An aero-engine thrust estimation method based on time domain convolution network

Abstract

The application belongs to the field of control system design, and discloses a thrust estimation method based on a time domain convolution network, and proposes a thrust estimator based on a multilayer residual time domain convolution network. On the one hand, a Gaussian error linear unit activation function is used to replace the correction linear unit in the residual block to solve the neuron death problem. On the other hand, the overall structure of the multilayer convolution network is adjusted by using residual connection, so that the network training speed and the thrust estimation effect are further improved. The comparison results with the traditional seven methods show that the method has significant advantages in convergence speed and estimation accuracy. In addition, the application also applies the method to an actual control system of an aero-engine, and further verifies the real-time performance and feasibility of the control scheme through a micro turbojet engine test bed test.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of neural network system identification, engine thrust control and estimation, and particularly relates to an aero-engine thrust estimation method based on a time domain convolution network. BACKGROUND

[0002] In an aero-engine control system, parameters related to thrust, such as rotation speed and oil supply, are often used as controlled variables to indirectly control thrust and thus achieve effective control of the aero-engine. However, considering changes in external environment and engine performance degradation, this method needs to reserve a large safety margin range in the design process to ensure normal operation of the engine, and it is difficult to achieve high control performance due to the problem of thrust control accuracy. Therefore, it is of great research significance to use a thrust estimator to directly use thrust as a controlled variable to achieve output feedback closed-loop control. Thrust estimation methods can be mainly divided into two types, namely model-based methods and data-driven methods. Model-based thrust estimation emphasizes the establishment of a physical model, and often uses Kalman filtering to achieve estimation of unmeasurable parameters. However, in cases where accurate system model structure parameters or unknown noise statistical characteristics are difficult to obtain, the effect of Kalman filtering will be significantly affected. In order to avoid problems caused by modeling errors and uncertain noise and ensure the stability and robustness of the thrust estimation value, a linear variable parameter robust H∞ filter related design is proposed and good results are achieved. However, for aero-engines, which are nonlinear, multivariable and complex control objects, the accuracy of modeling is very important, and with the addition of more engine parameters, the calculation of model parameters also faces challenges. Data-based methods do not use any direct information related to the model, but directly map the relationship between sensor measurement parameters and thrust, such as genetic neural networks, least squares support vector machines, and correlation analysis. Compared with model-based methods, data-based methods simplify the implementation process, reduce time investment, and are widely applicable to various control objects, but most intelligent methods still have room for improvement in terms of thrust estimation accuracy and computational cost, even if there are no difficulties in application.

[0003] In recent years, with the continuous development of neural networks, data-driven methods have improved their performance in processing large-capacity, high-dimensional time series model data, and their application in aero-engine thrust estimation has gradually become a research focus. However, due to differences in system structure design, different networks have their own characteristics and defects. For example, convolutional neural networks (CNN) have advantages in feature extraction but lack memory function. Recurrent neural networks (RNN) and long short-term memory networks (LSTM) can store historical information, but they still have problems such as gradient vanishing and long execution time.

[0004] Therefore, in 2018, Bai S, Frasconi P, Simard P, et al. proposed a time-domain convolutional network (TCN) in the paper "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling". The TCN network performs sequence modeling on the basis of the CNN framework and effectively solves all the above problems by adding causal convolution, dilated convolution and residual connection. It has the characteristics of parallel processing convolution, flexible change of receptive field and stable gradient. However, a single deep TCN requires a large memory space to receive the original sequence during training, and the TCN model may not perform well due to the difference in memory requirements in different applications. Therefore, the neural network-based thrust estimator still needs to be improved. SUMMARY

[0005] The present application aims at the defects of the above-mentioned existing method, and proposes a new network model improved based on TCN. The multi-layer lightweight TCN network stack is realized by the way of residual connection, which improves the flexibility of the network model. Compared with deep TCN, it can reach better thrust estimation accuracy with faster convergence speed. The real-time performance and feasibility of the control scheme are further verified by ground test bench test.

