A method for aircraft engine fault diagnosis based on transferable neural network
By combining domain adaptive ideas and RVFL algorithms in aircraft engine fault diagnosis, the transferable data feature representation is learned, which solves the limitations of data distribution assumptions in traditional methods, improves diagnostic accuracy, and realizes effective fault diagnosis of different models of engines .
Patent Information
- Application Number
- CN202111282685.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-01
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-11-01
AI Technical Summary
The traditional aero engine fault diagnosis method assumes that the training data set and the test data set obey the same data distribution, resulting in a reduced diagnostic accuracy in actual applications, and the need for a large amount of labeled data leads to high labor and time costs, making it difficult to implement in actual engineering applications.
Using a transfer learning method, combined with domain adaptive ideas and RVFL algorithms, the UD-RVFL and UJD-RVFL algorithms are proposed by learning transferable data feature representations to reduce the distribution differences between the source domain and the target domain data. For air circuit fault diagnosis of aircraft engines.
By reducing the distribution difference between the source domain and the target domain data, the fault diagnosis accuracy is improved, effective fault diagnosis of different models of engines is achieved, the edge and condition distribution differences are reduced, and the practical application capability of diagnosis is improved.
Smart Images

Figure CN114077867B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of aircraft engine fault diagnosis, and in particular relates to an aircraft engine fault diagnosis method based on a transferable neural network. Background Art
[0002] As one of the effective components of the engine health management system, the aircraft engine fault diagnosis system has always been a hot topic in the industry and academia. The probability of engine gas path component failure can account for more than 90% of the overall engine failure. Therefore, it is particularly important to establish an effective method for gas path component fault diagnosis. At present, the feasible methods for engine fault diagnosis are mainly concentrated on model-based methods and data-driven methods. The model-based method mainly establishes an engine mathematical model based on the actual engine operating conditions to make a judgment on the engine health status. This method requires researchers to be very familiar with the working principle of the engine. However, with the continuous innovation and improvement of the engine itself, the difficulty of establishing an accurate model is also increasing. The uncertainty in the model and the increasing nonlinear complexity of the system will affect the judgment accuracy of this method. In addition, it should be pointed out that this method requires different mathematical models to be established for different models of engines. The data-driven method can detect and isolate the target according to the real-time data of the engine sensor and the historical collection data. This method can overcome the difficulties of the above-mentioned methods. As long as an effective machine learning algorithm is selected and improved, the fault diagnosis task of different models of engines can be completed. The present invention adopts a data-driven approach to solve the problems existing in engine gas path fault diagnosis.
[0003] However, most studies mainly use traditional machine learning theory to solve fault diagnosis problems, which may be impractical in actual application scenarios. Usually, these diagnostic methods based on traditional machine learning (including popular deep learning) often assume that the training data set and the test data set obey the same data distribution, which is too idealistic in practical applications. In addition, for those diagnostic algorithms that require a large amount of data, the huge manpower and time costs of labeling data are often unbearable. Therefore, although these algorithms may have achieved satisfactory diagnostic accuracy, the underlying idealized assumptions are likely to keep them only in a carefully designed experimental environment, rather than landing in actual engineering application scenarios.
[0004] In order to overcome the above challenges, intelligent fault diagnosis based on transfer learning is expected to become one of the most promising solutions. At present, the transfer learning method has been adopted by many researchers in various fields. Its research focus is on how to adjust the machine learning model constructed in the source domain and apply it to the target domain that is different from the initial source domain but related. Usually, the data set with labeled information used for training is called the source data set, and the data in the target domain is used for testing. To this end, the present invention develops a technical means for fault diagnosis of aero-engine gas path based on transfer learning. By combining the domain adaptation idea and the RVFL algorithm, two algorithms, UD-RVFL and UJD-RVFL, are proposed, so that under the premise of retaining the main advantage of the simple topological structure of the original RVFL network, the distribution difference between the source domain and the target domain data can be reduced by learning the transferable data feature representation, while retaining the data attributes and feature structure of the source domain as much as possible. Among them, the UD-RVFL algorithm mainly deals with the situation where only the marginal distribution difference is considered, while the UJD-RVFL algorithm uses the reconstructed source domain data to train the classifier, iteratively learns the pseudo-label information of the data in the target domain, thereby trying to reduce the marginal distribution and conditional distribution differences at the same time. Summary of the invention
[0005] Purpose of the invention: In order to overcome the shortcoming that diagnostic methods based on traditional machine learning often assume that the training data set and the test data set obey the same data distribution, the present invention proposes an aircraft engine fault diagnosis method based on a transferable neural network. By combining the domain adaptation idea and the RVFL algorithm, the distribution difference between the source domain and the target domain data is reduced by learning the transferable data feature representation, thereby improving the fault diagnosis accuracy.
