A method for fault diagnosis of aircraft engine degradation state based on transfer learning

Abstract

The present invention provides a method for fault diagnosis of aircraft engine degradation state based on transfer learning. The method is implemented based on extreme learning machine (ELM) and incorporates the idea of ​​transfer learning. The method utilizes a large amount of operating data of the engine when it is not degraded and a small amount of data of a certain degradation state of the engine, and regards the data when it is not degraded as source domain data and the data after degradation as target domain data. The method is divided into two stages. In the first stage, the extreme learning machine model is trained with data from the source domain to extract information from the source domain; in the second stage, the target domain data is used for training to adapt the target domain. The present invention solves the problem of lack of aircraft engine fault diagnosis data through transfer learning, and combines it with ELM to ensure the real-time and accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention aims at fault diagnosis of aircraft engine degradation state. In order to solve the problem of small amount of fault data in a certain degradation state, the data collected in the non-degraded state is supplemented, and transfer learning is used based on the extreme learning machine to fully utilize the information in the two fields. Background Art

[0002] Fault diagnosis is an important part of aircraft engine health management and is of great significance to ensuring the healthy operation of aircraft and engines. Since 90% of aircraft engine failures are caused by gas path component failures, it is very necessary to diagnose gas path component failures. Changes in flow and efficiency can directly reflect the health status of engine components, but they cannot be directly observed. Some measurable parameters on the engine, such as temperature, pressure, speed, etc., cannot directly reflect the health status of aircraft engine components. This brings certain difficulties to aircraft engine fault diagnosis.

[0003] Intelligent fault diagnosis methods represented by machine learning methods can establish a connection between measurable parameters on aircraft engines and the health status of engine components. Due to the rapid development of machine learning methods, the effectiveness of aircraft engine fault diagnosis has been greatly improved. Such methods usually use labeled data from the engine's historical operation as training data and data collected under the current operating state as test data. These methods all assume that the data in the training set and the test set follow the same distribution. In other words, the training data and the test data are collected under the same conditions. However, due to different engine states, different operating conditions, and engine installation errors, it is difficult for the training data and the test data to follow the same distribution.

[0004] For example, during the life cycle of an aircraft engine, as the number of uses increases, the overall performance of the engine will decline. When this decline reaches a certain level, the engine reaches the end of its life. Under different degrees of degradation, the performance of aircraft engines will be different. Therefore, the data collected under different degrees of degradation will obey different distributions. Therefore, using data collected under different states to train the same fault diagnosis model will result in lower performance or even unusable models. In addition, it is not feasible to collect enough data from the same state to train a fault diagnosis model. This is because the acquisition of fault data is very difficult and usually costs a lot.

[0005] Transfer learning aims to solve the problem of data shortage in the field of machine learning. Its main idea is to extract information from other fields (source domains) and apply it to the target domain. The source domain and the target domain have certain similarities, and the source domain is rich in data and contains a lot of useful information. However, due to the differences between the source domain and the target domain, the information contained in the source domain cannot be used directly. Fortunately, the information can be transferred to the target domain through transfer learning, thereby solving the problem of insufficient data in the target domain. For aircraft engine fault diagnosis, transfer learning can be used to extract information contained in data collected from other states and use it to train the fault diagnosis model of the current engine. In particular, for engine fault diagnosis in a degraded state, the data collected by the engine in the current degraded state can be used as the target domain data, and the data collected in the non-degraded or other degraded states can be used as the source domain data. Transfer learning is used to solve the problem that the data in the current state is insufficient to train the fault diagnosis model.

[0006] Since aircraft engine fault diagnosis has high real-time requirements, a single-hidden-layer feedforward neural network (SLFN) has been widely used in the field of fault diagnosis. Methods containing SLFN structures can be roughly divided into three categories: algorithms based on gradient optimization, algorithms using standard optimization procedures, and algorithms using least squares methods. Algorithms based on gradient optimization, such as BP neural networks, are easy to fall into local optimal points. Algorithms using standard optimization procedures, such as support vector machines (SVM), usually have very complex steps. Algorithms based on least squares methods, such as ELM, have great advantages in training and testing speed. ELM was proposed in 2004 and is widely used in the field of aircraft engine fault diagnosis. The network weights of ELM are obtained by random generation and direct calculation. Therefore, ELM has a very fast training speed and good generalization performance. These advantages are consistent with the requirements of aircraft engine fault diagnosis.

