An intelligent prediction method for key state parameters of aviation equipment
By constructing a deep learning network consisting of CNN, Bi-LSTM, and Attention, the problem of nonlinear prediction of EGT in aero-engines was solved, achieving more accurate state parameter prediction and improving the maintenance and management level of aero-equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAVAL AVIATION UNIV
- Filing Date
- 2022-12-30
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to accurately predict the nonlinear changes in exhaust gas temperature (EGT) of aircraft engines, impacting the precision of engine health management.
A deep learning network consisting of a convolutional neural network (CNN), a bidirectional long short-term memory network (Bi-LSTM), and an attention mechanism is used. Combined with data preprocessing and model training, a nonlinear mapping model between key state parameters and multivariate parameters is established, and prediction is made using historical data of aircraft.
It improved the predictive accuracy of EGT, enhanced the preventive maintenance and support capabilities of aviation equipment, and improved the accuracy of health management.
Smart Images
Figure CN115936254B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aviation flight data analysis technology, and in particular to an intelligent prediction method for aviation equipment status parameters. Background Technology
[0002] Aviation equipment such as aero engines and flight control systems are crucial subsystems of aircraft, and their operational status significantly impacts flight safety. Key parameters indicating the condition of these subsystems are influenced by multiple parameters. For example, the exhaust gas temperature (EGT) of an aero engine is a vital parameter for monitoring its health; its changes can effectively reflect the engine's operating condition and performance degradation. EGT prediction allows for early detection of engine health conditions, providing data support for preventative maintenance. However, the change in EGT over time is a complex, non-linear process, difficult to describe with precise physical models. Furthermore, many factors contribute to EGT changes, resulting from the combined effects of various engine components. Therefore, accurately predicting EGT has become a hot research topic in the field of aero engine health management. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an intelligent prediction method for key performance parameters of aviation equipment. This method uses historical flight data of aircraft to train an algorithm to establish a nonlinear mapping model between key state parameters and multivariate parameters, and uses the constructed model to predict key state parameters.
[0004] The present invention provides an intelligent prediction method for the state parameters of aviation equipment, characterized by the following steps:
[0005] 1. Data Preprocessing Stage
[0006] Complete outlier correction, missing data filling, data normalization, and sliding window setup;
[0007] The specific steps for outlier correction are as follows: Analyze the statistical data distribution to identify outliers. Since time series data have a cumulative relationship over time, the outliers are corrected to the median of their nearest neighboring data.
[0008] The specific steps for filling the missing value are as follows: for a missing value, the median of its adjacent values is used to fill it;
[0009] The specific steps of the data normalization process are as follows: Since the numerical ranges of multiple observed variables such as engine speed, lubricating oil temperature, lubricating oil pressure and exhaust temperature are different, if the data volume is different, it may cause the trained network model to be inaccurate. Therefore, multiple observed variables need to be normalized. The three common data normalization methods are Z-score, Min-max and MaxAbs normalization methods.
[0010] The specific steps for setting up a sliding window are as follows: Setting up a sliding window involves dividing the sample data into multiple time series data windows with a window width of w and a step size of s.
[0011] 2. Model Training Phase
[0012] The training process utilizes a model composed of Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), and an Attention mechanism for weight allocation. CNN extracts features from a large amount of data, Bi-LSTM extracts temporal correlation patterns from these features to achieve data memory prediction, the Attention-Permute layer adjusts the data dimensionality to provide a suitable data structure for the next layer's computation, and the Attention-Merge layer calculates the Attention value to correct network weights and biases, thereby accurately constructing a nonlinear mapping model between predicted parameters and input multivariate parameters. To prevent overfitting, the Dropout layer discards some weights and biases, preventing them from participating in network iteration and updates. The Dense layer performs nonlinear transformations on the features extracted by previous layers to establish correlations between features, and the Flatten layer flattens the multidimensional data to provide a suitable data structure for the next layer's computation.
[0013] To improve prediction accuracy, the prediction process involves training three network models: CNN, Bi-LSTM, and Attention.
[0014] Specifically, the CNN is a deep feedforward neural network composed of convolutional layers, pooling layers, and fully connected layers. Convolution is used instead of fully connected layers to extract local features of the data. The convolution formula (1) with a convolution depth of D and a kernel size of F is:
[0015]
[0016] Among them, a i,j w is the element in the i-th row and j-th column of the feature matrix after convolution. d,m,n Let x be the weight of the convolution kernel in the m-th row and n-th column of the d-th layer. d,i+m,j+nw represents the element in the d-th layer, i-th row, j-th column of the matrix before convolution. b Let f be the bias term of the convolution kernel and f be the activation function. Each neuron in a layer connects to some neurons in the previous layer using the same convolution kernel. Feature extraction is achieved through the local connectivity and weight sharing of the convolutional network.