[0006] The technical scheme of the present application is as follows:

[0007] An aero-engine thrust estimation method based on a time-domain convolutional network, comprising the following steps:

[0008] Step A: Obtain the output data of various parameters through micro turbojet engine test, and the data should cover the steady state running process of the acceleration stage (50000rpm-150000rpm) and the deceleration stage (150000rpm-50000rpm) after the slow running state of the engine as much as possible;

[0009] Step B: Data preprocessing; screening the original data obtained in step A, extracting important feature parameters related to thrust estimation, and obtaining training, validation and test data sets of the neural network model through data segmentation and standardization processing of the selected features;

[0010] The important feature parameters related to thrust estimation include rotational speed, fuel, engine exhaust temperature and environmental humidity.

[0011] The segmentation is to divide the engine feature data into training set and validation set by using leave-one-out cross-validation method with a ratio of 0.75:0.25; the selection range of the training set should cover different stages of the engine acceleration, deceleration and the like as much as possible.

[0012] The standardization adopts the following mean variance standardization method, so that all features are subject to the distribution characteristics of mean 0 and standard deviation 1.

[0013]

[0014] Here, μ is the mean of the feature x to be processed, δ is the standard deviation of the feature x, x * is the feature after standardization;

[0015] Step C, the time series combination of the thrust outside features in the training data set is taken as input, and the thrust is taken as output. The improved time domain convolution network model is trained by using a self-defined callback function training algorithm. In the training process, the estimation accuracy of the current network model is calculated in real time by using the test data set, and the best estimation network model is obtained by selecting the algorithm;

[0016] The time series combination input of the thrust outside features includes the current m time, the historical (n-1) time fuel information W f (m),…,W f (m-n+1), the historical n time speed information N(m-1),…,N(m-n), and the current time exhaust port temperature and ambient humidity T(m) and H(m); the neural network model is represented as:

[0017] F(m)=g(W f (m),…,W f (m-n+1),N(m-1),…,N(m-n),T(m),H(m))

[0018] Wherein, g(·) represents the mapping function relationship of the neural network model, and F(m) is the output thrust at the current time;

[0019] The callback function training algorithm includes the following steps:

[0020] (1) Defining the neural network training process callback function according to the evaluation index;

[0021] The evaluation index includes the mean absolute error MAE, the root mean square error RMSE and the correlation coefficient R 2 , and the definition method is as follows:

[0022]

[0023]

[0024]

[0025] Where y i , actual, estimated, and average values, respectively;

[0026] The custom callback function inherits the Callback class, uses the test set to predict the current model output at the end of each epoch, obtains the above evaluation indicators, and the specific data during training will be returned as a dictionary as the log content; At the same time, the ModelCheckpoint class function is added to the training process to realize the printing output of various indicators and the saving function of the best model;

[0027] (2) After setting the network model parameters, start training;

[0028] The parameter settings include the following parts:

[0029] 2.1) Network part; filter number filter, convolution kernel size k, expansion factor d, and fully connected layer unit number units;

[0030] Among them, the determination of the best parameters is mainly obtained through a variety of combination tests, and the specific selection range is shown in the following table:

[0031] Table 1 Parameter selection

[0032]

[0033]

[0034] 2.2) Training part; single training sample number batch=100, network training number of times epoch=200, and test error initial value;

[0035] The improved time domain convolution network retains the original network parallelism, gradient stability and other advantages, and further improves the convergence speed of the training process, and can realize higher precision thrust estimation. The improvement includes two aspects of internal composition of residual block and overall structure of network:

[0036] The improvement of internal composition is to use Gaussian error linear unit activation function (GELU) instead of rectified linear unit (ReLU) in the residual block of the time domain convolution network; The Gaussian error linear unit activation function is expressed as:

[0037]

[0038] where, Z~N(0,1) is the cumulative distribution function of the standard normal distribution, and erf(·) is the error function; By introducing the idea of random regularization to describe the probability of neuron input, it not only ensures unilateral inhibition, but also solves the dead ReLU problem;

[0039] The improvement of the overall network structure is that the lightweight TCN blocks of multiple different memories (different k and d) are combined through residual connection, so that the network structure can be automatically adjusted according to the actual application, a certain degree of domain migration is realized, and the flexibility of the overall network is improved.

