Aviation starter generator position estimation method based on optimized neural network modeling
Through the combination of Bayesian optimized convolutional neural network and sliding mode observer, the accuracy and storage problems of aviation-starting generator position estimation in extreme environments are solved, and high-precision motor control is achieved.
Patent Information
- Application Number
- CN202510579013.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-08-15
AI Technical Summary
The traditional aero-starting generator position estimation method lacks accuracy in extreme environments and occupies a large amount of storage space, affecting the motor control cost and reliability.
The current-position-macro-link data table is modeled using a Bayesian optimization convolutional neural network, and a sliding mode observer is used to perform position estimation to reduce storage requirements and improve estimation accuracy.
It significantly reduces data storage requirements, improves the accuracy and control accuracy of aviation generator position estimation, and reduces storage costs.
Smart Images
Figure CN120493720A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of motor control technology, and specifically relates to a method for estimating the position of an aviation generator based on optimized neural network modeling, and more specifically to a method for estimating the position of an aviation generator based on a Bayesian optimized convolutional neural network (CNN). Background Art
[0002] With the rapid development of aviation electrification technology, more-electric aircraft (MEA) and all-electric aircraft (AEA) designs are gradually becoming mainstream in the industry. As a core power component, the aircraft starter integrates engine starting and power generation functions, offering significant advantages in simplifying aircraft power architecture and improving energy efficiency. It has become a key component of the next-generation aviation power system. However, aircraft starters must operate under harsh operating conditions such as a wide speed range, high vibration and shock, and extreme temperature fluctuations. The accuracy of their position estimation directly affects the efficiency and power output stability of both modes (motor / generator), placing stringent demands on aircraft starter control technology.
[0003] Traditional aircraft starter position estimation relies on rotor position sensors, which not only increases system cost and complexity but also reduces motor reliability in extreme environments. Consequently, rotor position estimation technology for aircraft starters has become widely used in recent years. Accurate modeling of the torque and current characteristics of aircraft starters is essential for precise control. However, nonlinearities caused by saturation, eddy currents, and hysteresis in the magnetic circuit of aircraft starters make accurate modeling of torque and current characteristics difficult using traditional mathematical methods. While finite element simulations can accurately obtain nonlinear electromagnetic characteristics, they cannot accurately determine torque and current characteristics. Traditional position estimation algorithms utilize a large number of data points stored from offline measurements to construct a "current-position-flux linkage" data table. However, these algorithms suffer from the following issues: First, significant nonlinear flux distortion occurs in high-speed power generation mode, increasing linear modeling errors. Second, this approach consumes a large amount of storage space, increasing motor control costs and limiting the large-scale application of aircraft starters. Summary of the Invention
[0004] In order to improve the control accuracy of aircraft starters, improve the accuracy of the model, and reduce the storage space occupied, the present invention provides an aircraft starter position estimation method based on optimized neural network modeling. The powerful global optimization ability and good generalization ability of the convolutional neural network based on Bayesian optimization are utilized to model the "current-position-flux" data table commonly used in aircraft starter control to obtain high-precision nonlinear characteristics of the aircraft starter. At the same time, the aircraft starter position is estimated based on the sliding mode observer (SMO), which significantly reduces the use of storage space and improves control accuracy, thereby promoting the promotion of aircraft starters.
