Thrust optimization control method of turbofan engine based on time domain convolutional network

Abstract

The present invention relates to a method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network. The method comprises the following steps: obtaining output data of various parameters during the open-loop operation of the turbofan engine; training a time-domain convolutional network model to obtain a regression model of the turbofan engine; applying a control variable to the time-domain convolutional network as a prediction model to design a nonlinear optimization controller; and, based on the nonlinear optimization controller, forming a thrust closed-loop control system together with the turbofan engine. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network proposed in the present invention utilizes the basic structural characteristics of the time-domain convolutional network to design a prediction model that fully reflects the nonlinear system characteristics of the engine. This prediction model is then combined with an optimization control method and applied to the actual control of the turbofan engine. This method solves the problems of long response time of dynamic prediction models and poor control effects of traditional neural network models, thereby achieving excellent tracking performance and control effects.

Description

Technical Field

[0001] The present invention relates to the field of engine control technology, and in particular to a turbofan engine thrust optimization control method based on a time domain convolutional network. Background Art

[0002] For complex nonlinear, multivariable control objects such as aircraft engines, even if traditional PID control algorithms can be used for control, the adjustment of PID parameters is relatively difficult due to the complex coupling relationship between the various control loops. In addition, the aircraft engine control process needs to consider various constraints such as temperature and speed, and a simple controller is difficult to achieve ideal control effects.

[0003] The model predictive control algorithm is a special optimization control method that considers the rolling solution of the optimization problem under constraints. The basic idea of ​​this method is to predict the system output over a period of time by solving the constrained open-loop optimal control problem in a finite time domain based on the control system model, and repeatedly perform this operation to achieve the control goal of minimizing the local tracking error. Model predictive control uses rolling optimization and feedback correction strategies to promptly compensate for the uncertainty caused by model mismatch and external disturbances, thereby achieving good dynamic performance and strong robustness. It is widely used in fields with multiple variables, constraints, and time delays. The algorithmic theory of achieving stable operation and optimal control of closed-loop systems through predictive control has always been a research focus. The establishment of the prediction model, as one of the key steps, is often based on the physical meaning and dynamic characteristics of the system. However, in the specific implementation process of rolling optimization, the computational complexity and calculation time of model predictive control iterations are directly affected by the model dimensions and the number of control variables, which places high demands on the computing performance of the system equipment. In addition, the actual control effect is highly dependent on the accuracy of the predictive model modeling. Therefore, predictive control methods are mostly used in slow-dynamic industrial processes that allow a certain control error and can use a larger sampling period, such as chemical industry and refining. However, how to widely apply them to fast-dynamic fields such as aerospace, which have strict requirements on control period and control accuracy, still faces challenges.

[0004] Data-driven approaches do not utilize any model-related physical information. Instead, they directly map the relationship between model inputs and outputs using a large amount of experimental data to establish the model. These methods, such as genetic neural networks, least squares support vector machines, and correlation analysis, simplify the implementation process compared to model-based approaches and are widely applicable to various control objects. In recent years, with the continuous development of neural networks, their application as predictive models in model predictive control (MPC) to replace traditional nonlinear models has gradually become a research focus. However, using traditional network models such as multi-layer feedforward neural networks, convolutional neural networks, deep neural networks, and recurrent neural networks to implement nonlinear MPC using unmeasurable thrust as a direct control variable still struggles to achieve good control results. Further design of thrust predictive controllers based on engine operating characteristics is needed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a turbofan engine thrust optimization control method based on a time domain convolutional network, which is used to solve the problems of long response time of dynamic prediction models and poor control effect of traditional neural network models.

[0006] The present invention is achieved through the following technical solutions:

[0007] A turbofan engine thrust optimization control method based on a time-domain convolutional network comprises the following steps:

[0008] S1. Obtain output data of various parameters during the open-loop operation of the turbofan engine;

[0009] S2. Train the time-domain convolutional network model to obtain a regression model for the turbofan engine.

[0010] S3. Applying the control variable to the time-domain convolutional network as a prediction model to design a nonlinear optimization controller;

[0011] S4. Based on the nonlinear optimization controller, a thrust closed-loop control system is formed together with the turbofan engine.