[0040] (3) calculating the test error of the current training model;

[0041] (4) if the error is smaller than the initial value, updating the initial value and saving the model, otherwise keeping the initial value unchanged;

[0042] (5) continuing the next training, repeating steps (3) and (4) until the training is completed;

[0043] (6) outputting the final network model.

[0044] The beneficial effects of the present application: the improved network model used in the thrust estimation method based on time domain convolution not only has the advantages of gradient stability, network flexibility and fast convergence, but also further improves the accuracy of thrust estimation, and the method can be applied to direct thrust control of an aero-engine, which has important practical significance in engineering application. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the control system structure diagram of the present application.

[0046] Figure 2 is the improved network model structure diagram of the present application.

[0047] Figure 3 is the network model training process flow chart of the present application.

[0048] Figure 4 (a) is the simulation comparison diagram of the training loss of eight methods of the present application.

[0049] Figure 4 (b) is the simulation comparison diagram of the estimation accuracy of eight methods of the present application.

[0050] Figure 5 is the experimental result diagram of the thrust estimation of the present application.

[0051] Figure 6 is the experimental result diagram of the thrust control of the present application. DETAILED DESCRIPTION

[0052] The specific embodiments of the present application are further described below in combination with the drawings and technical solutions.

[0053] In a specific implementation, the control system scheme described in the embodiment of the present application is mainly for direct thrust control of an aero-engine. In the following, a micro turbojet engine is taken as an example to specifically describe a thrust estimation method based on a time domain convolution network provided by the present application.

[0054] Referring to Figure 1 is a control system structure diagram of the embodiment of the present application. The specific steps of the thrust estimation method based on the time domain convolution network provided by the present application are as follows:

[0055] Step A: Obtain the output data of various parameters through micro turbojet engine test running, and the data should cover the steady running process of the acceleration stage and the deceleration stage after the slow running state of the engine as much as possible;

[0056] Step B: Data preprocessing; screen the original data obtained in step A, extract important feature parameters related to thrust estimation, and obtain the training, verification and test data sets of the neural network model through data segmentation and standardization processing after feature selection;

[0057] In a specific implementation, the important feature parameters related to thrust estimation described in the embodiment include the rotation speed, fuel, engine exhaust port temperature and environmental humidity;

[0058] In a specific implementation, the segmentation in the embodiment is to divide the engine feature data into a training set and a verification set in a ratio of 0.75:0.25 using the leave-one-out cross-validation method;

[0059] In a specific implementation, the basic idea of the standardization in the embodiment is to scale the data in proportion under the condition that the relationship between the overall data is unchanged, so as to determine it in a smaller specific interval. The mean variance standardization method is adopted in the present application, which has the distribution characteristics that all features are subject to the mean value of 0 and the standard deviation of 1;

[0060]

[0061] Here, μ is the mean value of the feature x to be processed, δ is the standard deviation of the feature x, and x * is the feature after standardization;

[0062] Step C: Combine the time series of the features other than the thrust in the training data set as input, and the thrust as output. The improved time domain convolution network model is trained by using a self-defined callback function training algorithm. In the training process, the estimation accuracy of the current network model is calculated in real time by using the test data set, and the best estimation network model is obtained by selecting the algorithm;

[0063] In a specific implementation, the time series combination input of the features other than the thrust in the embodiment includes the fuel information W f(m),…,W f (m-n+1), history n time speed information N(m-1),…,N(m-n) and current time exhaust port temperature and ambient humidity T(m) and H(m); the neural network model is represented as:

[0064] F(m)=g(W f (m),…,W f (m-n+1),N(m-1),…,N(m-n),T(m),H(m)) wherein, g(·) represents the mapping function relationship of the neural network model, and F(m) is the output thrust of the current time;

[0065] In a specific implementation, the callback function training algorithm includes the following steps, and the specific process is as shown in Figure 3 .