[0006] The present invention provides an aircraft engine fault diagnosis method based on a transferable neural network, which specifically comprises the following steps:
[0007] Step 1, collecting aircraft engine operation data as sample data, including normal data samples and fault data samples, and establishing a mathematical model of the original RVFL;
[0008] The objective function is:
[0009]
[0010] in is the engine operation data sample, N is the total number of samples, and d is the number of features. For the i-th instance, x i is a d-dimensional feature vector. is the output weight, is the jth row of β. j and b jis the randomly set weight and bias of the input to the enhancement layer. T=[t i ,…t N ] T is the label set of the sample, if x i belongs to the jth class, then t ij is 1, and the rest are 0. j (·) is the activation function of the j-th enhancement node,
[0011] For x i , which is defined as:
[0012]
[0013] Formula (1) can be further expressed in matrix form, and the objective function is written as:
[0014]
[0015] The matrix The definition of is shown in the following formula
[0016]
[0017] Then, the optimal output of RVFL is
[0018]
[0019] in, is the generalized inverse matrix of H.
[0020] Step 2, adding regularization terms to the original RVFL mathematical model to reduce the risk of overfitting and simplify the model complexity, and optimize the objective function;
[0021]
[0022] in is the error vector, C is the predefined balance factor, and further h(x i )=[h(w1,x i ,b1),…,h(w L ,x i ,b L ),x i ].
[0023] Step 3, solve the mathematical model to obtain the output weight;
[0024] Step 4: Calculate the reconstructed data representation based on the obtained output weights, and perform the fault diagnosis task through the reconstructed data representation.
[0025] Furthermore, step 2 also includes using the strategy of reducing marginal distribution in transfer learning and integrating engine data information to obtain the mathematical model of UD-RVFL, reconstruct formula (6), and the objective function can be written as:
[0026]
[0027] in is the reconstruction error of the i-th sample in the source domain, and the samples with superscript t represent the samples in the target domain. φ(·) consists of the connection weights of the network, both randomly initialized and optimized. is the output weight, λ and C are predefined balancing factors;
[0028] After simplified calculation, formula (9) can be changed to:
[0029]
[0030] in is a diagonally balanced matrix, and the matrix The definition of
[0031]
[0032] Among them, H s and H t The definition of can be known from formula (4). In addition, is the MMD matrix, and its calculation is shown as follows
[0033]
[0034] For ease of calculation, expand the two matrices to appropriate dimensions. Assume for
[0035]
[0036] and in A vector of all zeros with appropriate dimensions. Therefore, the objective function of the UD-RVFL algorithm in formula (10) can be further simplified as
[0037]
[0038] Step 3, solve the mathematical model and obtain the output weight, specifically:
[0039] Taking the derivative of the model in formula (14) with respect to β and setting it to zero, we can obtain
[0040] β+H T ΩHβ-H T ΩX+CH TMHβ=0. (13)
[0041] Then, when (L+d)≤(n s +n t ) holds true, the solution of formula (15) is
[0042] β=(I+H T (Ω+CM)H) -1 H T ΩX, (14)
[0043] Otherwise, the solution can be expressed as
[0044] β=H T (I+(Ω+CΜ)HH T ) -1 ΩX. (15)
[0045] Step 4, calculating the reconstructed data representation according to the obtained output weights;
[0046] For the data in the source domain, the transferable data representation can be calculated as
[0047] X' s =H s β, (16)
[0048] The reconstructed feature representation of the target domain data can be calculated by the following formula
[0049] X' t =H t β. (17)
[0050] From then on, the reconstructed target data can be recognized by the classifier trained by the reconstructed source labeled data.
