Method for predicting residual service life of aero-engine based on GRU-CNN residual network model

Abstract

The invention relates to an aero-engine residual service life prediction method based on a GRU-CNN residual network model, and the method comprises the following steps: obtaining multi-source sensing data of an aero-engine, carrying out the preprocessing and feature selection processing, obtaining a data set, dividing the data set into a training set, a verification set and a test set, and determining an RUL label of each sample; according to the method, a GRU-CNN residual network model is constructed, the GRU-CNN residual network model comprises a GRU layer 1, a GRU layer 2, a convolution layer, a pooling layer and a full connection layer, data is input into the GRU layer 1 and then enters the GRU layer 2, after residual connection is carried out on output of the GRU layer 2 and input of the GRU layer 1, dimension adjustment is carried out, the data sequentially passes through the convolution layer, a PReLU activation function, the pooling layer, flattening operation and the full connection layer, and a prediction result is output; and a trained GRU-CNN residual network model is obtained and is used for predicting the residual service life of the aero-engine. According to the method, the training stability and the calculation efficiency of the model are remarkably improved, and the accuracy and real-time requirements can be considered.

Description

Technical Field

[0001] The present invention relates to the field of aero-engines, and in particular to a method for predicting the remaining useful life of an aero-engine based on a GRU-CNN residual network model. Background Art

[0002] In order to improve the efficiency of aircraft engine maintenance, reduce the waste of component resources, reduce maintenance costs, and effectively ensure the reliability of engine operation, developed countries in aviation have carried out research and development of aircraft engine prognostics and health management (PHM) systems. The PHM system combines the latest research results in the fields of information technology and artificial intelligence. It mainly realizes the functions of aircraft engine fault diagnosis and isolation, real-time system performance monitoring and fault prediction, engine remaining service life prediction and health management through the mining of sensor monitoring data and the management of extracted feature information.

[0003] Remaining useful life prediction is the core technology of the PHM system, a key link between fault prediction and health management, and a core component of modern industrial intelligent manufacturing. Accurate RUL prediction technology can determine the state of engine performance degradation before the engine fails and issue failure warnings in advance, so as to carry out timely maintenance, reduce the probability of failure, change regular maintenance to proactive maintenance, and avoid accidents caused by untimely maintenance and waste caused by excessive maintenance in traditional maintenance methods. With the rapid development of PHM technology in the aviation field, technological progress has expanded from the remaining useful life prediction of single aircraft components to the prediction of the usable life of the entire aircraft engine. How to achieve accurate RUL prediction has become a major research hotspot in recent years.

[0004] The principle of RUL prediction is to evaluate the remaining service life of a component or the entire system based on the current health status, working links and loads, status monitoring information, etc. of the equipment, combined with physical failure models, historical performance degradation data, fault diagnosis information, etc.

[0005] In the field of aircraft engine remaining useful life (RUL) prediction, deep learning technology, especially convolutional neural network (CNN) and gated recurrent unit (GRU), has attracted widespread attention due to its powerful data processing and feature extraction capabilities. However, traditional CNN-based RUL prediction methods have some inherent limitations. The convolution kernel parameters of CNN are shared for all input data during the training process, which may limit the model's ability to deeply mine the characteristics of engine multi-source sensor data. In addition, although the learning ability of the model can be improved by expanding the number of convolution kernel channels or increasing the number of convolution layers, this will also lead to a significant increase in the number of model parameters, which may cause delays in real-time calculations of the model and affect its performance and efficiency in actual deployment. At the same time, a single GRU model may lack the ability to efficiently capture spatial features, especially when facing high-dimensional sensor data. For example, the aircraft engine remaining life prediction method based on a parallel CNN model disclosed in Chinese patent CN112100865A has advantages in extracting spatial features of multi-dimensional sensor data, but has inherent deficiencies in processing time series data and cannot capture the time dependency in the data. In addition, although simply increasing the number of CNN layers to improve model capabilities is effective, it will significantly increase the computational cost and affect the model's real-time prediction capabilities. For example, Wang Wenqing et al. (Wang Wenqing, Guo Heng, Fan Qifu. Prediction of Remaining Life of Aircraft Engines Based on CNN and GRU [C] / / Control Theory Committee of the Chinese Automation Society. Proceedings of the 37th China Control Conference (F). Key Laboratory of System Control and Information Processing, Ministry of Education, Department of Automation, Shanghai Jiaotong University;, 2018:6.) used CNN and GRU for RUL prediction. The network structure is that the sensor data enters the multi-layer CNN, and the CNN extracts the abstract feature sequence as the input of the GRU, and then connects several layers of CNN for filtering. Finally, the remaining life prediction result is obtained through a linear regressor. Although its RMSE and SCORE indicators are significantly improved, its generalization ability needs to be further improved. Summary of the invention