[0066] (1) Defining a neural network training process callback function according to evaluation indexes;

[0067] The evaluation indexes include mean absolute error MAE, root mean square error RMSE and correlation coefficient R 2 , and the definition method is as follows:

[0068]

[0069]

[0070]

[0071] Wherein y i , are actual value, estimated value and average value respectively; different indexes reflect different emphases, so it is necessary to combine multiple indexes to evaluate together, wherein the smaller the values of mean absolute error and root mean square error are, and the larger the value of correlation coefficient is, the better the overall performance of the network model is.

[0072] In a specific implementation, the self-defined callback function inherits the Callback class, uses the test set to predict the current model output at the end of each epoch, obtains the above evaluation indexes, and the specific data in the training process will be returned as a dictionary form as the log content; at the same time, the ModelCheckpoint class function is added in the training process to realize the printing output of various indexes and the saving function of the best model;

[0073] (2) Starting training after setting the network model parameters;

[0074] In a specific implementation, the parameter setting includes the following parts:

[0075] 2.1) Network part: filter number filter, convolution kernel size k, dilation factor d, fully connected layer unit number units;

[0076] The determination of the optimal parameters is mainly obtained through a variety of combination tests, and the specific selection range is shown in the following table:

[0077] Table 1 Parameter selection

[0078]

[0079]

[0080] 2.2) Training part; single training sample number batch = 100, network training times epoch = 200, test error initial value;

[0081] In the specific implementation, the improved time domain convolution network retains the original network parallelism, gradient stability and other advantages, and the training process convergence speed is further improved, and higher precision thrust estimation can be realized. The improvement includes two aspects of internal composition of residual block and overall network structure:

[0082] The improvement of the internal composition is to use a Gaussian error linear unit activation function instead of a rectified linear unit in the residual block of the time domain convolution network; the Gaussian error linear unit activation function can be specifically represented as:

[0083]

[0084] wherein, Z~N(0,1) is the cumulative distribution function of the standard normal distribution, and erf(·) is the error function;

[0085] The improvement of the overall network structure is to combine a variety of lightweight TCN blocks with different memories through residual connection;

[0086] In the specific implementation, the Gaussian error linear unit activation function introduces the idea of random regularization to describe the probability of neuron input, so that the model training process is more robust, which not only ensures the zero-center symmetric unilateral inhibition, but also solves the problem of neuron death;

[0087] In specific implementation, when the improved network model is performing back propagation, the error can be effectively propagated to the shallow layer even if the gradient of an intermediate network block is small, ensuring the stability of the overall network. Moreover, compared with a single deep network, the improved network model can automatically adjust the network structure according to the actual application during network learning, thereby realizing a certain degree of domain migration, improving the flexibility of the overall network, and realizing the identity mapping of the residual TCN structure. Compared with approximating the identity mapping through multiple nonlinear layers, the learning process is easier and the convergence speed is faster, effectively avoiding the network degradation problem.

[0088] (3) calculating the test error of the current training model;

[0089] (4) if the error is smaller than the initial value, updating the initial value and saving the model, otherwise keeping the initial value unchanged;

[0090] (5) continuing the next training, repeating steps (3) and (4) until the training is completed;

[0091] (6) outputting the final network model.