[0017] Based on the mathematical principles of CNNs, convolutional layers extract the feature space influencing key state parameters from flight data containing multiple observation parameters of aviation equipment through convolution operations. The convolution kernel is multiplied and summed with the corresponding data from the observed variables to obtain a single value in the feature space. Then, a sliding window is set according to the stride to move the observed variables, and after each movement, the value is multiplied again with the corresponding element of the convolution kernel to calculate the corresponding feature value. This process is repeated to obtain the convolutional feature space. To prevent overfitting, a Dropout layer is used to discard some weights and biases, preventing them from participating in network iteration and updates. The network output is then fed into a Bi-LSTM.
[0018] Specifically, in order to make full use of the contextual information of the neurons in the preceding and following nodes of the network and improve the prediction accuracy, the internal node neurons of the Bi-LSTM are the same as those of the LSTM, and three gates are constructed—the forget gate f, the input gate i, and the output gate o;
[0019] At time step t, the input vector of the LSTM hidden layer is x. t The output vector is h t The memory unit is c t Among them, the forget gate formula (2) is:
[0020] f t =σ(W xf x t +W hf h t-1 +b f (2)
[0021] Input gate formula (3):
[0022] i t =σ(W xi x t +W hi h t-1 +b i (3)
[0023]
[0024]
[0025] Output gate formula (4):
[0026] o t =σ(W xox t +W ho h t-1 +b o (4)
[0027]
[0028] in This represents the element-wise multiplication of a vector; W represents the weight. σ It is the sigmoid function, b is the bias term, and tanh is the hyperbolic tangent function;
[0029] The Bi-LSTM is based on the forward layer calculation of LSTM with the addition of the reverse layer calculation. The calculation method of each node is as shown in formulas (5)-(7).
[0030]
[0031]
[0032] h t =f(w4h) t R +w2h t L (7)
[0033] The Attention is the probability of selecting the i-th piece of information from information X given a query vector q, calculated by formula (8):
[0034]
[0035] Wherein, s(x i ,q) is the attention scoring function, which can be an additive, dot product, scaled dot product, or bilinear function. Probability α i The resulting probability vector is called the attention distribution, representing the relevance of the i-th piece of information in the input information vector X to the query q. Let probability α... i The input information is summarized using a weighted average, resulting in the Attention value formula (9):
[0036]
[0037] 3. Prediction and Visualization Stage
[0038] Using the trained network model, key state parameters are predicted. Then, the key state parameters are denormalized to obtain the predicted values, and their visualization curves are plotted.
[0039] This invention presents an intelligent prediction method for aviation equipment state parameters. It employs a deep learning network combining multiple layers: a feature extraction layer, a data memory layer, and an attention layer. The network is trained using historical flight training data from the aircraft to establish a mapping relationship between multiple state parameter information and prediction parameters. A convolutional neural network is introduced to extract features from diverse feature parameters. A long short-term memory network is used to effectively remember the time accumulation effect of flight data, and an attention mechanism is added to adjust network weights in real time, thereby improving prediction accuracy. The proposed intelligent prediction method has significant application value in the preventive maintenance and health management of important subsystems of aviation equipment, such as aero-engines and flight control systems, and is of great importance in improving the preventive maintenance capabilities of aviation equipment. Attached Figure Description
[0040] Figure 1 A flowchart of an intelligent prediction method for the state parameters of aviation equipment;
[0041] Figure 2 This is a schematic diagram of a sliding window;
[0042] Figure 3 This is a schematic diagram of sliding window prediction;
[0043] Figure 4 This is a diagram of the CNN architecture;
[0044] Figure 5 It uses an LSTM network structure;
[0045] Figure 6 It is a Bi-LSTM network structure;
[0046] Figure 7 A comparison chart of five MAPE models;
[0047] Figure 8 This is a comparison chart of the predicted and actual values for five models at 20 sampling time points. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] Example 1
[0050] This embodiment provides an intelligent prediction method for the state parameters of aviation equipment. (See appendix.) Figure 1-6 This includes the following steps:
[0051] 1. Data Preprocessing Stage
[0052] Complete outlier correction, missing data filling, data normalization, and sliding window setup;
[0053] The specific steps for outlier correction are as follows: Analyze the statistical data distribution to identify outliers. Since time series data have a cumulative relationship over time, the outliers are corrected to the median of their nearest neighboring data.