[0051] Furthermore, step 2 also includes utilizing the strategy of reducing both marginal distribution and conditional distribution in transfer learning and integrating engine data information to obtain the mathematical model of UJD-RVFL, reconstruct formula (6), and the objective function can be shown as:
[0052]
[0053] in represents the dataset with the kth true label category in the source domain, and Then it represents the dataset with the kth pseudo-label category in the target domain. After some simplification, the mathematical model of the above formula for the UJD-RVFL algorithm can be restated as follows
[0054]
[0055] in, is the MMD matrix, and its calculation is shown as follows
[0056]
[0057] In addition, M0=M, which is defined as shown in formula (12).
[0058] Step 3, solve the mathematical model and obtain the output weight, specifically:
[0059] Take the derivative of the model in formula (21) with respect to β and set it to zero. When the condition (L+d)≤(n s +n t ) is established, we can get
[0060]
[0061] Otherwise, there is
[0062]
[0063] The reconstructed data representation of UJD-RVFL cannot be directly obtained, and only the marginal distribution is aligned in the initialization stage. Then, the marginal distribution and conditional distribution are considered simultaneously in the iterative process, and at the same time, the pseudo-label information in the target domain is updated until the termination condition is reached.
[0064] Step 4, calculating the reconstructed data representation according to the obtained output weights;
[0065] For the data in the source domain, the transferable data representation can be calculated as
[0066] X' s =H s β, (23)
[0067] The reconstructed feature representation of the target domain data can be calculated by the following formula
[0068] X' t =H t β. (24)
[0069] From then on, the reconstructed target data can be recognized by the classifier trained by the reconstructed source labeled data.
[0070] Furthermore, aircraft engine failure refers to a low-pressure compressor LPC failure, a high-pressure compressor HPC failure, a high-pressure turbine HPT failure and a low-pressure turbine LPT failure in the engine.
[0071] Furthermore, by adjusting the control variables, environmental variables and performance degradation amount, these sample data have different distributions.
[0072] In aircraft engine fault diagnosis, the data samples are first normalized, and then the sample data and its corresponding sample labels are used as training samples to train the diagnosis model; the trained model is used to perform fault detection on various aircraft engine components.
[0073] Beneficial effects: By combining the domain adaptation idea and the RVFL algorithm, transferable data feature representation is learned, the distribution difference between source domain and target domain data is reduced, and the fault diagnosis accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0074] Figure 1 It is a block diagram of the main components of an aircraft engine.
[0075] Figure 2 This is a comparison chart of the overall experimental results of the first six groups of cases in the mixed migration experiment.
[0076] Figure 3 This is a comparison chart of the overall experimental results of six groups of cases after the mixed migration experiment.
[0077] Figure 4 This is a comparison chart of the overall experimental results of the first six groups of cases in the complete migration experiment.
[0078] Figure 5 This is a comparison chart of the overall experimental results of the six groups of cases after the complete migration experiment. DETAILED DESCRIPTION
[0079] In the case of multi-fault diagnosis of aircraft engines, the objective function of the original RVFL model is first established based on all the collected engine operation data (including normal data sample information and fault data sample information).
[0080]
[0081] in is a sample, N is the total number of samples, and d is the number of features. For the i-th instance, x i is a d-dimensional feature vector. is the output weight, is the jth row of β. j and b j is the randomly set weight and bias of the input to the enhancement layer. T=[t i ,…t N ] T is the label set of the sample, if x i belongs to the jth class, then t ij is 1, and the rest are 0. j (·) is the activation function of the jth enhancement node, for x i , which is defined as:
[0082]
[0083] Furthermore, formula (1) can be further expressed in matrix form, and its objective function is:
[0084]
[0085] The matrix The definition of is shown in the following formula
[0086]
[0087] Then, the optimal output of RVFL is
[0088]
[0089] in, is the generalized inverse matrix of H. Generally speaking, a regularization term is added to RVFL to reduce the risk of overfitting and simplify the model complexity. The objective function can be rewritten as follows
[0090]
[0091] in is the error vector, C is the predefined balance factor, and further h(x i )=[h(w1,x i ,b1),…,h(w L ,x i ,b L ),x i ]. According to the KKT condition, the solution of formula (6) can be expressed as
[0092]
[0093] Where I is the identity matrix of the corresponding dimension. For an unknown test sample x, its predicted label can be expressed as
[0094] g(x)=argmax{h(x)β} (8)
[0095] Then, reconstructing formula (6), we can get:
[0096]
[0097] in is the reconstruction error of the i-th sample in the source domain (the set where the training samples are located), and the samples with superscript t represent the samples in the target domain (test dataset). φ(·) consists of the connection weights of the network, including randomly initialized and optimized ones. is the output weight, λ and C are predefined balancing factors. In order to maintain the inter-class relationship of the source domain, the first constraint in formula (9) is set to force the output of the source domain to approach the input.