[0005] To achieve the above objectives, the technical solutions provided by the present invention are:
[0006] A method for estimating the position of an aviation generator based on optimized neural network modeling is provided, comprising the following steps:
[0007] Step 1: Obtain the magnetic flux characteristics, current characteristics and rotor position characteristics of the aviation generator and construct the data table Ψ ph (i ph ,θ), where θ is the rotor position, i ph is the phase current, Ψ ph is the phase flux linkage value;
[0008] Step 2: Construct a convolutional neural network model for estimating the position of the aviation generator, using the phase flux value Ψ ph and rotor position θ parameters as input layer data input, phase current i ph As the output of the output layer, after the input layer receives the data, it passes it to the subsequent layers, including the convolution layer, pooling layer and fully connected layer, and finally outputs it through the output layer;
[0009] Step 3: Based on the data table constructed in step 1, the hyperparameters of the constructed convolutional neural network model are adjusted using the Bayesian optimization algorithm, which includes the following sub-steps:
[0010] Step 3.1: Use a set of randomly selected hyperparameter combinations as the starting point for Bayesian optimization. Based on the sampled data in the data table constructed in step 1, use the Gaussian process model to obtain the predicted phase current of the model under the current hyperparameter combination. The mean μ and variance σ of 2 , and obtain the next hyperparameter point x to be evaluated by maximizing the acquisition function t , and evaluate the objective function value to form a new hyperparameter data set D 1:t ={D 1:t-1 ,(x t ,y t )}:
[0011]
[0012] y t =f(x t )+ε t
[0013] Where x t represents the hyperparameter vector of the neural network model selected at time t, x t The value range of is limited by the hyperparameter space χ, α represents the acquisition function, x represents the hyperparameter set of the model, and D 1:t-1represents the historical dataset from the first iteration to the t-1th iteration, including the hyperparameter points evaluated in all previous iterations and their corresponding objective function values, y t is the observed value of the neural network corresponding to time t, that is, at the hyperparameter x t The objective function value obtained by evaluating under t is the observation error, D 1:t is the observation data set from the first iteration to the t-1th iteration;
[0014] Step 3.2: Based on the data table constructed in step 1, retrain the constructed neural network model with the new hyperparameter data set formed, and update the mean μ and variance σ of the model. 2 , repeat this process until the stopping condition is triggered, the optimal hyperparameter combination is obtained, and the neural network model under the optimal hyperparameter is obtained;
[0015] Step 4: Calculate the motor phase winding flux at time k using the following formula:
[0016] Ψ ph (k)=[u ph (k)-i ph (k)R ph ]T s +Ψ ph (k-1)
[0017] Where, ph (k),u ph (k), i ph (k) is the flux linkage value, voltage value, and current value of the motor phase winding at time k, R ph is the resistance value of the motor phase winding, T s is the sampling time, Ψ ph (k-1) is the flux linkage value of phase winding at time k-1;
[0018] Step 5: Use the motor phase winding flux value at time k calculated in step 4 and the estimated motor rotor position fed back by the sliding mode observer at time k as the input of the neural network model to obtain the estimated phase current at time k output by the neural network model under the optimal hyperparameters.
[0019] Step 6: Based on the dynamic equations of the rotor angular velocity and rotor position of the aviation generator, a sliding mode observer is constructed to estimate the rotor angular velocity and rotor position of the motor:
[0020]
[0021] The sliding mode plane of the sliding mode observer is:
[0022]
[0023] Where, is the state estimate of the motor rotor angular velocity at time k, is the output estimate of the rotor position at time k, ω and k θ are the gains of rotor angular velocity estimation and rotor position estimation respectively, sgn() is the sign function, T e (k) and T L (k) are the electromagnetic torque and load torque of the motor at time k, D is the friction coefficient of the motor, J is the moment of inertia, e(k) is the current state error at time k, and i ph (k) are the estimated phase current state quantity and the measured phase current state quantity of the motor at time k, is the estimated motor rotor position fed back at time k, and m represents the number of motor winding phases.
[0024] Furthermore, in step 2, the convolution layer of the model uses the convolution kernel k to extract features and perform convolution operations, which can be expressed as in, and Represents the feature matrices of the nth layer and the previous layer, k i,j is the convolution kernel, is the bias term, f(·) is the activation function; the pooling layer of the model extracts higher-level features and further reduces the matrix size. The output of the pooling layer is expressed as in, Represents the downsampling pooling method used by the pooling layer, Represents the weight of the product; the fully connected layer of the model maps the extracted features to the prediction results by adjusting the connection weights and biases between neurons, and its output is expressed as Y = f(W T X+b), where X represents the input feature vector, b represents the bias vector, Y represents the output, and W represents the weight of the fully connected layer.
