A turbofan engine thrust real-time estimation method fusing UKF and neural network
Patent Information
- Application Number
- CN202310963903.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-02
- Publication Date
- 2026-10-23
- Estimated Expiration
- 2043-08-02
AI Technical Summary
[0005]本发明要解决的技术问题是涡扇发动机推力的机载自适应模型无法同时保证高精度和高实时性问题
[0042]本发明的效果和益处是：本发明能够充分结合已有的相关知识，将涡扇发动机自适应建模中的数据驱动方法和模型驱动方法有效地结合起来，并通过机载自适应模型的计算进行推力的实时估计，有效提高了机载自适应模型的实时性，为同时实现机载自适应模型的高精度和高实时性提供了新的解决思路。这种融合LM-IUKF方法和LSTM神经网络模型的推力估计过程，能够有效解决传统方法的精度低、收敛速度慢、易发散、计算速度慢等问题，在涡扇发动机气路性能出现退化的状态下，兼具推力估计精度和实时性，满足实际工程应用的25ms运行周期要求。同时该方法可适用于航空发动机的直接推力控制，在工程实践中具有重要意义。
Smart Images
Figure CN116933654B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of system identification, engine adaptive modeling and thrust estimation. It is a real-time thrust estimation method for turbofan engines that integrates UKF and neural networks. Background Technology
[0002] In turbofan engine control systems, to ensure safe and reliable engine operation, a more efficient approach than traditional indirect control using engine speed is direct control of thrust, a key characteristic of aero-engine performance. Under constant flight altitude and Mach number, a direct linear relationship exists between the throttle lever and thrust, meaning that internal engine performance degradation or changes in external environmental factors will not affect the control system. However, currently, there are no sensors capable of directly measuring thrust. Therefore, designing a high-precision and real-time thrust estimator is crucial for achieving direct performance control with thrust as the controlled variable, significantly improving engine control quality and service life.
[0003] Thrust estimation methods are mainly divided into two categories: model-driven methods and data-driven methods. Model-driven methods obtain the engine's thrust estimate by establishing an onboard adaptive model of the engine and using the output of the onboard adaptive model. As described in the paper "A comparison of filtering approaches for aircraft engine healthestimation" published in Aerospace Science and Technology by Dan Simon et al., Kalman filtering, as a commonly used model-based state estimation method, mainly includes two processes: state prediction and measurement update. It and its improved methods are widely used in the onboard adaptive modeling and thrust estimation process of turbofan engines by academic and industrial communities at home and abroad, including linear Kalman filtering, extended Kalman filtering, unscented Kalman filtering, and particle filtering methods. On the other hand, with the development of computing technology and the accumulation of flight data, data-driven methods have gained widespread attention in recent years. As described in the paper "A size-transferring radial basis function network for aero-engine thrustestimation" published in the journal *Engineering Applications of Artificial Intelligence* by Yong-Ping Zhao et al., commonly used methods include: neural network methods, which have the ability to model nonlinear and complex dynamic systems; machine learning methods, which have simpler network structures and faster training and deployment times; and deep learning methods, which have more complex network structures to achieve more accurate fitting and prediction of system dynamic characteristics. Data-driven methods rely solely on collected data and do not involve additional characteristic curves or thermodynamic theories for calculation, simplifying the parameter estimation design process.
