Turboshaft engine identification and predictive control method based on MRR-KELM

Abstract

The invention belongs to the technical field of helicopter control, and provides a turboshaft engine identification and prediction control method based on MRR-KELM in order to solve the problem that a prediction model constructed for a turboshaft engine is poor in generalization, an MRR-KELM algorithm is combined with a Huber loss function and a self-adaptive regularization strategy, noise interference and abnormal value influence are restrained, and the prediction model is optimized. The prediction error of the constructed turboshaft engine prediction model is greatly reduced, and adaptive adjustment can be performed through Gaussian kernel dynamic mapping and working condition parameters; according to the algorithm, a regularization item and a robust loss function are introduced, the kernel function mapping capability is combined, the generalization performance and the calculation efficiency of the model under the complex working conditions of small samples, noise interference and the like are remarkably improved, further, the turboshaft engine prediction model is embedded into a nonlinear model prediction control framework, and the prediction accuracy of the turboshaft engine is improved. And high-precision stable control over the rotating speed of the power turbine is achieved through rolling optimization and feedback correction.

Description

Technical Field

[0001] The present invention belongs to the technical field of helicopter control, and in particular relates to a turboshaft engine identification and predictive control method based on MRR-KELM. Background Art

[0002] As the core power unit of helicopters, turboshaft engines must maintain stable turbine speed under complex flight conditions, placing extremely high demands on control system modeling accuracy and real-time performance. Traditional component-level modeling methods rely on aerodynamic and thermodynamic equations, resulting in complex modeling and high computational costs. System identification technology, however, is gaining importance through data-driven modeling. While the least squares support vector machine (LS-SVM) currently performs well in small-sample regression, its solution involves the operation of a system of linear equations, resulting in computational complexity that increases significantly with data volume. The LS-SVM is also sensitive to noise, making it difficult to meet the requirements of real-time engine control. Furthermore, the LS-SVM model suffers from poor sparsity and empirical reliance on parameter adjustment, resulting in insufficient generalization in dynamic flight environments. Therefore, a new identification algorithm with high accuracy, low computational complexity, and strong robustness is needed to support efficient predictive control of turboshaft engines. Summary of the Invention

[0003] In order to solve at least one of the above-mentioned technical problems existing in the prior art, the present invention provides a turboshaft engine identification and predictive control method based on MRR-KELM.

[0004] The present invention is implemented by the following technical solution: a turboshaft engine identification and predictive control method based on MRR-KELM, comprising the following steps: Step S1: Based on the MRR-KELM algorithm, the Huber loss function is introduced to calculate the weighted error, and an improved MRR-KELM algorithm model is constructed. At the same time, an adaptive regularization strategy is combined to balance the complexity and generalization ability of the MRR-KELM algorithm model; Step S2: identifying the dynamic characteristics of the turboshaft engine based on the improved MRR-KELM algorithm model and establishing a turboshaft engine prediction model; Step S3: embedding the turboshaft engine prediction model into a nonlinear model predictive control framework, establishing a performance index function with constant power turbine speed and minimum fuel consumption as optimization objectives, setting a fuel flow upper limit and speed fluctuation constraints, and using a sequential quadratic programming method to solve the optimal fuel input sequence in the future time domain. Based on the optimal fuel input sequence, the predicted power turbine speed is obtained. Step S4: comparing the predicted speed with the actual speed of the turboshaft engine in real time to obtain a power turbine speed error, and correcting the control instruction based on a proportional-integral term of the power turbine speed error to dynamically update the model parameters of the turboshaft engine prediction model; Step S5: Apply the first control increment issued by the updated turboshaft engine prediction model to the actual control system to control the time domain rolling advance.

[0005] Preferably, the adaptive regularization strategy in step S1 includes: Dynamically adjust the Gaussian kernel width according to the flight altitude and Mach number; The regularization coefficient is determined jointly by cross-validation and Bayesian optimization; A support vector sparsification strategy is adopted to screen the dominant eigenvectors with cumulative energy accounting for more than 95%, and then the MRR-KELM algorithm model is compressed.