[0006] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a method for predicting the remaining useful life of an aircraft engine based on a GRU-CNN residual network model. The method can effectively combine the advantages of the two networks, GRU and CNN, comprehensively extract spatial and temporal features, significantly improve the training stability and computational efficiency of the model, overcome the limitations of the existing methods in feature extraction and computational complexity, and take into account both accuracy and real-time requirements.

[0007] The technical solution adopted by the present invention to solve the technical problem is: A method for predicting the remaining useful life of an aircraft engine based on a GRU-CNN residual network model, the prediction method comprising the following contents: Acquire multi-source sensor data of aircraft engines, perform preprocessing and feature selection to obtain a data set, divide the data set into a training set, a validation set, and a test set, establish a mapping relationship between the monitoring variable X and the remaining useful life RUL, and determine the RUL label of each sample; Construct the GRU-CNN residual network model, The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolution layer, pooling layer and fully connected layer. Data is input into GRU layer 1 and then into GRU layer 2. The output of GRU layer 2 is residually connected with the input of GRU layer 1, and then passes through the convolution layer, PReLU activation function, pooling layer, flattening operation and fully connected layer in sequence after dimension adjustment to output the prediction result. The GRU-CNN residual network model is trained using the data set, and the loss function MSE during training is:

[0008] In the formula, represents the predicted value of the i-th engine and the true value The difference between , b is the number of samples in the mini-batch used for training; Use RMSE and score functions to evaluate the performance of the model; The trained GRU-CNN residual network model is obtained to predict the remaining useful life of aircraft engines.

[0009] Furthermore, the process of preprocessing and feature selection of multi-source sensor data of aircraft engines is as follows: Step 1..1: Obtain historical aircraft engine failure data and form a data set ,in, represents the total running track length of N aircraft engine samples; N represents the number of aircraft engine samples; represents the monitoring trajectory length of the Nth aircraft engine sample; V represents the number of sensors in the aircraft engine; Step 1..2: Visualize the V different types of sensor signals collected, analyze the data change trend, and identify the variables that do not show obvious changes during the engine life cycle based on the visualization results; Step 1.3: Based on the observation results, remove those features that remain constant throughout the life cycle, perform feature selection on V monitoring variables, and obtain F monitoring variables. The original data set is reduced to ; The F monitoring variables are standardized according to the "minimum-maximum" method to obtain the data set.

[0010] Furthermore, the score function calculation formula is:

[0011] In the formula, represents the predicted value of the i-th engine and the true value The difference is , and P is the total number of engines for predicting RUL.

[0012] Furthermore, the hidden unit dimension of GRU layer 1 is 64, the hidden unit dimension of GRU layer 2 is the same as the original input feature dimension, the convolution kernel size of the convolution layer is 4, the step size is 2, the number of output channels of the convolution layer is 100, and the padding size of the convolution layer is 2; the pooling kernel size of the pooling layer is 3; the learning rate is 0.007, the dropout ratio of GRU layer 1 is 0.2, and the dropout ratio of GRU layer 2 is 0.3.