[0004] However, the above methods have the following shortcomings: Model-driven thrust estimation methods, such as linear filtering, have the advantage of high real-time performance, but there is a certain loss of accuracy. As described in the paper "A hybrid onboard adaptive model for aero-engine parameter prediction" published by Shuwei Pang et al. in Aerospace Science and Technology, nonlinear filtering algorithms can achieve more accurate parameter estimation results, but they are constrained by the limited computing power and storage resources of current airborne systems, and cannot simultaneously guarantee accuracy and real-time performance. Data-driven thrust estimation methods simplify the implementation process compared to model-based methods, but they are highly sensitive to data. As described in the paper "Performance-Based Gas Turbine Health Monitoring, Diagnostics, and Prognostics: A Survey" published by Houman Hanachi et al. in TRANSACTIONS ON RELIABILITY, they are highly dependent on reliable data sources and preprocessing processes. Furthermore, due to the complex and diverse characteristics of network structures, it is sometimes impossible to simultaneously achieve high accuracy and high real-time performance in practical applications of thrust estimation. Summary of the Invention
[0005] The technical problem this invention aims to solve is that the airborne adaptive model for turbofan engine thrust cannot simultaneously guarantee high accuracy and high real-time performance. To address this issue, this invention proposes a real-time thrust estimation method for turbofan engines that integrates UKF and neural networks, combining the high real-time performance advantages of data-driven methods with an airborne adaptive model based on a model-driven approach. First, an airborne adaptive model for a turbofan engine based on the LM-IUKF method is established using the Levenberg-Marquardt optimization algorithm combined with the Unscented Kalman Filter (UKF) method. This model exhibits high accuracy and rapid tracking and estimation capabilities for engine state under performance degradation. Second, an LSTM model is used to fit the turbofan engine's mechanistic model, and process data generated by the mechanistic model is collected. A corresponding data-driven engine model is then established through training. Finally, a real-time thrust estimation model based on the fusion of UKF and neural networks is proposed. The trained engine LSTM model is applied to the measurement and update process of the LM-IUKF method, replacing the nonlinear mechanistic model used in the LM-IUKF method, thus improving the real-time performance of the airborne adaptive model's thrust estimation process. The real-time performance and effectiveness of the proposed method are further verified through hardware-in-the-loop experiments with a runtime of 25ms.
[0006] The technical solution of this invention:
[0007] A real-time thrust estimation method for turbofan engines integrating UKF and neural networks, comprising the following steps:
[0008] Step 1: Establish an airborne adaptive model of the turbofan engine based on the LM-IUKF method;
[0009] When establishing an airborne adaptive model for a turbofan engine using the Unscented Kalman Filter (UKF), the calculation process is improved. The UKF mainly consists of two processes: state transition and measurement update. For the measurement update process, the state estimation problem is transformed into an optimization problem, as follows:
[0010]
[0011] in,
[0012]
[0013] In the formula, Let r(x) be the vector of engine health parameters to be estimated, and V(x) be the cost function; in the vector r(x), the first part... Characterizing the bias in sensor measurement noise, Part Two Characterizing the deviation of state prediction; y t Let and h(x) be the output and observation model of the engine system at time t, respectively, and R t To measure the noise covariance, P is the state vector covariance; t indicates that the engine is operating at time t.
[0014] The Levenberg-Marquart algorithm is used to solve equation (1) to obtain the LM-IUKF method with iterative form. Based on the LM-IUKF method, an onboard adaptive model of the turbofan engine is established. The measurable input and output of the engine are used as the inputs of the onboard adaptive model of the turbofan engine. The measurable input of the engine is the fuel supply to the main combustion chamber. The measurable output of the engine includes the low-pressure rotor speed, the high-pressure rotor speed, the internal pressure after the fan, the total pressure at the outlet of the high-pressure compressor, the total temperature at the outlet of the high-pressure compressor, and the total pressure at the outlet of the low-pressure turbine. The unmeasurable output of the onboard adaptive model of the turbofan engine is the thrust and the health parameters of the four rotating components of the air path. The health parameters include the efficiency factor of the high-pressure compressor, the efficiency factor of the fan, the efficiency factor of the high-pressure turbine, and the efficiency factor of the low-pressure turbine. When the engine performance degrades, the onboard adaptive model of the turbofan engine estimates the four health parameters and updates the onboard adaptive model of the turbofan engine to obtain the thrust value under the performance degradation state.
[0015] Step 2: Establish an LSTM neural network model for the turbofan engine
[0016] Step 2.1: Data Selection and Preprocessing
[0017] (1) Data selection
[0018] Operating data of the turbofan engine under performance degradation of rotating components in the gas path were obtained from the nonlinear mechanism model of the turbofan engine. The obtained operating data covers the entire steady-state operation process of the engine under various degradation conditions. Data sets were composed of important characteristic parameters related to engine thrust, including main combustion chamber fuel supply, low-pressure rotor speed, high-pressure rotor speed, fan back chamber pressure, high-pressure compressor outlet total pressure, high-pressure compressor outlet total temperature and low-pressure turbine outlet total pressure.
