Aero-engine remaining useful life prediction method based on GRU-CNN residual network model

Abstract

The application is a method for predicting the remaining useful life of an aero-engine based on a GRU-CNN residual network model, which comprises the following contents: obtaining multi-source sensing data of the aero-engine, and obtaining a data set after preprocessing and feature selection processing, dividing the data set into a training set, a validation set and a test set, and determining the RUL label of each sample; constructing a GRU-CNN residual network model, which comprises a GRU layer 1, a GRU layer 2, a convolution layer, a pooling layer and a full connection layer, the data is input into the GRU layer 1, then into the GRU layer 2, the output of the GRU layer 2 is connected in residual with the input of the GRU layer 1, then after dimension adjustment, it is sequentially subjected to the convolution layer, the PReLU activation function, the pooling layer, the flattening operation and the full connection layer, and the prediction result is output; obtaining the trained GRU-CNN residual network model for predicting the remaining useful life of the aero-engine. The method significantly improves the training stability and calculation efficiency of the model, and can balance the accuracy and real-time requirements.

Description

TECHNICAL FIELD

[0001] The application relates to the field of aero-engines, and in particular to an aero-engine residual service life prediction method based on a GRU-CNN residual network model. BACKGROUND

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

[0003] Residual service life prediction is the core technology of the PHM system, is a key link connecting fault prediction and health management, and is a core component of modern industrial intelligent manufacturing. Accurate RUL prediction technology can judge the performance degradation state of the engine before the engine fails and make failure warning in advance, so as to carry out timely repair, reduce the probability of failure, change periodic maintenance to active maintenance, and avoid the phenomenon of 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, technical progress has expanded from the residual service life prediction of a single component of an aircraft to the available life prediction of an aero-engine as a whole, and how to realize accurate RUL prediction has become a major research hotspot in recent years.

[0004] The RUL prediction principle is to evaluate the residual service life of a component or a system as a whole according to the current health state, working link and load, state monitoring information, etc. of the equipment, in combination with physical failure models, historical performance degradation data, fault diagnosis information, etc.

[0005] In the field of aero-engine remaining useful life (RUL) prediction, deep learning techniques, especially convolutional neural networks (CNN) and gated recurrent units (GRU), have received extensive attention due to their 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 features 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 cause a significant increase in the number of model parameters, which may cause delays in real-time calculations and affect the performance and efficiency of the model in actual deployment. At the same time, a single GRU model may lack efficient capture ability for spatial features, especially when facing high-dimensional sensor data. As disclosed in Chinese Patent CN112100865A, the aero-engine remaining life prediction method based on a parallel CNN model has advantages in extracting spatial features of multi-dimensional sensor data, but has inherent shortcomings in processing time series data, which cannot capture the time dependence in the data. In addition, simply relying on increasing the number of CNN layers to improve model capability is effective, but it significantly increases the computational cost and affects the real-time prediction ability of the model. For example, Wang Wenqing et al. (Wang Wenqing, Guo Heng, Fan Qifu. Aero-engine remaining life prediction based on CNN and GRU [C] / / China Automation Association Control Theory Committee. Proceedings of the 37th China Control Conference (F). Shanghai Jiaotong University Department of Automation Key Laboratory of System Control and Information Processing, Ministry of Education; 2018:6.) use CNN and GRU to combine RUL prediction, the network structure is that the sensor data enters the multi-layer CNN, the CNN extracts the abstract features of the sequence as the input of the GRU, and then connects several layers of CNN for filtering, and finally obtains the remaining life prediction result through the linear regressor. Although the RMSE and SCORE indicators are significantly improved, the generalization ability needs to be further improved. SUMMARY

[0006] To overcome the shortcomings of the prior art, the technical problem to be solved by the present application is to provide an aero-engine remaining useful life prediction method based on a GRU-CNN residual network model. This method can effectively combine the advantages of GRU and CNN networks, fully extract spatial and temporal features, significantly improve the training stability and computational efficiency of the model, overcome the limitations of existing methods in feature extraction and computational complexity, and balance the accuracy and real-time requirements.