[0006] Preferably, step S1 further includes: The output layer weights are quickly solved through random hidden layer weight initialization and closed-form analytical solution.

[0007] Preferably, the closed-form analytical solution is: Where, is the output weight matrix, which represents the connection weight between the hidden layer and the output layer in the improved MRR-KELM algorithm model; is the output matrix of the hidden layer, which contains the result of each training sample after being processed by the hidden layer; is the expected output matrix, which represents the expected output of each training sample.

[0008] Preferably, the turboshaft engine prediction model in step S2 is: Where, It is the output sequence of multiple future moments in the prediction time domain, used for multi-step prediction of rolling optimization; is the input data sequence to the prediction model.

[0009] Preferably, step S3 further includes: Sensor data corresponding to the current and historical fuel flow, required torque, gas turbine speed, power turbine speed, rotor torque, and turbine interstage temperature of the turboshaft engine are obtained; and the above sensor data are used as input to the turboshaft engine prediction model.

[0010] Preferably, step S5 further includes: The MRR-KELM algorithm model is updated using online incremental learning, while the training dataset is updated in real time through a sliding window strategy; The network structure of the MRR-KELM algorithm model is retained, and only the output weights are adjusted according to the updated model parameters and the kernel function width and regularization parameters are fine-tuned.

[0011] Preferably, before step S1, the method further includes: Multi-source sensor data of a turboshaft engine are collected and preprocessed through sliding window denoising and operating condition segmented normalization.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention adopts the MRR-KELM algorithm combined with the Huber loss function and the adaptive regularization strategy to suppress noise interference and the influence of outliers, greatly reducing the prediction error of the constructed turboshaft engine prediction model, and at the same time can realize Gaussian kernel dynamic mapping and adaptive adjustment of operating parameters; the algorithm introduces regularization terms and robust loss functions, combined with the kernel function mapping capability, significantly improving the generalization performance and computational efficiency of the model under complex working conditions such as small samples and noise interference. Furthermore, the turboshaft engine prediction model is embedded in the nonlinear model predictive control framework, and high-precision and stable control of the power turbine speed is achieved through rolling optimization and feedback correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0014] Figure 1 Schematic diagram of the simulation process for turboshaft engine system identification research using the MRR-KELM algorithm; Figure 2 Schematic diagram of the framework structure for embedding the nonlinear model predictive control into the turboshaft engine prediction model; Figure 3 A schematic diagram of the engine nonlinear model predictive control framework structure provided in this application; Figure 4 A recursive schematic diagram of an engine prediction model provided in this application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention are clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other implementations derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0016] It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size should fall within the scope of the technical content disclosed in the present invention without affecting the efficacy and purpose that can be achieved by the present invention. It should be noted that in this specification, relational terms such as first and second are only used to distinguish one entity from several other entities, and do not necessarily require or imply any actual relationship or order between these entities.

[0017] The present invention relates to a nonlinear model predictive engine control method based on an improved robust kernel extreme learning machine (MRR-KELM), aiming to solve the problems of traditional component-level modeling methods for turboshaft engines under complex operating conditions, which rely on aerodynamic thermodynamic equations, have complex modeling and high computational costs.

[0018] The core of the nonlinear model predictive engine control method is to embed the turboshaft engine prediction model into the nonlinear model predictive control framework to construct a prediction model, which performs rolling optimization control within a certain control time domain. However, in actual engineering, it is impossible to find a prediction model that is completely consistent with the controlled object.

[0019] Due to the complexity of aircraft engine component modeling, engine modeling has also been greatly developed through system identification. System identification is a scientific method that uses offline or online data of an unknown system to establish a mathematical model that describes the system.

