Turbofan engine RUL prediction method and system based on historical data fusion

Abstract

The invention discloses a turbofan engine RUL prediction method and system based on multi-scale historical information fusion, and the method comprises the steps: collecting various sensor data during the operation of a turbofan engine, and carrying out the preprocessing, and obtaining a plurality of pieces of time window data with RUL labels; processing data by using three sliding windows of different scales, superposing all historical time window data before the current moment on a third dimension, performing dimension reduction on the superposed data by using a principal component analysis method, and finally fusing the obtained data of three scales to obtain training and testing samples; an ISEformer network framework is adopted, and an RUL prediction model is obtained through training; and outputting a prediction result by the prediction model according to the operation data of the to-be-detected turbofan engine. Through the innovative method combining historical data fusion and the deep learning technology, the RUL prediction performance of the turbofan engine under the complex working condition is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft engine RUL prediction, and in particular relates to a turbofan engine RUL prediction method and system based on historical data fusion. Background Art

[0002] Turbofan engines are key components of aircraft, and the reliability and safety of their performance have a vital impact on the normal operation of aircraft. Since the engine is in complex and harsh working conditions such as high temperature and high pressure for a long time, it is prone to degradation or even failure, which not only affects the service life of the equipment, but may also cause serious safety accidents. Therefore, accurately predicting the remaining useful life (RUL) of turbofan engines is of great significance to improving the operation and maintenance efficiency of aircraft and ensuring flight safety.

[0003] At present, the remaining life prediction methods are mainly divided into prediction based on physical models and data-driven models. The physical model establishes a mathematical model of the degradation process by analyzing the failure mechanism of the equipment. This type of method relies on a detailed understanding of the internal physical characteristics and failure modes of the equipment. However, due to the complex and changeable operating environment of turbofan engines, the degradation law is difficult to accurately describe, and the application of physical models is often limited. In contrast, the data-driven method does not require in-depth modeling of the degradation mechanism of the equipment. Instead, it uses the historical monitoring data of the equipment and automatically constructs a prediction model through technologies such as deep learning, which has gradually become the mainstream method for RUL prediction.

[0004] The existing data-driven turbofan engine RUL prediction methods have obvious limitations in data processing methods, mainly manifested in that they only use the data of the current time window for prediction, while ignoring the historical data of the previous cycles. For example, conventional methods usually only rely on the time window of the current cycle, and do not make full use of the time window information in the previous cycle. This processing method fails to effectively capture the long-term dependencies in the time series, resulting in the important information contained in the previous data not being fully mined, which in turn affects the prediction accuracy and rationality of the model. In fact, historical data often contain rich degradation trends and dependency information. The rational use of this data helps to more comprehensively understand the operating status and degradation process of the equipment. However, existing methods generally lack the fusion of historical data, especially when processing complex multi-sensor time series data, and cannot effectively capture long-term dependencies, which ultimately limits the improvement of RUL prediction performance. Summary of the invention

[0005] The present invention aims to solve the problem in the prior art that the RUL prediction of a turbofan engine has insufficient performance under complex working conditions.

[0006] One object of the present invention is to provide a method for predicting the remaining useful life (RUL) of a turbofan engine based on historical data fusion. By combining historical data fusion with deep learning technology, the RUL prediction performance of a turbofan engine under complex working conditions is significantly improved.

[0007] Another object of the present invention is a turbofan engine RUL prediction system based on historical data fusion.

[0008] In order to achieve the above-mentioned object, the present invention provides a turbofan engine RUL prediction method based on historical data fusion, comprising:

[0009] Collect various sensor data when the turbofan engine is running, and pre-process them to obtain multiple time window data with RUL labels;

[0010] Three sliding windows of different scales are used to process historical data. All historical time window data before the current moment are superimposed in the time dimension, and the superimposed data are reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples.

[0011] According to the training and testing samples, the ISEformer network framework is used to train and obtain the RUL prediction model;

[0012] According to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

[0013] A further preferred technical solution of the present invention is that the data of various sensors during the operation of the turbofan engine are collected and preprocessed; including:

[0014] Collect various sensor data when the turbofan engine is running;

[0015] The acquired sensor data is screened to remove the data that remain unchanged during the cycle;

[0016] Normalize the filtered sensor data;

[0017] The normalized data is processed using a sliding window to obtain multiple time window data;

[0018] Perform RUL label processing on data in multiple time windows.