[0007] The technical solution adopted by the present application to solve the technical problem is,

[0008] An aero-engine remaining useful life prediction method based on a GRU-CNN residual network model, the prediction method comprising the following contents:

[0009] A plurality of source sensing data of an aero-engine is acquired, and after preprocessing and feature selection processing, a data set is obtained, the data set is divided into a training set, a validation set and a test set, a mapping relationship between a monitoring variable X and a remaining useful life RUL is established, and the RUL label of each sample is determined;

[0010] A GRU-CNN residual network model is constructed,

[0011] 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, then enters the GRU layer 2, the output of the GRU layer 2 is connected in residual with the input of the GRU layer 1, then after dimension adjustment, sequentially passes through the convolution layer, the PReLU activation function, the pooling layer, the flattening operation and the full connection layer, and outputs a prediction result;

[0012] The GRU-CNN residual network model is trained by using the data set, and the loss function MSE during training is:

[0013]

[0014] In the formula, represents the predicted value of the i th engine and the true value , and b is the number of samples in a small batch for training;

[0015] The performance of the model is evaluated by using the RMSE and the score function;

[0016] The trained GRU-CNN residual network model is obtained, which is used for predicting the remaining useful life of the aero-engine.

[0017] Further, the process of preprocessing and feature selection processing of the plurality of source sensing data of the aero-engine is:

[0018] Step 1..1: Obtain historical aero-engine failure data to form a data set In the formula, represents the total running track length of N aero-engine samples; N represents the number of aero-engine samples; represents the monitoring track length of the N th aero-engine sample; V represents the number of sensors in the aero-engine;

[0019] Step 1..2: Visualize the V different types of sensor signals collected, analyze the change trend of the data, and according to the visualization result, identify the variables that do not show obvious changes in the whole life cycle of the engine;

[0020] Step 1.3: Based on the observation, the features that remain constant throughout the life cycle are removed, feature selection is performed on the V monitoring variables, F monitoring variables are obtained, and the original data set is reduced to ; After the F monitoring variables are standardized in the "minimum-maximum" manner, the data set is obtained.

[0021] Further, the score function calculation formula is:

[0022]

[0023] In the formula, represents the predicted value of the i th engine and the true value , and P is the total number of engines for which RUL is predicted.

[0024] Further, the hidden unit dimension of the GRU layer 1 is 64, the hidden unit dimension of the GRU layer 2 is the same as the original input feature dimension, the convolution kernel size of the convolution layer is 4, the step is 2, the output channel number 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 the GRU layer 1 is 0.2, and the dropout ratio of the GRU layer 2 is 0.3.

[0025] Further, the SCORE value is controlled within 260, and the value of RMSE is controlled within 20.

[0026] Compared with the prior art, the beneficial effects of the present application are:

[0027] The model in the present application exhibits excellent performance, and compared with the traditional GRU, CNN and bidirectional LSTM model, the prediction accuracy of the present application is greatly improved. In terms of the two key performance indicators of RMSE and SCORE, the model of the present application achieves excellent performance of 22.013 and 257.114, which is significantly better than GRU (RMSE 33.594, SCORE 1511.451), CNN (RMSE 27.196, SCORE 1364.601) and bidirectional LSTM (RMSE 34.594, SCORE 1910.125). These data show that the present application 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 through the SCORE score function, further enhancing the robustness of the model.

[0028] By optimizing the feature fusion of the CNN and the GRU, the application can comprehensively extract the spatial and temporal features in the sensor data, and the integrated innovation significantly improves the prediction performance of the model, uses the PReLU function for lightweight design of the model, can more flexibly adapt to different data distribution, improves the applicability of the model, and further improves the model performance.

[0029] To sum up, the GRU-CNN residual neural network structure of the application can perform more accurate remaining useful life prediction in complex and high-dimensional aero-engine sensor data, and is suitable for various different working conditions, and has significant engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 C-MAPSS simplified engine simulation model;

[0031] Figure 2 Monitoring index visualization;

[0032] Figure 3 GRU-CNN residual network model architecture diagram;

[0033] Figure 4 GRU-CNN residual network model structure visualization;

[0034] Figure 5 Training curve: comparison diagram of training loss and verification loss;

[0035] Figure 6 Comparison diagram of model predicted RUL and actual RUL (training set sample);

[0036] Figure 7 Comparison diagram of predicted RUL and actual RUL;

[0037] Figure 8 Comparison diagram of predicted and actual RULs sorted by actual RUL. DETAILED DESCRIPTION

[0038] The application will be further explained in conjunction with the embodiments and the drawings, but it is not limited to the scope of protection of the application.