[0013] Furthermore, the SCORE value was controlled within 260 and the RMSE value was controlled within 20.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The model of the present invention shows excellent performance. Compared with the traditional GRU, CNN and bidirectional LSTM models, the prediction accuracy of the present invention is greatly improved. In terms of the two key performance indicators of RMSE and SCORE, the model of the present invention achieved excellent performance of 22.013 and 257.114 respectively, significantly better than GRU (RMSE 33.594, SCORE 1511.451), CNN (RMSE27.196, SCORE 1364.601) and bidirectional LSTM (RMSE 34.594, SCORE 1910.125). These data show that the present invention not only greatly reduces the root mean square error (RMSE) and improves the prediction accuracy, but also effectively reduces the error caused by overestimation or underestimation of the prediction through the SCORE scoring function, further enhancing the robustness of the model.

[0015] By optimizing the feature fusion of CNN and GRU, this application can comprehensively extract the spatial and temporal features in the sensor data. This integrated innovation significantly improves the prediction performance of the model. The PReLU function is used for lightweight model design, which can more flexibly adapt to different data distributions, improve the applicability of the model, and further improve the model performance.

[0016] In summary, the GRU-CNN residual neural network structure of the present invention can perform more accurate remaining service life prediction in complex, high-dimensional aircraft engine sensor data, and is applicable to a variety of different working conditions, with significant engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 C-MAPSS simplified engine simulation model; Figure 2 Visualization of monitoring indicators; Figure 3 Schematic diagram of the GRU-CNN residual network model architecture; Figure 4 Visualization of the GRU-CNN residual network model structure; Figure 5 Training curve: comparison chart of training loss and validation loss; Figure 6 Comparison chart of model predicted RUL and actual RUL (training set samples); Figure 7 Comparison chart of predicted RUL and actual RUL; Figure 8 Comparison chart of predictions and actual RUL sorted by actual RUL. DETAILED DESCRIPTION

[0018] The present invention is further explained below in conjunction with the embodiments and drawings, but this is not intended to limit the scope of protection of the present application.

[0019] Accurate prediction of the remaining useful life (RUL) of aircraft engines requires extracting effective degradation feature information from complex multi-source sensor data. These multi-source sensor data are usually high-dimensional and high-noise, making feature extraction a challenge. The GRU-CNN residual network model proposed in this paper not only improves the accuracy and efficiency of RUL prediction, but also enhances the adaptability of the model to different data sets, providing a powerful tool for aircraft engine health management and fault prediction maintenance systems. Its main contributions are reflected in the following three aspects: (1) Time series feature extraction: First, the GRU network is used to conduct in-depth analysis of time series data to capture the dynamic characteristics and long-term dependencies in the process of aircraft engine performance degradation. (2) Spatial feature enhancement: Based on the GRU layer, the spatial features of the data are further extracted through the CNN layer to enhance the model's sensitivity to local changes, and the feature weights are learned by simulating the interaction between the channels of the convolutional features (the convolution operation and the subsequent pooling operation can be regarded as a certain weight distribution of spatial features. Through the learning of the convolution kernel, the model can adaptively enhance important spatial features and suppress irrelevant features. This learning is implicit because the convolution kernel will automatically adjust during the training process, thereby giving different "weights" to different spatial features.), to achieve feature enhancement and suppression. (3) Residual connection optimization: In the model, the input and output of the GRU network are connected with residual connections to improve the stability and efficiency of model training, while reducing the number of model parameters, avoiding overfitting, and improving the real-time computing performance and practicality of the model.

[0020] The invention discloses a method for predicting the remaining useful life of an aircraft engine based on a GRU-CNN residual network model, comprising the following steps: Step 1: Obtain multi-source sensor data of aircraft engines, perform preprocessing and feature selection to obtain a data set, establish a mapping relationship between the monitoring variable X and the remaining useful life RUL, and determine the RUL label of each sample.