[0019] (2) Data preprocessing
[0020] The dataset is divided into training and validation sets proportionally, and normalization is performed based on the data range, calculated as follows:
[0021]
[0022] In the formula, x i and These represent the data components before and after normalization, x and x respectively. ma and x minThese are the maximum and minimum values of the data for the variable to be normalized, respectively.
[0023] The engine output parameters at time t are used as y data,t The engine output parameters include low-pressure rotor speed, high-pressure rotor speed, internal pressure after the fan, total pressure at the high-pressure compressor outlet, total temperature at the high-pressure compressor outlet, total pressure at the low-pressure turbine outlet, and thrust. The model input is a sequence combination of the engine's measurable inputs, four health parameters, and the engine's measurable outputs at time (t-1). The LSTM neural network model is then represented as:
[0024]
[0025] In the formula, W f For engine fuel flow; SE1, SE2, SE3, and SE4 are the high-pressure compressor efficiency factor, fan efficiency factor, high-pressure turbine efficiency factor, and low-pressure turbine efficiency factor, respectively, which are the four health parameters; dataModel(·) represents the LSTM neural network model;
[0026] Step 2.2: After setting the network parameters, start training. The parameter settings include the following:
[0027] (1) Model structure parameters
[0028] To obtain a lightweight network, the number of hidden layers was set to 3, the number of nodes in fully connected layers and LSTM layers was 30 each, and the number of nodes in the input and output layers was 14 and 10 respectively, depending on the dimension of the variables. Two ReLU activation functions were used to restrict the network input and output to positive values. The ReLU calculation method is as follows:
[0029] ReLU(x) = max(0,x) #(4)
[0030] (2) Model training parameters
[0031] The batch size was set to 128, and the number of training epochs was 20,000. The initial learning rate was 0.005. To improve model accuracy, the learning rate was changed linearly in a piecewise manner during training. After every 1,000 iterations, lr = lr * 0.6. The Adam learning rate adaptive algorithm was used for parameter optimization.
[0032] Step 3: Establish a real-time thrust estimation model for the engine based on the fusion of UKF and neural network methods;
[0033] To improve the real-time performance of the turbofan engine airborne adaptive model established in step 1, the LSTM neural network model from step 2 is integrated into the measurement update process of the LM-IUKF method, replacing the nonlinear mechanism model of the turbofan engine in step 1. The LSTM neural network model in step 2 iteratively updates its internal parameters during computation, which cannot match the requirement in step 1 to maintain the same initial conditions when calculating the Sigma point set. Therefore, the calculation process for the Sigma point set is improved, and the calculation steps for a measurement update process at time t are as follows:
[0034] ① Record the neuron states C of the LSTM neural network model t and neuron output H t ;
[0035] ② Calculate the output of the LSTM neural network model for the input Sigma point according to equation (3);
[0036] ③ Restore the neuron state C of the LSTM neural network model t and neuron output H t ;
[0037] ④ Proceed to the next step based on whether the Sigma point set calculation is complete. If the calculation is complete, proceed to step ⑤.
[0038] Otherwise, return to step ① and continue the calculation;
[0039] ⑤ Perform state and covariance updates based on the covariance matrix and Kalman gain of the LM-IUKF algorithm.
[0040] Obtain estimated values for health parameters;
[0041] ⑥ Update the neuron state of the LSTM neural network model to C t+1 and neuron output H t+1 Based on the health parameters obtained in step ⑤, the airborne adaptive model is updated to obtain the real-time thrust output by the model.
