Aero-engine service life monitoring method based on artificial intelligence

Abstract

The invention discloses an aero-engine service life monitoring method based on artificial intelligence, and relates to the field of equipment fault prediction and health management, and the aero-engine service life monitoring method comprises the steps: constructing a BiLSTM-GRU-Attention model based on multiple optimization, carrying out the multi-time sampling output through combining with a Dropout technology in a reasoning stage, directly carrying out the uncertainty quantification in a neural network frame, and carrying out the prediction of the service life of an aero-engine. According to the method, the global search capability of the model can be effectively improved, local optimum is avoided, so that the prediction precision of the model on time series data is improved, meanwhile, the model is more stable in performance and higher in convergence speed in complex tasks, the calculation complexity and memory requirements are reduced, the training and prediction efficiency of the model is improved, and the prediction efficiency of the model is improved. The method can better process high-dimensional multi-modal data, provides more reliable prediction results and risk assessment, and has stronger generalization ability and adaptability.

Description

Technical Field

[0001] The present invention relates to the field of equipment failure prediction and health management, and specifically to an aircraft engine life monitoring method based on artificial intelligence. Background Art

[0002] As the main power source of large aerospace vehicles, aircraft engines’ failures will seriously affect the safety and reliability of the equipment, so they need to be paid special attention in equipment maintenance. With the development of science and technology, prognostic and health management (PHM) technology is widely used in complex industrial systems, among which predicting the remaining useful life (RUL) of equipment is an important part of PHM. RUL prediction not only ensures the smooth operation of the engine and reduces the possibility of catastrophic failure, but also helps to formulate scientific maintenance strategies, reduce unnecessary maintenance expenses, realize the transformation from traditional shutdown maintenance to condition-based maintenance, and promote the sustainable development of the aviation industry. At present, scholars can roughly divide the prediction of the remaining useful life of aircraft engines into three categories: physics-based methods, data-driven methods, and fusion methods. The physical model-based method predicts the remaining useful life (RUL) of the equipment by constructing a mathematical or physical model of component degradation and combining it with observational data. Although this method has high prediction accuracy under specific conditions, it lacks universality. The data-driven method collects the input and output data of the system and can achieve RUL prediction without in-depth understanding of the working principle or failure mechanism of the equipment. Data-driven methods are mainly divided into two categories: mathematical statistics and machine learning, among which machine learning is more widely used. Machine learning can be divided into shallow and deep learning. The former has difficulty in identifying high-dimensional data features, while deep learning solves this problem through automatic feature extraction.

[0003] Although deep learning has performed well in time series data modeling, most methods have neglected uncertainty quantification, which is essential for safety-critical applications such as aircraft engine life prediction. Uncertainty quantification can evaluate the reliability of the model, especially in the face of unknown operating conditions, which helps to optimize risk management, formulate flexible maintenance strategies, avoid excessive or insufficient maintenance, and thus optimize resources and costs; on this basis, an aircraft engine life monitoring method based on artificial intelligence is proposed to improve the reliability of the engine remaining service life prediction results. Summary of the invention

[0004] The purpose of the present invention is to provide an aircraft engine life monitoring method based on artificial intelligence to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: an aircraft engine life monitoring method based on artificial intelligence, comprising the following steps:

[0006] S1. Preprocess the operating conditions in the aircraft engine data set collected by the multi-dimensional sensor and divide it into a training set and a test set;

[0007] S2. Construct a BiLSTM-GRU-Attention network structure model based on multiple optimizations;

[0008] S3. Use the training set to train the BiLSTM-GRU-Attention network model based on multiple optimizations to obtain a trained network model;

[0009] S4. Input the test set data into the trained network model to achieve real-time prediction of the remaining service life of the engine.

[0010] Furthermore, in step S1: data preprocessing includes the following steps:

[0011] S1-1, determining the correlation between the original sensor data of the collection device and the remaining service life, and if the value of the sensor data is constant, discarding the original sensor data of the device;

[0012] S1-2, perform max-min normalization on the filtered data and map all the original data to the interval [0,1];

[0013] S1-3. In the normalized data, use the sliding window method to generate several groups of multivariate time series, so as to construct the time series training set of different equipment. The step size of the sliding time window is 1, and the life label and label normalization are established to obtain the processed training set and test set.