[0021] In order to improve the accuracy of the remaining useful life (RUL) prediction, the multi-source sensor data of the engine is strictly preprocessed and feature selected. The specific process of data preprocessing and feature selection for aircraft engine monitoring data includes the following steps: Step 1..1: Obtain historical aircraft engine failure data and form a data set ,in, represents the total running trajectory length of N aircraft engine samples, N represents the number of aircraft engine samples, represents the monitoring trajectory length of the Nth aircraft engine sample, v=1, 2, , V represents the number of sensors in the aircraft engine.

[0022] Step 1..2: Visualize the V different types of sensor signals collected and analyze the trend of data changes. Based on the visualization results, identify the variables that do not show obvious changes during the engine life cycle. These variables cannot reflect the degradation process of engine performance due to their stability.

[0023] Step 1.3: Based on the observation results, remove those features that remain constant throughout the life cycle. Perform feature selection on V monitoring variables to obtain F monitoring variables. The original data set is reduced to ; After standardizing the F monitoring variables in a “minimum-maximum” manner, the data set is obtained. The calculation formula is as follows:

[0024] in, represents the original data of the nth engine signal j at the i-th moment, yes The standardized value, and Represent the maximum and minimum values ​​of the engine signal j respectively.

[0025] Step 1.4: Divide the data set into a training set, a test set, and a validation set; use the training set to obtain the mapping relationship between the monitoring variable X and the remaining useful life RUL, and determine the RUL label of each sample.

[0026] During the normal operation of the turbofan engine, the failure rate is constant and low, and the degradation is negligible; at the end of the equipment's use, the failure rate increases rapidly with time. The data monitored before the equipment begins to rapidly decline is not suitable for model training, so a mutation threshold is set.

[0027] Based on the set mutation threshold, the RUL label in the training set is set as a piecewise linear function. Use mathematical formulas to express the relationship between RUL and the monitoring variables in the training set. Assume that t represents the current operation cycle of the device, T represents the maximum operation cycle of the device, and R early Indicates the mutation threshold set according to the situation.

[0028]

[0029] Step 2: Build the GRU-CNN residual network model. The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolution layer, pooling layer and fully connected layer. Data is input into GRU layer 1 and then into GRU layer 2. The output of GRU layer 2 is residually connected with the input of GRU layer 1, and then passes through the convolution layer, PReLU activation function, pooling layer, flattening operation and fully connected layer in sequence to output the prediction result.

[0030] Step 2.1 GRU part Two Gated Recurrent Unit (GRU) layers are used, each with a hidden layer dimension of and The GRU layer is a recurrent neural network (RNN) that captures long-term dependencies in time series data through a gating mechanism.

[0031] Gating mechanism of GRU layer: Update gate: determines the state of the previous moment How to update to the current status , the formula is:

[0032] in, To update the gate, is the weight matrix, is the bias, is the sigmoid function.

[0033] Reset gate: control the state of the previous moment How to use the current input Combined, the formula is:

[0034] in, To reset the gate, are the weight matrices and biases.

[0035] New state: Calculate the candidate hidden state at the current moment , the formula is:

[0036] in, represents element-wise multiplication, , .

[0037] Final hidden state: Combine the update gate and the candidate state to get the final hidden state:

[0038] Therefore, in GRU layer 1, given an input X, the output of the GRU is:

[0039] In GRU layer 2, the output is:

[0040] in, and They are the parameters of the two-layer GRU, parameters and It usually includes all weights and biases inside the GRU unit, which also covers the relevant parameters of the activation function. The activation function uses the PReLU activation function.

[0041] Step 2.2 Residual Connection Adding the input X directly to the output of GRU layer 2 and performing a residual connection can enhance feature propagation and gradient flow. The formula for the residual connection is:

[0042] in, The output at time t after adding the residual can help alleviate the gradient vanishing problem in deep networks and improve the training effect of the model.