[0098] After some simplifications, formula (9) can be transformed into:
[0099]
[0100] in is a diagonally balanced matrix, and the matrix The definition of
[0101]
[0102] Among them, H s and H t The definition of can be known from formula (4). In addition, is the MMD matrix, and its calculation is shown as follows
[0103]
[0104] For ease of calculation, expand the two matrices to appropriate dimensions. Assume for
[0105]
[0106] and in A vector of all zeros with appropriate dimensions. Therefore, the objective function of the UD-RVFL algorithm in formula (10) can be further simplified as
[0107]
[0108] Taking the derivative of the model in formula (14) with respect to β and setting it to zero, we can obtain
[0109] β+H T ΩHβ-H T ΩX+CH T MHβ=0. (15)
[0110] Then, when (L+d)≤(n s +n t ) holds true, the solution of formula (15) is
[0111] β=(I+H T (Ω+CM)H) -1 H T ΩX, (16)
[0112] Otherwise, the solution can be expressed as
[0113] β=H T (I+(Ω+CΜ)HH T ) -1 ΩX. (17)
[0114] Using the solution of the obtained equation, the reconstructed feature representation can be calculated. For the data in the source domain, the transferable data representation can be calculated as
[0115] X' s =H s β, (18)
[0116] The reconstructed feature representation of the target domain data can be calculated by the following formula
[0117] X' t =H t β. (19)
[0118] From then on, the reconstructed target data can be recognized by the classifier trained by the reconstructed source labeled data.
[0119] As for the mathematical model of the UJD-RVFL algorithm, reconstructing formula (6), we can obtain:
[0120]
[0121] in represents the dataset with the kth true label category in the source domain, and Then it represents the dataset with the kth pseudo-label category in the target domain. After some simplification, the mathematical model of the above formula for the UJD-RVFL algorithm can be restated as follows
[0122]
[0123] in, is the MMD matrix, and its calculation is shown as follows
[0124]
[0125] In addition, M0=M, which is defined as shown in formula (12).
[0126] Take the derivative of the model in formula (21) with respect to β and set it to zero. When the condition (L+d)≤(n s +n t ) is established, we can get
[0127]
[0128] Otherwise, there is
[0129]
[0130] Different from UD-RVFL, the reconstructed data representation of UJD-RVFL cannot be directly obtained because the pseudo-label information of the target data is predicted in an iterative manner. Specifically, only the marginal distributions are aligned in the initialization stage. Then, both the marginal distributions and the conditional distributions are considered during the iteration process. At the same time, the pseudo-label information in the target domain is updated until the termination condition is reached.
[0131] Using the solution of the obtained equation, the reconstructed feature representation can be calculated. For the data in the source domain, the transferable data representation can be calculated as
[0132] X' s =H s β, (25)
[0133] The reconstructed feature representation of the target domain data can be calculated by the following formula
[0134] X' t =H t β. (26)
[0135] From then on, the reconstructed target data can be recognized by the classifier trained by the reconstructed source labeled data.
[0136] The implementation processes of the UD-RVFL and UJD-RVFL algorithms are given below:
[0137]
[0138]
[0139]
[0140] The above algorithms are evaluated by multi-classification algorithms. The evaluation index of the classification algorithm is accuracy. The definition of accuracy is as follows:
[0141]
[0142] Where y(x) is the true label information of sample x, f(x) is the predicted label information of sample x, and the value 1 is the best value it can take.