[0012] According to the above technical solution, preferably, in step S1, the output data is preprocessed, and the data preprocessing includes data screening and data normalization. Data screening is to extract important parameters that have a strong correlation with engine thrust and remove duplicate parameters with feature inclusion relationships to simplify the model and speed up the solution of subsequent control optimization problems. Data normalization uses the following minimum-maximum scaling method, which is mainly used to eliminate scale differences between different features and accelerate network training convergence.

[0013] According to the above technical solution, preferably, in step S2, the output data of the various parameters that have undergone data preprocessing in step S1 are used as input data for training the time domain convolutional network model, and the input data includes the main fuel information W at the current time k f (k), throat area A8(k), compressor outlet pressure P t3 (k), high pressure turbine outlet pressure and temperature P t5 (k), T t5 (k), low-pressure turbine outlet pressure and temperature P t6 (k), T t6 (k), high and low pressure rotor speed N H (k), N L (k).

[0014] According to the above technical solution, preferably, step S2 also includes: performing fitting verification on the regression model of the turbofan engine and testing its prediction accuracy.

[0015] According to the above technical solution, preferably, in step S3, the nonlinear optimization controller introduces a penalty term for the control input increment, specifically:

[0016]

[0017] sty min ≤y(k+i)≤y max

[0018] Δu min ≤Δu(k+j)≤Δu max

[0019] i=1,…,n y

[0020] j=0,…,n u -1,

[0021] Among them, y r is the expected reference trajectory of the engine output, y(k+i) and Δu(k+j) are the predicted future output of the engine and the control increment to be optimized, respectively. Q is the prediction error weight coefficient matrix, P is the control increment weight coefficient matrix, and n y represents the prediction time domain, n u represents the control time domain, and n y ≥n u ;y min ,y max Respectively represent the minimum and maximum values ​​of engine output; Δu min , Δu max They represent the minimum and maximum values ​​of the engine control increment respectively.

[0022] The beneficial effects of the present invention are:

[0023] The thrust optimization control method of a turbofan engine based on a time-domain convolutional network proposed in the present invention designs a prediction model that can fully reflect the nonlinear system characteristics of the engine according to the basic structural characteristics of the time-domain convolutional network, and combines it with the optimization control method and applies it to the actual control of the turbofan engine. It solves the problems of long response time of the dynamic prediction model and poor control effect of the traditional neural network model, and has good tracking performance and control effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural diagram of the nonlinear model predictive control system of the present invention.

[0025] Figure 2 This is the structure diagram of the time domain convolutional network model.

[0026] Figure 3 This is the schematic diagram of the thrust optimization control method based on the time domain convolutional network.

[0027] Figure 4 It is the thrust control curve of the engine under fixed flight altitude and Mach number conditions.

[0028] Figure 5 It is a partial enlarged view of the response curve after the target thrust increases.

[0029] Figure 6 It is a partial enlarged view of the response curve after the target thrust decreases. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and the best embodiment. Based on the embodiments of the invention, all other embodiments obtained by those skilled in the art without making any creative work shall fall within the scope of protection of the invention.

[0031] It should also be noted that the control system solution described in the embodiment of the present invention is mainly aimed at the direct thrust control of aircraft engines. The following will take a turbofan engine as an example to specifically describe the turbofan engine thrust optimization control method based on a time domain convolutional network provided by the present invention.

[0032] As shown in the figure, the present invention includes the following steps:

[0033] Step S1. Input a control signal to the dynamic model of an existing turbofan engine, obtain output data of various parameters during the open-loop operation of the turbofan engine, and perform data preprocessing.

[0034] Among them, the input control signal includes the main fuel W f, afterburner fuel W fa , front duct jet FVABI, rear duct jet RVABI, throat area A8, flight altitude H and Mach number Ma, where the afterburner fuel is always zero, the duct jet, altitude and Mach number remain constant, the main fuel and throat area are controlled, and a pseudo-random sequence is introduced within a reasonable range to ensure that the operation process covers all thrust stages of the engine.