[0092] Step D, simulating and comparing the improved model with various common methods to intuitively present the estimation accuracy of different methods in the form of data, thereby verifying the superiority of the improved network model;

[0093] In specific implementation, the common methods include traditional machine learning methods such as support vector regression (SVR) and K-nearest neighbor algorithm (KNN), and various neural network methods such as multilayer perceptron (MLP), gated recurrent unit (GRU), long short-term memory network (LSTM), one-dimensional convolutional network (1DCNN), and time domain convolutional network (TCN).

[0094] In specific implementation, the simulation comparison mainly includes two aspects of training loss during model training and prediction accuracy in the test stage, as shown in Figure 4 .

[0095] Overall, the model fitting result based on the neural network is better than that of the traditional method. As a new network, TCN solves the two main problems of gradient stability and parallelism of GRU and LSTM RNN networks, uses residual connection to ensure efficient propagation of information, and realizes parallel computing through causal convolution to reduce memory consumption in the training process, so the effect is remarkable in processing time series data. At the same time, TCN and 1DCNN are both convolutional network structure frameworks, but the introduction of causal convolution and residual connection of TCN effectively improves the estimation effect.

[0096] M-RTCN is a network model proposed on the basis of improvement of TCN, which further increases the network depth and adjusts the network structure on the premise of retaining the original advantages of TCN. In theory, the network depth will directly affect the model performance, therefore, in order to verify the effectiveness of the M-RTCN model, the overall network depth of TCN and M-RTCN is set to the same limit in the experiment. On this basis, training is carried out, and according to Figure 4 It can be seen that the M-RTCN proposed in the application has the fastest convergence speed and the accuracy is further improved compared with other network models.

[0097] Step E, build a ground test bench for physical test, verify the real-time performance and feasibility of the thrust estimation control scheme.

[0098] In specific implementation, the ground test bench mainly includes a control system (main controller, thrust estimator, auxiliary controller, console and measurement and control system), oil pump, force measuring bench, oil tank and the like.

[0099] In specific implementation, the control scheme can be further described as follows:

[0100] The engine is first started from a static state in an open loop, and then the given throttle lever angle is converted into a thrust target command, and the thrust error is output to the oil pump motor through the controller to control the engine fuel supply, and then the speed parameter is obtained, wherein the engine fuel and speed instructions are input into the thrust estimator to obtain the estimated thrust, and a closed-loop control structure of thrust feedback is formed. The estimated results obtained based on the M-RTCN neural network thrust estimator are compared with the measurement results fed back by the force measuring bench, and the dynamic characteristic curve of direct thrust control is shown in Figure 5 、 6 .

[0101] It can be seen that the thrust estimation method has good effect, and the estimation error in the system control process can be basically controlled within 2%. The phenomenon of larger error at some time is mainly affected by the physical factors of the test site, such as the vibration of the engine during high-speed running in the large-thrust stage, which causes the test bench to be in a non-horizontal state, and then affects the thrust estimation effect.

[0102] In addition, according to the test results shown in the figure, the tracking performance of the thrust control is good, and there is almost no overshoot, and the steady-state error accuracy is about 1%, and the regulation time is 2-3s, which shows that the thrust controller proposed in the application achieves the design expectation, and further verifies the feasibility of the control scheme based on the neural network thrust estimator.

[0103] In summary, the aero-engine thrust estimation method based on the time domain convolution network provides a new thrust estimation design idea. The improved model is proposed according to the basic structure characteristics of the time domain convolution network, the Gaussian error linear unit is used to replace the activation function of the internal residual block, and the overall structure of the network is adjusted by using the residual connection, and the training convergence speed and thrust estimation accuracy of the network model are obviously improved through comparison with various methods. Meanwhile, the method is applied to the actual control system of the aero-engine, and the real-time performance and feasibility of the control scheme are further verified through the physical test of the micro turbojet engine test bed. The method provided by the application can realize the thrust estimation based on the neural network, is suitable for the direct thrust control of the aero-engine, and has important practical significance in engineering application.