[0007] Due to the excellent performance of ELM in aircraft engine fault diagnosis, the present invention chooses to use transfer learning based on ELM to solve the problem of aircraft engine fault diagnosis in a degraded state. Summary of the invention

[0008] Purpose of the invention:

[0009] In order to solve the problem that when an aircraft engine degrades to a certain stage, a good fault diagnosis model cannot be trained due to insufficient fault data, the present invention adopts the idea of ​​transfer learning, uses the data collected when it is not degraded as the source domain data, and uses the data collected in the current degraded state as the target domain data, and designs an aircraft engine fault diagnosis model based on ELM to improve the effectiveness of engine fault diagnosis.

[0010] Technical solution:

[0011] A method for fault diagnosis of aircraft engine degradation state based on transfer learning, namely a two-stage transfer learning extreme learning machine algorithm (Two-stage-transfer-learning ELM, referred to as TSTL-ELM), comprises the following steps:

[0012] Step 1: Collect historical flight data at the nominal engine state as source domain data;

[0013] Step 2: Collect flight data at the current engine degradation level as target domain data;

[0014] Step 3: Use the source domain data to train the regularized extreme learning machine model to obtain the intermediate model, and then proceed to step 4; or use the combination of source domain data and target domain data to train the regularized extreme learning machine model to obtain the fault diagnosis model, and then proceed to step 5;

[0015] Step 4: Use the target domain data to continue training the intermediate model to obtain the fault diagnosis model;

[0016] Step 5: Input the real-time flight data into the fault diagnosis model to diagnose whether the fault occurs and the component where it occurs.

[0017] Preferably, in steps 1 and 2, the collected flight data comes from 10 sensors, including: high pressure rotor speed N H , low pressure rotor speed N L , fan outlet temperature T 22 , compressor outlet pressure T3, compressor outlet pressure P3, low-pressure turbine inlet temperature T 45 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 , mixing chamber inlet temperature T 65 , and the main fuel flow WFB.

[0018] Preferably, the nominal state historical operation data collected in step 1 includes fault data and non-fault data, wherein the data volume of each type of data is more than 100, and the flight data collected at the current degradation level in step 2 includes fault data and non-fault data, wherein the data volume of each type of data exceeds 2; the data volume of the source domain data is greater than the data volume of the target domain data, and the difference is at least two orders of magnitude.

[0019] Preferably, in step 3, the specific steps of using the source domain data to train the regularized extreme learning machine model to obtain the intermediate model are:

[0020] The objective function of the regularized extreme learning machine model is:

[0021]

[0022] Among them, β is the output weight of the model, T is the actual output of the data, H is the output of the hidden layer, and the calculation method of H is as follows:

[0023]

[0024] The w in the formula i 、b i denote the input weight and bias corresponding to the i-th hidden layer neuron, i = 1, 2, ..., L, j = 1, 2, ..., N, L denotes the number of neurons in the hidden layer, N denotes the number of samples, g(·) denotes the activation function of the hidden layer, h(x j ) represents the hidden layer output corresponding to the jth sample;

[0025] The optimal solution of the objective function is:

[0026]

[0027] Among them, I is the identity matrix of the corresponding dimension, and C is the regularization constant;

[0028] The regularized extreme learning machine model is trained using source domain data, and the output weight of the intermediate model is obtained according to formula (3):

[0029]

[0030] X s represents the source domain data, T s represents the source domain label vector, I L represents the identity matrix of dimension L.

[0031] Preferably, the specific steps of continuing to train the intermediate model using the target domain data in step 4 include:

[0032] According to the target domain data and the output weights of the intermediate model obtained in step 3 Construct the objective function:

[0033]

[0034] Among them, C1>0, C2>0 are two constants;

[0035] By taking the derivative of formula (5) and setting it equal to 0, we get the optimal solution:

[0036]

[0037] X t represents the target domain data, T t represents the target domain label vector;

[0038] Finally, the target domain data is used for testing.