[0035] Data preprocessing includes data screening and data normalization. Data screening aims to extract important parameters with a strong correlation with engine thrust and remove duplicate parameters with feature inclusion relationships to simplify the model and accelerate the solution of subsequent control optimization problems. Data normalization uses the following minimum-maximum scaling method to eliminate scale differences between different features and accelerate network training convergence.

[0036]

[0037] where x min and x max They are the minimum and maximum values, x * It is the characteristic after standardization.

[0038] Step S2. Use the output data of various parameters that have undergone data preprocessing in step S1 as input data for training the time domain convolutional network model, train the time domain convolutional network model, and obtain a regression model of the turbofan engine.

[0039] The input data includes the main fuel information W at the current k moment f (k), throat area A8(k), compressor outlet pressure P t3 (k), high pressure turbine outlet pressure and temperature P t5 (k), T t5 (k), low-pressure turbine outlet pressure and temperature P t6 (k), T t6 (k), high and low pressure rotor speed N H (k), N L (k) and other sensor information; the predicted output data includes the engine thrust F(k+1) at the next moment and the above sensor information.

[0040] Specifically, the time domain convolutional network model can be expressed as:

[0041] y k+1 =g(x k )

[0042] Among them, x k =[W f (k),P t3 (k),P t5 (k),Tt5 (k),P t6 (k),T t6 (k),N H (k),N L (k)],y k+1 =[F(k+1),P t3 (k+1),P t5 (k+1),T t5 (k+1),P t6 (k+1),T t6 (k+1),N H (k+1),N L (k+1)], g(·) represents the mapping function relationship of the time domain convolutional network model. In this example, the network model uses three residual blocks. The kernel size of the dilated causal convolution layer in each residual block is 3, and the number of filters is 32. The overall structure is as follows Figure 2 shown.

[0043] In addition, the regression model of the turbofan engine is fitted and verified to test its prediction accuracy. The model testing method in this example is to compare the prediction output of the time domain convolutional network model on the test set with any given input, and then judge the accuracy of the fitting model. The evaluation indicators used include mean square absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and determination coefficient (R) 2 , the definition method is as follows:

[0044]

[0045]

[0046]

[0047]

[0048] where y i , They are actual value, estimated value and average value respectively. The specific test results are shown in the following table.

[0049]

[0050] Step S3. Apply the control quantity to the time domain convolutional network as a prediction model to design a nonlinear optimization controller. Figure 3 As shown in the figure, it mainly includes three parts: a nonlinear prediction model that can reflect the engine's large-scale steady-state and dynamic response, a nonlinear optimization algorithm that realizes input and output limit management, and a feedback correction link.

[0051] Among them, the time-domain convolutional network model is used as a prediction model. It is necessary to unify the network model storage format and use the AI ​​open source toolkit OpenVINO provided by Intel to perform inference deployment on the model to achieve experimental verification of the overall control solution. OpenVINO supports the optimization and compression of models in TensorFlow, TensorFlow Lite, PaddlePaddle and ONNX formats into an intermediate representation format (IR) or directly read and used by the runtime library (Runtime). The specific deployment process can be divided into the following steps: Model initialization: Create a kernel object, read the model through the kernel object and load the model to the CPU, and create an inference request based on the model; Pre-processing: Normalize and pre-process the input data, and use the data to fill the input tensor; Inference: Perform calculations by calling the dynamically loaded plug-in of the corresponding hardware device; Post-processing: Obtain the inference results and post-process the data.

[0052] The control objective of the nonlinear optimization controller is to control the thrust to the desired value under the input and output constraints while ensuring good dynamic performance. In order to make the system smoothly adjust the control input during the control process and slow down the rapid change or oscillation of the output, a penalty term Δu for the control input increment is introduced. The control increment is defined as:

[0053] Δu(k)=u(k)-u(k-1)

[0054] The cost function can be specifically described as:

[0055]

[0056] sty min ≤y(k+i)≤y max

[0057] Δu min ≤Δu(k+j)≤Δu max

[0058] i=1,…,n y

[0059] j=0,…,n u -1,

[0060] Among them, y r is the expected reference trajectory of the engine output, y(k+i) and Δu(k+j) are the predicted future output of the engine and the control increment to be optimized, respectively. Q is the prediction error weight coefficient matrix, P is the control increment weight coefficient matrix, and n y represents the prediction time domain, n u represents the control time domain, and n y ≥n u ;y min,y max Respectively represent the minimum and maximum values ​​of engine output; Δu min , Δu max They represent the minimum and maximum values ​​of the engine control increment respectively.