[0039] The accurate prediction of the remaining useful life (RUL) of an aero-engine requires the extraction of effective degradation feature information from complex multi-source sensor data. These multi-source sensor data are usually high-dimensional and noisy, making feature extraction a challenge. The GRU-CNN residual network model proposed by the invention 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 aero-engine health management and fault prediction maintenance systems. The main contributions are reflected in the following three aspects: (1) Time series feature extraction: first, the GRU network is used to analyze the time series data in depth, capturing the dynamic features and long-term dependencies in the performance degradation process of the aero-engine. (2) Spatial feature enhancement: based on the GRU layer, the CNN layer is further used to extract the spatial features of the data, enhancing the sensitivity of the model to local changes, and learning feature weights by simulating the interaction between each channel of the convolutional features (the convolution operation and the subsequent pooling operation can be regarded as a certain weight allocation to the 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, as the convolution kernel will automatically adjust during training, thus giving different "weights" to different spatial features.), achieving feature enhancement and suppression. (3) Residual connection optimization: the input and output of the GRU network are connected by a residual connection in the model, improving the stability and efficiency of the model training, reducing the number of model parameters, avoiding overfitting, and improving the real-time computing performance and practicality of the model.

[0040] The aero-engine remaining useful life prediction method based on the GRU-CNN residual network model of the invention includes the following steps:

[0041] Step 1: Obtain multi-source sensor data of the aero-engine, and after preprocessing and feature selection processing, obtain a data set, establish a mapping relationship between the monitoring variables X and the remaining useful life RUL, and determine the RUL label of each sample.

[0042] In order to improve the accuracy of the remaining useful life (RUL) prediction, the multi-source sensor data of the engine are strictly preprocessed and feature selected. The specific process of data preprocessing and feature selection processing of the aero-engine monitoring data includes the following steps:

[0043] Step 1.1: Obtain historical aero-engine failure data to form a data set , wherein, represents the total running track length of N aero-engine samples, N represents the number of aero-engine samples, represents the monitoring track length of the Nth aero-engine sample, v = 1, 2, V represents the number of sensors in the aero-engine.

[0044] Step 1..2: Visualize the collected V different types of sensor signals, analyze the trend of data changes. According to the visualization results, identify the variables that do not show obvious changes in the whole life cycle of the engine. These variables cannot reflect the degradation process of engine performance due to their stability.

[0045] Step 1.3: Based on the observation results, eliminate those features that remain constant throughout the life cycle. Feature selection is performed on the V monitoring variables to obtain F monitoring variables, and the original data set is reduced to ; After standardizing the F monitoring variables in the "minimum-maximum" way, the data set is obtained, and the calculation formula is as follows:

[0046]

[0047] Where, represents the original data of the nth engine signal j at the ith time, is the standardized value, and and represent the maximum and minimum values of engine signal j, respectively.

[0048] Step 1.4: Divide the data set into training set, test set and validation set; use the training set to obtain the mapping relationship between the monitoring variables X and the remaining useful life RUL, and determine the RUL label of each sample.

[0049] The failure rate of turbofan engine is constant and low during normal operation stage, and the degradation can be ignored; only at the end of the device life stage, the failure rate increases rapidly with time. The data monitored before the device begins to rapidly decline is not suitable for model training, so a mutation threshold is set.

[0050] Based on the set mutation threshold, the remaining useful life RUL label in the training set is set as a piecewise linear function,

[0051] Use mathematical formula to express the relationship between RUL and monitoring variables in the training set, assuming t represents the current running cycle of the device, T represents the maximum running cycle of the device, R early represents the mutation threshold set according to the situation.

[0052]

[0053] Step 2: Construct a GRU-CNN residual network model,

[0054] The GRU-CNN residual network model comprises a GRU layer 1, a GRU layer 2, a convolutional layer, a pooling layer and a fully connected layer, data is input into the GRU layer 1, then enters the GRU layer 2, the output of the GRU layer 2 is connected in residual with the input of the GRU layer 1, and then sequentially passes through the convolutional layer, the PReLU activation function, the pooling layer, the flattening operation and the fully connected layer, and outputs a prediction result.