[0025] Furthermore, in step 2, the pooling method used by the pooling layer of the model is maximum pooling or average pooling.
[0026] Furthermore, in step 3.1, the root mean square error (RMSE) is used as the objective function value, and the calculation formula is:
[0027]
[0028] Where N is the total number of data, is the phase current prediction value of the jth data point, i ph,j is the actual value of the phase current at the jth data point.
[0029] Furthermore, in step 3.2, the stopping condition includes reaching a preset number of iterations or the objective function converges.
[0030] Furthermore, in step 6, k ω Set the value to be much larger than In this case As the interference term can be ignored, the dynamic equation is simplified to:
[0031]
[0032]
[0033] The advantages of the present invention are:
[0034] The present invention proposes an aviation generator position estimation method based on optimized neural network modeling, which uses a Bayesian optimized convolutional neural network to model the electromagnetic characteristics of the motor in the data table "current-position-flux", so that the network model can more accurately reflect the actual behavior of the motor, thereby accurately fitting the nonlinear characteristic curve of the motor, while significantly reducing data storage requirements and lowering storage costs; in addition, the present invention estimates the motor position based on a sliding mode observer, in which the phase current estimation error of each phase output by the model is always approached to zero, so that more accurate motor rotor angular velocity and rotor position can be obtained, thereby improving motor control accuracy and performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The above and / or other features and advantages of the present invention will become more readily understood through the following description with reference to the accompanying drawings, in which:
[0036] Figure 1 It is a flow chart of the aviation generator position estimation method based on optimized neural network modeling of the present invention;
[0037] Figure 2 This is a block diagram of the position-free control of the aviation generator based on the sliding mode observer in the present invention;
[0038] Figure 3 It is a schematic diagram of the convolutional neural network structure in the present invention;
[0039] Figure 4 It is a flow chart of the neural network optimized based on the Bayesian algorithm in the present invention;
[0040] Figure 5 It is the motor position estimation result in the example of the present invention. DETAILED DESCRIPTION
[0041] The present invention will be described in detail below with reference to the accompanying drawings by means of exemplary embodiments of the present invention. It should be noted that the following detailed description of the present invention is only for the purpose of illustration and is not intended to limit the present invention.
[0042] In response to the problems of large modeling errors and large data storage space occupied by the lookup table method commonly used in aviation generator control, the present invention proposes a method for modeling the electromagnetic characteristics of aviation generators based on Bayesian optimization convolutional neural network. The powerful global optimization ability and good generalization ability of the convolutional neural network based on Bayesian optimization are utilized to model the data table "current-position-flux linkage". At the same time, a sliding mode observer is used for position-free control to obtain high-precision nonlinear electromagnetic characteristics of the aviation generator, while significantly reducing the use of storage space and improving control accuracy.
[0043] Reference Figure 1 and Figure 2 The method for estimating the position of an aviation generator based on optimized neural network modeling provided by the present invention comprises the following steps:
[0044] Step S1, obtain the magnetic flux characteristics, current characteristics and rotor position characteristics of the aviation generator and construct a data table Ψ ph (i ph ,θ), where θ is the rotor position, i ph is the phase current, Ψ ph is the phase flux linkage value;
[0045] Step S2, constructing a convolutional neural network model for estimating the position of the aviation generator, using the phase flux value Ψ ph and rotor position θ parameters as input layer data input, phase current i ph As the output of the output layer, after the input layer receives the data, it passes it to the subsequent layers, including the convolution layer, pooling layer and fully connected layer, and finally outputs it through the output layer;
[0046] Step S3, based on the data table constructed in step S1, using a Bayesian optimization algorithm to adjust the hyperparameters of the constructed convolutional neural network model to obtain a neural network model with optimal hyperparameters;
[0047] Step S4, calculate the motor phase winding flux value at time k using the following formula:
[0048] Ψ ph (k)=[u ph (k)-i ph (k)R ph ]T s +Ψ ph (k-1)
[0049] Where, ph (k),u ph(k), i ph (k) is the flux linkage value, voltage value, and current value of the motor phase winding at time k, R ph is the resistance value of the motor phase winding, T s is the sampling time, Ψ ph (k-1) is the flux linkage value of phase winding at time k-1;
[0050] Step S5: Using the motor phase winding flux value at time k calculated in step 4 and the estimated motor rotor position fed back by the sliding mode observer at time k as the input of the neural network model, the estimated phase current at time k output by the neural network model under the optimal hyperparameters is obtained.