Claims

1. A time-domain convolution network-based aero-engine thrust estimation method, characterized in that, The steps are as follows: Step A: Obtain the output data of various parameters through micro turbojet engine test, and the data should cover the steady-state operation process of the engine from the slow-speed state to the acceleration and deceleration stages as far as possible; Step B: Data preprocessing; screening the original data obtained in step A, extracting important feature parameters related to thrust estimation, and obtaining training, validation and test data sets of the neural network model through data segmentation and standardization processing after feature selection; The important feature parameters related to thrust estimation include speed, fuel, engine exhaust temperature and environmental humidity; The segmentation is to divide the engine feature data into training set and validation set in the proportion of 0.75 and 0.25 by using the leave-one-out cross-validation method; The standardization adopts the following mean-variance standardization method, so that all features conform to the distribution characteristics with mean value of 0 and standard deviation of 1; Here, μ is the mean of the feature x to be processed, δ is the standard deviation of the feature x, and x * is the feature after standardization. Step C: Combine the time series of features other than thrust in the training data set as input, and use the self-defined callback function training algorithm to train the improved time-domain convolution network model for regression, and calculate the estimation accuracy of the current network model in real time during the training process using the test data set, and select the algorithm to obtain the best estimation network model; The time series combination input of the thrust outside characteristics includes fuel information W(m-1),…,W(m-n+1) of the current m moment and the history (n-1) moments f (m),…,W f (m-n+1), the speed information N(m-1),…,N(m-n) of the history n moments and the exhaust port temperature and the environment humidity T(m) and H(m) of the current moment; the neural network model is represented as: F(m) = g(W f (m),…,W f (m-n+1),N(m-1),…,N(m-n),T(m),H(m)) Wherein, g(·) represents the mapping function relationship of the neural network model, and F(m) is the output thrust at the current time; The callback function training algorithm includes the following steps: (1) Define the neural network training process callback function according to the evaluation index; Evaluation metrics, including mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R) 2 defined as follows: where y i , are the actual, estimated and average values, respectively; The self-defined callback function inherits the Callback class, and uses the test set to predict the current model output at the end of each epoch, obtains the above evaluation index, and the specific data in the training process will be returned as a dictionary form as the log content; At the same time, the ModelCheckpoint class function is added to realize the printing output of various indexes and the saving function of the best model; (2) Start training after setting the network model parameters; The parameter setting includes the following parts: 2.1) Network part: filter number, convolution kernel size k, expansion factor d, and full connection layer unit number units; 2.2) Training part; single training sample number batch, network training number epoch, and test error initial value; The improved time-domain convolution network model includes improvements in internal composition and overall network structure: The improvement in internal composition is to use a Gaussian error linear unit activation function instead of a rectified linear unit in the residual block of the time-domain convolution network; The Gaussian error linear unit activation function is represented as: wherein Z ~ N(0, 1) is the cumulative distribution function of the standard normal distribution, erf(·) is the error function; The improvement in the overall network structure is to combine multiple lightweight TCN blocks with different memories through residual connection; (3) Calculate the test error of the current training model; (4) If the error is smaller than the initial value, update the initial value and save the model, otherwise keep the initial value unchanged; (5) Continue the next training, repeat steps (3), (4) until the training is completed; (6) Output the final network model.

2. The time-domain convolution network based aero-engine thrust estimation method according to claim 1, characterized in that, The acceleration stage in step A is the process from 50000 rpm to 150000 rpm, and the deceleration stage is the process from 150000 rpm to 50000 rpm.

3. The time-domain convolution network based aero-engine thrust estimation method according to claim 1, characterized in that, The optimal parameters of the network part in step C are mainly obtained through a variety of combination tests, and the specific selection range is shown in the following table: Table 1 Parameter selection 4. The time-domain convolution network based aero-engine thrust estimation method according to claim 1, characterized in that, The number of single training samples of the training part in step C is batch = 100, and the network training times are epoch = 200.

Images