[0039] Preferably, the specific steps of training the regularized extreme learning machine model using a combination of source domain data and target domain data in step 3 include:

[0040] The objective function of the regularized extreme learning machine model is:

[0041]

[0042] Where C1 and C2 are two constants greater than 0. By taking the derivative of equation (7) and setting the derivative equal to 0, the solution is:

[0043]

[0044] Finally, the target domain data is used for testing.

[0045] The specific steps of the two-stage transfer learning extreme learning machine algorithm include:

[0046] Step 1: Given the source domain data, target domain data, hidden layer node type, number of hidden layer nodes, regularization constant C, and balance parameters C1 and C2.

[0047] 1) Source domain data: data collected when the engine is not degraded.

[0048] 2) Target domain data: data collected at the current engine degradation level,

[0049] 3) Hidden layer node type: generally using sigmoid function, tanh function, ReLU function, etc.

[0050] 4) The number of hidden layer nodes L: It is generally determined through experience or simulation experiments. This parameter has a great impact on the performance of the model.

[0051] 5) Regularization constant C: The regularization constant of the first stage, which ensures the generalization ability of the model.

[0052] 6) Balance parameters C1, C2: These two parameters are used to prevent the model from overfitting and control the degree of retention of source domain information.

[0053] Step 2: Randomly generate the network’s input weights w and hidden layer bias b.

[0054] Step 3: According to formula (2), use the data from the source domain and the target domain to calculate H(X s ) and H(X t ).

[0055] Step 4: Calculate the output weight of the intermediate model according to formula (4)

[0056] Step 5: Calculate the output weight of the second stage of the model according to formula (6)

[0057] Step 6: Input real-time flight data and perform real-time engine fault diagnosis.

[0058] Another solution of the present invention: a one-stage transfer learning extreme learning machine algorithm (OSTL-ELM), the algorithm selects a part of the source domain data as training data (ud_train), selects a part of the data in the target domain as training data (d_train), and selects a part as test data (d_test); ud_train and d_train are used for training at the same time, but different parameters are assigned to the two parts. The specific steps include:

[0059] Step 1: Given the source domain data, target domain data, hidden layer node type, number of hidden layer nodes, and balance parameters C1 and C2.

[0060] Step 2: Randomly generate the network’s input weights w and hidden layer bias b.

[0061] Step 3: According to formula (2), use the data from the source domain and the target domain to calculate H(X s ) and H(X t ).

[0062] Step 4: According to formula (8), calculate the output weight

[0063] Step 6: Input real-time flight data and perform real-time engine fault diagnosis.

[0064] Beneficial effects:

[0065] This method of aircraft engine degradation state fault diagnosis based on transfer learning can extract information from data collected in other states and apply it to the engine in the current state, thereby solving the problem of insufficient engine fault data under the current degradation degree and improving the effectiveness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the present invention;

[0067] Figure 2 The calculation process of TSTL-ELM;

[0068] Figure 3 It is a schematic diagram of the structure of a twin-rotor turbofan engine;

[0069] Figure 4 This is the diagnostic effect curve of changing the number of hidden layer nodes under working condition 1;

[0070] Figure 5 This is the diagnostic effect curve of changing the number of hidden layer nodes under working condition 2;

[0071] Figure 6 This is the diagnostic effect curve of changing the number of hidden layer nodes under working condition 3;

[0072] Figure 7 This is the diagnostic effect curve of changing the number of hidden layer nodes under working condition 4;

[0073] Figure 8 is the diagnostic effect curve of changing the ratio of target domain to source domain data under working condition 1;

[0074] Fig. 9 It is the diagnostic effect curve of changing the ratio of target domain to source domain data under working condition 2;

[0075] Fig.10 Diagnostic effect curve of changing the ratio of target domain to source domain data under working condition 3;

[0076] Fig.11 This is the diagnostic effect curve when changing the ratio of the target domain to the source domain data under working condition 4. DETAILED DESCRIPTION

[0077] This experiment is a simulation experiment. The engine model used in the experiment is a turbofan engine model. The main components include fan, compressor (HPC), combustion chamber, high pressure turbine (HPT), low pressure turbine (LPT), and micro nozzle. The schematic diagram is as follows Figure 3As shown, in order to clarify the meaning of the internal parameters of the engine, some sections in the figure are numbered. Section 2 represents the fan inlet section, section 22 represents the fan outlet section, section 3 represents the compressor outlet section, section 42 represents the high-pressure turbine outlet section, and section 46 represents the low-pressure turbine outlet section.