[0019] Preferably, the filtered sensor data is normalized; the specific method is:

[0020] The maximum-minimum normalization is used to scale the sensor data of each channel to the range [0,1]. The normalization formula is:

[0021]

[0022] In the formula, represents the normalized data of the sth sensor at the tth time step, max{x s} and min{x s} respectively represent the maximum and minimum values ​​in the sth sensor; eps = 10 -12 .

[0023] Preferably, the normalized data is processed using a sliding window to obtain multiple time window data; specifically:

[0024] For the entire cycle of a turbofan engine from operation to failure, assuming that the total number of engine operation cycles is T, the window size is W, and the sliding step is p, multiple samples are obtained, each of which is W×F in size, where F is the number of features and one sample is a time window data.

[0025] Preferably, the performing RUL label processing on the data in multiple time windows is specifically as follows:

[0026] For the n+1th time window data, the corresponding RUL label calculation formula is:

[0027] RUL=TWn*p;

[0028] The labels of the time window data are processed using a piecewise linear function, and the formula is:

[0029]

[0030] The obtained multiple time window data with RUL labels are used as the training data set for the RUL prediction model.

[0031] As a preferred method, three sliding windows of different scales are used to process historical data, all historical time window data before the current moment are superimposed according to the time dimension, and the superimposed data is reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples; specifically:

[0032] Use windows with step sizes of 1, 2, and 3 for sliding, so that the temporal characteristics of the data can be captured from different time scales, improving the model's ability to perceive multi-scale information.

[0033] When the step size is 1, a forward sliding window is used. Specifically, the sliding window starts from the beginning of the data sequence and slides in sequence according to the step size of 1 until it reaches the current moment to be predicted. Each time it slides, the data in the window will be collected and used as the input data of a time window.

[0034] When the step size is 2, a reverse sliding window is used. Specifically, the sliding window goes back from the current moment, slides through a window with a step size of 2, and collects data before the current moment. Reverse sliding can ensure that it can slide to the current prediction window, while if forward sliding is used, it may not slide to the current prediction window.

[0035] When the step size is 3, it is similar to the reverse sliding window with a step size of 2. It also traces back from the current moment and slides through a window with a step size of 3 to collect historical data in a larger range.

[0036] For example, for the nth time window, there are n-1 time window data at this moment. When the step size is 1, a total of n time windows can be slid out, and the shape of each time window is (W, F). Each time window is added with one dimension, and the shape becomes (W, F, 1). Then, each time window data is superimposed along the third dimension, and the shape becomes (W, F, n); when the step size is 2, a total of n time windows can be slid out. time windows, the shape becomes Similarly, when the step length is 3, a total of time windows, the shape becomes

[0037] The shape obtained is (W, F, n), and The fusion time window is reconstructed into (W×F,n), and Then PCA is used to reduce the dimension, and the shapes of the three windows become (W×F,1). Then the three windows are reintegrated into (W,F). Finally, the three information containing different scales are fused into (W,F,3) as the input of the ISEformer network.

[0038] Preferably, the specific method of reducing dimension by PCA is:

[0039] Taking the step size of 1 as an example, let the input data X∈R W×F×n , first reconstruct the three-dimensional data into a two-dimensional matrix X r ∈R (W ×F)×n , calculate the covariance matrix C of the two-dimensional matrix, the calculation formula is:

[0040]

[0041] Perform eigenvalue decomposition on the covariance matrix C and obtain the eigenvalue λ i and the corresponding eigenvector v i , the calculation formula is:

[0042] Cv i =λ i v i ;

[0043] Select the eigenvector v corresponding to the largest eigenvalue i Projecting onto the principal component, we get the dimension-reduced data Z, whose shape is (W×F,1) and the calculation formula is:

[0044] Z=X r v1;

[0045] Finally, Z∈R (W×F)×1 The shape is reconstructed into (W×F), generating the fused data Z'∈R W×F .

[0046] Preferably, the ISEformer network framework includes a 1×1 convolutional layer; an input embedding layer and multiple stacked encoder layers; each of the encoder layers is composed of a multi-head self-attention mechanism, a feedforward neural network (in which the ISE attention mechanism is introduced) and layer normalization; the ISE attention mechanism is particularly applied to the feedforward neural network part to enhance the model's ability to capture complex dependencies in time series data.

[0047] First, the training samples are reduced in dimension through 1×1 convolution, and then the input embedding layer maps the dimension of the input vector to the predefined model dimension, and then the encoder layer performs feature extraction.