[0020] like Figure 1 As shown, an embodiment of the present invention provides a turboshaft engine identification and predictive control method based on MRR-KELM, comprising the following steps: Step S1: Based on the MRR-KELM algorithm, the Huber loss function is introduced to calculate the weighted error, and an improved MRR-KELM algorithm model is constructed. At the same time, an adaptive regularization strategy is combined to balance the complexity and generalization ability of the MRR-KELM algorithm model.

[0021] Optionally, before step S1, the method further includes: collecting sensor data of the turboshaft engine, and preprocessing the data through sliding window noise reduction and working condition segmented normalization.

[0022] This embodiment collects and preprocesses data from a turboshaft engine. The present invention uses sensor data from a GE-T700 turboshaft engine under various flight conditions as an example. This data includes parameters such as altitude, forward flight speed, gas turbine speed, power turbine speed, fuel flow, and output torque. High-precision sensors are used for data collection, covering the entire turboshaft engine's flight envelope, encompassing various flight phases such as hover, forward flight, climb, and descent. The sampling frequency is set at 10 Hz to ensure that the data reflects the engine's rapid dynamic changes.

[0023] In this embodiment, the data preprocessing steps are as follows: 1. Use median filtering to remove noise points in the data, and the filter window size is dynamically adjusted according to the severity of data changes.

[0024] 2. Normalize all parameters and uniformly map the numerical range to the [0,1] interval to improve the convergence speed and stability of model training.

[0025] 3. The processed data is divided into training set and test set in a ratio of 7:3. A stratified sampling method is used to ensure that both the training set and the test set can reflect the operating characteristics of the turboshaft engine under different working conditions.

[0026] Optionally, the adaptive regularization strategy includes: dynamically adjusting the Gaussian kernel width according to the flight altitude and Mach number; jointly determining the regularization coefficient through cross-validation and Bayesian optimization; adopting the support vector sparsification strategy to screen the dominant eigenvectors with a cumulative energy share of more than 95%, and compressing the MRR-KELM algorithm model.

[0027] In this example, the training data set is fed into the MRR-KELM algorithm. The input data includes historical sequences of controlled variables such as fuel flow and required torque, while the output data is the corresponding sequences of controlled variables such as power turbine speed and output torque. The data sequence length is set to 20 sampling points to ensure that the model can capture the dynamic characteristics of the engine.

[0028] During training, the kernel width and regularization parameters are optimized and adjusted through cross-validation. The model is trained using batch gradient descent. At each training iteration, the model's prediction error on the current training data is calculated, and the network's output weights are adjusted through backpropagation. Simultaneously, the kernel width and regularization parameters are dynamically adjusted to minimize the model's prediction error and improve its generalization performance.

[0029] In this embodiment, the scale of the MRR-KELM algorithm model is compressed to a range that can be supported by the embedded controller.

[0030] Optionally, the method further includes: quickly solving the output layer weights through random hidden layer weight initialization and closed-form analytical solution.

[0031] In this embodiment, random hidden layer weight initialization refers to the process of randomly assigning hidden layer weights to values ​​before training begins in the MRR-KELM algorithm. This initialization method helps break the symmetry of the neural network, preventing different neurons from learning the same features, thereby improving the model's expressiveness and generalization capabilities.

[0032] After the hidden layer weights are initialized, a closed-form analytical solution can be used to quickly solve the output layer weights.

[0033] The closed-form analytical solution is expressed as follows: in: is the output weight matrix, which represents the connection weight between the hidden layer and the output layer; H is the output matrix of the hidden layer, which contains the result of each training sample after being processed by the hidden layer; T is the expected output matrix, which represents the expected output of each training sample; In this embodiment, the Huber loss function, also known as the smoothed L1 loss function, is a loss function commonly used in regression tasks. It combines the advantages of the mean square error (MSE) and the mean absolute error (MAE).