[0051] Step S6: Based on the dynamic equations of the rotor angular velocity and rotor position of the aviation generator, a sliding mode observer is constructed to estimate the rotor angular velocity and rotor position of the motor:
[0052]
[0053]
[0054] The sliding mode plane of the sliding mode observer is:
[0055]
[0056] Where, is the state estimate of the motor rotor angular velocity at time k, is the output estimate of the rotor position at time k, ω and k θ are the gains of rotor angular velocity estimation and rotor position estimation respectively, sgn() is the sign function, T e (k) and T L (k) are the electromagnetic torque and load torque of the motor at time k, D is the friction coefficient of the motor, J is the moment of inertia, e(k) is the current state error at time k, and i ph (k) are the estimated phase current state quantity and the measured phase current state quantity of the motor at time k, is the estimated motor rotor position fed back at time k, and m represents the number of motor winding phases.
[0057] Reference Figure 3 According to the present invention, in step S2, the convolution layer of the model uses the convolution kernel k to extract features and performs a convolution operation, which is expressed as:
[0058]
[0059] Where, and Represents the feature matrices of the nth layer and the previous layer, k i,j is the convolution kernel, is the bias term, and f(·) is the activation function.
[0060] After convolution, the model uses a pooling layer to extract higher-level features and further reduce the matrix size. The output of the pooling layer can be expressed as:
[0061]
[0062] Where, Represents the downsampling pooling method used by the pooling layer, usually maximum pooling or average pooling, Represents the weight of the product.
[0063] After extracting high-level features through multiple convolutional layers and pooling layers, it is connected to the fully connected layer. By adjusting the connection weights and biases between neurons, the extracted features are mapped to the prediction results. The output is expressed as:
[0064] Y=f(W T X+b)
[0065] Where X represents the input feature vector, b represents the bias vector, Y represents the output, and W represents the weight of the fully connected layer.
[0066] Reference Figure 4 , step S3 specifically includes the following sub-steps:
[0067] Step S3.1, using a set of randomly selected hyperparameter combinations as the starting point of Bayesian optimization, based on the sampled data in the data table constructed in step S1, obtain the predicted phase current of the model under the current hyperparameter combination through the Gaussian process model The mean μ and variance σ of 2 , and obtain the next hyperparameter point x to be evaluated by maximizing the acquisition function t , and evaluate the objective function value to form a new hyperparameter data set D 1:t ={D 1:t-1 ,(x t ,y t )}:
[0068]
[0069] y t =f(x t )+ε t
[0070] Where x t represents the hyperparameter vector of the neural network model selected at time t, x tThe value range of is limited by the hyperparameter space χ, α represents the acquisition function, x represents the hyperparameter set of the model, and D 1:t-1 represents the historical dataset from the first iteration to the t-1th iteration, including the hyperparameter points evaluated in all previous iterations and their corresponding objective function values, y t is the observed value of the neural network corresponding to time t, that is, at the hyperparameter x t The objective function value obtained by evaluating under t is the observation error, D 1:t is the observation data set from the first iteration to the t-1th iteration;
[0071] Step 3.2: Based on the data table constructed in step S1, retrain the constructed neural network model with the new hyperparameter data set, and update the mean μ and variance σ of the model. 2 , repeating this process until the stopping condition is triggered, the optimal hyperparameter combination is obtained, and the neural network model under the optimal hyperparameter is obtained. The stopping condition may include reaching a preset number of iterations or the objective function converges.