[0043] Step 2.3 CNN part Output after residual connection It is resized, and a channel dimension is added before the output of the GRU part is input to the CNN part. The CNN part includes the following key steps: Convolutional layer: extract features from input time series data through two-dimensional convolution operation. Represents the output feature map of the convolutional layer, the formula is:

[0044] in, Represents the parameters of the convolutional layer; Represents a two-dimensional convolution operation, which can capture local features in the input data and enhance the model's ability to recognize time series patterns.

[0045] Activation function: The output of the convolutional layer is processed by the Parametric ReLU (PReLU) activation function and the output is , the formula is:

[0046] PReLU can improve the expressiveness of the model by learning the parameterized negative semi-axis slope.

[0047] Pooling layer: The feature map after convolution activation is reduced in dimension through the Max Pooling operation to reduce computational complexity and prevent overfitting. The formula of the pooling layer is:

[0048] in, is the output of the pooling layer, is the convolution kernel size of the convolution layer; Represents the maximum pooling operation, which can extract the most significant features and enhance the model's anti-noise ability.

[0049] Flatten operation: Flatten the pooled output into a one-dimensional vector for input into the fully connected layer. for:

[0050] Step 2.4 Fully connected layer and output The flattened vector after the flattening operation Perform linear transformation through the fully connected layer to output the final remaining useful life prediction value The formula for the fully connected layer is:

[0051] in, is the parameter of the fully connected layer, and the final output is the model's predicted value for the engine's remaining useful life (RUL).

[0052] Step 3: Train the model: The training set is input into the GRU-CNN residual network model for training the model, and after training, the aircraft engine remaining service life prediction model is obtained.

[0053] In each training cycle, the model is trained based on small batches of data. Training with small batches of data helps speed up model training and reduce memory usage. Each step of the training process includes the following key operations: Step 3.1 Forward propagation: Input the current mini-batch data into the model and calculate the model's predicted output.

[0054] Step 3.2 Calculate the loss: Calculate the error between the model's predicted output and the true label through the specified loss function (MSE Loss) .

[0055]

[0056] In the formula, represents the predicted value of the i-th engine and the true value The difference is , and b is the number of samples in the mini-batch used for training.

[0057] Step 3.3 Backpropagation: Calculate the gradient of the loss with respect to the model parameters. The purpose of gradient backpropagation is to adjust the model parameters to minimize the loss function.

[0058] Step 3.4 Parameter update: Use the calculated gradient to update the model parameters through an optimizer (such as the Adam optimizer). After each iteration, the optimizer adjusts the model parameters based on the current gradient so that the loss function gradually decreases.

[0059] Step 3.5 Validation process: In each training cycle, the model is validated using the validation set data. By calculating the loss on the validation set, you can monitor whether the model has overfitting during the training process. The validation set loss can be calculated using a custom score function, and different metrics (such as RMSE) can be used to evaluate the performance of the model. The training loss and validation loss will be recorded for subsequent analysis.

[0060] RMSE is used to evaluate the performance of the model. The RMSE calculation formula is:

[0061] In the formula, represents the predicted value of the i-th engine and the true value The difference between them is , and K is the total number of engines predicting RUL in the validation set.

[0062] During the entire training process, the model will output loss at a fixed period (such as every 10 epochs) to monitor the model's training status in real time. By regularly outputting training loss and validation loss, you can determine the model's convergence speed and whether there are overfitting or underfitting problems. In addition, the loss value during training can help determine the optimal number of training epochs, that is, when the validation loss is minimal, the model performs best.

[0063] Through the above steps, the model continuously adjusts its parameters to minimize the loss on the training set while maintaining good performance on the validation set.

[0064] Step 3.6: Test set prediction: The model will gradually predict the remaining useful life (RUL) of each sample based on the input test data.

[0065] Step 3.7: Result evaluation: Use appropriate evaluation indicators (such as mean square error, root mean square error, etc.) to evaluate the prediction results of the model to quantify the performance of the model on the test set. The present invention uses the score function to judge the prediction level of the final prediction result. By comparing the predicted RUL with the actual RUL of the test set, the prediction accuracy of the model can be judged.