[0034] The Huber loss function is a commonly used loss function in regression tasks. It combines the advantages of the mean squared error (MSE) and the mean absolute error (MAE). Specifically, the Huber loss function uses the mean squared error (MSE) when the error is small, making it insensitive to noise; when the error is large, it uses the mean absolute error (MAE), which effectively suppresses the influence of outliers. Combining it with the MRR-KELM algorithm model can better handle noisy data and outliers during training, improving the model's robustness and prediction accuracy.

[0035] The improved MRR-KELM algorithm model structure is as follows: Input layer: The input layer receives multi-source sensor data from the turboshaft engine, including current and historical fuel flow, required torque, gas turbine speed, power turbine speed, rotor torque, and turbine interstage temperature. After preprocessing, this sensor data serves as the input feature vector for the improved MRR-KELM algorithm model.

[0036] Hidden layer: The hidden layer consists of multiple neurons, each of which has its input connected to the feature vector of the input layer. The weights of the hidden layer are randomly initialized before training begins, and the activation function uses a nonlinear function (such as the sigmoid function) to introduce nonlinear mapping capabilities.

[0037] Output layer: The output layer uses a closed-form analytical solution to quickly calculate the output weight matrix based on the output of the hidden layer and the expected output of the training data. The output layer's predictions include controlled variables such as turbine speed and output torque.

[0038] Model training and optimization: During the MRR-KELM algorithm model training process, the Huber loss function is introduced to calculate the weighted error and construct an improved MRR-KELM algorithm model. The kernel function width and regularization parameters are optimized and adjusted through cross-validation, and the MRR-KELM algorithm model is trained using batch gradient descent. The specific steps are as follows: ① Initialize the hidden layer weights; ②Calculate the hidden layer output matrix; ③Use Huber loss function to calculate prediction error; ④ Update the output weights through batch gradient descent; ⑤Dynamically adjust the kernel function width and regularization parameters to minimize the model's prediction error and improve its generalization performance.

[0039] Through this model structure and training strategy, the improved MRR-KELM algorithm model can better handle the dynamic characteristics of the turboshaft engine and improve the accuracy and robustness of the turboshaft engine prediction model.

[0040] To further optimize the model, we employed an adaptive regularization strategy. The Gaussian kernel width was dynamically adjusted based on flight altitude and Mach number to accommodate data distribution under varying operating conditions. Cross-validation and Bayesian optimization were used to determine the regularization coefficient, ensuring that the improved MRR-KELM algorithm model fits the data well while avoiding overfitting during training, thereby improving the model's generalization capabilities. Furthermore, a support vector sparsification strategy was employed to select dominant eigenvectors with cumulative energy exceeding 95%, and the MRR-KELM algorithm model was compressed to reduce model complexity and computational complexity, making it more suitable for embedded controller applications.

[0041] During model training, we also employed an online incremental learning approach. Using a sliding window strategy, we updated the training dataset in real time, preserving the network structure of the MRR-KELM model. We then adjusted the output weights based on the updated model parameters and fine-tuned the kernel width and regularization parameters. This ensured that the MRR-KELM model could quickly adapt to new data, maintaining good predictive performance while also reducing the computational resources and time required to update the MRR-KELM model.

[0042] The improved MRR-KELM model combines the Huber loss function with the MRR-KELM algorithm. Through this model structure and training strategy, the accuracy and robustness of the turboshaft engine prediction model constructed based on the improved MRR-KELM algorithm are effectively improved. In practical applications, the turboshaft engine prediction model can better meet the control requirements of turboshaft engines under complex operating conditions and provide reliable support for the nonlinear model predictive control framework.

[0043] Step S2: Identify the dynamic characteristics of the turboshaft engine based on the MRR-KELM algorithm model and establish a turboshaft engine prediction model.