[0042] The effects and benefits of this invention are as follows: This invention can fully integrate existing relevant knowledge, effectively combining data-driven and model-driven methods in adaptive modeling of turbofan engines. It achieves real-time thrust estimation through the calculation of the airborne adaptive model, effectively improving the real-time performance of the airborne adaptive model and providing a new solution for simultaneously achieving high accuracy and high real-time performance of the airborne adaptive model. This thrust estimation process, which integrates the LM-IUKF method and the LSTM neural network model, effectively solves the problems of low accuracy, slow convergence speed, easy divergence, and slow calculation speed of traditional methods. Even when the turbofan engine's aerodynamic performance degrades, it maintains both thrust estimation accuracy and real-time performance, meeting the 25ms cycle requirement for practical engineering applications. Furthermore, this method is applicable to direct thrust control of aero-engines, which is of great significance in engineering practice. Attached Figure Description
[0043] Figure 1 This is a structural diagram of the airborne adaptive model system based on the LM-IUKF method in this invention.
[0044] Figure 2 This is a structural diagram of the LSTM data-driven model in this invention.
[0045] Figure 3(a) is a schematic diagram of the fusion process in this invention.
[0046] Figure 3(b) is a flowchart of the measurement and update process of the LM-IUKF algorithm after fusion according to the present invention.
[0047] Figure 4 This is a structural diagram of the airborne adaptive model based on the fusion data-driven method of this invention.
[0048] Figure 5 This is a comparison chart of the accuracy of the LSTM model and the fused data-driven model in this invention.
[0049] Figure 6(a) is a structural diagram of the hardware-in-the-loop experimental system of the fusion data-driven model of the present invention.
[0050] Figure 6(b) shows the accuracy results of thrust estimation using the fusion data-driven model of the present invention.
[0051] Figure 6(c) is a statistical chart of the time consumption for thrust estimation using the data-driven model of the present invention. Detailed Implementation
[0052] The following description, in conjunction with the accompanying drawings and technical solutions, uses a certain type of dual-rotor high-bypass turbofan engine as an example to further illustrate the specific implementation of the present invention.
[0053] This invention proposes a real-time thrust estimation method for turbofan engines that integrates UKF and neural networks. The specific implementation steps are as follows:
[0054] Step 1: Establish an airborne adaptive model of the turbofan engine based on the LM-IUKF method;
[0055] When establishing an adaptive model for a turbofan engine using the Unscented Kalman Filter (UKF), problems such as low accuracy, slow convergence speed, and easy divergence exist. Therefore, the computational process is improved. The UKF mainly consists of two processes: state transition and measurement update. For the measurement update process, the state estimation problem is transformed into an optimization problem, as follows:
[0056]
[0057] in,
[0058]
[0059] In the formula, The vector represents the engine health parameters to be estimated, specifically including the high-pressure compressor efficiency factor, fan efficiency factor, high-pressure turbine efficiency factor, and low-pressure turbine efficiency factor [SE1 SE2 SE3 SE4]. T V(x) is the cost function; in the vector r(x), the first part Characterizing the bias in sensor measurement noise, Part Two Characterizing the deviation description of the state prediction; y t Let and h(x) be the output and observation model of the engine system at time t, respectively, and R t To measure the noise covariance, P is the state vector covariance; t indicates that the engine is operating at time t.
[0060] The Levenberg-Marquart algorithm is used to solve equation (1) to obtain the LM-IUKF method with iterative form; a system is established based on the LM-IUKF method. Figure 1 The airborne adaptive model of the turbofan engine shown is based on measurable engine inputs (main combustion chamber fuel supply W). fm ) and output (including low-pressure rotor speed, high-pressure rotor speed, internal pressure after the fan, total pressure at the high-pressure compressor outlet, total temperature at the high-pressure compressor outlet, and total pressure at the low-pressure turbine outlet y=[N L N H P 25 P 31 T 31 P6] TAs input to the turbofan engine adaptive model, the unmeasurable output of the model is thrust and the health parameters of four rotating air path components. When the engine performance degrades, the airborne adaptive model of the turbofan engine estimates the four health parameters and updates the airborne adaptive model to obtain the thrust value under the performance degradation state. The results are verified by hardware simulation and real-time testing of the health parameter estimation process on the NVIDIA Jetson series hardware platform.