[0066] The score function calculation formula used to judge the generalization ability of the model is:

[0067] In the formula, represents the predicted value of the i-th engine and the true value The difference between the two is , and P is the total number of engines for which RUL is predicted in the test set. The smaller the RMSE and Score values ​​are, the smaller the prediction error is; the difference between the two is that when the absolute values ​​of the errors are the same, RMSE is unbiased, while Score will get a larger penalty score when the RUL predicted value is greater than the actual value, which is more in line with the engineering practice of taking conservative estimates when predicting the remaining service life in aircraft engine maintenance. Output the prediction results of the model and compare and analyze them with the true values ​​of the test set. Through visualization means, such as drawing a comparison curve between the true RUL and the predicted RUL, the prediction effect of the model can be intuitively displayed.

[0068] Step 4: Obtain the trained GRU-CNN residual network model for predicting the remaining useful life of aircraft engines.

[0069] The implementation process and prediction effect of the present invention are described below in conjunction with specific application examples: Example 1 In this specific implementation, the C-MAPSS simulation data set for aircraft engines provided by NASA is used. C-MAPSS is a modular aviation propulsion system simulation software developed by NASA's research center in the United States. It aims to simulate the entire degradation process of the aircraft from normal to failure and provide a data basis for the prediction model. The simulation experiment was created under the pycharm tool, simulating an engine model with a thrust of 90,000 pounds. The program includes an atmospheric model and an electrical management system, involving five component modules: fan (Fan), low-pressure compressor (LPC), high-pressure compressor (HPC), high-pressure turbine (HPT) and low-pressure turbine (LPT). Figure 1 The figure shows the logical structural relationship of the five modules in the aircraft engine simulation experiment and other main components: nozzle, combustion chamber, low-pressure rotor speed (N1) and high-pressure rotor speed (N2).

[0070] The open source data contains a total of four sets of simulation data. In the specific implementation process of the present invention, "train_FD001" and "test_FD001" are selected as the training set and the test set respectively. Each sub-dataset contains 26 columns, including unit number, time (unit is cycle), condition setting 1 (flight altitude), condition setting 2 (Mach number), condition setting 3 (throttle lever angle) and 21 monitoring indicators. The 21 monitoring data are used to output signal data in the engine degradation process in the simulation experiment. The specific meanings represented are described in Table 1.

[0071]

[0072] The specific process of using the method of the present invention to predict life span is as follows: Step 1: From the train_FD001.txt file in the first set of simulation data sets, we can obtain the aircraft engine failure data X20631×26, which refers to the full process data from a certain starting moment to the final failure. The 20631 rows are the total duration of the 100 engine operation cycles, and the 26 columns include unit number, time (in cycles), condition setting 1 (flight altitude), condition setting 2 (Mach number), condition setting 3 (throttle lever angle) and 21 monitoring indicators. Among them, the visualization results of the 3 conditions and 21 monitoring indicators are as follows Figure 2 shown.

[0073] Based on the observation results, those features that remain constant throughout the life cycle are eliminated. Feature selection is performed on a total of 24 variables, resulting in 17 monitoring variables numbered as conditions 1, 2 and indicators 2, 3, 4, 6, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, 21. The original data set is reduced to ; The 17 monitoring variables are standardized according to the "minimum-maximum" method, and the calculation formula is as follows: in, represents the original data of the nth engine signal j at the i-th moment, yes The standardized value, and Respectively represent the maximum and minimum values ​​of the engine signal j; Step 2: Establish the mapping relationship between the monitoring variable X and the remaining useful life RUL in the training set.

[0074] Use mathematical formulas to express the relationship between RUL and the monitoring variables in the training set. Assume that t represents the current operation cycle of the device, T represents the maximum operation cycle of the device, and R earlyIt indicates the threshold value set according to the situation. In order to better simulate the degradation behavior of the equipment during actual use, studies have shown that setting the mutation threshold to 130 operating cycles is the best effect.