[0061] After clarifying the specific optimization form, the optimal control problem in the finite control time domain is solved online. The control variable is applied to the prediction model, and only the first control variable is applied to the actual engine model to update the state parameters. This process is repeated online as the optimization time domain progresses, continuously optimizing the control input. At the same time, the weight coefficients of each item are further appropriately adjusted to balance different objectives and constraints, completing the design of the nonlinear optimization controller.

[0062] Step S4. Based on the nonlinear optimization controller, a thrust closed-loop control system is formed together with the turbofan engine, and a thrust tracking instruction is set to verify the tracking performance and control effect of the overall control scheme.

[0063] Under the condition of fixed altitude and Mach number, a thrust target is given for verification. The thrust command signal includes acceleration and deceleration stages of different amplitudes, and multiple steady-state points are set in the process. The specific simulation results are as follows: Figure 4-6 As shown in the figure, the time-domain convolutional network model clearly demonstrates excellent thrust prediction, with its prediction error remaining within 2% during system control. Furthermore, the test results in the figure demonstrate excellent tracking performance, with an overshoot of less than 5%, a steady-state error of less than 2%, and a settling time of approximately 3 seconds. The proposed thrust prediction controller meets design expectations, further validating the superiority of this proposed thrust control scheme based on the time-domain convolutional network.

[0064] The thrust optimization control method for a turbofan engine based on a time-domain convolutional network proposed in the present invention designs a prediction model that can fully reflect the nonlinear system characteristics of the engine according to the basic structural characteristics of the time-domain convolutional network, and combines it with the optimization control method and applies it to the actual control of the turbofan engine, solving the problems of long response time of the dynamic prediction model and poor control effect of the traditional neural network model. The simulation test results also show that the thrust optimization control method has good tracking performance and control effect.

[0065] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A turbofan engine thrust optimization control method based on time domain convolutional network, characterized in that: The steps include: S1. Obtain output data of various parameters during the open-loop operation of the turbofan engine; S2. Train the time-domain convolutional network model to obtain a regression model for the turbofan engine. S3. Applying the control variable to the time-domain convolutional network as a prediction model to design a nonlinear optimization controller; S4. Based on the nonlinear optimization controller, a thrust closed-loop control system is formed together with the turbofan engine. In step S3, the nonlinear optimization controller introduces a penalty term for the control input increment, specifically: s.t.y min ≤y(k+i)≤y max Δu min ≤Δu(k+j)≤Δu max i=1,…,n y j=0,…,n u -1, Among them, y r is the expected reference trajectory of the engine output, y(k+i) and Δu(k+j) are the predicted future output of the engine and the control increment to be optimized, respectively. Q is the prediction error weight coefficient matrix, P is the control increment weight coefficient matrix, and n y represents the prediction time domain, n u represents the control time domain, and n y ≥n u ;y min ,y max Respectively represent the minimum and maximum values ​​of engine output; Δu min , Δu m a x They represent the minimum and maximum values ​​of the engine control increment respectively.

2. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network according to claim 1, characterized in that: In step S1, data preprocessing is performed on the output data.

3. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network according to claim 2, characterized in that: The data preprocessing in step S1 includes data screening and data normalization.

4. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network according to claim 2 or 3, characterized in that: In step S2, the output data of the various parameters that have undergone data preprocessing in step S1 are used as input data for training the time domain convolutional network model.

5. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network according to claim 4, characterized in that: In step S2, the input data includes the main fuel information W at the current time k f (k), throat area A8(k), compressor outlet pressure P t3 (k), high pressure turbine outlet pressure and temperature P t5 (k), / t5 (k), low-pressure turbine outlet pressure and temperature P t6 (k), / t6 (k), high and low pressure rotor speed 0 / (k), 00(k).

6. The method for optimizing thrust control of a turbofan engine based on a time-domain convolutional network according to claim 4, characterized in that: Step S2 also includes: performing fitting verification on the regression model of the turbofan engine.

Images