[0055] Step 2.1 GRU part

[0056] Two Gated Recurrent Unit (GRU) layers are adopted, each having a hidden layer dimension of and units. The GRU layer is a kind of recurrent neural network (RNN) which captures long-term dependencies in time series data through a gating mechanism.

[0057] Gating mechanism of the GRU layer:

[0058] Update gate: decides how to update the state of the previous moment to the current state , and the formula is:

[0059]

[0060] wherein, is the update gate, is the weight matrix, is the bias, and sigmoid is the sigmoid function.

[0061] Reset gate: controls how the state of the previous moment is combined with the current input , and the formula is:

[0062]

[0063] wherein, is the reset gate, is the weight matrix and bias.

[0064] New state: calculates the candidate hidden state of the current moment , and the formula is:

[0065]

[0066] wherein, represents element-wise multiplication, , .

[0067] Final hidden state: combines the update gate and the candidate state to obtain the final hidden state: ​​

[0068]

[0069] Thus, in the GRU layer 1, given the input X, the output of the GRU is:

[0070]

[0071] In the GRU layer 2, the output is:

[0072]

[0073] where, and are the parameters of the two layers of GRU, respectively, the parameters and usually include all the weights and biases inside the GRU unit, which also covers the related parameters of the activation function, and the activation function adopts the PReLU activation function.

[0074] Step 2.2 Residual Connection

[0075] The input X is directly added to the output of the GRU layer 2 for residual connection, which can enhance feature propagation and gradient flow. The formula of residual connection is:

[0076]

[0077] where, is the output at time t after adding the residual, which helps to alleviate the gradient vanishing problem in deep networks and improve the training effect of the model.

[0078] Step 2.3 CNN Part

[0079] The output after residual connection is re-adjusted, and a channel dimension is added before inputting the output of the GRU part into the CNN part. The CNN part includes the following key steps:

[0080] Convolutional layer: the input time series data is extracted by two-dimensional convolution operation. Let represent the output feature map of the convolutional layer, the formula is:

[0081]

[0082] where, represents the parameters of the convolutional layer; represents the two-dimensional convolution operation, which can capture local features in the input data and enhance the model's ability to recognize time series patterns.

[0083] Activation function: The output of the convolutional layer is processed by the Parametric ReLU (PReLU) activation function and output as , the formula is:

[0084]

[0085] PReLU can improve the expression ability of the model by learning the parameterized negative half-axis slope.

[0086] Pooling layer: The feature map after convolution and activation is reduced in dimension by the Max Pooling operation to reduce computational complexity and prevent overfitting. The formula of the pooling layer is:

[0087]

[0088] where, is the output of the pooling layer, is the convolution kernel size of the convolutional layer; represents the max pooling operation, which can extract the most significant features and enhance the noise resistance of the model.

[0089] Flatten operation: The output after pooling is flattened into a one-dimensional vector to be input into the fully connected layer. The flattened output is:

[0090]

[0091] Step 2.4 Fully connected layer and output

[0092] The flattened vector after the flattening operation is linearly transformed by the fully connected layer to output the final remaining useful life prediction value . The formula of the fully connected layer is:

[0093]

[0094] where, is the parameter of the fully connected layer, and the final output is the predicted value of the model for the remaining useful life (RUL) of the engine.

[0095] Step 3: Train the model

[0096] Input the training set into the GRU-CNN residual network model to train the model, and obtain the remaining useful life prediction model of the aero-engine after training.

[0097] In each training cycle, the model is trained on a small batch of data. Training on small batches helps speed up the training process and reduces memory usage. Each step of the training process includes the following key operations:

[0098] Step 3.1 Forward Propagation: The current small batch of data is input into the model, and the model's predicted output is calculated.

[0099] Step 3.2 Calculate Loss: The error between the model's predicted output and the true labels is calculated using the specified loss function (MSE Loss) .

[0100]

[0101] where, represents the difference between the predicted value and the true value of the i-th engine, and b is the number of samples in the small batch used for training.