[0054] The specific steps for filling the missing value are as follows: for a missing value, the median of its adjacent values is used to fill it;
[0055] The specific steps of the data normalization process are as follows: select one of the three common data normalization methods, Z-score, Min-max and MaxAbs, for normalization;
[0056] The calculation methods and applicable scope of the three methods are shown in Table 1.
[0057] Table 1 Comparison of Data Standardization Methods
[0058]
[0059] The specific steps for setting up a sliding window are as follows: Setting up a sliding window involves dividing the sample data into multiple time series data windows with a window width of w and a step size of s. (Appendix) Figure 2 The sample data from the six time points were divided into three time series data windows with a window width of 4 and a step size of 1.
[0060] 2. Model Training Phase
[0061] The prediction model takes data from each window as input and outputs the predicted data for the next time step, as shown in the attached figure. Figure 3 As shown, input the data from the first four time points, and use the model to predict the data from the fifth time point.
[0062] The prediction model consists of a Convolutional Neural Network (CNN), a Bidirectional Long Short-Term Memory (Bi-LSTM) network, and an Attention mechanism weight allocation network. The CNN extracts features from a large amount of data, the Bi-LSTM extracts temporal correlation patterns from these features to achieve data memory prediction, the Attention-Permute layer adjusts the data dimensionality to provide a suitable data structure for the next layer's computation, and the Attention-Merge layer calculates the Attention value to correct the network weights and biases, thereby accurately constructing a nonlinear mapping model between the prediction parameters and the input multivariate parameters. To prevent overfitting, the Dropout layer discards some weights and biases, preventing them from participating in network iteration and updates. The Dense layer performs nonlinear transformations on the features extracted by previous layers to establish correlations between features, and the Flatten layer flattens the multidimensional data to provide a suitable data structure for the next layer's computation.
[0063] To improve prediction accuracy, the prediction process involves training three network models: CNN, Bi-LSTM, and Attention.
[0064] Specifically, the CNN is a deep feedforward neural network composed of convolutional layers, pooling layers, and fully connected layers. Convolution is used instead of fully connected layers to extract local features of the data. The convolution formula (1) with a convolution depth of D and a kernel size of F is:
[0065]
[0066] Among them, a i,j w is the element in the i-th row and j-th column of the feature matrix after convolution. d,m,n Let x be the weight of the convolution kernel in the m-th row and n-th column of the d-th layer. d,i+m,j+n w represents the element in the d-th layer, i-th row, j-th column of the matrix before convolution. b Let f be the bias term of the convolution kernel and f be the activation function. Each neuron in a layer connects to some neurons in the previous layer using the same convolution kernel. Feature extraction is achieved through the local connectivity and weight sharing of the convolutional network.
[0067] Based on the mathematical principles of CNNs, convolutional layers extract the feature space that influences the values of key state parameters from flight data containing multiple observation parameters of aviation equipment through convolution operations, as shown in the attached figure. Figure 4As shown, the convolution kernel is multiplied and summed with the corresponding data of the observed variable to obtain a single value in the feature space. Then, a sliding window is set according to the stride to move the observed variable. After moving, the kernel element is multiplied again to calculate the corresponding feature value. This process is repeated to obtain the convolutional feature space. To prevent overfitting, a Dropout layer is used to discard some weights and biases, preventing them from participating in the network iteration calculation and update. The network output is then fed into a Bi-LSTM.
[0068] Specifically, to fully utilize the contextual information of neurons in the preceding and following nodes and improve prediction accuracy, the internal neurons of the Bi-LSTM are the same as those of the LSTM, and three gates are constructed—the forget gate f, the input gate i, and the output gate o; (Appendix) Figure 5 The diagram shows the internal node structure of an LSTM.
[0069] At time step t, the input vector of the LSTM hidden layer is x. t The output vector is h t The memory unit is c t Among them, the forgetting gate formula (2) is:
[0070] f t =σ(W xf x t +W hf h t-1 +b f (2)
[0071] Input gate formula (3):
[0072] i t =σ(W xi x t +W hi h t-1 +b i (3)
[0073]
[0074]
[0075] Output gate formula (4):
[0076] o t =σ(W xo x t +W ho h t-1 +b o (4)
[0077]
[0078] in σ represents element-wise multiplication of a vector; W represents weights; σ is the sigmoid function, b is the bias term, and tanh is the hyperbolic tangent function.
[0079] The Bi-LSTM is based on the forward layer computation of LSTM, with the addition of a backward layer computation, as shown in the attached figure. Figure 6 As shown, the calculation methods for each node are as shown in formulas (5)-(7).
