Deep learning-based turbofan engine residual service life prediction method and system

Abstract

The invention discloses a deep learning model-based turbofan engine remaining service life prediction method and system, and the method comprises the steps: obtaining operation data of a turbofan engine during operation, carrying out the preprocessing, and constructing a training data set; constructing a dual-channel initial learning framework, and performing training by adopting a TCN-BiLSTM network framework in which a CBAM attention mechanism is introduced according to the data set; according to the other path, an average value and a degradation trend coefficient of operation data are extracted through manual features to serve as auxiliary input of the prediction model; obtaining a trained prediction model; and according to operation data of a to-be-detected turbofan engine, predicting the remaining service life of the turbofan engine by using the prediction model. According to the method, the TCN-BiLSTM hybrid model based on the improved CBAM attention mechanism is adopted to process complex time sequence data, and meanwhile, the average value and the degradation trend coefficient which are extracted manually serve as auxiliary input, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial detection, and in particular to a method and system for predicting the remaining useful life of a turbofan engine based on deep learning. Background Art

[0002] As a core equipment in the aviation field, the reliability of turbofan engines is directly related to flight safety and economic benefits. Therefore, accurate prediction of the remaining useful life (RUL) of turbofan engines is crucial for airlines, which not only helps to avoid risks but also optimizes aircraft maintenance and operation plans.

[0003] In recent years, with the rapid development of data processing technology, especially the advancement of machine learning technology, data-driven methods have become a hot topic of research. Data-driven methods collect a large amount of operating data by installing sensors on equipment, and then analyze historical operating data to identify fault modes and predict RUL. Data-driven methods include statistical methods and machine learning methods. Statistical methods use mathematical and statistical theories to analyze and process data. However, statistical methods require a large amount of high-quality data for accurate modeling, and have high requirements for data integrity and accuracy. In contrast, machine learning methods have greater flexibility and adaptability, can automatically learn and improve from a large amount of data, and are the mainstream methods for current fault diagnosis and RUL prediction.

[0004] The remaining useful life prediction methods based on machine learning are mainly divided into traditional machine learning and deep learning. The remaining useful life prediction methods based on traditional machine learning mainly include neural networks, support vector machines, random forests, etc. However, since the degradation data collected by sensors is often complex and nonlinear time domain data, compared with traditional machine learning methods, deep learning can automatically learn the potential features in the data, thereby more effectively processing nonlinear data.

[0005] Although the application of deep learning in the RUL prediction of turbofan engines has made significant progress, there are still some problems. For example, the working environment of turbofan engines is harsh, and the data generated during their operation is very complex, involving various parameters monitored by multiple sensors at the same time. It is difficult for a single network model to extract effective features from them, resulting in the inability to accurately predict the remaining service life of turbofan engines. In time series data, the data consists of time steps and features, and it is necessary to pay attention to the contribution of time steps and features to RUL prediction at the same time. Although deep learning models can automatically extract and learn useful features, some feature engineering processing of the data is still required in practical applications. Summary of the invention

[0006] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0007] One object of the present invention is to provide a method for predicting the remaining useful life of a turbofan engine based on a deep learning model, which uses a TCN-BiLSTM hybrid model based on an improved CBAM attention mechanism to process complex time series data, and also uses manually extracted average values ​​and degradation trend coefficients as auxiliary inputs to improve the accuracy of the prediction.

[0008] Another object of the present invention is to provide a turbofan engine remaining service life prediction system based on a deep learning model.

[0009] In order to achieve the above-mentioned purpose, the present invention provides a method for predicting the remaining service life of a turbofan engine based on a deep learning model, comprising:

[0010] Obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0011] Construct a dual-channel initial learning framework, one of which is trained based on the data set using the TCN-BiLSTM network framework that introduces the CBAM attention mechanism; the other path uses manual feature extraction to extract the average value and degradation trend coefficient of the operating data as auxiliary input of the prediction model; and obtain a trained prediction model;

[0012] According to the operating data of the turbofan engine to be tested, the remaining service life of the turbofan engine is predicted using the prediction model.