[0102] Step 3.3 Backpropagation: The gradient of the loss with respect to the model's parameters is calculated. The purpose of backpropagation is to adjust the model's parameters to minimize the loss function.

[0103] Step 3.4 Parameter Update: The calculated gradients are used to update the model's parameters using an optimizer (such as the Adam optimizer). After each iteration, the optimizer adjusts the model's parameters based on the current gradients to gradually reduce the loss function.

[0104] Step 3.5 Validation Process: During each training cycle, the model is validated using data from the validation set. By calculating the loss on the validation set, we can monitor whether the model is overfitting during training. The validation set loss can be calculated using a custom score function, and different metrics (such as RMSE) can be used to evaluate the model's performance. The loss during training and the validation loss are recorded for subsequent analysis.

[0105] The RMSE is used to evaluate the model's performance, and the RMSE calculation formula is:

[0106]

[0107] where, represents the difference between the predicted value and the true value of the i-th engine, and K is the total number of engines in the validation set for which RUL is predicted.

[0108] Throughout the training process, the model will output the loss at fixed intervals (e.g., every 10 cycles) to monitor the training status of the model in real time. By regularly outputting the training loss and validation loss, we can determine the convergence speed of the model and whether there is overfitting or underfitting. In addition, the loss value during training can help determine the optimal number of training epochs (Epochs), i.e., when the validation loss is the smallest, the model performs best.

[0109] 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.

[0110] Step 3.6: Test set prediction: The model will predict the remaining useful life (RUL) of each sample step by step based on the input test data.

[0111] Step 3.7: Result evaluation: Evaluate the model's prediction results using appropriate evaluation metrics (such as mean squared error, root mean squared error, etc.) to quantify the model's performance on the test set. The score function is used to evaluate the prediction level of the final prediction results. By comparing the predicted RUL with the true RUL of the test set, we can determine the prediction accuracy of the model.

[0112] The score function for evaluating the generalization ability of the model is calculated as follows:

[0113]

[0114] In the formula, represents the predicted value of the i-th engine and the true value , P is the total number of engines in the test set for which RUL is predicted. The smaller the RMSE and Score values, the smaller the prediction error; the difference between the two is that RMSE is unbiased when the absolute error is the same, while Score gives a larger penalty score when the predicted RUL is greater than the actual value, which is more consistent with the engineering practice of preferring conservative estimates in predicting the remaining useful life of an aero-engine. Output the prediction results of the model and compare them with the true values of the test set. Through visualization methods such as plotting the comparison curve of true RUL and predicted RUL, the prediction effect of the model can be intuitively displayed.

[0115] Step 4: Obtain the trained GRU-CNN residual network model for predicting the remaining useful life of an aero-engine.

[0116] The implementation process and prediction effect of the present application will be described below in conjunction with specific application examples:

[0117] Example 1

[0118] In this embodiment, the C-MAPSS simulation data set provided by NASA about the aero-engine is adopted. The C-MAPSS is a modular aero-propulsion system simulation software developed by the research center of NASA in the United States, which aims to simulate the entire degradation process of the aircraft from normal to failure, and provide data basis for the prediction model. The simulation experiment is created under the pycharm tool, which simulates an engine model with a thrust of 90000 pounds. The program includes an atmospheric model and an electrical management system, involving five component modules of fan, low-pressure compressor (LPC), high-pressure compressor (HPC), high-pressure turbine (HPT) and low-pressure turbine (LPT). Figure 1 The logical structure relationship of the five modules in the aero-engine simulation experiment and other main components such as nozzle, combustor, low-pressure rotor speed (N1) and high-pressure rotor speed (N2) are shown in FIG.

[0119] The open source data contains four groups of simulation data. In the embodiment of the present application, "train_FD001" and "test_FD001" are selected as the training set and the test set respectively, wherein each sub-data set 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 indexes. The 21 monitoring data are used to output signal data in the engine degradation process in the simulation experiment, and the specific meaning is described as shown in Table 1.