[0048] As a preferred method, after the training is completed, the RUL prediction model is evaluated by the root mean square error RMSE and the score function Score;

[0049] The calculation formula of the root mean square error RMSE is:

[0050]

[0051] The calculation formula of the score function Score is:

[0052]

[0053] Where N is the total number of samples, j represents the jth sample, and y j and They represent the actual RUL value and predicted RUL value of the jth sample respectively,

[0054] Another aspect of the present invention provides a turbofan engine RUL prediction system based on historical data fusion, comprising:

[0055] The data acquisition module is used to collect various sensor data when the turbofan engine is running, and perform preprocessing to obtain multiple time window data with RUL tags;

[0056] The data processing module uses three sliding windows of different scales to process historical data, superimposes all historical time window data before the current moment according to the time dimension, and uses the principal component analysis method to reduce the dimension of the superimposed data. Finally, the data of the three scales are fused to generate training and test samples;

[0057] The model building module is used to train and obtain the RUL prediction model based on the training and test samples using the ISEformer network framework;

[0058] The RUL prediction module is used to input the operating data of the turbofan engine to be tested and output the RUL prediction result of the turbofan engine using the prediction model.

[0059] Yet another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions enabling a computer to execute the above-mentioned turbofan engine RUL prediction method based on historical data fusion.

[0060] Yet another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned turbofan engine RUL prediction method based on historical data fusion.

[0061] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned turbofan engine RUL prediction method based on historical data fusion.

[0062] Beneficial effects: Three sliding windows of different scales are used to process historical data, all historical time window data before the current moment are superimposed according to the time dimension, and the superimposed data is reduced in dimension using principal component analysis. Finally, the data of the three scales are fused to generate training and test samples. The fused data is input into the improved ISEformer network framework. ISEformer introduces the ISE attention mechanism in the feedforward neural network part to enhance the model's ability to capture complex dependencies in time series, thereby achieving more accurate RUL prediction. Through this innovative method of combining historical data fusion with deep learning technology, the present invention significantly improves the RUL prediction performance of turbofan engines under complex working conditions.

[0063] The present invention solves the problem of insufficient utilization of historical data in existing data-driven methods by using a method for predicting the remaining useful life of a turbofan engine based on historical data fusion. Three sliding windows of different scales are used to process historical data, and all historical time window data before the current moment are superimposed according to the time dimension. The superimposed data is reduced in dimension using the principal component analysis method, which effectively reduces data redundancy and extracts key features. Finally, the data of the three scales are fused to generate training and test samples. Furthermore, the improved ISEformer network framework introduces the ISE attention mechanism in the feedforward neural network part to enhance the model's ability to capture complex dependencies in time series. Experimental results show that the accuracy of ISEformer in RUL prediction is significantly better than existing methods, especially under complex working conditions, showing better robustness. This method provides an innovative technical path for the prediction of the remaining useful life of a turbofan engine and has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a flow chart of a turbofan engine RUL prediction method based on historical data fusion of the present invention;

[0065] Figure 2 It is a display diagram of 21 sensor data of FD001 data set in an embodiment of the present invention;

[0066] Figure 3 This is an example diagram of data sliding window processing in an embodiment of the present invention;

[0067] Figure 4 This is an example diagram of historical data fusion with a step size of 1 in an embodiment of the present invention;

[0068] Figure 5 is a structural diagram of ISEformer in an embodiment of the present invention;

[0069] Figure 6 is a structural diagram of a multi-head attention mechanism in an embodiment of the present invention;

[0070] Figure 7 It is a structural diagram of an ISE module in an embodiment of the present invention;

[0071] Figure 8 is a flow chart of model training in an embodiment of the present invention;

[0072] Fig. 9 FD001 is a prediction result diagram of the data set in the embodiment of the present invention;

[0073] Fig.10 This is a diagram of the prediction results of the FD004 data set in an embodiment of the present invention;

[0074] Fig.11This is a prediction result diagram of engine No. 32 in the FD001 data set in an embodiment of the present invention;

[0075] Fig.12 This is a prediction result diagram of engine No. 135 in the FD004 data set in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0077] Combine the following Figures 1 to 12 The present invention describes a turbofan engine RUL prediction method and system based on historical data fusion.