[0143] All experiments were performed on a desktop computer with IntelR CoreTM, i7-9750 CPU, 2.60GHz main frequency, 8G memory, Windows 10 system and MATLAB 2018a version. It is worth noting that since the training set and the test set come from different distributions, the cross-validation technique cannot be used directly to find the best parameter combination. Therefore, we directly use the grid search method to repeat the experiment 20 times for each parameter combination and give the best average result as the final performance comparison, which is also the most commonly used strategy in transfer learning. In order to fully verify the actual performance of the algorithm proposed in this chapter, the fault diagnosis results of the two proposed methods are compared with other widely used methods, mainly including RVFL, CORAL, TCA, GFK and ETL. For the CORAL algorithm, it has no parameters that need to be adjusted in the transfer learning stage, and the ordinary ELM algorithm is selected as its benchmark classifier. For the TCA algorithm, the dimension parameter is searched from the alternative set {6,7,8,9,10}, and the MMD is calculated using the RBF kernel function. λ and γ are selected from the set {2 -5 , 2 -4 , …, 2 4 , 2 5} and {2 -5 , 2 -5 , …, 2 4 , 2 5}, and then the ELM is also used as its benchmark classifier. For the GRK algorithm, the parameters of the geodesic flow kernel are searched from the candidate set {1,2,3,4,5}. For the ETL algorithm, the cosine distance is adopted, and the ELM is used as its benchmark classifier. For the UD-RVFL and UJD-RVFL algorithms, the parameters of the geodesic flow kernel are searched from the candidate set {1,2,3,4,5}. -4 , 10 -3 , …, 10 3 , 10 4} and {2 -4 , 2 -3 , …, 2 3 , 2 4}, and searched the balance parameters C and λ from {50,100,…,350,400}, and finally used RVFL as its benchmark classifier.
[0144] The present invention is tested with a twin-rotor turbofan engine. Figure 1As shown, the main components of the engine include an intake duct, a low pressure compressor (LPC), a high pressure compressor (HPC), a combustion chamber, a high pressure turbine (HPT), a low pressure turbine (LPT) and a tail nozzle. The air flows into the compressor through the intake duct, and the gas becomes high pressure gas after passing through the low pressure compressor and the high pressure compressor. In the combustion chamber, the fuel is injected and mixed with the high pressure gas to form a mixed gas. When the mixed gas flows through the high pressure turbine and the low pressure turbine, the high pressure compressor and the low pressure compressor connected to the high pressure shaft and the low pressure shaft are driven. Finally, the hot gas is discharged into the atmosphere at high speed.
[0145] The LPC, HPC, HPT, and LPT connected to the aircraft engine rotor are prone to failure at high speeds, so only the failures of these four components are considered. Before the experiment, simulation data of the full flight envelope were collected, and the details of the four data sets are listed in Table 1. For each data set, they contain an equal number of health conditions, and each health condition has the same number of data samples.
[0146] Table 1 Detailed information of the engine datasets collected under different working conditions
[0147]
[0148] As for the environmental variables of each data set, several typical operating points within the engine flight range are selected here to characterize the simulation tests under different working conditions. Specifically, the operating points of each data set are listed in Table 2. It can be seen that data sets A and B share the same operating points. As for the control variables, the data acquisition tests of different data sets are carried out in different dynamic ways.
[0149] Table 2 Environmental variable specifications involved in different data sets
[0150]
[0151] The present invention first designs a hybrid test scheme, which divides each data set equally and randomly into two parts: a training set and a test set. Taking data set A as an example, the training set is represented as A0 and the test set is represented as A1. For traditional fault diagnosis, the learning task is usually A0→A1. In fault diagnosis based on transfer learning, the learning tasks mainly include A0→B1, A0→C1 and A0→D1, etc. It is worth noting that since the training set and the test set come from different distributions, the cross-validation technique cannot be used directly to find the best parameter combination. Therefore, we directly use the grid search method to repeat the test 20 times for each parameter combination, and give the best average result as the final performance comparison, which is also the most commonly used strategy in transfer learning.
