulation dataset published by NASA, we randomly selected engine No. 100 in the FD001 sub-dataset, engine No. 236 in the FD002 sub-dataset, engine No. 99 in the FD003 sub-dataset, and engine No. 65 in the FD004 sub-dataset. We used 90% of the sample data of each engine to construct a training set and 10% to construct a test set.

[0130] According to steps 1 to 4 given above, a spatiotemporal degradation dataset is constructed, and the measurement values ​​of each sensor in the training set samples are substituted into the spatiotemporal degradation dataset to obtain an offline spatiotemporal degradation dataset; the spatiotemporal graph convolutional neural network model is trained using the offline spatiotemporal degradation dataset; the mean square error is used as the loss function, and the model parameters are iteratively optimized using the gradient descent optimization algorithm. When the number of training iterations reaches the set threshold (for example, 50 times), the iteration is stopped to obtain a trained spatiotemporal graph convolutional neural network model.

[0131] Then, the sensor measurements in the test set samples are substituted into the spatiotemporal degradation dataset to obtain the online spatiotemporal degradation dataset. The online spatiotemporal degradation dataset is input into the trained spatiotemporal graph convolutional neural network model to obtain the predicted life of the aircraft engine. The prediction results and prediction errors are shown in Figure 2. Figure 5 As shown in the figure, it can be seen that in all sub-datasets, the predicted values ​​of the remaining life are very consistent with the actual values, thus proving the effectiveness and accuracy of the method of the present invention.

[0132] The method of the present invention is applicable to the remaining life prediction of various types of aircraft engines. In practical applications, implementers can use this method to establish an aircraft engine sensor knowledge graph and construct a knowledge-embedded spatiotemporal degradation data set, and use the spatiotemporal graph convolutional neural network model proposed by the method of the present invention to predict the remaining life of the aircraft engine, thereby reducing the operation and maintenance costs. It should be pointed out that without departing from the concept of the present invention, the adjustments and deformations made to the method of the present invention should also be regarded as the protection scope of the present invention.

[0133] Those skilled in the art will appreciate that the embodiments herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network, characterized in that: The following steps are involved: Step 1: construct an aircraft engine sensor knowledge graph; the aircraft engine sensor knowledge graph includes a number of engine components, a number of sensors for monitoring each engine component, connection relationships between different engine components, and monitoring relationships between engine components and sensors; Step 2: Convert the knowledge graph into an embedding vector; this step includes the following sub-steps, Step 2.1: construct a mathematical representation of the knowledge graph using triples; each triple includes a head entity, a tail entity, and a relationship between the head entity and the tail entity; the head entity and the tail entity are selected from a set of transmitter components and sensors; all triples constitute a triple set; Step 2.2: Construct a set of replacement triples; Step 2.3: Iteratively update each triple in the triple set to obtain the optimal embedding vector for each entity and relationship in the triple; Step 3: Based on the embedding vectors of each sensor obtained in step 2, a spatiotemporal degradation dataset for knowledge embedding is constructed; Step 4: Substitute the sensor measurement history data into the spatiotemporal degradation dataset to obtain the offline spatiotemporal degradation dataset; And use the offline spatiotemporal degradation dataset to train the spatiotemporal graph convolutional neural network model; Step 5: Substitute the online data measured by the sensor into the spatiotemporal degradation dataset to obtain the online spatiotemporal degradation dataset; input the online spatiotemporal degradation dataset into the trained spatiotemporal graph convolutional neural network model to obtain the aircraft engine life prediction result.

2. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 1 is characterized in that: The step 1 comprises the following steps: Step 1.1: Establish connection relationships between different engine components according to the structure of the aircraft engine and the connection types of each engine component; Step 1.2: Establish the monitoring relationship between the engine components and the sensors according to the location layout of the sensors and the monitoring objects; Step 1.3: Construct an aviation engine sensor knowledge graph based on the connection relationship between different engine components and the monitoring relationship between engine components and sensors.

3. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 2 is characterized in that: The connection relationship includes mechanical connection and air flow connection.

4. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 1, characterized in that: The specific method of step 2.1 is: The triple (e h ,r,e t ) form to construct the mathematical representation of the knowledge graph, where e h ,e t ∈E is an entity, which takes values ​​from the set E of sensors and engine components; r represents a relationship; all triples constitute a set S; all triples are initialized, e h 、r、e t From uniform distribution Random sampling is performed in, where U represents uniform distribution and M is the dimension of the embedding vector; and r is normalized. Here, ||·|| represents the 2-norm.

5. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 4 is characterized in that: The specific method of step 2.2 is: Using the formula For each original triple in the set S, e h or e t Randomly replace with other entities in set E and construct a replacement triple set S′.

6. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 5 is characterized in that: The step 2.3 comprises the following sub-steps: Step 2.3.1: Use the formula The entity is normalized; where ||·|| represents the 2-norm; Step 2.3.2: Randomly extract a batch of triplets S of size b from the set S batch ; For S batch For each triple in , a replacement triple is extracted from the set S ′ to form a triple pair set T batch ; Step 2.3.3: Construct a loss function using the margin-based ranking criterion Where d represents the 1-norm, [·] + Indicates taking the positive part; Step 2.3.4: Based on minimization Update the embedding vectors of entities and relations in triples in batches; Repeat steps 2.3.2-2.3.4 until all triples in set S are updated, and this round of iteration is completed; then repeat step 2.3 until the iteration requirements are met.

7. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 1, characterized in that: The step 3 comprises the following sub-steps: Step 3.1: Extract the embedding vector of each sensor from the triplet updated in step 2; Step 3.2: Use the formula Calculate the edge d between the i-th sensor and the j-th sensor ij ; Among them, s i,m 、s j,m Represent the mth value of the i-th sensor and the j-th sensor respectively; Step 3.3: Construct the adjacency matrix W of each sensor; Step 3.4: Construct a knowledge-embedded aerospace engine spatiotemporal degradation dataset Specifically: in, represents the sensor set, s i represents the i-th sensor, E represents the set of edges between sensors, (s i ,s j ) represents the edge between the i-th sensor and the j-th sensor, represents the set of sensor measurements, x i represents the measurement value of the i-th sensor.

8. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to any one of claims 1 to 7, characterized in that: The spatiotemporal graph convolutional neural network model includes a spatial module, a temporal module and an output module arranged in sequence; the spatial module is used to extract features containing spatial information from a spatiotemporal degradation data set to obtain a first feature; the temporal module is used to extract features containing time series information from the first feature to obtain a second feature; and the output module is used to obtain an aircraft engine life prediction result based on the second feature.

9. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 8, characterized in that: The spatial module uses the GCN layer of the multi-head attention mechanism; the temporal module uses the LSTM network; and the output module uses the fully connected layer.

10. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 9, characterized in that: The spatial module takes the spatiotemporal degradation dataset as input, and the convolution operation formula is: Among them, K represents the number of attention heads, and Φ (k) They represent the adjacency matrix elements and weight matrix under the k-th attention head, and σ represents the sigmoid function; represents the output features of the GCN layer, n represents the number of sensors, represents a set of real numbers, and F represents the dimension of each sensor output feature; represents the first-order neighborhood of the ith sensor.

11. The method for predicting the life of an aircraft engine based on a knowledge embedded graph neural network according to claim 10, characterized in that: In step 4, before training the spatiotemporal graph convolutional neural network model, the adjacency matrix of each sensor is constructed as follows: Using the formula and Substitute the edge d ij Calculate the corresponding adjacency matrix element w ij , and construct the adjacency matrix W: During the training of the spatiotemporal graph convolutional neural network model, the attention mechanism is used to update the adjacency matrix. Specifically, the formula e is used. ij =a(Φx i ,Φx j ), calculate the attention coefficient e ij ; Among them, the hyperparameter a(·,·) is the attention mechanism vector; The attention coefficient e ij Regularization processing obtains the updated adjacency matrix element w ij And the updated adjacency matrix W is calculated as:

Images