[0013] A further preferred technical solution of the present invention is that the operation data of the turbofan engine is obtained and preprocessed to construct a data set; including:

[0014] Use various sensors to collect various sensor data during the operation of the turbofan engine;

[0015] Perform data screening on the sensor data to filter out the data that remain unchanged during the cycle;

[0016] Normalize the remaining data;

[0017] The sliding window is used to process the data, and the RUL label is processed at the same time to construct the data set.

[0018] Preferably, the normalizing the remaining data comprises:

[0019] The maximum-minimum normalization is used to scale the sensor data of each channel to the range [0,1]. The normalization formula is:

[0020]

[0021] 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 .

[0022] Preferably, the method of processing the data by using a sliding window and processing the URL labels to construct a data set comprises:

[0023] 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 S, multiple samples are obtained, each of which is w×n in size, where n is the number of features;

[0024] For the i+1th sample, the corresponding RUL label calculation formula is:

[0025] RUL=Twi*S

[0026] The sample labels are processed using a piecewise linear function, the formula is:

[0027]

[0028] A data set is constructed using the obtained samples and the labels corresponding to each sample.

[0029] Preferably, the method for constructing the dual-channel initial learning framework is:

[0030] TCN and BiLSTM are used for feature extraction, and after the BiLSTM network, a network framework of the CBAM module is added as one of the paths of the dual-channel initial learning framework;

[0031] The average value and degradation trend coefficient of the running data are extracted by manual features, and the average values ​​and degradation trend coefficients of all features are combined to form a feature vector as another path;

[0032] The feature maps obtained from the two paths are flattened, fully connected and fused respectively, and finally passed through a fully connected layer to output the predicted RUL value.

[0033] Preferably, the method of extracting the average value and degradation trend coefficient of the operating data by manual feature extraction, combining the average values ​​and degradation trend coefficients of all features to form a feature vector comprises:

[0034] For a sliding window of data X = [x1, x2…x n], its shape is (w,n), w and n represent the size of the sliding window and the number of sensors, that is, the number of features, and the average value is calculated as:

[0035]

[0036] Where x n is the nth feature in a time step; a set of average data is obtained by calculation

[0037] The degradation trend coefficient is obtained by linear regression of each feature, and the calculation formula is:

[0038] y n =k n x n +b

[0039] Where k is the regression coefficient, which is defined as the regression trend coefficient; y n is the predicted value, which is the nth feature in a time step;

[0040] By calculating, we can get a set of degradation trend coefficients K = [k1, k2…k n ];

[0041] Finally, the average values ​​and degradation trend coefficients of all features are combined to form a feature vector h = [μ1, μ2…μ n ,k1,k2…k n ], as auxiliary input of the prediction model.

[0042] As a preferred method, the training results of the trained prediction model are evaluated using two evaluation indicators: root mean square error RMSE and score function Score;

[0043] The calculation formula of the root mean square error RMSE is:

[0044]

[0045] The calculation formula of the score function Score is:

[0046]

[0047] Where N is the total number of samples, i represents the i-th sample, and y i and Respectively represent the actual RUL value and the predicted RUL value of the i-th sample,

[0048] Another aspect of the present invention provides a system for predicting the remaining useful life of a turbofan engine based on a deep learning model, comprising:

[0049] A data processing module is used to obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0050] The prediction model building module is used to build a dual-channel initial learning framework and train it according to the training data set to obtain a trained prediction model;

[0051] The RUL prediction module is used to input the operating data of the turbofan engine to be tested and use the prediction model to predict the remaining service life of the turbofan engine.

[0052] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned method for predicting the remaining useful life of a turbofan engine based on a deep learning model.

[0053] 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 through the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned method for predicting the remaining useful life of a turbofan engine based on a deep learning model.

[0054] 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 method for predicting the remaining useful life of a turbofan engine based on a deep learning model.

[0055] Beneficial effects: The method and system for predicting the remaining useful life of a turbofan engine based on a deep learning model of the present invention proposes a turbofan engine remaining useful life (RUL) prediction model based on a time-domain convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM) combined with a convolutional block attention module (CBAM). By processing complex time series data, the model performs well in extracting data features and capturing degradation processes. The TCN-BiLST-CBAM hybrid model designed by the present invention performs well in capturing local patterns and long-term dependencies in time series data. TCN can effectively process time series data, while BiLSTM simultaneously obtains past and future information through its bidirectional structure.

[0056] The improved convolutional block attention module (CBAM) is introduced after the BiLSTM network to further enhance the model's ability to capture key degradation information in time series. Experimental results show that the model outperforms multiple existing methods on NASA's C-MAPSS dataset, especially showing excellent performance when processing complex turbofan engine degradation data. The channel and spatial attention mechanism of CBAM helps the model identify important feature channels and key time periods, thereby improving the accuracy of prediction and the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a method for predicting the remaining useful life of a turbofan engine based on a deep learning model of the present invention;

[0058] Figure 2 is a schematic diagram of a residual module in TCN in Embodiment 1 of the present invention;

[0059] Figure 3 This is a schematic diagram of the dilated causal convolution in Example 1 of the present invention;

[0060] Figure 4 It is a BiLSTM structure diagram in Example 1 of the present invention;

[0061] Figure 5 is a schematic diagram of the CBAM structure in Example 1 of the present invention;

[0062] Figure 6 is a structural diagram of a channel attention module in Embodiment 1 of the present invention;

[0063] Figure 7 is a structural diagram of a spatial attention module in Embodiment 1 of the present invention;

[0064] Figure 8 This is an example diagram of manual feature extraction in Example 1 of the present invention;

[0065] Fig. 9 is a structural diagram of a hybrid model in Example 1 of the present invention;

[0066] Fig.10 It is the 21 sensor data of the FD001 data set in Example 1 of the present invention;

[0067] Fig.11 This is an example of data sliding window processing in Embodiment 1 of the present invention;

[0068] Fig.12 are the prediction results of the four subsets in Example 1 of the present invention;

[0069] in, Fig.12 (a) is the prediction result of the sub-dataset FD001. Fig.12(b) is the prediction result of the sub-dataset FD002. Fig.12 (c) is the prediction result of sub-dataset FD003. Fig.12 (d) is the prediction result of sub-dataset FD004;

[0070] Fig.13 is the error distribution of the four subsets in Example 1 of the present invention;

[0071] Fig.14 is the RUL prediction result of the tests FD001 to FD004 in Example 1 of the present invention;

[0072] in, Fig.14 (a) is the prediction result of the test engine unit 76 in FD001, Fig.14 (b) is the prediction result of 176 in FD002. Fig.14 (c) is the prediction result of 99 in FD003. Fig.14 (d) is the prediction result of 213 in FD004. DETAILED DESCRIPTION

[0073] 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.

[0074] Combine the following Figure 1-Figure 14 The present invention describes a method and system for predicting the remaining useful life of a turbofan engine based on a deep learning model.

[0075] Embodiment 1: This embodiment provides a method for predicting the remaining service life of a turbofan engine based on a deep learning model. Figure 1 As shown, the specific steps are:

[0076] S1. Use various sensors to collect various sensor data when the turbofan engine is running.

[0077] S2. Preprocess the sensor data and build a training data set:

[0078] S21. Filter the sensor data. Some sensor data remain unchanged during the cycle, which is meaningless for RUL prediction and needs to be filtered out.

[0079] S22, normalize the remaining data; specifically:

[0080] The maximum-minimum normalization is used to scale the sensor data of each channel to the range [0,1]. The normalization formula is:

[0081]

[0082] 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 of the design is to prevent the denominator from being zero and avoid calculation errors;

[0083] S23, using a sliding window to process the data, and at the same time perform RUL label processing to construct a data set; specifically:

[0084] 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 S, multiple samples are obtained, each of which is w×n in size, where n is the number of features;

[0085] For the i+1th sample, the corresponding RUL label calculation formula is:

[0086] RUL=Twi*S

[0087] Since the degradation characteristics of the engine are not obvious in the early stage of operation, the label of the sample is processed using a piecewise linear function, the formula is:

[0088]

[0089] A data set is constructed using the obtained samples and the labels corresponding to each sample.

[0090] S3. Build a dual-channel initial learning framework:

[0091] S31. The data is divided into two paths and input into the network. The network framework of one path uses a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM) for feature extraction. TCN takes advantage of the dilated convolution structure and can effectively expand the receptive field when processing time series data, while avoiding the gradient vanishing problem in traditional recurrent neural networks. The output of the TCN layer is then fed into the BiLSTM network. BiLSTM can capture both past and future information of the time series through its bidirectional structure, which is particularly important for predicting the future state of turbofan engines. After the BiLSTM network, an improved convolutional block attention module (CBAM) is added to further enhance the model's understanding and utilization of the global dependency of the time series.

[0092] The TCN network consists of a series of residual modules, and the core parts of the residual module are causal convolution and dilated convolution. In causal convolution, the output at a time point only depends on the input data at that time point and before, and does not use the data after that time point. However, simply increasing the number of network layers or expanding the size of the convolution kernel to broaden the receptive field of information will significantly increase the computational cost of the network. To solve this problem, a strategy of inserting blank values ​​in the convolution kernel is introduced. This method is called dilated convolution. Dilated convolution can effectively expand the receptive field of the network layer, so that nodes in higher layers can cover a wider range of input information, thereby adding more historical data without additional computational burden. For time series data x∈R n And a filter f:{0,1…,k-1}→R, the dilated convolution calculation formula is as follows:

[0093]

[0094] In the formula, d is the dilation factor, k is the convolution kernel size, and sd·i represents the past direction of the time series. Figure 3 As shown, it is an example diagram of a three-layer dilated causal convolution, where the convolution kernel size d is 3, and the dilation factor sizes d are 1, 2, and 4 respectively.

[0095] The structure of the residual block in TCN is as follows Figure 2As shown. The residual block consists of dilated causal convolution, weight normalization, Relu activation function, and Dropout. The input first passes through the dilated causal convolution to ensure that the model's prediction depends only on past and current inputs. The convolution kernel size k and dilation factor size d of the dilated causal convolution in each residual block are the same. This is followed by weight normalization, which helps the network converge faster during training. Dropout reduces the model's dependence on specific training samples by randomly discarding a part of the neurons in the neural network during training, which can effectively prevent the overfitting problem of deep neural networks. At the same time, the input passes through a 1x1 convolution layer, which is mainly used to adjust the dimension of the signal to match the input and output to ensure that residual connections can be made. Finally, the output of this 1x1 convolution is added to the original input to form a residual connection, and the output is the result of the entire residual module. The TCN network is composed of many residual blocks stacked together. Figure 3 The left side shows a TCN network consisting of n residual modules stacked together, where the convolution kernel size in each residual block is k and the dilation factor size in the nth residual block is 2. n-1 .

[0096] BiLSTM combines two independent LSTM networks, one for processing the forward sequence and the other for processing the reverse sequence, so as to obtain both past and future context information. Figure 4 shown.

[0097] CBAM is a simple and effective feed-forward convolutional neural network attention module. Figure 5 The structure diagram of CBAM shown in the figure contains a channel attention module and a spatial attention module. This structural design enables CBAM to be quickly combined with existing models, saving parameters and computing resources.

[0098] Channel attention module such as Figure 6 As shown, the channel attention module obtains the corresponding degraded feature vector from the input degraded feature map through average pooling and maximum pooling. The feature vector after global maximum pooling and average pooling is then input into a shared fully connected layer, which is used to learn the attention weights of each channel. Since the fully connected layer disperses the input, spatial information is lost. In contrast, the 1×1 convolutional layer retains the spatial structure of the input feature map while performing inter-channel feature fusion, thereby maintaining the integrity of the spatial information. This embodiment replaces the two fully connected layers with a 1×1 convolutional layer. Through learning, the network can adaptively determine which channels are more important for the current task. The specific calculation formula of the channel attention module is as follows:

[0099]

[0100] Among them, F represents the input feature map, P max and P avg represents the result after global maximum pooling and global average pooling, C represents a 1×1 convolution operation, and σ is a sigmoid activation function. c (F) is multiplied with the original feature map F to obtain the weighted feature map F'.

[0101] The spatial attention module is as follows Figure 7 As shown in the figure, the spatial attention module first uses global maximum pooling (MaxPooling) and global average pooling (Average Pooling) to process the input feature map along the channel axis. Next, the two feature maps are spliced ​​together along the spatial dimension, and then the convolution kernel is used to extract features to generate a spatial attention map. The final spatial attention map is processed by the sigmoid activation function to obtain the attention weight map, and finally the attention weight map is multiplied with the original feature map to enhance the model's attention to important spatial areas. The calculation formula of the spatial attention module is as follows:

[0102] A s (F′)=σ(f conv ([maxpool(F′),avgpool(F′)]))

[0103] Where F' represents the feature map after CAM, where f conv Convolution operation, maxpool and avgpool are maximum pooling and average pooling respectively. Finally, F' and A are combined s (F') is multiplied to obtain the final feature map.

[0104] S32, another path uses the mean value and degradation trend coefficient of the data extracted by manual features as auxiliary input to enhance the prediction ability of the model; Figure 8 As shown, the normalized raw data is manually extracted for features, and the extracted features are standardized. The features extracted by this method will be used as auxiliary input to improve the learning ability of the model. Specifically:

[0105] For a sliding window of data X = [x1, x2…x n ], its shape is (w,n), w and n represent the size of the sliding window and the number of sensors, that is, the number of features, and the average value is calculated as:

[0106]

[0107] Where x n is the nth feature in a time step; a set of average data is obtained by calculation:

[0108]

[0109] The degradation trend coefficient is obtained by linear regression of each feature, and the calculation formula is:

[0110] y n =k n x n +b

[0111] Where k is the regression coefficient, which is defined as the regression trend coefficient; y n is the predicted value, which is the nth feature in a time step;

[0112] By calculating, we can get a set of degradation trend coefficients K = [k1, k2…k n ];

[0113] Finally, the average values ​​and degradation trend coefficients of all features are combined to form a feature vector h = [μ1, μ2…μ n ,k1,k2…k n ], as auxiliary input of the prediction model.

[0114] S33, build a dual-channel hybrid model, flatten the feature maps obtained by the two paths, and then fuse them after full connection, and finally output the predicted RUL value after passing through a fully connected layer, specifically:

[0115] like Fig. 9 The figure shows the structure of the hybrid model. First, the raw data is processed by a sliding window operation to form a shape suitable for the model input. Then, the data is divided into two paths to enter the network. One path extracts features through a temporal convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM). TCN takes advantage of the dilated convolution structure and can effectively expand the receptive field when processing time series data, while avoiding the gradient vanishing problem in traditional recurrent neural networks. The output of the TCN layer is then fed into the BiLSTM network. BiLSTM can capture both past and future information of the time series through its bidirectional structure, which is particularly important for predicting the future state of turbofan engines. After the BiLSTM network, an improved convolutional block attention module (CBAM) is applied to further enhance the model's understanding and utilization of the global dependency of the time series. The other path extracts the mean value and degradation trend coefficient of the data through manual features, which are fused with the features extracted by the previous path, and finally output the predicted remaining useful life (RUL) value through the fully connected layer.

[0116] S4. Use the training data set to train the model, and use two evaluation indicators, root mean square error RMSE and score function Score, to evaluate the performance of the model.

[0117] The first is the root mean square error RMSE, which can intuitively reflect the average level of prediction error. The calculation formula of the root mean square error RMSE is:

[0118]

[0119] The second is the score function Score, which imposes a greater penalty on lagged predictions. The calculation formula of the score function Score is:

[0120]

[0121] Where N is the total number of samples, i represents the i-th sample, and y i and Respectively represent the actual RUL value and the predicted RUL value of the i-th sample,

[0122] In order to verify the technical effect that the present invention can produce, the prediction method of Example 1 was experimentally verified. The present invention 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 7 6800H, NVIDIA RTX 3060, 16GB RAM.

[0123] During the experiment, the training process of the TCN-BiLSTM-CBAM model is as follows:

[0124] 1) Prepare the experimental dataset

[0125] This embodiment uses the widely recognized Commercial Modular Aerospace Propulsion System Simulation (C-MAPSS) dataset to evaluate the newly proposed remaining useful life (RUL) prediction model. The dataset consists of four sub-datasets, namely FD001, FD002, FD003 and FD004 datasets. The information of the four sub-datasets is shown in Table 1, where the training set includes the complete cycle of the aircraft engine from normal state to fault state, and records the time series data of multiple state parameters in the whole process. The test set contains the data records of the state parameters before the fault occurs, as well as the corresponding remaining useful life values.

[0126] Table 1 Dataset

[0127]

[0128] 2) Data preprocessing

[0129] In the C-MAPSS dataset, not all sensor data are meaningful for lifespan prediction. Fig.10The following table shows the 21 sensor data of all engines in the FD001 training set. Obviously, some sensor data remain unchanged during the cycle, which is meaningless for RUL prediction. Therefore, the sensor data numbered 1, 5, 6, 10, 16, 18, and 19 are removed, and the remaining 14 sensor data and 3 operating conditions are used as input features for the experiment.

[0130] An example of a sliding window of data is Fig.11 As shown. For the entire cycle from engine operation to failure, assume that the total number of engine operation cycles is T, the window size is w, and the sliding step is S. The size of each sample is w×n, where n is the number of features. For the i+1th sample, the corresponding RUL label calculation formula is as follows:

[0131] RUL=Twi*S

[0132] Since the degradation characteristics of the engine are not obvious in the early stage of operation, the label of the sample is processed using a piecewise linear function. The formula is as follows:

[0133]

[0134] 3) Model training parameter setting

[0135] Table 2 shows the parameter settings of the FD001 network model. The shape of the input data is (B, 30, 17), and the shape of the data input to the manual feature extraction is (B, 34), where B represents the batch size. Output represents the shape of the data after passing through the network layer. The batch size is set to 128, the epoch is set to 50, the optimizer is adam, and the learning rate is set to 0.001. The activation function uses Relu, and the probability of Dropout randomly discarding neurons is set to 0.2.

[0136] Table 2 Network parameter settings

[0137]

[0138]

[0139] 4) Results and analysis

[0140] The prediction results of the four sub-datasets are as follows Fig.12 As shown, Fig.12 (a) is the prediction result of the sub-dataset FD001. Fig.12 (b) is the prediction result of the sub-dataset FD002. Fig.12 (c) is the prediction result of sub-dataset FD003. Fig.12(d) is the prediction result of sub-dataset FD004; the red dotted line represents the predicted RUL, and the blue solid line represents the actual RUL. Fig.13 It is the error distribution diagram of the true value and the predicted value. It can be seen from the figure that the prediction results of FD001 and FD003 are better than those of FD002 and FD004, and are closer to the real RUL. The error range of FD001 and FD003 is concentrated in [-20,20], and the error range of FD002 and FD004 is concentrated in [-40,40]. This is because the working conditions of FD002 and FD004 are more complicated, which is in line with the actual situation.

[0141] Fig.14 (a) shows the prediction results for the FD001 test engine unit 76, Fig.14 (b) shows the prediction results of 176 in FD002. Fig.14 (c) shows the prediction results of 99 in FD003. Fig.14 (d) shows the prediction results of 213 in FD004. It can be seen that in the early stage of operation, the RUL of the engine tends to be stable, indicating that the engine is in a healthy state. At the beginning of the degradation stage, the predicted RUL and the actual RUL are quite different. This is because the degradation characteristics in the early stage are not obvious. The prediction accuracy of the model is higher in the later stage of engine degradation. This is because the degradation information of the engine is more obvious in the later stage of degradation. The prediction accuracy of FD001 and FD003 is much higher than that of FD002 and FD004, which is also consistent with the actual situation.

[0142] In order to prove the feasibility of the present invention, comparative experiments were carried out. In order to avoid the influence of randomness, each experiment was repeated 10 times and the average value was taken. According to the experimental results, it can be seen that the RMSE and score of the present invention on different data sets perform well, which is better than most existing methods. This shows that compared with these methods, the proposed method can better capture long-term dependencies from aircraft engine degradation data. For RMSE, the proposed model can be reduced by 2.6%, 4.1%, and 1.3% on the FD001~FD003 data sets. For score, the proposed model can be reduced by 2.72% and 8.45% on the FD001 and FD004 data sets. From the average values ​​of the four data sets, the average RMSE of the proposed model is 15.14, and the average score is 987.52, which is the lowest among all the comparison methods. In general, the method proposed in the present invention can produce better RUL prediction results than other methods. This further proves the superiority and reliability of the proposed model and demonstrates its potential in practical applications.

[0143] Table 4 Comparison results of this method with other methods

[0144]

[0145]

[0146] Table 5 Comparison of the average values ​​of this method and other methods

[0147]

[0148] in conclusion:

[0149] The present invention proposes a turbofan engine remaining useful life (RUL) prediction model based on a time-domain convolutional network (TCN) and a bidirectional long short-term memory network (BiLSTM) combined with a convolutional block attention module (CBAM). By processing complex time series data, the model performs well in extracting data features and capturing degradation processes. The TCN-BiLST-CBAM hybrid model designed by the present invention performs well in capturing local patterns and long-term dependencies in time series data. TCN can effectively process time series data, while BiLSTM simultaneously obtains past and future information through its bidirectional structure.

[0150] The improved convolutional block attention module (CBAM) is introduced after the BiLSTM network to further enhance the model's ability to capture key degradation information in time series. Experimental results show that the model outperforms multiple existing methods on NASA's C-MAPSS dataset, especially showing excellent performance when processing complex turbofan engine degradation data. The channel and spatial attention mechanism of CBAM helps the model identify important feature channels and key time periods, thereby improving the accuracy of prediction and the interpretability of the model.

[0151] Embodiment 2: A system for predicting the remaining useful life of a turbofan engine based on a deep learning model, comprising:

[0152] A data processing module is used to obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0153] The prediction model building module is used to build a dual-channel initial learning framework and train it according to the training data set to obtain a trained prediction model;

[0154] The RUL prediction module is used to input the operating data of the turbofan engine to be tested and use the prediction model to predict the remaining service life of the turbofan engine.

[0155] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute a method for predicting the remaining useful life of a turbofan engine based on a deep learning model, the method comprising the following steps:

[0156] Obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0157] Construct a dual-channel initial learning framework, one of which is trained based on the data set using the TCN-BiLSTM network framework that introduces the CBAM attention mechanism; the other path uses manual feature extraction to extract the average value and degradation trend coefficient of the operating data as auxiliary input of the prediction model; and obtain a trained prediction model;

[0158] According to the operating data of the turbofan engine to be tested, the remaining service life of the turbofan engine is predicted using the prediction model.

[0159] Embodiment 4: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call logic instructions in the memory to execute a method for predicting the remaining useful life of a turbofan engine based on a deep learning model, the method comprising the following steps:

[0160] Obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0161] Construct a dual-channel initial learning framework, one of which is trained based on the data set using the TCN-BiLSTM network framework that introduces the CBAM attention mechanism; the other path uses manual feature extraction to extract the average value and degradation trend coefficient of the operating data as auxiliary input of the prediction model; and obtain a trained prediction model;

[0162] According to the operating data of the turbofan engine to be tested, the remaining service life of the turbofan engine is predicted using the prediction model.

[0163] 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.

[0164] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a method for predicting the remaining useful life of a turbofan engine based on a deep learning model. The method includes the following steps:

[0165] Obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set;

[0166] Construct a dual-channel initial learning framework, one of which is trained based on the data set using the TCN-BiLSTM network framework that introduces the CBAM attention mechanism; the other path uses manual feature extraction to extract the average value and degradation trend coefficient of the operating data as auxiliary input of the prediction model; and obtain a trained prediction model;

[0167] According to the operating data of the turbofan engine to be tested, the remaining service life of the turbofan engine is predicted using the prediction model.

[0168] 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.

[0169] 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.

[0170] 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 method for predicting the remaining useful life of a turbofan engine based on a deep learning model, characterized in that: include: Obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set; Construct a dual-channel initial learning framework, one of which uses the TCN-BiLSTM network framework that introduces the CBAM attention mechanism for training based on the data set; the other path uses manual feature extraction to extract the average value and degradation trend coefficient of the operating data as auxiliary input of the prediction model; Obtain the trained prediction model; According to the operating data of the turbofan engine to be tested, the remaining service life of the turbofan engine is predicted using the prediction model.

2. The method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 1, characterized in that: The method of obtaining the operating data of the turbofan engine during operation, performing preprocessing, and constructing a data set comprises: Use various sensors to collect various sensor data during the operation of the turbofan engine; Perform data screening on the sensor data to filter out the data that remain unchanged during the cycle; Normalize the remaining data; The sliding window is used to process the data, and the RUL label is processed at the same time to construct the data set.

3. The method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 2, characterized in that: The normalization of the remaining data includes: 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 method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 2, characterized in that: The method uses a sliding window to process the data and simultaneously processes the RUL labels to construct a data set; including: 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 S, multiple samples are obtained, each of which is w×n in size, where n is the number of features; For the i+1th sample, the corresponding RUL label calculation formula is: RUL=Twi*S The sample labels are processed using a piecewise linear function, the formula is: A data set is constructed using the obtained samples and the labels corresponding to each sample.

5. The method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 2, characterized in that: The construction method of the dual-channel initial learning framework is: TCN and BiLSTM are used for feature extraction, and after the BiLSTM network, a network framework of the CBAM module is added as one of the paths of the dual-channel initial learning framework; The average value and degradation trend coefficient of the running data are extracted by manual features, and the average values ​​and degradation trend coefficients of all features are combined to form a feature vector as another path; The feature maps obtained from the two paths are flattened, fully connected and fused respectively, and finally passed through a fully connected layer to output the predicted RUL value.

6. The method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 5, characterized in that: The method of extracting the average value and degradation trend coefficient of the operating data by manual feature extraction, merging the average values ​​and degradation trend coefficients of all features to form a feature vector comprises: For a sliding window of data X = [x1, x2…x n ], whose shape is (w,n), w and n represent the size of the sliding window and the number of sensors, that is, the number of features, respectively, and the average value is calculated as: Where x n is the nth feature in a time step; a set of average data is obtained by calculation The degradation trend coefficient is obtained by linear regression of each feature, and the calculation formula is: y n =k n x n +b Where k is the regression coefficient, which is defined as the regression trend coefficient; y n is the predicted value, which is the nth feature in a time step; By calculating, we can get a set of degradation trend coefficients K = [k1, k2…k n ]; Finally, the average values ​​and degradation trend coefficients of all features are combined to form a feature vector h = [μ1, μ2…μ n ,k1,k2…k n ], as auxiliary input of the prediction model.

7. The method for predicting the remaining useful life of a turbofan engine based on a deep learning model according to claim 1, characterized in that: The training results of the trained prediction model are evaluated using two evaluation indicators: root mean square error (RMSE) and 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, i represents the i-th sample, and y i and Respectively represent the actual RUL value and the predicted RUL value of the i-th sample, 8. A turbofan engine remaining service life prediction system based on a deep learning model, characterized in that: include: A data processing module is used to obtain the operating data of the turbofan engine during operation, perform preprocessing, and construct a training data set; The prediction model building module is used to build a dual-channel initial learning framework and train it according to the training data set to obtain a trained prediction model; The RUL prediction module is used to input the operating data of the turbofan engine to be tested and use the prediction model to predict the remaining service life of the turbofan engine.

9. A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions enable a computer to execute the method for predicting the remaining useful life of a turbofan engine based on a deep learning model as described in any one of claims 1 to 7.

10. 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 through the communication bus, and the processor calls the logic instructions in the memory to execute the method for predicting the remaining useful life of a turbofan engine based on a deep learning model as described in any one of claims 1 to 7.

Images