[0078] When the performance of an aircraft engine degrades or a failure occurs, the health parameters of the engine will change. Therefore, the present invention simulates the performance degradation and failure of the engine by changing the health parameters of the engine. Performance degradation means that the overall performance of the engine has decreased, so the health parameters of the engine components are reduced by 1% (some parameters are increased) to simulate the performance degradation of the engine. The engine failure is simulated by changing the health parameters of a single component. Four types of failures were set in this experiment, namely fan failure, compressor failure, high-pressure turbine failure, and low-pressure turbine failure. The collected data include: high-pressure rotor speed N H , low pressure rotor speed N L , fan outlet temperature T 22 , compressor outlet pressure T3, compressor outlet pressure P3, low-pressure turbine inlet temperature T 45 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 , mixing chamber inlet temperature T 65 , and the main fuel flow WFB.

[0079] In order to test the generalization ability of the model, the experiment used data collected from different working conditions. The working conditions were changed by changing the altitude and the flight Mach number. The four working conditions in the experiment are shown in Table 1.

[0080] Table 1 Four working conditions

[0081]

[0082] In the experiment, the data collected in the nominal state (non-degraded) is used as the source domain data, and the data collected in the degraded state is used as the target domain data. A part of the source domain data is selected as the training data (ud_train), a part of the target domain data is selected as the training data (d_train), and a part of the target domain data is selected as the test data (d_test). The data volume of each data set in the experiment is shown in Table 2.

[0083] Table 2 Data volume of each data set

[0084]

[0085] Note: #normal state indicates the number of normal data, #fan fault indicates the number of fan fault data, #HPC fault indicates the number of compressor fault data, #HPT fault indicates the number of high-pressure turbine fault data, and #LPT fault indicates the number of low-pressure turbine fault data.

[0086] In order to prove the effectiveness of the present invention, some other methods were used in the experiment for comparative experiments. The specific steps of each method are as follows:

[0087] Baseline1: Use ud_train to train the RELM model and d_test to test it.

[0088] Baseline2: Use ud_train and d_train to train the RELM model, and then use d_test for testing.

[0089] Baseline3: Use d_train to train the RELM model and d_test to test it.

[0090] OSTL-ELM: One-stage transfer learning extreme learning machine algorithm, which uses ud_train and d_train for training at the same time, but assigns different parameters to the two parts. Its objective function is:

[0091]

[0092] Where C1 and C2 are two constants greater than 0. By taking the derivative of this formula and setting the derivative equal to 0, we can get the solution:

[0093]

[0094] Finally, use d_test to test.

[0095] TSTL-ELM: The first stage uses ud_train for training, and the second stage uses d_train for training.

[0096] Finally, d_test is used for testing.

[0097] Table 3 shows the experimental results of the five methods, including the parameters used by each method and the diagnostic accuracy under different working conditions. In order to ensure the real-time performance of the model, the number of hidden layer nodes of all algorithms is 50.

[0098] Table 3 Experimental results

[0099]

[0100] As can be seen from Table 3, TSTL-ELM has the highest diagnostic accuracy, followed by OSTL-ELM. Among the three baseline methods, Baseline2 has the highest diagnostic accuracy, followed by Baseline1 and Baseline2. This is because Baseline2 uses data from both the source domain and the target domain for training. The Baseline2 method can also be regarded as a special case of OSTL-ELM, that is, the case where C1=C2. The OSTL-ELM method can adjust the ratio of information in the two domains, so it can obtain higher diagnostic accuracy. However, since TSTL-ELM is divided into two stages, the first stage extracts information from the source domain, and the second stage adapts the target domain alone, the TSTL-ELM method has the highest diagnostic accuracy.

[0101] Figures 4 to 7 The influence of changing the number of hidden layer nodes on the diagnostic effect of each algorithm under different working conditions is shown. As can be seen from the figure, the diagnostic accuracy of each algorithm will increase with the increase of the number of hidden layer nodes. Under different numbers of hidden layer nodes, TSTL-ELM has the highest diagnostic accuracy.