[0061] Step 2: Establish an LSTM neural network model for the turbofan engine
[0062] Step 2.1: Data Selection and Preprocessing
[0063] (1) Data selection
[0064] Operating data of the turbofan engine under performance degradation of its gas path components were obtained from a nonlinear mechanism model, covering the entire steady-state operation of the engine under various degradation conditions. A training dataset was compiled using key characteristic parameters related to engine thrust, including main combustion chamber fuel supply, low-pressure rotor speed, high-pressure rotor speed, fan-backfill internal pressure, high-pressure compressor outlet total pressure, high-pressure compressor outlet total temperature, and low-pressure turbine outlet total pressure. Furthermore, at the ground operating point, the W input parameter was analyzed... fm Step-type signal excitations were applied to SE1, SE2, SE3 and SE4 to obtain training set data consisting of engine input and output data under a wide range of health parameter changes. The model sampling period was 0.025 seconds.
[0065] (2) Data preprocessing
[0066] The dataset is divided into training and validation sets according to a certain ratio, and normalization is performed based on the data range. The calculation method is as follows:
[0067]
[0068] In the formula, x i and These represent the data components before and after normalization, respectively, x ma and x min These are the maximum and minimum values of the data for the variable to be normalized, respectively.
[0069] The engine output parameters at time t (including low-pressure rotor speed, high-pressure rotor speed, fan backfill pressure, high-pressure compressor outlet total pressure, high-pressure compressor outlet total temperature, low-pressure turbine outlet total pressure, and thrust) are used as y. data,tThe model input consists of a sequence of conventional control variables (main combustion chamber fuel supply), four health parameter variables, and the output measurable parameters at time (t-1) (including low-pressure rotor speed, high-pressure rotor speed, fan backfill pressure, high-pressure compressor outlet total pressure, high-pressure compressor outlet total temperature, and low-pressure turbine outlet total pressure). The LSTM network model is then represented as:
[0070]
[0071] In the formula, W f The engine fuel flow rate is represented by SE1, SE2, SE3, and SE4, which are the high-pressure compressor efficiency factor, fan efficiency factor, high-pressure turbine efficiency factor, and low-pressure turbine efficiency factor, respectively, representing the four estimated health parameters; dataModel(·) represents the LSTM network model.
[0072] Step 2.2: After setting the network parameters, start training. The parameter settings include the following:
[0073] (1) Model structure parameters
[0074] To obtain a lightweight network, establish such a network as Figure 2 The LSTM model shown has 3 hidden layers, 30 nodes in both the fully connected layers and the LSTM layers, and the number of nodes in the input and output layers is related to the variable dimension, being 14 and 10 respectively. Two ReLU activation functions are used to restrict the network input and output to positive values. The ReLU calculation method is as follows:
[0075] ReLU(x) = max(0,x) #(4)
[0076] (2) Model training parameters
[0077] The batch size was set to 128, and the number of training epochs was 20,000. The initial learning rate was 0.005. To improve model accuracy, the learning rate was changed linearly in a piecewise manner during training, with lr = lr * 0.6 after every 1,000 iterations. The Adam adaptive learning rate algorithm was used for parameter optimization. After 20,000 training epochs, the RMSE value on the validation set was 0.0024, indicating that the established data-driven model has high accuracy.
[0078] Step 3: Establish a real-time thrust estimation model for the engine based on the fusion of UKF and neural network methods;
[0079] To improve the real-time performance of the turbofan engine airborne adaptive model established in step 1, as shown in Figure 3, the LSTM model from step 2 is integrated into the measurement update process of the LM-IUKF method, replacing the engine nonlinear mechanism model in step 1. However, the LSTM model in step 2 iteratively updates its internal parameters during calculation, which cannot match the requirement in step 1 to maintain the same initial conditions when calculating the Sigma point set. Therefore, the calculation process of the Sigma point set is improved, and the calculation steps of a measurement update process at time t are as follows:
[0080] ① Record the neuron state C of the LSTM model t and neuron output H t ;
[0081] ② Calculate the LSTM model output for the input Sigma point according to equation (3);
[0082] ③ Restore the neuron state C of the LSTM model t and neuron output H t ;
[0083] ④ Proceed to the next step based on whether the Sigma point set has been calculated. If the calculation is complete, proceed to step ⑤; otherwise, return to step ① to continue the calculation.