[0120]

[0121] The specific process of life prediction by applying the method of the present application is as follows:

[0122] Step 1: The aero-engine failure data X20631x26 can be obtained from the train_FD001.txt file in the first group of simulation data set, which refers to the whole process data from a certain starting time to the final failure. The 20631 rows are the total duration of 100 engine operation cycles, and the 26 columns include 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 indexes. The visualization results of the 3 conditions and 21 monitoring indexes are shown in FIG. Figure 2

[0123] Based on the observation results, the features that remain constant throughout the life cycle are removed. Feature selection is performed on a total of 24 variables, and 17 monitoring variables are obtained, which are numbered as condition 1, 2 and index 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 in the "min-max" manner, and the calculation formula is as follows:

[0124]

[0125] wherein, denotes the original data of the nth engine signal j at the i th moment, is the standardized value, and and denote the maximum value and the minimum value of the engine signal j, respectively;

[0126] Step 2: Establish the mapping relationship between the monitoring variables X and the remaining useful life RUL in the training set.

[0127] The relationship between RUL and the monitoring variables in the training set is expressed by a mathematical formula. Assuming that t represents the current running cycle of the equipment, T represents the maximum running cycle of the equipment, R early represents a threshold value set according to the situation, in order to better simulate the degradation behavior of the equipment in the actual use process, research shows that the best setting of the mutation critical value is 130 running cycles.

[0128] When the remaining running cycle of the equipment is greater than or equal to 130 running cycles, the RUL label will be set to a constant value of 130. This setting reflects the life characteristics of the equipment in the normal running stage, because the failure rate is low and the degradation of the equipment can be ignored. When the remaining running cycle of the equipment is less than 130, the RUL label will monotonically decrease with the increase of the running cycle. This setting simulates the running state of the equipment after entering the rapid decline period, which helps the model to more accurately capture the behavior characteristics of the equipment when the failure is about to occur.

[0129]

[0130] The sequences in the training set are used to train the GRU-CNN model, and the test set is used to evaluate the prediction performance of the model.

[0131] Step 3: Construct the GRU-CNN residual network model

[0132] The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolution layer, pooling layer and full connection layer. The data is input into GRU layer 1, then processed by prelu activation function to enter GRU layer 2, and then processed by prelu activation function to perform residual connection with the input of GRU layer 1, and then sequentially pass through the convolution layer, the prelu activation function, the pooling layer, the flattening operation and the full connection layer, and output the prediction result. The model structure is as shown in Figure 3 The actual architecture is as follows:Figure 4 As shown in the figure, the whole model architecture diagram shows the complete process from input data, through GRU processing time features, and then through CNN to extract spatial features, and finally get the model output. The model first extracts time series features through two layers of GRU, and enhances feature preservation through residual connection, then passes the GRU output to the CNN layer to further extract spatial features, and finally outputs the prediction result through the fully connected layer.

[0133] Step 4: Train the model using the training set samples. In each training cycle, the model will be trained according to the batch data. The training of the batch data helps to speed up the training of the model and reduce the memory occupation. The batch-size selected in this embodiment is batch-size=190.

[0134] The test set is input into the GRU-CNN residual network model to obtain the remaining useful life prediction value of the in-service aero-engine. Through continuous forward propagation and back propagation, the trained model is obtained by updating the parameters. The RMSE is used to evaluate the performance of the model. The RMSE calculation formula is:

[0135]

[0136] In the formula, represents the predicted value of the i th engine and the true value The difference, K is the total number of engines for which RUL is predicted.

[0137] In addition, the loss value in the training process can help to determine the optimal number of training cycles (Epochs), and the number of cycles used in this embodiment is epochs=110.

[0138] The remaining hyperparameters of the model in this embodiment are shown in Table 2.

[0139]

[0140] The loss function is visualized, and the result is shown in Figure 5 As the iteration of training, the RUL prediction error evaluation becomes smaller and tends to be stable, which can be used as a trained model. Then the validation set data is predicted, and the result is visualized, as shown in Figure 6 .

[0141] Step 5: Output the prediction result

[0142] The "test_FD001" is pre-processed by the method of step 1, and then the processed data is input into the trained model, and then the model outputs the prediction result, which is compared with the actual value, and is brought into the score function to calculate the loss, which is used to evaluate the generalization of the model, and the result is visualized. The results are shown in Figure 7

[0143] The score function used in the application is calculated according to the following formula:

[0144]

[0145] In the formula, represents the predicted value of the i th engine and the real value The difference between P is the total number of engines in the test set for predicting RUL.