[0075] When the remaining operating cycle of the device is greater than or equal to 130 operating cycles, the RUL tag will be set to a constant value of 130. This setting reflects the life characteristics of the device during the normal operation stage, because the failure rate is low and the degradation of the device is negligible during this stage. When the remaining operating cycle of the device is less than 130, the RUL tag will decrease monotonically as the operating cycle increases. This setting simulates the operating state of the device after entering the rapid decline period, which helps the model more accurately capture the behavioral characteristics of the device when it is about to fail.

[0076]

[0077] The sequences in the training set are used to train the GRU-CNN model, while the test set is used to evaluate the prediction performance of the model.

[0078] Step 3: Build the GRU-CNN residual network model The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolution layer, pooling layer and fully connected layer. Data is input into GRU layer 1, and then processed by prelu activation function to enter GRU layer 2. After being processed by prelu activation function, residual connection is performed with the input of GRU layer 1, and then it passes through convolution layer, prelu activation function, pooling layer, flattening operation and fully connected layer in sequence to output the prediction result. The model structure is as follows Figure 3 The actual architecture is as shown in Figure 4 As shown in the figure, the entire model architecture diagram shows the complete process from input data, processing temporal features through GRU, extracting spatial features through CNN, and finally obtaining the model output. The model first extracts time series features through two layers of GRU, and enhances feature retention through residual connections, then passes the GRU output to the CNN layer to further extract spatial features, and finally outputs the prediction results through the fully connected layer.

[0079] Step 4: Train the model using the training set samples. In each training cycle, the model is trained based on small batches of data. Training with small batches of data helps speed up the training of the model and reduce memory usage. The batch size selected in this embodiment is batch-size=190.

[0080] The test set is input into the GRU-CNN residual network model to obtain the remaining useful life prediction value of the in-service aircraft engine. Through continuous forward propagation and back propagation, the parameters are updated to obtain the trained model. The present invention uses RMSE to evaluate the model performance. The RMSE calculation formula is:

[0081] In the formula, represents the predicted value of the i-th engine and the true value The difference is , and K is the total number of engines for predicting RUL.

[0082] In addition, the loss value during the training process can help determine the optimal number of training epochs. The number of epochs used in the present invention is 110.

[0083] The remaining hyper parameters of the model in this embodiment are shown in Table 2.

[0084]

[0085] Visualize the loss function, the results are as follows Figure 5 As shown in the figure, with the iteration of training, the RUL prediction error evaluation becomes smaller and smaller and tends to be stable, which can be used as a trained model. Then the validation set data is predicted and the results are visualized, as shown in the figure. Figure 6 shown.

[0086] Step 5: Output prediction results Use the method in step 1 to preprocess the data of "test_FD001", then input the processed data into the trained model, and then the model outputs the predicted result, compares the result with the actual value, and inputs it into the score function to calculate the loss, which is used to evaluate the generalization of the model and visualize the result. Figure 7 shown.

[0087] The present invention uses the score function calculation formula as follows:

[0088] In the formula, represents the predicted value of the i-th engine and the true value The difference is , and P is the total number of engines that predict RUL in the test set.

[0089] Table 3 shows the prediction results of the prediction method of the present invention and other methods on the C-MAPSS data set, that is, a comparison of the remaining useful life prediction methods of aircraft engines on the C-MAPSS data set of the embodiments of the present invention.

[0090]

[0091] The CNN-GRU-CNN model in Table 3 is the model structure proposed by Wang Wenqing et al. (Wang Wenqing, Guo Heng, Fan Qifu. Prediction of Remaining Life of Aircraft Engines Based on CNN and GRU [C] / / Control Theory Committee of Chinese Automation Society. Proceedings of the 37th China Control Conference (F). Key Laboratory of System Control and Information Processing, Ministry of Education, Department of Automation, Shanghai Jiaotong University;, 2018:6.) in the background. After testing, it was found that although the model has strong feature extraction and time-dependent processing capabilities in theory, the CNN-GRU-CNN model first extracts spatial features through the convolution layer, then inputs the extracted features into the GRU layer for time-dependent processing, and finally further processes them through the CNN layer. This process involves multiple data transformations and hierarchical interactions, which increases the complexity and parameter quantity of the model. Although the structure is complex, the actual effect of the model is not ideal, especially the overall prediction performance is poor, and the generalization performance is relatively poor.