[0084] ⑤ Perform state and covariance updates based on the covariance matrix and Kalman gain of the LM-IUKF algorithm.
[0085] Obtain estimated values for health parameters;
[0086] ⑥ Update the neuron state of the LSTM model to C t+1 and neuron output H t+1 Based on the health parameters obtained in step ⑤, the airborne adaptive model is updated to obtain the real-time thrust output by the model.
[0087] By combining the above six steps, we can obtain Figure 4 The airborne adaptive model of the turbofan engine shown is a fusion of UKF and LSTM neural network models. Closed-loop testing was performed on the hardware platform shown in Figure 6(a). Figure 5 The results show that both the parameter estimation results from the fusion of UKF and neural network methods and the parameter estimation results from the LSTM neural network-based method have high accuracy under multi-component performance degradation. However, the estimation results of the LSTM method deviate with increasing sequence length, especially in... Figure 5 The T shown 25The results show that the airborne adaptive model proposed in this invention has higher accuracy and also has a certain corrective effect on the estimation bias of LSTM. Figure 6(b) shows the thrust estimation results of the model of the established fusion data-driven method under different performance degradation modes. It can be seen that the thrust estimation model of this invention can accurately estimate the thrust under different performance degradation modes, with high accuracy and numerical stability, and can meet the thrust direct control requirements of actual engineering. Furthermore, in the single-step calculation time results shown in Figure 6(c), it can be seen that the parameter estimation model of the original LM-IUKF method has a single-step calculation time of close to 25ms in the hardware simulation from the 10th to the 20th second, which is close to the sampling period. At the same time, the calculation time in the 11th second reaches 46ms, which exceeds the sampling period limit. Considering that the direct control of thrust also requires the controller's solution process, the remaining time is difficult to meet the thrust control requirements. In contrast, the single-step calculation time of the parameter estimation model based on the fusion neural network model is stable within 1ms, with a maximum of only 3ms. No time limit was exceeded under a sampling period of 25ms, which meets the real-time requirements of thrust estimation.
[0088] Therefore, it can be seen that the method of the present invention meets the thrust control requirements of practical engineering in terms of accuracy and numerical stability. Comparison with parameter estimation models based on the LM-IUKF method and the LSTM model shows that the invented method can effectively improve the calculation speed of the thrust estimation process, reducing the average calculation time per step from 15ms to no more than 1ms, thus combining high accuracy and high real-time performance.
[0089] Figure 1 The airborne adaptive model established for this invention, based on the LM-IUKF method, uses the measurable output y of the turbofan engine and the estimated output of the airborne adaptive model. The deviation between them is used as input for health parameters. State estimation is performed, and then the parameters of the airborne model are adaptively adjusted to follow the actual engine state, thereby obtaining an estimate of the unmeasurable output thrust signal of the actual engine.
[0090] Figure 2 This diagram shows the network structure of the LSTM neural network model established based on turbofan engine operating data in this invention. The value of n in the diagram depends on the length of the model sequence input. A lightweight data-driven model is implemented using a single LSTM layer. Two ReLU activation functions are used to restrict the network input and output, preventing negative values and reflecting the meaning of the actual data. Simultaneously, during training, some neurons can be set to output 0, introducing sparsity, reducing the interdependence between parameters, and mitigating overfitting. Finally, a fully connected layer is used to obtain the model's predicted output, with the output dimension matching the output dimension of the engine model to be fitted.
[0091] Figure 3(a) illustrates the fusion process of the data-driven model in this invention. To improve the overall computational speed of the airborne adaptive model, an LSTM neural network model is used to replace the nonlinear mechanism model of the turbofan engine, thereby accelerating the calculation of the sampled point measurements. Therefore, the airborne adaptive model of the turbofan engine based on the fusion of UKF and neural network possesses both the accuracy advantage of the model-driven method and the speed advantage of the data-driven method.
[0092] Figure 3(b) illustrates the core technical aspect of this invention, depicting a flowchart of the measurement update process of the fused LM-IUKF algorithm. The iterative LM-IUKF algorithm calls the LSTM data-driven model to calculate the output Sigma point set during the measurement update process, thereby accelerating the calculation speed of the measurement update process and improving the overall real-time performance of the model.