[0072] In step S3.1, the acquisition function can use the expected improvement algorithm. In addition, the root mean square error (RMSE) can be used as the objective function value. The calculation formula is:
[0073]
[0074] Where N is the total number of data, is the phase current prediction value of the jth data point, i ph,j is the actual value of the phase current at the jth data point.
[0075] In step S6, It can ensure that the error e converges correctly to e=0, The estimated rotor position fed back has an initial value of 0 when it is first calculated.
[0076] As described in the above steps, the present invention uses a Bayesian-optimized convolutional neural network to model the electromagnetic characteristics of the motor in the data table "current-position-flux", so that the network model can more accurately reflect the actual behavior of the motor, thereby accurately fitting the nonlinear characteristic curve of the motor, while significantly reducing data storage requirements and reducing storage costs; in addition, the present invention estimates the motor position based on a sliding mode observer, in which the phase current estimation error of each phase output by the model is always approached to zero, and a more accurate motor rotor angular velocity and rotor position can be obtained, thereby improving the motor control accuracy and performance.
[0077] In particular, the gain k ω and k θ The value of can be set according to the motor characteristics and control requirements, kω The value is usually set to be much larger than In this case As the interference term can be ignored, the dynamic equation can be simplified to:
[0078]
[0079] This simplified formula is therefore used to achieve position-free control of aircraft generators. By ignoring minor factors such as friction torque and load torque, the model focuses more on the main dynamic characteristics, which can further improve the motor control accuracy. At the same time, the simplified formula reduces the parameters and variables that need to be calculated, lowering the computing power requirements of the processor, helping to better track the target position and improve the real-time and accuracy of control.
[0080] Next, the aviation generator position estimation method based on optimized neural network modeling provided by the present invention is further explained with reference to examples.
[0081] In this example, a three-phase 12 / 8 switched reluctance motor is used to verify the position estimation accuracy of the proposed method at a speed of 600r / min. Figure 5 The motor position estimation result is shown in FIG. As can be seen from the figure, the rotor position estimated by the proposed method is basically consistent with the actual position waveform, with a small error, which verifies that the method proposed in the present invention can estimate the rotor position of the motor more accurately.
[0082] Finally, it should be noted that the features mentioned and / or illustrated in the above description of the exemplary embodiments of the present invention may be incorporated into one or more other embodiments in the same or similar manner, combined with features in other embodiments, or substituted for corresponding features in other implementations. The technical solutions obtained by such combination or substitution shall also be deemed to be included in the scope of protection of the present invention.
Claims
1. A method for estimating the position of an aviation generator based on optimized neural network modeling, characterized in that: The following steps are involved: Step 1: Obtain the magnetic flux characteristics, current characteristics and rotor position characteristics of the aviation generator and construct the data table Ψ ph (i ph ,θ), where θ is the rotor position, i ph is the phase current, Ψ ph is the phase flux linkage value; Step 2: Construct a convolutional neural network model for estimating the position of the aviation generator, using the phase flux value Ψ ph and rotor position θ parameters as input layer data input, phase current i ph As the output of the output layer, after the input layer receives the data, it passes it to the subsequent layers, including the convolution layer, pooling layer and fully connected layer, and finally outputs it through the output layer; Step 3: Based on the data table constructed in step 1, the hyperparameters of the constructed convolutional neural network model are adjusted using the Bayesian optimization algorithm, which includes the following sub-steps: Step 3.1: Use a set of randomly selected hyperparameter combinations as the starting point for Bayesian optimization. Based on the sampled data in the data table constructed in step 1, use the Gaussian process model to obtain the predicted phase current of the model under the current hyperparameter combination. The mean μ and variance σ of 2 , and obtain the next hyperparameter point x to be evaluated by maximizing the acquisition function t , and evaluate the objective function value to form a new hyperparameter data set D 1:t ={D 1:t-1 ,(x t ,y t )}: y t =f(x t )+ε t Where x t represents the hyperparameter vector of the