[0092] In Table 3, the GRU-CNN model is a structure in which two networks are directly connected in series, and the double-layer GRU residual connection-CNN model is a structure in which a residual connection is set between two layers of GRU layers and then connected in series with the CNN part. As can be seen from the data in the table, the effects of different combinations of GRU and CNN are quite different. The present invention creatively places the residual connection between two layers of GRU and between GRU and CNN, and organically combines the two together, which can have a smaller SCORE value while ensuring that the RMSE is relatively low. In this embodiment, the SCORE value can be controlled within 260, and the RMSE value can be controlled within 20, which can significantly improve the prediction effect of the model, making the model more excellent in processing time dependency and feature extraction, while maintaining a high prediction accuracy.

[0093] The present invention improves the real-time computing performance and practicability of the model, the computing process is simple and effective, and the prediction accuracy is very high.

[0094] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for predicting the remaining useful life of an aircraft engine based on a GRU-CNN residual network model, characterized in that: The prediction method includes the following contents: Acquire multi-source sensor data of aircraft engines, perform preprocessing and feature selection to obtain a data set, divide the data set into a training set, a validation set, and a test set, establish a mapping relationship between the monitoring variable X and the remaining useful life RUL, and determine the RUL label of each sample; Construct the GRU-CNN residual network model, The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolution layer, pooling layer and fully connected layer. Data is input into GRU layer 1 and then into GRU layer 2. The output of GRU layer 2 is residually connected with the input of GRU layer 1, and then passes through the convolution layer, PReLU activation function, pooling layer, flattening operation and fully connected layer in sequence after dimension adjustment to output the prediction result. The GRU-CNN residual network model is trained using the data set, and the loss function MSE during training is: In the formula, represents the predicted value of the i-th engine and the true value The difference between , b is the number of samples in the mini-batch used for training; Use RMSE and score functions to evaluate the performance of the model; The trained GRU-CNN residual network model is obtained to predict the remaining useful life of aircraft engines.

2. The prediction method according to claim 1, characterized in that: The process of preprocessing and feature selection of multi-source sensor data of aircraft engines is: Step 1..1: Obtain historical aircraft engine failure data and form a data set ,in, represents the total running track length of N aircraft engine samples; N represents the number of aircraft engine samples; represents the monitoring trajectory length of the Nth aircraft engine sample; V represents the number of sensors in the aircraft engine; Step 1..2: Visualize the V different types of sensor signals collected, analyze the data change trend, and identify the variables that do not show obvious changes during the engine life cycle based on the visualization results; Step 1.3: Based on the observation results, remove those features that remain constant throughout the life cycle, perform feature selection on V monitoring variables, and obtain F monitoring variables. The original data set is reduced to ; The F monitoring variables are standardized according to the "minimum-maximum" method to obtain the data set.

3. The prediction method according to claim 1, characterized in that: The score function calculation formula is: In the formula, represents the predicted value of the i-th engine and the true value The difference is , and P is the total number of engines for predicting RUL.

4. The prediction method according to claim 1, characterized in that: The hidden unit dimension of GRU layer 1 is 64, the hidden unit dimension of GRU layer 2 is the same as the original input feature dimension, the convolution kernel size of the convolution layer is 4, the step size is 2, the number of output channels of the convolution layer is 100, and the padding size of the convolution layer is 2; the pooling kernel size of the pooling layer is 3; the learning rate is 0.007, the dropout ratio of GRU layer 1 is 0.2, and the dropout ratio of GRU layer 2 is 0.

3.

5. The prediction method according to claim 4, characterized in that: The SCORE value is controlled within 260 and the RMSE value is controlled within 20.

Images