[0080]
[0081]
[0082] h t =f(w4h) t R +w2h t L (7)
[0083] The Attention is the probability of selecting the i-th piece of information from information X given a query vector q, calculated by the formula (8) as follows:
[0084]
[0085] Wherein, s(x i ,q) is the attention scoring function, which can be an additive, dot product, scaled dot product, or bilinear function. Probability α i The resulting probability vector is called the attention distribution, representing the relevance of the i-th piece of information in the input information vector X to the query q. Let probability α... i The input information is summarized using a weighted average, resulting in the Attention value formula (9):
[0086]
[0087] 3. Prediction and Visualization Stage
[0088] Using the trained network model, key state parameters are predicted. Then, the key state parameters are denormalized to obtain the predicted values, and their visualization curves are plotted.
[0089] Example 2
[0090] This embodiment uses eight observation parameters—exhaust temperature, engine speed, lubricating oil pressure, air box pressure, lubricating oil temperature, cylinder head temperature, air box temperature, and throttle displacement—as input to the training network. The exhaust temperature of the aero-engine after five time steps is used as the predicted output parameter. The network model constructs a mapping model between future exhaust temperature and the eight parameters. 13,157 flight data samples from typical flights of a certain type of aero-engine were selected from historical flight data, with 90% used as training samples and 10% as test samples. The model predicts the trend of parameter changes and monitors the engine exhaust temperature status.
[0091] Includes the following steps:
[0092] 1. Data Preprocessing Stage
[0093] Complete outlier correction, missing data filling, data normalization, and sliding window setup;
[0094] The specific steps for outlier correction are as follows: Analyze the statistical data distribution to identify outliers. Since time series data have a cumulative relationship over time, the outliers are corrected to the median of their nearest neighboring data.
[0095] The specific steps for filling the missing value are as follows: for a missing value, the median of its adjacent values is used to fill it;
[0096] The specific steps of the data normalization process are as follows: Since the numerical ranges of multiple observed variables such as engine speed, lubricating oil temperature, lubricating oil pressure and exhaust temperature are different, if the data volume is different, it may cause the trained network model to be inaccurate. Therefore, multiple observed variables need to be normalized, and Min-max is used to normalize them.
[0097] 2. Model Training Phase
[0098] The predictive model aims to establish a mapping relationship between each window of data and its EGT value at the next time step, as shown in the attached figure. Figure 3 As shown in Table 2, to improve prediction accuracy, the prediction model training includes training three network models: CNN, Bi-LSTM, and Attention. The parameter settings for each network layer are shown in Table 2.
[0099] Table 2 Network Parameter Settings
[0100]
[0101] Depending on the iteration number EPOCHS, the five model prediction performance metrics, such as MAPE, are used. Figure 7As shown in the MAPE comparison chart, when the number of iterations (EPOCHS) is 10, Bi-LSTM has the smallest MAPE value among the five models. When the EPOCHS is 20 and 30, CNN has the smallest MAPE value among the five models. In these three cases, the MAPE value of the Attention-CNN-Bi-LSTM model proposed in this paper is not the smallest. However, when the EPOCHS changes from 40 to 100, the MAPE value of the proposed model is the smallest among the five models. When the number of iterations is 90, the MAPE value of the Attention-CNN-Bi-LSTM model decreases from 3.2% of Bi-LSTM, 3.9% of CNN, 5.1% of CNN-Bi-LSTM, and 4.1% of CNN-Attention to 1.2%.
[0102] 3. Prediction Visualization Stage
[0103] Using a trained network model, flight data test samples of aviation equipment are read to predict key state parameters. Then, the key state parameters are denormalized to obtain the predicted values, and their visualization curves are plotted.
[0104] Experimental sample data from aero-engines were used to predict data using Bi-LSTM, CNN, CNN-Bi-LSTM, CNN-Attention, and Attention-CNN-Bi-LSTM models, respectively. The network learning rate was set to 0.001, and the number of epochs was increased from 10 to 100 in increments of 10. The prediction results of the five models at 20 time points were obtained as follows: Figure 8 As shown in the figure, the proposed Attention-CNN-Bi-LSTM model has a better EGT curve fitting effect than Bi-LSTM, CNN, and CNN-Bi-LSTM, and its predicted values are closer to the true values.