[0078] Embodiment 1: This embodiment provides a turbofan engine RUL prediction method based on historical data fusion, such as Figure 1 As shown, it includes: collecting various sensor data when the turbofan engine is running, and preprocessing them to obtain multiple time window data with RUL labels; using three sliding windows of different scales to process historical data, superimposing all historical time window data before the current moment according to the time dimension, and using the principal component analysis method to reduce the dimension of the superimposed data, and finally, fusing the data of the three scales to generate training and test samples; according to the fused and reduced-dimensional data, using the ISEformer network framework to train and obtain the RUL prediction model; testing through the test set to obtain the predicted RUL result; using RMSE and Score two evaluation indicators to evaluate the performance of the model; according to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

[0079] The above method is described in detail in combination with the data set. The steps of the above method are:

[0080] S1. Collect various sensor data when the turbofan engine is running. In this embodiment, the C-MAPSS simulation data set disclosed by NASA is directly used. The C-MAPSS data set contains four sub-data sets, namely FD001, FD002, FD003, and FD004. In order to verify the applicability of the method of the present invention under different failure modes and operating conditions, the FD001 and FD004 data sets were selected in this embodiment. The FD001 data set contains one failure mode and one operating condition, which is suitable for RUL prediction under simple operating conditions; while the FD004 data set contains two failure modes and six operating conditions, which is suitable for RUL prediction under complex operating conditions. Detailed information of the two data sets is shown in Table 1.

[0081] Table 1 Dataset

[0082] Dataset FD001 FD004 Number of engines in the training set 100 249 Number of test set engines 100 248 Training samples 20631 61249 Test samples 13096 41214 Failure Mode 1 2 Working mode 1 6

[0083] S2. Preprocess the acquired sensor data, including:

[0084] S21. Filter the acquired sensor data and remove the data that remain unchanged during the cycle, because these unchanged features have limited contribution to the RUL prediction of the turbofan engine and cannot provide meaningful degradation information. Figure 2 As shown in the figure, some of the 21 sensor data remain unchanged during the cycle. Figure 2 The sensor data numbered 1, 5, 6, 10, 16, 18, and 19 are used, and the remaining 14 sensor data and 3 operating conditions are used as input features.

[0085] S22, normalizing the filtered sensor data; the specific method is:

[0086] The maximum-minimum normalization is used to scale the sensor data of each channel to the range [0,1]. The normalization formula is:

[0087]

[0088] In the formula, represents the normalized data of the sth sensor at the tth time step, max{x s} and min{x s} respectively represent the maximum and minimum values ​​in the sth sensor; eps = 10 -12 The main purpose is to prevent the denominator from being zero and avoid calculation errors.

[0089] S23, using a sliding window to process the normalized data to obtain multiple time window data; specifically:

[0090] like Figure 3As shown, for the entire cycle of a turbofan engine from operation to failure, assuming that the total number of engine operation cycles is T, the window size is W, and the sliding step size is p, multiple samples are obtained, and the size of each sample is W×F, where F is the number of features, and one sample is a time window data.

[0091] S24, perform RUL label processing on multiple time window data; for the n+1th time window data, the corresponding RUL label calculation formula is:

[0092] RUL=TWn*p;

[0093] Since the degradation characteristics of the engine are not obvious in the early stage of operation, the sample labels are processed using piecewise linear functions. When the RUL of the turbofan engine is greater than 125, its value is fixed to 125, indicating that the engine is in a healthy state. The formula is:

[0094]

[0095] The obtained multiple time window data with RUL labels are used as the training data set for the RUL prediction model.

[0096] S3, use three sliding windows of different scales to process historical data, superimpose all historical time window data before the current moment according to the time dimension, and use principal component analysis to reduce the dimension of the superimposed data, effectively reducing data redundancy and extracting key features. Finally, the data of the three scales are fused to generate training and test samples. Including:

[0097] S31. When the step size is 1, a forward sliding window is used. Specifically, the sliding window starts from the beginning of the data sequence and slides in sequence according to the step size 1 until it reaches the current time to be predicted. Each time it slides, the data in the window will be collected and used as the input data of a time window.

[0098] When the step size is 2, a reverse sliding window is used. Specifically, the sliding window goes back from the current moment, slides through a window with a step size of 2, and collects data before the current moment. Reverse sliding can ensure that it can slide to the current prediction window, while if forward sliding is used, it may not slide to the current prediction window.

[0099] When the step size is 3, it is similar to the reverse sliding window with a step size of 2. It also traces back from the current moment and slides through a window with a step size of 3 to collect historical data in a larger range.