[0102] Since the target domain data is very scarce, the size of the data has a great impact on the performance of the engine fault diagnosis model. Table 4 lists the different target domain data volumes. Figures 8 to 11 The curves showing the changes in diagnostic accuracy of each algorithm when the ratio of target domain data to source domain data changes are shown. When the proportion of target domain data is at a low level, the TSTL-ELM method has the highest diagnostic accuracy. As the ratio gradually increases, the diagnostic accuracy of algorithms other than Baseline1 gradually increases, among which Baseline3 increases the fastest, because this method directly uses the target domain data. Overall, TSTL-ELM still has the highest diagnostic accuracy, especially when there is less target domain data. Therefore, the method used in the present invention is more suitable for fault diagnosis during the degradation process of aircraft engines.

[0103] Table 4 Details of each data set

[0104]

Claims

1. A method for fault diagnosis of aircraft engine degradation state based on transfer learning, characterized in that: The following steps are involved: Step 1: Collect historical flight data at the nominal engine state as source domain data; Step 2: Collect flight data at the current engine degradation level as target domain data; Step 3: Use the source domain data to train the regularized extreme learning machine model to obtain the intermediate model, and then proceed to step 4; or use the combination of source domain data and target domain data to train the regularized extreme learning machine model to obtain the fault diagnosis model, and then proceed to step 5; Step 4: Use the target domain data to continue training the intermediate model to obtain the fault diagnosis model; Step 5: Input the real-time flight data into the fault diagnosis model to diagnose whether the fault occurs and the component where the fault occurs; The specific steps of using the source domain data to train the regularized extreme learning machine model to obtain the intermediate model in step 3 are: The objective function of the regularized extreme learning machine model is: Among them, β is the output weight of the model, Τ is the actual output of the data, H is the output of the hidden layer, and the calculation method of H is as follows: The w in the formula i 、b i denote the input weight and bias corresponding to the i-th hidden layer neuron, i = 1, 2, ..., L, j = 1, 2, ..., N, L denotes the number of neurons in the hidden layer, N denotes the number of samples, g(·) denotes the activation function of the hidden layer, h(x j ) represents the hidden layer output corresponding to the jth sample; The optimal solution of the objective function is: Among them, I is the identity matrix of the corresponding dimension, and C is the regularization constant; The regularized extreme learning machine model is trained using source domain data, and the output weight of the intermediate model is obtained according to formula (3): X s represents the source domain data, T s represents the source domain label vector, I L represents the identity matrix of dimension L; The specific steps of using the target domain data to continue training the intermediate model in step 4 include: According to the target domain data and the output weights of the intermediate model obtained in step 3 Construct the objective function: Among them, C1>0, C2>0 are two constants; By taking the derivative of formula (5) and setting it equal to 0, we get the optimal solution: X t represents the target domain data, T t represents the target domain label vector; Finally, the target domain data is used for testing; The specific steps of using the combination of source domain data and target domain data to train the regularized extreme learning machine model in step 3 include: The objective function of the regularized extreme learning machine model is: Where C1 and C2 are two constants greater than 0. By taking the derivative of equation (7) and setting the derivative equal to 0, the solution is: β is the output weight of the model, H is the output of the hidden layer, X s represents the source domain data, T s represents the source domain label vector, I L represents the identity matrix of dimension L, X t represents the target domain data, T t represents the target domain label vector; Finally, the target domain data is used for testing.

2. The method for diagnosing aircraft engine degradation state faults based on transfer learning according to claim 1, characterized in that: In steps 1 and 2, the flight data collected comes from 10 sensors, including: high pressure rotor speed N H , low pressure rotor speed N L , fan outlet temperature T 22 , compressor outlet pressure T3, compressor outlet pressure P3, low-pressure turbine inlet temperature T 45 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 , mixing chamber inlet temperature T 65 , and the main fuel flow WFB.

3. The method for diagnosing aircraft engine degradation state faults based on transfer learning according to claim 1, characterized in that: The nominal state historical operation data collected in step 1 includes fault data and non-fault data, where the amount of data for each category is more than 100. The flight data collected at the current degradation level in step 2 includes fault data and non-fault data, where the amount of data for each category is more than 2. The amount of data in the source domain is greater than that in the target domain, and the difference is at least two orders of magnitude.

Images