[0146] Table 3 is the prediction result of the prediction method of the application and other methods on the C-MAPSS data set, that is, the comparison of the aircraft engine remaining useful life prediction method of the C-MAPSS data set of the embodiment of the application.

[0147]

[0148] The CNN-GRU-CNN model in Table 3 is the model structure proposed by Wang Wenqing et al. (Wang Wenqing, Guo Heng, Fan Qifu. Aviation Engine Remaining Life Prediction Based on CNN and GRU [C] / / China Automation Association Control Theory Professional Committee. 37th China Control Conference Proceedings (F). Shanghai Jiaotong University Department of Automation Key Laboratory of System Control and Information Processing, Ministry of Education; 2018:6.) in the background. After testing, it is found that although the model has strong feature extraction and time dependence processing ability 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 dependence processing, and finally further processes through the CNN layer. This process involves multiple data transformations and level interactions, increasing 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.

[0149] ​The GRU-CNN model in Table 3 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 arranged between two GRU layers and then connected with the CNN part in series. As can be seen from the data in the table, the effects of different combinations between GRU and CNN are quite different. The application creatively arranges a residual connection between two GRU layers and between GRU and CNN, and organically combines the two together, so that the SCORE value can be controlled within 260 and the RMSE value can be controlled within 20 in the embodiment, the prediction effect of the model can be significantly improved, the model performs more excellent in processing time dependence and feature extraction, and meanwhile, a higher prediction accuracy is maintained.

[0150] The application 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.

[0151] The application does not cover the prior art.

Claims

1. A method for predicting the remaining service life of an aero-engine based on a GRU-CNN residual network model, characterized in that, The prediction method includes the following: Multi-source sensor data of aero-engines are acquired, and after preprocessing and feature selection, a dataset is obtained. The dataset is divided into training set, validation set and test set. A mapping relationship between monitoring variable X and remaining service life RUL is established, and the RUL label of each sample is determined. Construct a GRU-CNN residual network model. The GRU-CNN residual network model includes GRU layer 1, GRU layer 2, convolutional 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 to the input of GRU layer 1. After dimensionality adjustment, the data passes through convolutional layer, PReLU activation function, pooling layer, flattening operation and fully connected layer in sequence to output the prediction result. The GRU-CNN residual network model is trained using the dataset, and the loss function MSE during training is: In the formula, This represents the predicted value of the i-th engine. and the true value RUL i The difference, where b is the number of samples in the mini-batch used for training; Use the RMSE and score functions to evaluate the model's performance; A trained GRU-CNN residual network model is obtained and used to predict the remaining service life of aero engines; The process of preprocessing and feature selection of multi-source sensor data for aero-engines is as follows: Step 1.1: Obtain historical aero-engine failure data to form a dataset X. L×V Where L = L1 + L2 + ... + L N L represents the total trajectory length of N aero-engine samples; N represents the number of aero-engine samples; L N V represents the length of the monitoring trajectory for the Nth aero-engine sample; V represents the number of sensors in the aero-engine. Step 1.2: Visualize the collected V different types of sensor signals, analyze the data change trends, and identify variables that have not shown significant changes throughout the engine's entire life cycle based on the visualization results; Step 1.3: Based on the observation results, remove those features that remain constant throughout the entire life cycle, perform feature selection on the V monitoring variables to obtain F monitoring variables, and reduce the dimensionality of the original dataset to X. L×F The dataset is obtained by standardizing the F monitoring variables in a "min-max" manner.

2. The prediction method according to claim 1, characterized in that, The formula for calculating the score function is as follows: In the formula, This represents the predicted value of the i-th engine. and the true value RUL i The difference is P, where P is the total number of engines used to predict RUL.

3. The prediction method according to claim 1, characterized in that, The hidden unit dimension of GRU layer 1 is 64, and the hidden unit dimension of GRU layer 2 is the same as the original input feature dimension. The convolutional layer has a kernel size of 4, a stride of 2, an output channel count of 100, and a padding size of 2. The pooling layer has a pooling kernel size of 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.

4. The prediction method according to claim 3, characterized in that, The score should be kept below 260, and the RMSE value should be kept below 20.

Images