[0093] Figure 4 This is a structural diagram of the airborne adaptive model based on the fusion of UKF and neural networks. The fused LM-IUKF method is used for real-time estimation of engine health parameters, enabling the airborne model to adjust in real time to track the actual engine state. This allows for real-time estimation of the actual engine thrust through a component-level model. This model building method can be extended to other data-driven methods for use in the fused airborne adaptive model.
[0094] Figure 5 This is a comparison chart of the accuracy of the LSTM model and the fused data-driven model in this invention. It can be seen that this invention can effectively reduce the tracking deviation of the LSTM model and achieve high-precision tracking of the measurable engine output when the simulation time is extended.
[0095] Figure 6(a) is a structural diagram of the hardware-in-the-loop test of the fusion neural network model of the present invention; Figure 6(b) is a diagram of the accuracy results of thrust estimation; Figure 6(c) is a statistical diagram of the time consumption of the hardware-in-the-loop test of thrust estimation using the fusion neural network model of the present invention. It can be seen that the method of the present invention can achieve high accuracy and numerical stability in thrust estimation while meeting the real-time requirement of 25ms, providing good feedback estimation for direct thrust control and possessing practical engineering significance.
[0096] In summary, this invention proposes a real-time thrust estimation method for turbofan engines based on the fusion of UKF and neural networks, offering a novel design approach for real-time thrust estimation. First, based on the turbofan engine's mechanistic model, an airborne adaptive model of the engine using the LM-IUKF method is established, resulting in a model with high-precision thrust estimation. Second, an LSTM neural network model of the turbofan engine is established, presenting a lightweight data-driven model that uses ReLU activation units to obtain data constraints consistent with actual data characteristics. Third, the established LSTM model is integrated into the measurement update process of the LM-IUKF method, improving the overall computational speed of the adaptive model, and the improved accuracy is verified through simulation. Furthermore, this invention applies this method to hardware-in-the-loop testing of a turbofan engine, further validating its real-time performance and feasibility. The method provided by this invention can achieve high-precision, high-real-time thrust estimation, is suitable for direct thrust control of aero-engines, and has significant practical implications for engineering applications.
Claims
1. A real-time thrust estimation method for turbofan engines integrating UKF and neural networks, characterized in that, The steps are as follows: Step 1: Establish an airborne adaptive model of the turbofan engine based on the LM-IUKF method; When establishing an airborne adaptive model for a turbofan engine using the Unscented Kalman Filter (UKF), its calculation process is improved. The UKF mainly consists of two processes: state transition and measurement update. For the measurement update process, the state estimation problem is transformed into an optimization problem, as follows: in, In the formula, The vector of engine health parameters to be estimated includes the high-pressure compressor efficiency factor, fan efficiency factor, high-pressure turbine efficiency factor, and low-pressure turbine efficiency factor [SE1 SE2 SE3 SE4]. T V(x) is the cost function; in the vector r(x), the first part Characterizing the bias in sensor measurement noise, Part Two Characterizing the bias in state prediction; t t Let and h(x) be the output and observation model of the engine system at time t, respectively, and R t To measure the noise covariance, P is the state vector covariance; t indicates that the engine is operating at time t. The Levenberg-Marquart algorithm is used to solve equation (1) to obtain the LM-IUKF method with iterative form. Based on the LM-IUKF method, an onboard adaptive model of the turbofan engine is established. The measurable input and output of the engine are used as the inputs of the onboard adaptive model of the turbofan engine. The measurable input of the engine is the fuel supply to the main combustion chamber. The measurable output of the engine includes the low-pressure rotor speed, the high-pressure rotor speed, the internal pressure after the fan, the total pressure at the outlet of the high-pressure compressor, the total temperature at the outlet of the high-pressure compressor, and the total pressure at the outlet of the low-pressure turbine. The unmeasurable output of the onboard adaptive model of the turbofan engine is the thrust and the health parameters of the four rotating components of the air path. The health parameters include the efficiency factor of the high-pressure compressor, the efficiency factor of the fan, the efficiency