[0014] Furthermore, in step S1-2: the filtered data is subjected to max-min normalization processing, and all the original data are mapped to the interval [0, 1]; the formula for the normalization processing is as follows:

[0015]

[0016] Among them, X′ i，j (t) is X i，j (t) The value obtained by data normalization; X i，j (t) is the raw data obtained by the jth sensor of the i-th engine at time t; min(X ，j ) and max(X ，j ) are the minimum and maximum values ​​obtained by the jth sensor for all engines at all times.

[0017] Furthermore, in step S2: the network model of BiLSTM-GRU-Attention based on multiple optimizations specifically includes:

[0018] S2-1. Construct the BiLSTM layer and abstract its parameters into particles in motion and optimize them using algorithms to obtain the optimal parameter combination of the BiLSTM layer hyperparameters.

[0019] S2-2. Construct a GRU layer based on Bayesian parameter optimization to further extract complex time series features and enhance feature expression capabilities;

[0020] S2-3. Construct an attention mechanism layer to perform weighted processing on the outputs of the BiLSTM and GRU layers. By calculating the importance weight of each time step, the most important features are selected to form a new feature representation.

[0021] S2-4, the features processed by the attention layer are input to the fully connected layer for further dimension reduction, the remaining service life of the engine is output, and the Dropout layer is inserted to prevent overfitting. The Dropout value is set to r;

[0022] S2-5. During the training process of predicting the fitting curve, the entire training data set is randomly divided into several small batches, each of which contains m samples. The model is trained with a learning rate of p. The training process of the FD001 sub-dataset is configured to run a epoch at most, and the training process of the FD003 sub-dataset is configured to run b epochs at most.

[0023] S2-6. In the inference stage, the Dropout technique is used to quantify the uncertainty of the model's prediction results.

[0024] Further, in step S2-1: construct a BiLSTM layer, abstract the parameters of the BiLSTM layer into particles in motion, and the position of particle i in one-dimensional space is represented as a vector X i =(x1, x2, ...x N ), the velocity is represented by vector V i =(v1, v2, ...v N ); Each particle has a target position X determined by the objective function, and can determine the current individual best position pbest and the current position x i , and can also determine the current group best position gbest, which is the best value in pbest; the particle updates its position by tracking two extreme values, namely the individual best position pbest and the group best position gbest, and finally obtains the optimal parameter combination of the BiLSTM layer corresponding to the group best position gbest. This method has self-organization, evolution and memory functions, and can enhance the feature capture ability of the BiLSTM layer, thereby improving the prediction accuracy of the network prediction model BiLSTM-GRU-Attention.

[0025] Furthermore, in step S2-1, the steps included are:

[0026] S2-1-1, initialize a group of particles with a population size of N, the initialization information includes the random position and speed of the particles, the initial number of iterations T = 0, and prepare to enter the optimization process;

[0027] S2-1-2, define the target position X of the particle, and evaluate the excellence of the current position, the individual best position and the group best position according to the target position. The closer the position is to the target position, the better it is.

[0028] S2-1-3, the current position x of each particle i Compare it with the individual optimal position pbest it has passed. If the current position is better, take it as the current individual optimal position pbest;

[0029] S2-1-4, the current position x of each particle i Compare it with the group optimal position gbest it has passed. If the current position is better, take it as the current group optimal position gbest;

[0030] S2-1-5. Adjust particle speed and position according to the following formula:

[0031] v i ′=ω×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i );

[0032] x i ′=x i +v i Δt;

[0033] Among them, v i ′ is the velocity of the particle after iterative update, v i is the current velocity of the particle, x i ′ is the position value of the particle after iterative update, x i is the current position value of the particle, Δt is the unit time interval, ω is the inertia factor, i=1, 2, 3, ..., N, N is the total number of particles in this group, rand() is a random number between (0, 1), pbest i is the current individual optimal position, gbest i is the optimal position of the current group, c1 and c2 are learning factors;

[0034] The inertia factor ω is a dynamically changing value, and the changing rule is:

[0035]

[0036] Among them, G k represents the maximum number of iterations, ω ini represents the initial inertia weight, ω gnd Indicates the inertia weight when iterating to the maximum evolutionary generation; the inertia factor ω affects the optimization ability of this method. When the ω value is large, the global optimization ability is strong and the local optimization ability is weak; when the ω value is small, the global optimization ability is weak and the local optimization ability is strong; dynamic ω can obtain better optimization results than fixed values. Dynamic ω can change linearly during the search process of the particle swarm optimization algorithm, and can also change dynamically according to a certain measurement function of the particle swarm optimization algorithm performance;

[0037] S2-1-6: If the iteration termination condition is not reached, continue to execute step S2-1-2, and set the number of iterations T = T + 1, until the number of iterations T = G k The iteration is terminated, and the optimal parameter combination of the BiLSTM layer corresponding to the optimal position gbest of the group is output.