[0152] The experimental test results are listed in Table 3. It can be seen that there are 16 hybrid verification test combinations, among which the results in bold represent the results of the traditional fault diagnosis test with the best performance, while the rest are the results of the fault diagnosis test using transfer learning.
[0153] Table 3 Experimental results based on RVFL algorithm
[0154]
[0155] It can be clearly seen that the diagnosis results of the traditional method (bold mark) are higher than those of the migration method (diagonal mark). Taking dataset C as an example, the diagnostic performance of the C0→C1 task is 97.28%, while the diagnostic performance of C0 on A1, B1 and D1 is 78.91%, 80.86% and 81.83%, respectively. Due to the distribution difference, the performance degradation is obvious. In addition, the experimental results of the other three datasets also show similar phenomena. It is worth noting that the performance degradation of data C in the migration experiment is significantly higher than that of other datasets, which may be caused by the large distribution difference between C and other datasets. Therefore, the performance degradation indicates that it is necessary to adopt a domain adaptation strategy to match the distribution difference.
[0156] Subsequently, extensive experimental verification was carried out based on each comparison algorithm. The detailed results of the domain adaptation experiment are shown in Table 4. The overall comparison results of the 12 diagnosis cases are plotted in Figure 2 and Figure 3As the basic classifier, RVFL has an average diagnostic performance of 86.97%. Unfortunately, both TCA and GFK show significant negative transfer effects, as their diagnostic performance is inferior to RVFL. Negative transfer phenomenon is relatively common in transfer learning and needs to be avoided. Then, ETL and CORAL outperform RVFL. For the proposed UD-RVFL, its average diagnostic performance of 12 cases is 88.53%, which is better than the original RVFL. The proposed UJD-RVFL is slightly better than UD-RVFL, but the advantage is not particularly large. This may be due to the small number of total labels in the dataset. In general, the diagnostic results of the two proposed methods are significantly better than those of the other compared algorithms, demonstrating the feasibility of the proposed domain adaptation method, especially the adopted unilateral alignment strategy. With the help of this strategy, the inter-class relationship of the source domain is maintained. As for the time complexity, TCA has the highest computational time. The computational cost of UJD-RVFL is significantly higher than that of UD-RVFL because the former method introduces an iterative process. The time complexity of the remaining methods is basically maintained at a similar level.
[0157] Table 4 Diagnostic performance of all compared methods (%)
[0158]
[0159] Subsequently, the collected complete dataset was used to further validate the proposed method. Again, there were 12 cases, and their diagnosis results are listed in detail in Table 5. Figure 4 and Figure 5 The overall performance comparison is described in . In general, the overall results are similar to those of the hybrid transfer experiments. Specifically, RVFL, as the basic classifier, shows similar classification accuracy as before. TCA and GFK still show obvious negative transfer effects. For ETL and CORAL, their performance is worse than the basic classifiers, but not as good as our proposed method. The proposed UJD-RVFL is slightly better than UD-RVFL and still achieves the best average diagnostic performance in the 12 transfer experiments. As for the computational cost, the computational time of all algorithms increases with the increase in the number of training datasets. It is worth noting that the computational time of TCA grows nonlinearly with the expansion of training samples.
[0160] Table 5 Diagnostic performance of all compared methods under the full dataset condition (%)
[0161]
Claims
1. An aircraft engine fault diagnosis method based on transferable neural network, characterized in that: The specific steps include: Step 1, collecting aircraft engine operation data as sample data, including normal data samples and fault data samples, and establishing a mathematical model of the original RVFL. The aircraft engine fault refers to a low-pressure compressor LPC fault, a high-pressure compressor HPC fault, a high-pressure turbine HPT fault, and a low-pressure turbine LPT fault in the engine; Step 2, adding a regularization term to the original RVFL mathematical model to optimize the objective function; Step 3, solve the mathematical model and obtain the output weight; Step 4: Calculate the reconstructed data representation according to the obtained output weights, and perform the fault diagnosis task through the reconstructed data representation; The objective function of the mathematical model of the original RVFL in step 1 is: in is the engine operation data sample, N is the total number of samples, and d is the number of features; For the i-th instance, x i is a d-dimensional feature vector; is the output weight, is the jth row of β; w j and b j is the randomly set weight and bias of the input to the enhancement layer; T = [t i ,…t N ] T is the label set of the sample, if x i belongs to the jth class, then t ij is 1, and the rest are 0; h j (·) is the activation function of the jth enhancement node; For x i , which is defined as: Formula (1) is further expressed in matrix form, and the objective function is written as: The matrix The definition of is shown in the following formula: Then, the optimal output of RVFL is: in, is the generalized inverse matrix of H; Step 2: Add a regular term to the original RVFL mathematical model to reduce the risk of overfitting and simplify the model complexity. The objective function is optimized as follows: in is the error vector, C is the predefined balance factor, and further h(x i )=[h(w1,x i ,b1),…,h(w L ,x i ,b L ),x i ]; Step 2 also includes using the strategy of reducing marginal distribution in transfer learning and integrating engine data information to obtain the mathematical model of UD-RVFL, reconstruct formula (6), and the objective function is written as: in is the reconstruction error of the i-th sample in the source domain, and the samples with superscript t represent the samples in the target domain; φ(·) consists of the connection weights of the network, including randomly initialized and optimized; is the output weight, λ and C are predefined balancing factors; After simplified calculation, formula (9) becomes: in is a diagonally balanced matrix, and the matrix The definition is as follows: also, is the MMD matrix, and its calculation is shown as follows: For ease of calculation, expand the two matrices to appropriate dimensions; assuming for: and in A vector of all zeros with appropriate dimensions; therefore, the objective function of the UD-RVFL algorithm in formula (10) is further simplified as follows: Step 3, solve the mathematical model and obtain the output weight, specifically: Taking the derivative of the model in formula (14) with respect to β and setting it to zero, we obtain: β+H T ΩHβ-H T ΩX+CH T MHβ=0 (15) Then, when (L+d)≤(n s +n t ) holds true, the solution of formula (15) is: β=(I+H T (Ω+CM)H) -1 H T ΩX (16) Otherwise, the solution is expressed as: β=H T (I+(Ω+CM)HH T ) -1 ΩX (17) Step 4, calculating the reconstructed data representation according to the obtained output weights; For the data in the source domain, the migratable data representation is calculated as: X s '=H s b (18) The reconstructed feature representation of the target domain data is calculated by the following formula: X t '=H t b (19) From then on, the reconstructed target data is recognized by the classifier trained by the reconstructed source labeled data.
2. The method for diagnosing aircraft engine faults based on a transferable neural network according to claim 1, characterized in that: Step 2 also includes using the strategy of reducing both marginal distribution and conditional distribution in transfer learning and integrating engine data information to obtain the mathematical model of UJD-RVFL and reconstruct formula (6). The objective function is: in represents the dataset with the kth true label category in the source domain, and Then it represents the dataset with the kth pseudo-label category in the target domain; after some simplification, the above formula is reformulated as follows for the mathematical model of the UJD-RVFL algorithm: in, is the MMD matrix, and its calculation is shown as follows: In addition, M0=M, which is defined as shown in formula (12); Step 3, solve the mathematical model and obtain the output weight, specifically: Take the derivative of the model in formula (21) with respect to β and set it to zero; when the condition (L+d)≤(n s +n t ) is established, it may: Otherwise, we have: The reconstructed data representation of UJD-RVFL cannot be directly obtained, and only the marginal distribution is aligned in the initialization stage; then, the marginal distribution and conditional distribution are considered simultaneously in the iterative process, and at the same time, the pseudo-label information in the target domain is updated until the termination condition is reached; Step 4, calculating the reconstructed data representation according to the obtained output weights; For the data in the source domain, the migratable data representation is calculated as: X s '=H s b (25) The reconstructed feature representation of the target domain data is calculated by the following formula: X t '=H t b (26) From then on, the reconstructed target data is recognized by the classifier trained by the reconstructed source labeled data.
3. The method for aircraft engine fault diagnosis based on transferable neural network according to claim 1, characterized in that: By adjusting the control variables, environmental variables and performance degradation, these sample data have different distributions.