[0044] In this embodiment, the KELM provides a solution for systematically identifying engine prediction models due to its strong generalization capabilities, fast computational speed, and high algorithmic stability. To accurately identify a turboshaft engine KELM model, a large amount of training data is required, resulting in complex data redundancy. When training the model using the standard KELM algorithm, the number of KELM hidden layers corresponds one-to-one with the training sample data, making it difficult to guarantee the KELM's generalization capabilities. Therefore, it is necessary to select sample data from the training sample that is beneficial for model identification, eliminate redundant sample data, and streamline the model structure to improve model generalization capabilities.

[0045] In this embodiment, when constructing a turboshaft engine prediction model based on the MRR-KELM algorithm model, the turboshaft engine prediction model structure is first designed. The input layer receives preprocessed turboshaft engine sensor data, including a historical sequence of sensor data corresponding to fuel flow, required torque, gas turbine speed, power turbine speed, rotor torque, and turbine interstage temperature, as input feature vectors to capture the dynamic characteristics of the engine. The hidden layer uses a Gaussian kernel function as a nonlinear mapping function to map the input data into a high-dimensional feature space. The number of nodes is determined based on cross-validation. The hidden layer weights are randomly initialized before training, and the activation function uses a nonlinear function. The output layer outputs the predicted controlled variables such as power turbine speed and output torque. The output weights are quickly solved using a closed-form analytical solution.

[0046] During the model training and optimization phase, the hidden layer weights are randomly initialized before training begins. The hidden layer output matrix is ​​then calculated based on the input data and the randomly initialized hidden layer weights. When constructing the improved MRR-KELM model, the Huber loss function is introduced to calculate the prediction error. Cross-validation and Bayesian optimization are combined to determine the regularization coefficient, dynamically adjusting the kernel function width and regularization parameter. Furthermore, a support vector sparsification strategy is used to filter dominant eigenvectors, remove redundant sample data, and streamline the model structure. This reduces the size of the MRR-KELM model to within the capacity of an embedded controller.

[0047] Then, the model is validated and optimized. The trained MRR-KELM model is validated using the test data, and the root mean square error and coefficient of determination are calculated. The model is updated every 100 hours using an online incremental learning method. When adding new data for retraining, the network structure and most parameters of the original model are retained, and only the output weights are adjusted, and the kernel function width and regularization parameters are fine-tuned. If the MRR-KELM model's prediction accuracy is found to be low under specific operating conditions, a parameter optimization method based on the particle swarm optimization algorithm is used for local optimization.

[0048] Finally, there is model embedding and control. Figure 2 、 Figure 3 As shown in the figure, the trained MRR-KELM model is embedded in the nonlinear model predictive control (NMPC) framework to construct the engine nonlinear model predictive control framework. The optimization objective and constraints are set, and the sequential quadratic programming (SQP) method is used to solve the optimal fuel input sequence. A feedback correction process is introduced into the NMPC. External disturbances are applied to the controlled object (i.e., the target helicopter) to obtain the actual target speed. The predicted speed is then compared with the actual speed in real time. Control commands are corrected based on the error, and model parameters are dynamically updated to ensure high-precision and stable control of the power turbine speed.

[0049] Optionally, it also includes: obtaining a historical sequence of sensor data corresponding to the fuel flow, required torque, gas turbine speed, power turbine speed, rotor torque and turbine interstage temperature of the turboshaft engine; and using the above sensor data as input to the turboshaft engine prediction model.

[0050] In this embodiment, the MRR-KELM algorithm model is used to identify the turboshaft engine prediction model. The model input of the turboshaft engine prediction model is the fuel flow rate. , required torque, gas turbine speed , power turbine speed , rotor torque and turbine interstage temperature The historical sequence of parameters such as the current moment is output as the corresponding sensor data 、 、 、 , the un-iterated turboshaft engine prediction model can be described as: Where, is the single-step output at the current moment, ; The input of the engine nonlinear model predictive control framework is: ; The time index of the current moment; is the length of the time window, that is, the number of time steps of historical data; Represents the transpose of a matrix.

[0051] In this embodiment, in order to ensure that the turboshaft engine prediction model has good model accuracy and reduce the complexity of the model structure, it is necessary to select a suitable turboshaft engine prediction model structure. Since the engine dynamic response can be simplified to a second-order response process, the control time window length can be selected. Since the controller solves the optimization problem of the power turbine speed constant as the objective function through rolling optimization, so that the input at the next moment is always the optimal value, the prediction time domain , so we can follow Figure 4 The scheme shown in the recursive diagram of the prediction model is based on the current fuel , future prediction time domain Fuel sequence , and the corresponding historical input and output information, to obtain the iterative prediction model of the turboshaft engine without iteration The turboshaft engine prediction model after the step: Where, It is the output sequence of multiple future moments in the prediction time domain, used for multi-step prediction of rolling optimization; is the input sequence to the turboshaft engine prediction model.

[0052] A turboshaft engine prediction model, built based on the MRR-KELM algorithm, is used to predict turboshaft engine output over a specific time horizon. Inputs to the turboshaft engine prediction model include current and historical fuel flow and required torque, and outputs include key parameters such as turbine speed and output torque. This prediction model allows for early estimation of engine behavior under varying control inputs, providing a foundation for rolling optimization.

[0053] Step S3: The turboshaft engine prediction model is embedded in a nonlinear model predictive control framework, and a performance index function is established with constant power turbine speed and minimum fuel consumption as optimization objectives. At the same time, a fuel flow upper limit and speed fluctuation constraints are set. The sequential quadratic programming method is used to solve the optimal fuel input sequence in the future time domain, and the predicted speed is obtained based on the optimal fuel input sequence.

[0054] In this embodiment, the MRR-KELM model is embedded in the nonlinear model predictive control framework, and multivariable coordinated control of fuel flow and power turbine speed is achieved by rolling optimization of the fuel input sequence in the future time domain. An integrated feedback correction mechanism compares the predicted output with the actual value in real time, and dynamically updates the model parameters to suppress cumulative errors, ensuring that the power turbine speed tracking accuracy is better than ±0.5%, the dynamic response delay is shortened to less than 10ms, and the fuel consumption is reduced by more than 15% by optimizing the control strategy when the load suddenly changes.

[0055] In this embodiment, rolling optimization is the core link of NMPC. Figure 4 As shown, in each control time domain, according to the current system state and the turboshaft engine prediction model, the optimal control action sequence is found to enable the system to achieve the predetermined performance indicators and meet the constraints within a period of time in the future. The specific steps are as follows: 1. Define performance indicators: Set control objectives, such as maintaining a constant power turbine speed or reducing fuel consumption, and convert them into mathematical performance indicator functions, which usually include minimizing tracking error and minimizing control variable changes. The specific formula of the performance indicator function is as follows: in: is The actual output of the system at that moment, here representing the power turbine speed of the turboshaft engine; It's time The desired output of the system, i.e. the target power turbine speed; express The increment of the control input at any moment, that is, the change in fuel flow; is the weight coefficient, which is used to balance the relationship between tracking error and control quantity change; is the length of the forecast horizon, i.e. the number of future time steps considered in the rolling optimization.

[0056] 2. Constraint setting: Determine various constraints during system operation, such as the upper limit of fuel flow rate and speed fluctuation range, to ensure that the engine operates under safe and stable conditions.

[0057] 3. Optimization algorithm selection: The Sequential Quadratic Programming (SQP) optimization algorithm is used to solve the rolling optimization problem. While satisfying the constraints, the control input sequence that minimizes the performance index is found.

[0058] 4. Rolling Implementation: At each sampling moment, predictions and optimization are re-performed based on the latest state information of the current nonlinear model predictive control framework. The first control increment obtained from the optimization is then applied to the actual control system. The control horizon is then rolled forward by one sampling period, and the above process is repeated to achieve real-time optimized control of the engine.

[0059] Step S4: Compare the predicted speed with the actual speed of the turboshaft engine in real time, and correct the control instruction based on the proportional integral term of the power turbine speed error to dynamically update the model parameters of the turboshaft engine prediction model.

[0060] In this embodiment, because the predictive control of the turboshaft engine prediction model relies on the accuracy of the model parameters, and actual systems are subject to factors such as model uncertainty and external interference, the introduction of a feedback correction process into the NMPC is crucial. This feedback correction feeds back the error between the actual system output and the turboshaft engine prediction model output to the control system, adjusting the turboshaft engine prediction model parameters and control inputs in real time, thereby improving control accuracy and system stability.

[0061] In this embodiment, the NMPC feedback correction module is designed based on a correction strategy for correcting the control command based on the model estimation error. The strategy for correcting the controller command does not change the NMPC structure. At the same time, in order to eliminate the steady-state error of the control system, the control command is corrected by a proportional integral term based on the error between the power turbine speed output corresponding to the prediction model and the actual engine speed.

[0062] Step S5: Apply the first control increment issued by the updated turboshaft engine prediction model to the actual control system to control the time domain rolling advance.

[0063] Optionally, it also includes: using online incremental learning to update the MRR-KELM algorithm model, and updating the training data set in real time through a sliding window strategy; retaining the network structure of the MRR-KELM algorithm model, and only adjusting the output weights and fine-tuning the kernel function width and regularization parameters according to the updated model parameters.

[0064] In this embodiment, the MRR-KELM algorithm is used for parameter setting. The Gaussian kernel function is selected as the kernel function of the MRR-KELM algorithm, and its width parameter is initially set to 1.5. The regularization parameter is initially set to 0.1 to balance the model's fitting ability and generalization ability. The number of hidden layer nodes in the MRR-KELM network is set to 50, and the sigmoid function is used as the activation function.

[0065] The training data was fed into the MRR-KELM algorithm. The input data included historical sequences of controlled variables such as fuel flow and required torque, while the output data consisted of corresponding sequences of controlled variables such as power turbine speed and output torque. The data sequence length was set to 20 sampling points to ensure that the model could capture the dynamic characteristics of the engine.

[0066] During training, the kernel width and regularization parameters are optimized and adjusted through cross-validation. The model is trained using batch gradient descent. At each training iteration, the model's prediction error on the current training data is calculated, and the network's output weights are adjusted through backpropagation. Simultaneously, the kernel width and regularization parameters are dynamically adjusted to minimize the model's prediction error and improve its generalization performance.

[0067] After about 2-3 hours of training, the MRR-KELM algorithm model achieved a root mean square error of 0.048 on the training set. The kernel width and regularization parameter were optimized to 1.8 and 0.05, respectively.

[0068] The trained MRR-KELM algorithm model was verified using the test data. The root mean square error of the prediction of the MRR-KELM algorithm model on the test set was calculated to be 0.042, and the coefficient of determination was The value reaches 0.983, indicating that the MRR-KELM algorithm model has high prediction accuracy and good generalization ability.

[0069] During the experimental simulation, it was found that the prediction accuracy of the MRR-KELM algorithm model was relatively low under low-speed and high-load conditions. Local optimization of the MRR-KELM algorithm model was performed, focusing on adjusting the kernel function width and the value of the regularization parameter under this condition. A parameter optimization method based on the particle swarm optimization algorithm was used to retrain the MRR-KELM algorithm model on the training data under this condition. After optimization, the root mean square error of the prediction of the MRR-KELM algorithm model on the test set was reduced to 0.038. The value is increased to 0.987, and the model performance is significantly improved.

[0070] During actual turboshaft engine operation, new operating data is collected every 100 hours. The MRR-KELM model is updated using an incremental learning approach. The new data is added to the training set and the MRR-KELM algorithm model is retrained. During the incremental learning process, the network structure and most parameters of the original MRR-KELM algorithm model are retained, with only the output weights adjusted. The kernel function width and regularization parameters are fine-tuned based on the characteristics of the new data. This significantly reduces training time, typically completing within 30 minutes to one hour.

[0071] The foregoing description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be readily conceived by a person skilled in the art within the technical scope disclosed herein should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A turboshaft engine identification and predictive control method based on MRR-KELM, characterized in that: The steps include: Step S1: Based on the MRR-KELM algorithm, the Huber loss function is introduced to calculate the weighted error, and an improved MRR-KELM algorithm model is constructed. At the same time, an adaptive regularization strategy is combined to balance the complexity and generalization ability of the MRR-KELM algorithm model; Step S2: identifying the dynamic characteristics of the turboshaft engine based on the improved MRR-KELM algorithm model and establishing a turboshaft engine prediction model; Step S3: embedding the turboshaft engine prediction model into a nonlinear model predictive control framework, establishing a performance index function with constant power turbine speed and minimum fuel consumption as optimization objectives, setting a fuel flow upper limit and speed fluctuation constraints, and using a sequential quadratic programming method to solve the optimal fuel input sequence in the future time domain. Based on the optimal fuel input sequence, the predicted power turbine speed is obtained. Step S4: comparing the predicted speed with the actual speed of the turboshaft engine in real time to obtain a power turbine speed error, and correcting the control instruction based on a proportional-integral term of the power turbine speed error to dynamically update the model parameters of the turboshaft engine prediction model; Step S5: Apply the first control increment issued by the updated turboshaft engine prediction model to the actual control system to control the time domain rolling advance.

2. The method for identifying and predicting a turboshaft engine based on MRR-KELM according to claim 1, characterized in that: The adaptive regularization strategy in step S1 includes: Dynamically adjust the Gaussian kernel width according to the flight altitude and Mach number; The regularization coefficient is determined jointly by cross-validation and Bayesian optimization; A support vector sparsification strategy is adopted to screen the dominant eigenvectors with cumulative energy accounting for more than 95%, and then the MRR-KELM algorithm model is compressed.

3. The method for identifying and predicting a turboshaft engine based on MRR-KELM according to claim 1, characterized in that: Step S1 further includes: The output layer weights are quickly solved through random hidden layer weight initialization and closed-form analytical solution.

4. The method for identifying and predicting a turboshaft engine based on MRR-KELM according to claim 3, characterized in that: The closed-form analytical solution is: Where, is the output weight matrix, which represents the connection weight between the hidden layer and the output layer in the improved MRR-KELM algorithm model; is the output matrix of the hidden layer, which contains the result of each training sample after being processed by the hidden layer; is the expected output matrix, which represents the expected output of each training sample.

5. According to the MRR-KELM-based turboshaft engine identification and predictive control method of claim 1, the turboshaft engine prediction model in step S2 is: Where, It is the output sequence of multiple future moments in the prediction time domain, used for multi-step prediction of rolling optimization; is the input data sequence to the prediction model.

6. According to the MRR-KELM-based turboshaft engine identification and predictive control method of claim 1, step S3 further comprises: Obtain sensor data corresponding to the turboshaft engine's current and historical fuel flow, required torque, gas turbine speed, power turbine speed, rotor torque, and turbine interstage temperature; The above sensor data is used as input to the turboshaft engine prediction model.

7. According to the MRR-KELM-based turboshaft engine identification and predictive control method of claim 1, step S5 further comprises: The MRR-KELM algorithm model is updated using online incremental learning, while the training dataset is updated in real time through a sliding window strategy; The network structure of the MRR-KELM algorithm model is retained, and only the output weights are adjusted according to the updated model parameters and the kernel function width and regularization parameters are fine-tuned.

8. The method for identifying and predicting a turboshaft engine based on MRR-KELM according to claim 1, before step S1, further comprising: The sensor data of the turboshaft engine is collected and preprocessed through sliding window denoising and operating condition segmented normalization.

Images