factor of the high-pressure turbine, and the efficiency factor of the low-pressure turbine. When the engine performance degrades, the onboard adaptive model of the turbofan engine estimates the four health parameters and updates the onboard adaptive model of the turbofan engine to obtain the thrust value under the performance degradation state. Step 2: Establish an LSTM neural network model for the turbofan engine Step 2.1: Data Selection and Preprocessing (1) Data selection Operating data of the turbofan engine under performance degradation of rotating components in the gas path were obtained from the nonlinear mechanism model of the turbofan engine. The obtained operating data covers the entire steady-state operation process of the engine under various degradation conditions. Data sets were composed of important characteristic parameters related to engine thrust, including main combustion chamber fuel supply, low-pressure rotor speed, high-pressure rotor speed, fan back chamber pressure, high-pressure compressor outlet total pressure, high-pressure compressor outlet total temperature and low-pressure turbine outlet total pressure. (2) Data preprocessing The dataset is divided into training and validation sets proportionally, and normalization is performed based on the data range, calculated as follows: In the formula, x i and These represent the data components before and after normalization, respectively, x max and x min These are the maximum and minimum values of the data for the variable to be normalized, respectively. The engine output parameters at time t are used as y data,t The engine output parameters include low-pressure rotor speed, high-pressure rotor speed, internal pressure after the fan, total pressure at the high-pressure compressor outlet, total temperature at the high-pressure compressor outlet, total pressure at the low-pressure turbine outlet, and thrust. The model input is a sequence combination of the engine's measurable inputs, four health parameters, and the engine's measurable outputs at time (t-1). The LSTM neural network model is then represented as: In the formula, W f For engine fuel flow; SE1, SE2, SE3, and SE4 are the high-pressure compressor efficiency factor, fan efficiency factor, high-pressure turbine efficiency factor, and low-pressure turbine efficiency factor, respectively, which are the four health parameters; dataModel(·) represents the LSTM neural network model; Step 2.2: After setting the network parameters, start training. The parameter settings include the following: (1) Model structure parameters To obtain a lightweight network, the number of hidden layers was set to 3, the number of nodes in fully connected layers and LSTM layers was 30 each, and the number of nodes in the input and output layers was 14 and 10 respectively, depending on the dimension of the variables. Two ReLU activation functions were used to restrict the network input and output to positive values. The ReLU calculation method is as follows: ReLU(x) = max(0,x) #(4) (2) Model training parameters The batch size was set to 128, and the number of training epochs was 20,000. The initial learning rate was 0.
005. To improve model accuracy, the learning rate was changed linearly in a piecewise manner during training. After every 1,000 iterations, lr = lr * 0.
6. The Adam learning rate adaptive algorithm was used for parameter optimization. Step 3: Establish a real-time thrust estimation model for the engine based on the fusion of UKF and neural network methods; To improve the real-time performance of the turbofan engine airborne adaptive model established in step 1, the LSTM neural network model from step 2 is integrated into the measurement update process of the LM-IUKF method, replacing the nonlinear mechanism model of the turbofan engine in step 1. The LSTM neural network model in step 2 iteratively updates its internal parameters during computation, which cannot match the requirement in step 1 to maintain the same initial conditions when calculating the Sigma point set. Therefore, the calculation process for the Sigma point set is improved, and the calculation steps for a measurement update process at time t are as follows: ① Record the neuron states C of the LSTM neural network model t and neuron output H t ; ② Calculate the output of the LSTM neural network model for the input Sigma point according to equation (3); ③ Restore the neuron state C of the LSTM neural network model t and neuron output H t ; ④ Proceed to the next step based on whether the Sigma point set has been calculated. If the calculation is complete, proceed to step ⑤; otherwise, return to step ① to continue the calculation. ⑤ Based on the covariance matrix and Kalman gain of LM-IUKF, state update and covariance update are performed to obtain the estimated values of health parameters; ⑥ Update the neuron state of the LSTM neural network model to C t+1 and neuron output H t+1 Based on the health parameters obtained in step ⑤, the airborne adaptive model is updated to obtain the real-time thrust output by the model.