[0100] For the current nth time step, the current time window shape is (W, F), and one dimension is added to change the shape to (W, F, 1);

[0101] S32, for the nth time step of step S31, there are n-1 time window data at this moment. When the step length is 1, a total of n time windows can be slid out, and the shape of each time window is (W, F). Each time window is added with one dimension, and the shape becomes (W, F, 1). Then, each time window data is superimposed along the third dimension, and the shape becomes (W, F, n); when the step length is 2, a total of n time windows can be slid out. time windows, the shape becomes Similarly, when the step length is 3, a total of time windows, the shape becomes

[0102] S33, the shape obtained is (W, F, n), and The fusion time window is reconstructed into (W×F,n), and Then PCA is used to reduce the dimension and change the shape to (W×F,1). Then the three windows are reintegrated into (W,F). Finally, the three windows containing information of different scales are fused into (W,F,3) as the input of the ISEformer network.

[0103] Take the step size of 1 as an example, Figure 4 As shown, the specific method of dimensionality reduction through PCA is:

[0104] S331, let input data X∈R W×F×n , first reconstruct the three-dimensional data into a two-dimensional matrix X r ∈R (W×F)×n , calculate the covariance matrix C of the two-dimensional matrix, the calculation formula is:

[0105]

[0106] S332, perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalue λ i and the corresponding eigenvector v i , the calculation formula is:

[0107] Cv i =λ i v i ;

[0108] S333, select the eigenvector v corresponding to the largest eigenvalue i Projecting onto the principal component, we get the dimension-reduced data Z, whose shape is (W×F,1) and the calculation formula is:

[0109] Z=X r v1;

[0110] S334, finally Z∈R (W×F)×1 The shape is reconstructed into (W×F), generating the fused data Z'∈R W×F .

[0111] S4. Input the processed samples into the constructed ISEformer model for training, and then test it with the test set to obtain the predicted RUL results.

[0112] The ISEformer model proposed in the present invention is as follows: Figure 5 As shown in the figure, it includes a 1×1 convolutional layer, an input embedding layer, and multiple stacked encoder layers; each encoder layer consists of a multi-head self-attention mechanism, a feedforward neural network (in which the ISE attention mechanism is introduced), and layer normalization. The ISE attention mechanism is particularly applied to the feedforward neural network part to enhance the model's ability to capture complex dependencies in time series data. The input embedding layer maps the input state monitoring data to a d through a fully connected layer. model dimensional vector, preparing for the subsequent feature extraction process.

[0113] (1) 1×1 convolution

[0114] By using a 1×1 convolution kernel, the number of channels of the input feature map can be reduced from 3 to 1 while keeping the spatial dimension unchanged, thereby achieving compression of the number of channels.

[0115] (2) Positional encoding

[0116] Since ISEformer does not contain a sequence structure like a recurrent neural network, but relies entirely on the attention mechanism to analyze sequence data, it cannot directly capture the temporal information in the sequence. For this reason, in addition to the input embedding layer, a position encoder is usually attached to provide the relative position information of the sequence by using sine and cosine functions of different frequencies to make up for this deficiency.

[0117] (3) Encoder layer

[0118] The encoder layer consists of a multi-head attention mechanism (Multi-HeadAttention), a feedforward neural network (FFN) and layer normalization (LayerNormalization).

[0119] The self-attention mechanism is an improvement based on the traditional attention mechanism. It no longer relies on external information, but captures the internal dependencies in the sequence through the characteristics of the model itself. It measures the correlation between features by calculating the similarity between feature vectors, which can effectively solve the problem of long-distance dependency. The calculation formula of the self-attention mechanism is as follows:

[0120]

[0121] Q=X f W Q ;

[0122] K=X f W K ;

[0123] V=X f W V ;

[0124] Among them, Q is the query matrix; K is the key matrix; V is the value matrix. These three matrices are composed of the input matrix X f , respectively, with the corresponding weight matrix W Q , W K , W V Multiply them together to get d k is the dimension of Q, K, and V.

[0125] The multi-head attention mechanism is an improvement on the self-attention mechanism. It divides the model into multiple independent attention heads to process information in different subspaces in parallel, such as Figure 6 As shown. Each attention head can focus on different aspects of the input features and capture dependencies in more dimensions. Ultimately, the multi-faceted information obtained by these attention heads will be spliced ​​together and merged into the final attention output through linear transformation. The calculation formula of the multi-head attention mechanism is as follows:

[0126] MulitHead(Q,K,V)=Concat(h1,h2,h3)·W;

[0127]

[0128] in, and are the weight matrices of the Kth attention head Q, K and V respectively; W is the multi-head attention weight matrix; m is the number of attention heads; is the calculation result of the Kth attention head; the splicing function represents the calculation results used to splice each attention head.