[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for intelligent prediction of key state parameters of aviation equipment, characterized in that... Includes the following steps: 1) Data preprocessing stage Complete outlier correction, missing data filling, data normalization, dataset partitioning, and sliding window setup; 2) Model training phase The prediction model maps each window of data to its EGT value at the next time step. It trains a model consisting of a Convolutional Feature Extraction Network (CNN), a Bidirectional Long Short-Term Memory (Bi-LSTM) network, and an Attention mechanism weight allocation network. The CNN extracts features from a large amount of time-varying parameters, the Bi-LSTM extracts temporal correlation patterns from these features to achieve data memory prediction, the Attention-Permute layer adjusts the data dimensionality to provide a suitable data structure for the next layer's computation, and the Attention-Merge layer calculates the Attention value, correcting the network weights and biases, thereby accurately constructing the relationship between the predicted parameters and the multivariate input parameters. A nonlinear mapping model is proposed. This model takes eight observation parameters (exhaust temperature, engine speed, lubricating oil pressure, air box pressure, lubricating oil temperature, cylinder head temperature, air box temperature, and throttle displacement) as input, and outputs the exhaust temperature after multiple time steps. It constructs a nonlinear mapping relationship between the future exhaust temperature and these eight observation parameters. To prevent overfitting, the Dropout layer discards some weights and biases, preventing them from participating in network iteration and updates. The Dense layer performs nonlinear transformations on the features extracted by previous layers to establish correlations between features. The Flatten layer transforms multidimensional data into one dimension, providing a suitable data structure for the next layer's calculations. 3) Prediction visualization stage Using the trained network model, key state parameters are predicted, and then the key state parameters are denormalized to obtain the predicted values, and their visualization curves are plotted. In the data preprocessing stage mentioned in 1): The specific steps for outlier correction are as follows: Analyze the statistical data distribution to identify outliers. Since time series data have a cumulative relationship over time, the outliers are corrected to the median of their nearest neighboring data. The specific steps for filling the missing value are as follows: for a missing value, the median of its adjacent values is used to fill it; The specific steps of the data normalization process are as follows: select one of the three common data normalization methods, Z-score, Min-max and MaxAbs, for normalization. The specific steps for setting up a sliding window are as follows: setting up a sliding window involves dividing the sample data into multiple time series data windows with a window width of w and a step size of s. In step 2) of the model training phase: The CNN is a deep feedforward neural network composed of convolutional layers, pooling layers, and fully connected layers. Convolution is used instead of fully connected layers to extract local features of the data. The convolution formula (1) is as follows: Among them, a i,j w is the element in the i-th row and j-th column of the feature matrix after convolution. d,m,n Let x be the weight of the convolution kernel in the m-th row and n-th column of the d-th layer. d,i+m,j+n w represents the element in the d-th layer, i-th row, j-th column of the matrix before convolution. b The bias term of the convolution kernel is denoted by f, and the activation function is denoted by f. Each neuron in the layer uses the same convolution kernel to connect with some neurons in the previous layer. Feature extraction is achieved through the local connectivity and weight sharing of the convolutional network. Based on the mathematical principles of CNN, the convolutional layer extracts the feature space that affects the values of key state parameters from the flight data of multi-dimensional observation parameters of aviation equipment through convolution operations. The convolution kernel is multiplied and added with the data at the corresponding position of the observation parameter to obtain one value of the feature space. Then, according to the stride, the observation variable is moved in a sliding window. After moving, it is multiplied with the corresponding element of the convolution kernel to calculate the corresponding feature value. This process is repeated to obtain the convolutional feature space. To fully utilize the contextual information of neurons at the beginning and end of the network and improve prediction accuracy, the internal neurons of the Bi-LSTM are the same as those of the LSTM, and three gates are constructed: the forget gate f, the input gate i, and the output gate o. At time step t, the input vector of the LSTM hidden layer is x. t The output vector is h t The memory unit is c t Among them, the forgetting gate formula (2) is: f t =σ(W xf x t +W hf h t-1 +b f ) (2) Input gate formula (3): i t =σ(W xi x t +W hi h t-1 +b i ) (3) Output gate formula (4): o t =σ(W xo x t +W ho h t-1 +b o ) (4) in σ represents element-wise multiplication of a vector; W represents weights; σ is the sigmoid function, b is the bias term, and tanh is the hyperbolic tangent function. The Bi-LSTM is based on the forward layer calculation of LSTM with the addition of the reverse layer calculation. The calculation method of each node is as shown in formulas (5)-(7). The Attention is the probability of selecting the i-th piece of information from information X given a query vector q, calculated by the formula (8) as follows: Wherein, s(x i (q) represents the attention scoring function, which can be an additive, dot product, scaled dot product, or bilinear function, with probability α. i The resulting probability vector is called the attention distribution, representing the relevance of the i-th piece of information in the input information vector X to the query q; the probability α i The input information is summarized using a weighted average, resulting in the Attention value formula (9):