[0038] Further, in step S2-2: the following steps are included:

[0039] S2-2-1, according to formula r t =σ(W r ·[h t-1 , χ t ]+b r ) Calculate the reset gate of GRU, where the hidden state of the previous moment is h t-1 , χ t is the input of the current time step, σ represents the Sigmoid activation function, W r and b r They are the weight matrix and bias vector of the forget gate respectively;

[0040] S2-2-2, according to formula z t =σ(W z ·[h t-1 , X t ]+b z ) Calculate the update gate of GRU, where W z and b z In particular, reset the weights and biases of the gates;

[0041] S2-2-3, according to the formula Calculate the candidate hidden state, where W and b are the weight and bias of the candidate hidden state, respectively, and tanh is the hyperbolic tangent activation function;

[0042] S2-2-4. Define the hyperparameter space, determine the hyperparameters to be tuned for the GRU layer, and set a reasonable search range for each hyperparameter;

[0043] S2-2-5. Introduce the Bayesian optimization algorithm and use the mean square error as the objective function:

[0044] S2-2-6. Surrogate model fitting: Use Gaussian process to fit the objective function. The fitting relationship is: Among them, μ is the mean function, which represents the expectation of the objective function; K is the covariance function, also known as the kernel function, which is used to estimate the similarity between different hyperparameter points; D is the evaluated hyperparameter combination;

[0045] S2-2-7. Bayesian optimization selects the next hyperparameter combination for evaluation by maximizing expectation: Among them, EI(χ) represents the maximization expectation, χ + is the objective function value corresponding to the current optimal hyperparameter point, is the predicted value of the new hyperparameter combination on the objective function;

[0046] S2-2-8. Update the proxy model, use the new evaluation point to update the Gaussian process proxy model, and iterate continuously until the preset number of evaluations is reached or the convergence condition is met.

[0047] Further, in step S2-3: the following steps are included:

[0048] S2-3-1. According to the formula h1=W1·X+b1, the linear layer of the attention mechanism layer is established, and the score of the Attention mechanism is calculated using linear transformation, where X=[X1, X2,…, X T ] is the input sample, h1 is the output of the linear layer, which is the attention score, W1 and b1 represent the weight and bias of the linear layer;

[0049] S2-3-2, according to formula W x = Softmax(h1) to establish the weight generation layer of the attention mechanism layer, where W x Represents the generated weights. The calculation function uses the Softmax() function to normalize the attention scores to ensure that the sum of the weights in the time dimension is 1.

[0050] S2-3-3, according to formula X att =W x ⊙X weights the initial input sample and the generated weights element by element, where X att To finally generate the output of the attention mechanism layer, ⊙ represents the element-by-element multiplication operation.

[0051] Further, in step S2-6: the following steps are included:

[0052] S2-6-1. In the inference phase, the activation state of the Dropout layer is maintained for each input sample. During each forward propagation, the Dropout layer randomly selects some neurons and discards them to form a new sub-model structure.

[0053] S2-6-2. Perform multiple forward propagations for each input sample, using different random inactivation modes each time to obtain multiple prediction values. Where N represents the number of forward propagations;

[0054] S2-6-3. Basic formula Calculate the predicted mean as the final life expectancy prediction value, where Mean Prediction represents the average prediction result under various model structures. is the output result of the i-th forward propagation, and N is the number of forward propagations;

[0055] S2-6-4, according to the formula Calculate the variance to quantify the uncertainty of the prediction process, where Uncertaint y It indicates the degree of change of the prediction results under different model structures. The larger the variance, the greater the difference in the prediction results under different model structures, and the higher the prediction uncertainty of the model for the sample. Conversely, the smaller the variance, the more stable the prediction results, and the lower the prediction uncertainty of the model for the sample.

[0056] Furthermore, in step S4: the built prediction model is trained using the training set, and the trained model is used to infer the validation set, and the root mean square error RMSE and the scoring function Scor e Two evaluation indicators are used to verify model performance, and the test set is used to evaluate the performance of the best model; the formulas of the two evaluation indicators are as follows:

[0057]

[0058] in, is the predicted value of the model, yi is the actual value, n is the number of engines in the test set, S is the calculated score, and h = R prediction -R actual , R prediction is the predicted value of the remaining useful life, R actual is the true value of the remaining useful life; the lower the RMSE and Score values ​​are, the better the prediction effect of the model used in the prediction method is.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The present invention reduces computational complexity and memory requirements, improves model training and prediction efficiency, can better process high-dimensional multimodal data, provide more reliable prediction results and risk assessment, and has stronger generalization ability and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a method flow of an aircraft engine life monitoring method based on artificial intelligence according to the present invention;

[0062] Figure 2 A schematic diagram of the BiLSTM layer structure of an aircraft engine life monitoring method based on artificial intelligence of the present invention;

[0063] Figure 3 A schematic diagram of the GRU layer structure of an aircraft engine life monitoring method based on artificial intelligence according to the present invention;

[0064] Figure 4 This is a GRU layer algorithm improvement flow chart of an aircraft engine life monitoring method based on artificial intelligence of the present invention;

[0065] Figure 5 This is a schematic diagram of the attention mechanism layer structure of an aircraft engine life monitoring method based on artificial intelligence in the present invention;

[0066] Figure 6 Uncertainty quantification confidence interval diagram for all engines in the FD001 and FD003 single-operating condition sub-datasets. DETAILED DESCRIPTION

[0067] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] Example: Figure 1-Figure 6 As shown, the present invention provides a technical solution, an aircraft engine life monitoring method based on artificial intelligence. Due to the confidentiality of information in the aerospace industry and the high cost of obtaining it, the data in this example selects the C-MAPSS turbofan engine degradation simulation data set disclosed by NASA in Table 1. This data set is widely used in aircraft engine RUL prediction.

[0069] Table 1 Introduction of sensors in CMAPSS dataset

[0070]

[0071]

[0072] The CMAPSS dataset contains four sub-datasets, namely FD001, FD002, FD003, and FD004. In this example, the single-condition FD001 and FD003 datasets are used, and both the training set and the test set contain 100 engine data.

[0073] S1: Preprocess various working conditions in the aircraft engine data set collected by multi-dimensional sensors and divide it into training set and test set;

[0074] S2: Construct a BiLSTM-GRU-Attention network structure model based on multiple optimizations;

[0075] S3: Use the training set to train the BiLSTM-GRU-Attention network model based on multiple optimizations to obtain the trained network model;

[0076] S4: Input the test set data into the trained network model to achieve real-time prediction of the remaining service life of the engine.

[0077] In the aforementioned step S1, data preprocessing specifically includes:

[0078] S1-1: Determine the correlation between the raw sensor data of the collected equipment and the remaining service life. If the value of the sensor data is constant, the raw sensor data of the equipment is discarded;

[0079] Step S1-2: Perform max-min normalization processing on the filtered data, and map all the original data to the interval [0, 1];

[0080] The normalization formula is as follows:

[0081]

[0082] Among them, X i，j (t) is X i，j (t) The value obtained by data normalization; X i，j (t) is the raw data obtained by the jth sensor of the i-th engine at time t; min(X ，j ) and max(X ，j ) are the minimum and maximum values ​​obtained by the jth sensor for all engines at all times;

[0083] Step S1-3: In the normalized data, use the sliding window method to generate several groups of multivariate time series, so as to construct a time series training set of different equipment. The step size of the sliding time window is 1, and the life label and label normalization are established to obtain the processed training set and test set.

[0084] In step S2, the network model of BiLSTM-GRU-Attention based on multiple optimizations specifically includes:

[0085] S2-1. Construct the BiLSTM layer and abstract its parameters into particles in motion and optimize them using algorithms to obtain the optimal parameter combination of the BiLSTM layer hyperparameters.

[0086] S2-2: Construct a GRU layer based on Bayesian parameter optimization to further extract complex time series features and enhance feature expression capabilities;

[0087] S2-3: Construct an attention mechanism layer to perform weighted processing on the outputs of the BiLSTM and GRU layers. By calculating the importance weight of each time step, the most important features are selected to form a new feature representation.

[0088] S2-4: The features processed by the attention layer are input to the fully connected layer for further dimensionality reduction, and the remaining service life of the engine is output. A Dropout layer is inserted to prevent overfitting. The Dropout value is set to r = 0.2;

[0089] S2-5: During the training of the prediction fitting curve, the entire training data set is randomly divided into several small batches, each of which contains m = 100 samples; the model is trained with a learning rate of p = 0.01, and the training process of the FD001 sub-dataset is configured to run a = 32 epochs at most, and the training process of the FD003 sub-dataset is configured to run b = 64 epochs at most;

[0090] Step S2-6: In the inference stage, the Dropout technique is used to quantify the uncertainty of the prediction results of the model;

[0091] In step S2-1: construct a BiLSTM layer, abstract the parameters of the BiLSTM layer into particles in motion, and represent the position of particle i in one-dimensional space as vector X i =(x1, x2, ...x N ), the velocity is represented by vector V i =(v1, v2, ...v N ); Each particle has a target position X determined by the objective function, and can determine the current individual best position pbest and the current position x i, and can also determine the current group best position gbest, which is the best value in pbest; the particle updates its position by tracking two extreme values, namely the individual best position pbest and the group best position gbest, and finally obtains the optimal parameter combination of the BiLSTM layer corresponding to the group best position gbest. This method has self-organization, evolution and memory functions, and can enhance the feature capture ability of the BiLSTM layer, thereby improving the prediction accuracy of the network prediction model BiLSTM-GRU-Attention.

[0092] In step S2-1: the steps included are:

[0093] S2-1-1, initialize a group of particles with a population size of N, the initialization information includes the random position and speed of the particles, the initial number of iterations T = 0, and prepare to enter the optimization process;

[0094] S2-1-2, define the target position X of the particle, and evaluate the excellence of the current position, the individual best position and the group best position according to the target position. The closer the position is to the target position, the better it is.

[0095] S2-1-3, the current position x of each particle i Compare it with the individual optimal position pbest it has passed. If the current position is better, take it as the current individual optimal position pbest;

[0096] S2-1-4, the current position x of each particle i Compare it with the group optimal position gbest it has passed. If the current position is better, take it as the current group optimal position gbest;

[0097] S2-1-5. Adjust particle speed and position according to the following formula:

[0098] v i ′=ω×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i );

[0099] x i ′=x i +v i Δt;

[0100] Among them, v i ′ is the velocity of the particle after iterative update, v i is the current velocity of the particle, x i′ is the position value of the particle after iterative update, x i is the current position value of the particle, Δt is the unit time interval, ω is the inertia factor, i=1, 2, 3, ..., N, N is the total number of particles in this group, rand() is a random number between (0, 1), pbest i is the current individual optimal position, gbest i is the optimal position of the current group, c1 and c2 are learning factors;

[0101] The inertia factor ω is a dynamically changing value, and the changing rule is:

[0102]

[0103] Among them, G k represents the maximum number of iterations, ω ini represents the initial inertia weight, ω gnd Indicates the inertia weight when iterating to the maximum evolutionary generation; the inertia factor ω affects the optimization ability of this method. When the ω value is large, the global optimization ability is strong and the local optimization ability is weak; when the ω value is small, the global optimization ability is weak and the local optimization ability is strong; dynamic ω can obtain better optimization results than fixed values. Dynamic ω can change linearly during the search process of the particle swarm optimization algorithm, and can also change dynamically according to a certain measurement function of the particle swarm optimization algorithm performance;

[0104] S2-1-6: If the iteration termination condition is not reached, continue to execute step S2-1-2, and set the number of iterations T = T + 1, until the number of iterations T = G k The iteration is terminated, and the optimal parameter combination of the BiLSTM layer corresponding to the optimal position gbest of the group is output.

[0105] The step S2-2 is specifically as follows:

[0106] S2-2-1: According to formula r t =σ(W r ·[h t-1 , x t ]+b r ) Calculate the reset gate of GRU, where the hidden state of the previous moment is h t-1 , χ t is the input of the current time step, σ represents the Sigmoid activation function, W r and b r They are the weight matrix and bias vector of the forget gate respectively;

[0107] S2-2-2: According to formula z t =σ(W z ·[h t-1 , x t ]+bz ) Calculate the update gate of GRU, where W z and b z In particular, reset the weights and biases of the gates;

[0108] S2-2-3: According to the formula Calculate the candidate hidden state, where W and b are the weight and bias of the candidate hidden state, respectively, and tanh is the hyperbolic tangent activation function;

[0109] S2-2-4: Define the hyperparameter space, determine the hyperparameters to be tuned for the GRU layer, and set a reasonable search range for each hyperparameter;

[0110] S2-2-5: Introduce Bayesian optimization algorithm and use mean square error (MSE)

[0111] As the objective function:

[0112] S2-2-6: Surrogate model fitting, using Gaussian process to fit the objective function, the fitting relationship is Where μ is the mean function, which represents the expectation of the objective function; K is the covariance function (kernel function), which is used to estimate the similarity between different hyperparameter points; D is the evaluated hyperparameter combination;

[0113] S2-2-7: Bayesian optimization selects the next hyperparameter combination for evaluation by maximizing expectation (EI): Among them, χ + is the objective function value corresponding to the current optimal hyperparameter point, is the predicted value of the new hyperparameter combination on the objective function;

[0114] Step S2-2-8: Update the proxy model, use the new evaluation point to update the Gaussian process proxy model, and iterate continuously until the preset number of evaluations is reached or the convergence condition is met.

[0115] The step S23 is specifically as follows:

[0116] S2-3-1: Establish the linear layer of the attention mechanism layer according to the formula h1=W1·X+b1, where X=[X1, X2, …, X T ] is the input sample, h1 is the output of the linear layer, W1 and b1 represent the weight and bias of the linear layer;

[0117] S2-3-2: According to formula W x = Softmax(h1) to establish the weight generation layer of the attention mechanism layer, where W x Represents the generated weights. The calculation function uses the Softmax() function to ensure that the weights are normalized in the time dimension so that their sum is 1.

[0118] S2-3-3: According to formula X att =W x ⊙X weights the initial input sample and the generated weights element by element, where X att To finally generate the output of the attention mechanism layer, ⊙ represents the element-by-element multiplication operation.

[0119] The step S2-6 is specifically as follows:

[0120] S2-6-1: In the inference phase, the activation state of the Dropout layer is maintained for each input sample. During each forward propagation, the Dropout layer randomly selects some neurons and discards them to form a new sub-model structure.

[0121] S2-6-2: Perform 10 forward propagations for each input sample, using a different random dropout pattern each time to obtain multiple predictions Where N represents the number of forward propagations;

[0122] S2-6-3: According to the formula Calculate the predicted mean as the final life expectancy prediction value, where Mean Prediction represents the average prediction result under various model structures. is the output result of the i-th forward propagation, and N is the number of forward propagations;

[0123] S2-6-4: According to the formula The variance is calculated to quantify the uncertainty of the prediction process. Uncertainty indicates the degree of change in the prediction results under different model structures. The larger the variance, the greater the difference in the prediction results under different model structures, and the higher the prediction uncertainty of the model for the sample. Conversely, the smaller the variance, the more stable the prediction results, and the lower the prediction uncertainty of the model for the sample.

[0124] In step S4, the built prediction model is trained using the training set, and the trained model is used to infer the validation set. The root mean square error RMSE and the scoring function Scor e Two evaluation metrics are used to verify model performance, and the performance of the best model is evaluated using the test set;

[0125] The formulas for the two evaluation indicators are as follows:

[0126]

[0127] in, is the predicted value of the model, y i is the actual value, n is the number of engines in the test set, S is the calculated score, and h = Rprediction -R actual , R prediction is the predicted value of the remaining useful life, R actual is the true value of the remaining useful life; the lower the RMSE and Score values ​​are, the better the prediction effect of the model used in the prediction method is.

[0128] Table 2 shows the RMSE and Scor of the prediction method proposed in the present invention on the single working condition data set of C-MAPSS in the ablation experiment. e The results show that the model used in the article can improve the accuracy of prediction and further verify the effectiveness of its method.

[0129] Table 2 RMSE and Scor of single working condition data set e result

[0130]

[0131]

[0132] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An aircraft engine life monitoring method based on artificial intelligence, characterized in that: The following steps are involved: S1. Preprocess the operating conditions in the aircraft engine data set collected by the multi-dimensional sensor and divide it into a training set and a test set; S2. Construct a BiLSTM-GRU-Attention network structure model based on multiple optimizations; S3. Use the training set to train the BiLSTM-GRU-Attention network model based on multiple optimizations to obtain a trained network model; S4. Input the test set data into the trained network model to achieve real-time prediction of the remaining service life of the engine.

2. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 1, characterized in that: In step S1: data preprocessing includes the following steps: S1-1, determining the correlation between the original sensor data collected from the sensor and the remaining service life, and if the value of the sensor data is constant, discarding the original sensor data of the sensor; S1-2, perform max-min normalization on the filtered data and map all the original data to the interval [0,1]; S1-3. In the normalized data, use the sliding window method to generate several groups of multivariate time series, so as to construct the time series training set of different equipment. The step size of the sliding time window is 1, and the life label and label normalization are established to obtain the processed training set and test set.

3. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 2, characterized in that: In step S1-2: the filtered data is normalized by max-min, and all the original data is mapped to the interval [0,1]. The formula for normalization is as follows: Among them, X′ i，j (t) is X i，j (t) The value obtained by data normalization; X i，j (t) is the raw data obtained by the jth sensor of the i-th engine at time t; min(X ，j ) and max(X ，j ) are the minimum and maximum values ​​obtained by the jth sensor for all engines at all times.

4. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 1, characterized in that: In step S2: the network model of BiLSTM-GRU-Attention based on multiple optimizations specifically includes: S2-1. Construct the BiLSTM layer and abstract its parameters into particles in motion and optimize them using algorithms to obtain the optimal parameter combination of the BiLSTM layer hyperparameters. S2-2. Construct a GRU layer based on Bayesian parameter optimization to further extract complex time series features and enhance feature expression capabilities; S2-3. Construct an attention mechanism layer to perform weighted processing on the outputs of the BiLSTM and GRU layers. By calculating the importance weight of each time step, the most important features are selected to form a new feature representation. S2-4, the features processed by the attention layer are input to the fully connected layer for further dimension reduction, the remaining service life of the engine is output, and the Dropout layer is inserted to prevent overfitting. The Dropout value is set to r; S2-5. During the training process of predicting the fitting curve, the entire training data set is randomly divided into several small batches, each of which contains m samples. The model is trained with a learning rate of p. The training process of the FD001 sub-dataset is configured to run a epoch at most, and the training process of the FD003 sub-dataset is configured to run b epochs at most. S2-6. In the inference stage, the Dropout technique is used to quantify the uncertainty of the model's prediction results.

5. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 4, characterized in that: In step S2-1: construct a BiLSTM layer, abstract the parameters of the BiLSTM layer into particles in motion, and represent the position of particle i in one-dimensional space as vector X i =(x1, x2, ...x N ), the velocity is represented by vector V i =(v1, v2, ...v N ); Each particle has a target position X determined by the objective function, and can determine the current individual best position pbest and the current position x i , and can also determine the current group best position gbest, which is the best value in pbest; particles update their positions by tracking two extreme values, namely the individual best position pbest and the group best position gbest.

6. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 4, characterized in that: In step S2-1: the steps included are: S2-1-1, initialize a group of particles with a population size of N. The initialization information includes the random position and speed of the particles. The initial number of iterations T = 0, and prepare to enter the optimization process; S2-1-2, define the target position X of the particle, and evaluate the excellence of the current position, the individual best position and the group best position according to the target position. The closer the position is to the target position, the better it is. S2-1-3, the current position x of each particle i Compare it with the individual optimal position pbest it has passed. If the current position is better, take it as the current individual optimal position pbest; S2-1-4, the current position x of each particle i Compare it with the group optimal position gbest it has passed. If the current position is better, take it as the current group optimal position gbest; S2-1-5. Adjust particle speed and position according to the following formula: v i ′=ω×v i +c1×rand()×(pbest i -x i )+c2×rand()×(gbest i -x i ); x i ′=x i +v i Δt; Among them, v i ′ is the velocity of the particle after iterative update, v i is the current velocity of the particle, x i ′ is the position value of the particle after iterative update, x i is the current position value of the particle, Δt is the unit time interval, ω is the inertia factor, i=1, 2, 3, ..., N, N is the total number of particles in this group, rand() is a random number between (0, 1), pbest i is the current individual optimal position, gbest i is the optimal position of the current group, c1 and c2 are learning factors; The inertia factor ω is a dynamically changing value, and the changing rule is: Among them, G k represents the maximum number of iterations, ω ini represents the initial inertia weight, ω gnd Represents the inertia weight when iterating to the maximum evolutionary generation; S2-1-6: If the iteration termination condition is not reached, continue to execute step S2-1-2, and set the number of iterations T = T + 1, until the number of iterations T = G k The iteration is terminated, and the optimal parameter combination of the BiLSTM layer corresponding to the optimal position gbest of the group is output.

7. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 4, characterized in that: In step S2-2: the following steps are included: S2-2-1, according to formula r t =σ(W r ·[h t-1 , χ t ]+b r ) Calculate the reset gate of GRU, where the hidden state of the previous moment is h t-1 , χ t is the input of the current time step, σ represents the Sigmoid activation function, W r and b r They are the weight matrix and bias vector of the forget gate respectively; S2-2-2, according to formula z t =σ(W z ·[h t-1 , χ t ]+b z ) Calculate the update gate of GRU, where W z and b z In particular, reset the weights and biases of the gates; S2-2-3, according to the formula Calculate the candidate hidden state, where W and b are the weight and bias of the candidate hidden state, respectively, and tanh is the hyperbolic tangent activation function; S2-2-4. Define the hyperparameter space, determine the hyperparameters to be tuned for the GRU layer, and set a reasonable search range for each hyperparameter; S2-2-5. Introduce the Bayesian optimization algorithm and use the mean square error as the objective function: S2-2-6. Surrogate model fitting: Use Gaussian process to fit the objective function. The fitting relationship is: Where μ is the mean function; K is the covariance function; D is the evaluated hyperparameter combination; S2-2-7. Bayesian optimization selects the next hyperparameter combination for evaluation by maximizing expectation: Among them, EI(χ) represents the maximization expectation, χ + is the objective function value corresponding to the current optimal hyperparameter point, is the predicted value of the new hyperparameter combination on the objective function; S2-2-8. Update the proxy model, use the new evaluation point to update the Gaussian process proxy model, and iterate continuously until the preset number of evaluations is reached or the convergence condition is met.

8. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 4, characterized in that: In step S2-3: the following steps are included: S2-3-1. According to the formula h1=W1·X+b1, the linear layer of the attention mechanism layer is established, and the score of the Attention mechanism is calculated using linear transformation, where X=[X1, X2, …, X T ] is the input sample, h1 is the output of the linear layer, which is the attention score, W1 and b1 represent the weight and bias of the linear layer; S2-3-2, according to formula W x = Softmax(h1) to establish the weight generation layer of the attention mechanism layer, where W x Represents the generated weights. The calculation function uses the Softmax() function to normalize the attention scores to ensure that the sum of the weights in the time dimension is 1. S2-3-3, according to formula X att =W x ⊙X weights the initial input sample element by element with the generated weights, where X att To finally generate the output of the attention mechanism layer, ⊙ represents the element-by-element multiplication operation.

9. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 4, characterized in that: In step S2-6: the following steps are included: S2-6-1. In the inference phase, the activation state of the Dropout layer is maintained for each input sample. During each forward propagation, the Dropout layer randomly selects neurons and discards them to form a new sub-model structure. S2-6-2. Perform forward propagation for each input sample, using a different random inactivation mode each time to obtain the predicted value Where N represents the number of forward propagations; S2-6-3. Basic formula Calculate the predicted mean as the final life expectancy prediction value, where Mean Prediction represents the average prediction result under the model structure. is the output result of the i-th forward propagation; S2-6-4, according to the formula The variance is calculated to quantify the uncertainty of the prediction process, where uncertainty represents the degree of variation in the prediction results under different model structures.

10. The method for monitoring the life of an aircraft engine based on artificial intelligence according to claim 1, characterized in that: In step S4: the training set is used to train the built prediction model, and the trained model is used to infer the validation set. The model performance is verified by two evaluation indicators, the root mean square error RMSE and the score function Score, and the performance of the best model is evaluated using the test set. The formulas of the two evaluation indicators are as follows: in, is the predicted value of the model, y i is the actual value, n is the number of engines in the test set, S is the calculated score, and h = R prediction -R actual , R prediction is the predicted value of the remaining useful life, R actual is the actual value of the remaining useful life.

Images