[0129] The main function of FFN is to further learn and process features through nonlinear transformation. FFN usually consists of two linear layers and a nonlinear activation function, and the formula is as follows:

[0130] F FC =Relu(0,xw1+b1)w2+b2;

[0131] Where: w1, w2 and b1, b2 are the weight and bias parameters trained in FFN.

[0132] Layer normalization is introduced before and after FFN. The role of layer normalization is to normalize the activation values ​​of all neurons in the middle layer of the network, thereby stabilizing the training process, accelerating model convergence, and improving model performance. The specific calculation formula is as follows:

[0133]

[0134] Where: is the i-th element in the input data of the l-th layer; ε is a small positive value, and γ and β are learnable scaling and translation parameters.

[0135] The present invention improves the SE attention module, which is called the ISE module. Figure 7 As shown in the figure, on the basis of the original Global Average Pooling (GAP), Global Max Pooling (GMP) is added. Combining GAP and GMP can more comprehensively retain the effective information in the feature map. GAP captures the overall situation of the feature map by calculating the average value, while GMP highlights the significant features in the feature map by extracting the maximum value. The results of the two poolings are respectively subjected to the excitation operation, and finally fused, and the obtained results are multiplied with the original feature map to obtain the weighted feature map

[0136] (4) Model training parameter setting

[0137] Table 2 shows the network model parameter settings for processing the FD001 dataset. For the FD004 dataset, the main difference is the way the sliding window is handled. In the FD004 dataset, the size of the sliding window is set to 15, so the parameter of the final fully connected layer is 480 / 1.

[0138] Table 2 Parameter settings

[0139]

[0140] (5) Training process

[0141] The flowchart of the model training in this embodiment is as follows Figure 8 As shown. This example uses Pycharm to train and test the model, and the experimental environment is Pytorch 2.0.1. All programs are run on a computer with the following configuration: Windows 11, AMD Ryzen 76800H, NVIDIARTX 3060, 16GB RAM.

[0142] (6) Results and analysis

[0143] The prediction results of the two sub-datasets are as follows Fig. 9 and Fig.10 As shown. In order to better observe the prediction results, the engine numbers are rearranged in descending order of remaining life. The horizontal axis represents the re-arranged engine test sample number, and the vertical axis represents the RUL value. The red dotted line represents the predicted RUL, and the blue solid line represents the actual RUL. It can be seen that on the two sub-datasets, the model makes reasonable predictions on the RUL of most samples. However, on some samples, there is a large difference between the predicted RUL and the actual RUL, especially in the area with large RUL values. For these samples, the RUL predicted by the model fluctuates greatly, which may be due to the insufficient capture of the early degradation characteristics by the model, resulting in increased errors when the remaining life of the sample is longer. In addition, as the RUL decreases, the prediction accuracy of the model improves, showing a good trend capture ability. In most cases, the red dotted line and the blue solid line are basically consistent, indicating that the model can effectively predict the remaining service life under different fault conditions.

[0144] In order to verify the effectiveness of the method proposed in this invention, the RUL prediction model is evaluated by the root mean square error RMSE and the score function Score;

[0145] The calculation formula of the root mean square error RMSE is:

[0146]

[0147] The calculation formula of the score function Score is:

[0148]

[0149] Where N is the total number of samples, j represents the jth sample, and y j and They represent the actual RUL value and predicted RUL value of the jth sample respectively,

[0150] To eliminate the randomness in the experimental process, each experiment was repeated ten times, and the average index was recorded. Table 3 provides the RMSE and Score comparison results of the method of the present invention on the FD001 and FD004 data sets. The present invention can better solve the problem of aircraft engine remaining life prediction in complex scenarios and has high application value.

[0151] Table 3 Evaluation results

[0152]

[0153] Fig.11 and Fig.12The comparison between the RUL prediction results and the actual values ​​of engine id32 in the FD001 dataset and engine id135 in the FD004 dataset is shown respectively. In the piecewise linear approximation method, when the life value in the early stage is set to be constant, the prediction results of the model are also relatively stable. When the remaining useful life of the engine begins to show a downward trend, the predicted value of the model also decreases. This intuitively verifies the effectiveness of the prediction model. However, when the engine life transitions from a constant value to a linear decline, the prediction results will have a certain degree of deviation. However, as the engine life gradually decreases, it can be observed that the model's prediction of RUL gradually improves, especially at the end of the life. This trend is more obvious. This is because when the engine is close to failure, the degradation characteristics become more significant, and the model can better capture these characteristics, thereby providing more accurate predictions. This phenomenon is also in line with expectations, that is, when the remaining life is short, the accuracy of the prediction should be improved.

[0154] in conclusion:

[0155] The present invention proposes a method for predicting the remaining useful life of a turbofan engine based on historical data fusion, which solves the problem of insufficient utilization of historical data in existing data-driven methods. By using three sliding windows of different scales to process historical data, all historical time window data before the current moment are superimposed according to the time dimension, and the superimposed data is reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples. By using sliding windows of different time scales, the dynamic change characteristics of the system in the short and long term can be effectively captured, and the operating status of the engine can be fully reflected. At the same time, the ISEformer network uses its self-attention mechanism to accurately capture the dependencies in the time series. Experimental results show that the method of the present invention has high accuracy in RUL prediction, and shows obvious advantages compared with existing methods. This method provides an innovative technical path for the prediction of the remaining useful life of a turbofan engine, and has important application value.

[0156] Embodiment 2: This embodiment provides a turbofan engine RUL prediction system based on historical data fusion, including:

[0157] The data acquisition module is used to collect various sensor data when the turbofan engine is running, and perform preprocessing to obtain multiple time window data with RUL tags;

[0158] The data processing module uses three sliding windows of different scales to process historical data, superimposes all historical time window data before the current moment according to the time dimension, and uses the principal component analysis method to reduce the dimension of the superimposed data. Finally, the data of the three scales are fused to generate training and test samples;

[0159] The model building module is used to train and obtain the RUL prediction model based on the training and test samples using the ISEformer network framework;

[0160] The RUL prediction module is used to input the operating data of the turbofan engine to be tested and output the RUL prediction result of the turbofan engine using the prediction model.

[0161] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium, on which computer instructions are stored, the computer instructions causing the computer to execute a turbofan engine RUL prediction method based on historical data fusion, the method comprising the following steps:

[0162] Collect various sensor data when the turbofan engine is running, and pre-process them to obtain multiple time window data with RUL labels;

[0163] Three sliding windows of different scales are used to process historical data. All historical time window data before the current moment are superimposed in the time dimension, and the superimposed data are reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples.

[0164] According to the training and testing samples, the ISEformer network framework is used to train and obtain the RUL prediction model;

[0165] According to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

[0166] Embodiment 4: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor may call the logic instructions in the memory to execute a turbofan engine RUL prediction method based on historical data fusion, the method comprising the following steps:

[0167] Collect various sensor data when the turbofan engine is running, and pre-process them to obtain multiple time window data with RUL labels;

[0168] Three sliding windows of different scales are used to process historical data. All historical time window data before the current moment are superimposed in the time dimension, and the superimposed data are reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples.

[0169] According to the training and testing samples, the ISEformer network framework is used to train and obtain the RUL prediction model;

[0170] According to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

[0171] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0172] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored in a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a turbofan engine RUL prediction method based on historical data fusion, which includes the following steps:

[0173] Collect various sensor data when the turbofan engine is running, and pre-process them to obtain multiple time window data with RUL labels;

[0174] Three sliding windows of different scales are used to process historical data. All historical time window data before the current moment are superimposed in the time dimension, and the superimposed data are reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples.

[0175] According to the training and testing samples, the ISEformer network framework is used to train and obtain the RUL prediction model;

[0176] According to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

[0177] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0178] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A turbofan engine RUL prediction method based on historical data fusion, characterized in that: include: Collect various sensor data when the turbofan engine is running, and pre-process them to obtain multiple time window data with RUL labels; Three sliding windows of different scales are used to process historical data. All historical time window data before the current moment are superimposed in the time dimension, and the superimposed data are reduced in dimension using the principal component analysis method. Finally, the data of the three scales are fused to generate training and test samples. According to the training and testing samples, the ISEformer network framework is used to train and obtain the RUL prediction model; According to the operating data of the turbofan engine to be tested, the prediction model is used to output the RUL prediction result of the turbofan engine.

2. The turbofan engine RUL prediction method based on historical data fusion according to claim 1, characterized in that: The method collects various sensor data when the turbofan engine is running and performs preprocessing; including: Collect various sensor data when the turbofan engine is running; The acquired sensor data is screened to remove the data that remain unchanged during the cycle; Normalize the filtered sensor data; The normalized data is processed using a sliding window to obtain multiple time window data; Perform RUL label processing on data in multiple time windows.

3. The turbofan engine RUL prediction method based on historical data fusion according to claim 2 is characterized in that: The screened sensor data is normalized; the specific method is: The maximum-minimum normalization is used to scale the sensor data of each channel to the range [0,1]. The normalization formula is: In the formula, represents the normalized data of the sth sensor at the tth time step, max{x s } and min{x s } respectively represent the maximum and minimum values ​​in the sth sensor; eps = 10 -12 .

4. The turbofan engine RUL prediction method based on historical data fusion according to claim 2 is characterized in that: The normalized data is processed using a sliding window to obtain multiple time window data; specifically: For the entire cycle of a turbofan engine from operation to failure, assuming that the total number of engine operation cycles is T, the window size is W, and the sliding step is p, multiple samples are obtained, each of which is W×F in size, where F is the number of features and one sample is a time window data.

5. The turbofan engine RUL prediction method based on historical data fusion according to claim 4 is characterized in that: The RUL label processing of the data in multiple time windows is specifically as follows: For the n+1th time window data, the corresponding RUL label calculation formula is: RUL=TWn*p; The labels of the time window data are processed using a piecewise linear function, and the formula is: The obtained multiple time window data with RUL labels are used as the training data set for the RUL prediction model.

6. The turbofan engine RUL prediction method based on historical data fusion according to claim 4, characterized in that: The method uses three sliding windows of different scales to process historical data, superimposes all historical time window data before the current moment according to the time dimension, and uses the principal component analysis method to reduce the dimension of the superimposed data. Finally, the data of the three scales are fused to generate training and test samples; specifically: For the nth time window, there are n-1 time window data at this moment, and the shape of each time window is (W, F). Each time window is increased by one dimension, and the shape becomes (W, F, 1). Then, each time window data is superimposed along the third dimension, and the shape becomes (W, F, n). The obtained fusion time window with the shape of (W, F, n) is reconstructed into (W×F, n), and then the dimension is reduced by principal component analysis to change the shape to (W×F, 1). Finally, the window is reintegrated into (W, F) as the input of the ISEformer network.

7. The turbofan engine RUL prediction method based on historical data fusion according to claim 6, characterized in that: The specific method of performing dimensionality reduction by principal component analysis is as follows: Assume input data X∈R W×F×n , first reconstruct the three-dimensional data into a two-dimensional matrix X r ∈R (W×F)×n , calculate the covariance matrix C of the two-dimensional matrix, the calculation formula is: Perform eigenvalue decomposition on the covariance matrix C and obtain the eigenvalue λ i and the corresponding eigenvector v i , the calculation formula is: Cv i =λ i v i ; Select the eigenvector v corresponding to the largest eigenvalue i Projecting onto the principal component, we get the dimension-reduced data Z, which has a shape of (W×F,1) and is calculated as: Z=X r v1; Finally, Z∈R (W×F)×1 The shape is reconstructed into (W×F), generating the fused data Z'∈R W×F .

8. The turbofan engine RUL prediction method based on historical data fusion according to claim 1, characterized in that: The ISEformer network framework includes an input embedding layer and multiple stacked encoder layers; each encoder layer consists of a multi-head self-attention mechanism, a feedforward neural network introducing the ISE attention mechanism, and layer normalization; First, the input embedding layer maps the dimension of the input vector to the predefined model dimension, and then the encoder layer performs feature extraction.

9. The turbofan engine RUL prediction method based on historical data fusion according to claim 8, characterized in that: After the training is completed, the RUL prediction model is evaluated by the root mean square error RMSE and the score function Score; The calculation formula of the root mean square error RMSE is: The calculation formula of the score function Score is: Where N is the total number of samples, j represents the jth sample, and y j and They represent the actual RUL value and predicted RUL value of the jth sample respectively, 10. A turbofan engine RUL prediction system based on historical data fusion, characterized in that: include: The data acquisition module is used to collect various sensor data when the turbofan engine is running, and perform preprocessing to obtain multiple time window data with RUL tags; The data processing module is used to process historical data using sliding windows of three different scales, superimpose all historical time window data before the current moment according to the time dimension, and use the principal component analysis method to reduce the dimension of the superimposed data. Finally, the data of the three scales are fused to generate training and test samples; The model building module is used to train and obtain the RUL prediction model based on the training and test samples using the ISEformer network framework; The RUL prediction module is used to input the operating data of the turbofan engine to be tested and output the RUL prediction result of the turbofan engine using the prediction model.

Images