neural network model selected at time t, x t The value range of is limited by the hyperparameter space χ, α represents the acquisition function, x represents the hyperparameter set of the model, and D 1:t-1 represents the historical dataset from the first iteration to the t-1th iteration, including the hyperparameter points evaluated in all previous iterations and their corresponding objective function values, y t is the observed value of the neural network corresponding to time t, that is, at the hyperparameter x t The objective function value obtained by evaluating under t is the observation error, D 1:t is the observation data set from the first iteration to the t-1th iteration; Step 3.2: Based on the data table constructed in step 1, retrain the constructed neural network model with the new hyperparameter data set formed, and update the mean μ and variance σ of the model. 2 , repeat this process until the stopping condition is triggered, the optimal hyperparameter combination is obtained, and the neural network model under the optimal hyperparameter is obtained; Step 4: Calculate the motor phase winding flux at time k using the following formula: P ph (k)=[u ph (k)-i ph (k)R ph ]T s +Ψ ph (k-1) Where, ph (k),u ph (k), i ph (k) is the flux linkage value, voltage value, and current value of the motor phase winding at time k, R ph is the resistance value of the motor phase winding, T s is the sampling time, Ψ ph (k-1) is the flux linkage value of phase winding at time k-1; Step 5: Use the motor phase winding flux value at time k calculated in step 4 and the estimated motor rotor position fed back by the sliding mode observer at time k as the input of the neural network model to obtain the estimated phase current at time k output by the neural network model under the optimal hyperparameters. Step 6: Based on the dynamic equations of the rotor angular velocity and rotor position of the aviation generator, a sliding mode observer is constructed to estimate the rotor angular velocity and rotor position of the motor: The sliding mode plane of the sliding mode observer is: Where, is the state estimate of the motor rotor angular velocity at time k, is the output estimate of the rotor position at time k, ω and k θ are the gains of rotor angular velocity estimation and rotor position estimation respectively, sgn() is the sign function, T e (k) and T L (k) are the electromagnetic torque and load torque of the motor at time k, D is the friction coefficient of the motor, J is the moment of inertia, e(k) is the current state error at time k, and i ph (k) are the estimated phase current state quantity and the measured phase current state quantity of the motor at time k, is the estimated motor rotor position fed back at time k, and m represents the number of motor winding phases.
2. The method for estimating the position of an aviation generator based on optimized neural network modeling according to claim 1, characterized in that: In step 2, the convolution layer of the model uses the convolution kernel k to extract features and perform convolution operations, which are expressed as in, and Represents the feature matrices of the nth layer and the previous layer, k i,j is the convolution kernel, is the bias term, f(·) is the activation function; The pooling layer of the model extracts higher-level features and further reduces the matrix size. The output of the pooling layer is represented as in, Represents the downsampling pooling method used by the pooling layer, represents the weight of the product; The fully connected layer of the model maps the extracted features to the prediction results by adjusting the connection weights and biases between neurons. Its output is expressed as Y = f(W T X+b), where X represents the input feature vector, b represents the bias vector, Y represents the output, and W represents the weight of the fully connected layer.
3. The method for estimating the position of an aviation generator based on optimized neural network modeling according to claim 2, characterized in that: In step 2, the pooling method used by the pooling layer of the model is maximum pooling or average pooling.
4. The method for estimating the position of an aviation generator based on optimized neural network modeling according to claim 1 or 2, characterized in that: In step 3.1, the root mean square error (RMSE) is used as the objective function value, and the calculation formula is: Where N is the total number of data, is the phase current prediction value of the jth data point, i ph,j is the actual value of the phase current at the jth data point.
5. The method for estimating the position of an aviation generator based on optimized neural network modeling according to claim 1 or 2, characterized in that: In step 3.2, the stopping conditions include reaching a preset number of iterations or the objective function convergence.
6. The method for estimating the position of an aviation generator based on optimized neural network modeling according to claim 1 or 2, characterized in that: In step 6, k ω Set the value to be much larger than In this case As the interference term can be ignored, the dynamic equation is simplified to: