A circular induction synchronizer error prediction method and system based on hybrid model

Abstract

This invention discloses a method and system for predicting circular induction synchronizer errors based on a hybrid model. The method relates to the field of precision measurement technology and includes obtaining target position-error data; constructing a circular induction synchronizer error prediction model based on the normalization principle, first harmonic linear terms, radial basis function neural network terms, dynamic parameter optimization mechanisms, and the principle of linear superposition; and inputting the target position-error data into the circular induction synchronizer error prediction model to obtain a predicted error value. This invention addresses the issues of positioning misalignment, poor real-time performance, and weak generalization capabilities in high-precision servo systems caused by temperature drift and complex interference. It achieves efficient and reliable embedded real-time correction of circular induction synchronizers, improving the accuracy of circular induction synchronizer error prediction.

Description

Technical Field

[0001] The present invention relates to the field of precision measurement technology, and in particular to a circular induction synchronizer error prediction method and system based on a hybrid model. Background Art

[0002] Circular induction synchronizers, as core components for high-precision angle measurement, are widely used in CNC machine tools, aerospace servo systems, and precision instruments. Their operating principle is based on electromagnetic induction, analyzing the coupled signals between the stator and rotor windings to obtain angle information. However, due to factors such as installation eccentricity, temperature drift, and electromagnetic interference, the actual output has zero-position errors, which seriously restricts the system's positioning accuracy. Especially at low speeds, in a wide temperature range (-40°C to 85°C), and in complex electromagnetic environments, the nonlinear characteristics of the error are significant, posing a significant challenge to traditional compensation methods.

[0003] In existing technologies, mainstream compensation schemes can be divided into three categories: the first is manual compensation, which directly measures physical deviations using external high-precision optical devices and performs manual compensation. However, this method is labor-intensive and not universally applicable. The second is harmonic analysis, which extracts error harmonic components through Fourier decomposition. However, this method can only correct periodic linear errors and is sensitive to random interference and high-order nonlinear components, resulting in high residual errors after compensation. The third is neural network, which uses a static back-propagation neural network to fit the error curve. Although it can partially capture nonlinear relationships, it suffers from central point fixation and poor generalization. It also fails to separate linear and nonlinear error sources, resulting in insufficient model interpretability. More seriously, these methods often rely on laboratory calibration data and fail to consider dynamic interference such as temperature gradients and mechanical vibrations under actual operating conditions. This results in insufficient field adaptability, and the compensation accuracy degrades by more than 30% after long-term use.

[0004] Furthermore, traditional compensation models have large parameter counts and high computational latency, making them difficult to meet the real-time requirements of low-speed servo systems (≤1ms). Furthermore, existing dynamic weight update algorithms suffer from high storage resource usage and weak noise immunity when deployed in embedded systems, limiting their application in low-cost controllers.

[0005] Therefore, there is an urgent need for a circular induction synchronizer error prediction method and system based on a hybrid model that integrates physical mechanisms and data-driven methods to solve the problems of positioning inaccuracy, poor real-time performance, and weak generalization ability caused by temperature drift and complex interference in high-precision servo systems. Summary of the Invention

[0006] To solve the above problems, this application proposes a circular induction synchronizer error prediction method and system based on a hybrid model, which not only realizes efficient and reliable embedded real-time correction of the circular induction synchronizer, but also improves the error prediction accuracy of the circular induction synchronizer.

[0007] The present invention provides a circular induction synchronizer error prediction method based on a hybrid model, comprising the following steps:

[0008] S1. Obtain target position-error data;

[0009] S2. Constructing a circular induction synchronizer error prediction model based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism, and the linear superposition principle; the circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch, and an error fusion layer;

[0010] S3. Input the target position-error data into the circular induction synchronizer error prediction model to obtain an error prediction value.

[0011] Preferably, in S2, a circular induction synchronizer error prediction model is constructed according to the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle, specifically including:

[0012] S201, obtaining a historical position-error data set; the historical position-error data set includes: historical position-error data and historical position-error data prediction results;

[0013] S202, constructing a training set and a test set based on the historical position-error dataset;

[0014] S203. Based on the normalization principle, construct an input coding layer;

[0015] S204, constructing a parallel processing branch based on the first harmonic linear term, the radial basis function neural network term, and the dynamic parameter optimization mechanism;

[0016] S205. Based on the linear superposition principle, construct an error fusion layer;

[0017] S206, splicing the input coding layer, the parallel processing branch and the error fusion layer to obtain an initial circular induction synchronizer error prediction model;

[0018] S207, inputting the training set into the initial circular induction synchronizer error prediction model for training, to obtain a trained initial circular induction synchronizer error prediction model;

[0019] S208, inputting the test set into the trained initial circular induction synchronizer error prediction model, adjusting the trained initial circular induction synchronizer error prediction model, and obtaining a trained initial circular induction synchronizer error prediction model;

[0020] S209: Determine the trained initial circular induction synchronizer error prediction model as the circular induction synchronizer error prediction model.

[0021] Preferably, in S203, based on the normalization principle, an input coding layer is constructed, which specifically includes:

[0022] The historical position-error data is normalized to obtain the sine feature component and the cosine feature component, and the input coding layer is obtained.

[0023] Preferably, in S204, a parallel processing branch is constructed based on the first harmonic linear term, the radial basis function neural network term, and the dynamic parameter optimization mechanism, specifically including:

[0024] Based on the sine and cosine characteristic components, the first harmonic prediction term is constructed through the least squares dynamic fitting;

[0025] According to the K-means algorithm and Gaussian kernel function, the radial basis function neural network term is constructed and the nonlinear error term is obtained;

[0026] Grid search and ridge regression algorithms are used to optimize radial basis function neural network terms to obtain optimized radial basis function neural network terms;

[0027] Calculate the nonlinear error according to the optimized radial basis function neural network term to obtain the nonlinear error term;

[0028] The first harmonic prediction term and the optimized radial basis function neural network term are spliced ​​together to obtain a parallel processing branch.

[0029] Preferably, grid search and ridge regression algorithms are used to optimize the radial basis function neural network terms to obtain optimized radial basis function neural network terms, specifically including:

[0030] Using grid search to optimize hyperparameters in the radial basis function neural network term to obtain an initial optimized radial basis function neural network term; the hyperparameters include: the number of radial basis function centers and the regularization coefficient;

[0031] The ridge regression algorithm is used to constrain the complexity of the initial optimized radial basis function neural network term and obtain the optimized radial basis function neural network term.

[0032] Preferably, the expression of the first harmonic prediction term is:

[0033] ;

[0034] in, For the first harmonic prediction, sin θ is the sinusoidal characteristic component, cos θ is the cosine eigencomponent, is the sinusoidal eigencomponent coefficient, is the cosine eigencomponent coefficient.

[0035] Preferably, the expression of the Gaussian kernel function is:

[0036] ;

[0037] in, is the input of the network, is the i-th node, is the Gaussian kernel function, for e The exponential function with base , is the center vector of the ith node in the network, is the base width parameter of the node;

[0038] The expression of the nonlinear error term is:

[0039] ;

[0040] in, is the nonlinear error, For the summation symbol, is the RBF output weight, is the Gaussian kernel function.

[0041] Preferably, in S205, an error fusion layer is constructed based on the linear superposition principle, specifically including:

[0042] Based on the first harmonic prediction term and the nonlinear error term, the error fusion layer is obtained by adopting the linear superposition principle, and the prediction error is obtained.

[0043] Preferably, the prediction error is expressed as:

[0044] ;

[0045] in, is the prediction error.

[0046] The present invention provides a circular induction synchronizer error prediction system based on a hybrid model, comprising:

[0047] A data acquisition module is used to obtain target position-error data;

[0048] A model construction module is used to construct a circular induction synchronizer error prediction model based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle; the circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch and an error fusion layer;

[0049] The prediction module is used to input the target position-error data into the circular induction synchronizer error prediction model to obtain the error prediction value.

[0050] In summary, the hybrid model-based circular induction synchronizer error prediction method and system of the present invention have the following advantages over traditional technologies:

[0051] (1) The circular induction synchronizer error prediction model constructed by the present invention adopts a physical-data dual-driven modeling mechanism and a collaborative compensation mechanism, which solves the problem that traditional error prediction methods usually adopt a single modeling method of pure data drive or pure physical model harmonic analysis, which is prone to overfitting and lacks interpretability.

[0052] (2) The circular induction synchronizer error prediction model constructed in this invention adopts a deep fusion architecture of theoretical modeling and data learning. Guided by physical priors, based on the circular induction synchronizer error generation mechanism, a first harmonic term is constructed, which not only directly corrects the fundamental frequency error caused by installation eccentricity, but also ensures the physical interpretability of the linear error component.

[0053] (3) The circular induction synchronizer error prediction model constructed by the present invention adopts a dynamic adaptive optimization algorithm, which not only avoids overfitting and improves the robustness in a noisy environment, but also realizes the dynamic adaptation of the circular induction synchronizer error prediction model structure and performance.

[0054] (4) The circular induction synchronizer error prediction model constructed by the present invention solves the problems of positioning inaccuracy, poor real-time performance and weak generalization ability caused by temperature drift and complex interference in high-precision servo systems, realizes efficient and reliable embedded real-time correction of circular induction synchronizers, and improves the error prediction accuracy of circular induction synchronizers.

[0055] The technical method of the present invention is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of a flow chart of a circular induction synchronizer error prediction method based on a hybrid model of the present invention;

[0057] Figure 2 This is a structural diagram of the circular induction synchronizer error prediction model of the present invention;

[0058] Figure 3 Input data graph for inventing historical position-error dataset;

[0059] Figure 4 This is a prediction flow chart of the circular induction synchronizer error prediction model of the present invention;

[0060] Figure 5 A comparison diagram of the measured error and the predicted error of the circular induction synchronizer error prediction model of the present invention;

[0061] Figure 6A physical diagram of the circular induction synchronizer error prediction model of the present invention;

[0062] Figure 7 This is a structural block diagram of a digital demodulation module of the circular induction synchronizer error prediction model of the present invention;

[0063] Figure 8 This is an oscilloscope measurement diagram of the induction signal of the circular induction synchronizer error prediction system of the present invention;

[0064] Figure 9 This is an oscilloscope measurement diagram of the power ripple of the circular induction synchronizer error prediction system of the present invention. DETAILED DESCRIPTION

[0065] The technical method of the present invention is further described below through the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and values ​​described in these embodiments do not limit the scope of this application.

[0066] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0067] Technologies, systems, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0068] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0069] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0070] The present invention provides a circular induction synchronizer error prediction method based on a hybrid model, such as Figure 1 As shown, specifically including:

[0071] Step S1, obtaining target position-error data. The present invention obtains target position-error data by circular grating calibration.

[0072] Step S2: Construct a circular induction synchronizer error prediction model based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle. Figure 2 As shown, the circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch and an error fusion layer.

[0073] Step S2 can be replaced by the following steps S201 to S209, as follows:

[0074] Step S201, obtain the historical position-error data set. The position-error data set obtained by circular grating calibration in this application includes: historical position-error data and historical position-error data prediction results. The input data diagram of the historical position-error data set of the present invention is as follows: Figure 3 shown.

[0075] When the present invention is put into practical application, a digital demodulation system is used in combination with a circular grating calibration system with a resolution of 24 bits to drive the permanent magnet synchronous motor to rotate at a relatively low speed, for example, at a constant speed of 0.065° / s, and 180 sets of historical position-error data are collected with a tolerance of 2°. The data format is a triplet. 、 and ,in The original output value of the circular induction synchronizer (14 bits, 0~16384). is the reference angle of the circular grating (calibration value), It is the zero error angle encoding.

[0076] Step S202: construct a training set and a test set based on the historical position-error dataset.

[0077] Step S203: construct an input coding layer based on the normalization principle.

[0078] Step S203 specifically normalizes the historical position-error data to obtain the sine feature component and the cosine feature component, and obtains the input coding layer. In the actual application process, the angle coding is first performed to convert the original angle Convert to a periodic feature:

[0079] ;

[0080] in, is the decimal angle value after period conversion, After the period is converted to the sinusoidal component, After the period conversion, it is the cosine component. is the normalized sine and cosine tensor matrix.

[0081] Then the error is normalized, specifically Perform sliding window Z-score normalization:

[0082] ;

[0083] in, is the standardized error, and are the mean and standard deviation of the most recent 50 samples, respectively.

[0084] Step S204: construct a parallel processing branch based on the first harmonic linear term, the radial basis function neural network term and the dynamic parameter optimization mechanism.

[0085] Step S204 specifically includes the following contents:

[0086] Based on the sine and cosine characteristic components, the first harmonic prediction term is constructed through the least squares method dynamic fitting. The expression of the constructed first harmonic prediction term is:

[0087] ;

[0088] in, For the first harmonic prediction, sin θ is the sinusoidal characteristic component, cos θ is the cosine eigencomponent, is the sinusoidal eigencomponent coefficient, is the cosine eigencomponent coefficient.

[0089] According to the K-means algorithm and Gaussian kernel function, the radial basis function neural network term is constructed and the nonlinear error term is obtained. Among them, the expression of the Gaussian kernel function is:

[0090] ;

[0091] in, is the input of the network, is the i-th node, is the Gaussian kernel function, for e The exponential function with base , is the center vector of the ith node in the network, is the base width parameter of the node;

[0092] The expression of the nonlinear error term is:

[0093] ;

[0094] in, is the nonlinear error, For the summation symbol, is the RBF output weight, is the Gaussian kernel function.

[0095] Grid search and ridge regression algorithms are used to optimize the radial basis function neural network term to obtain the optimized radial basis function neural network term. Specifically, grid search is first used to optimize the hyperparameters in the radial basis function neural network term to obtain the initial optimized radial basis function neural network term. Then, ridge regression is used to constrain the complexity of the initial optimized radial basis function neural network term to obtain the optimized radial basis function neural network term. The hyperparameters include the number of radial basis function centers and the regularization coefficient.

[0096] The nonlinear error is calculated according to the optimized radial basis function neural network term to obtain the nonlinear error term.

[0097] The first harmonic prediction term and the optimized radial basis function neural network term are spliced ​​together to obtain a parallel processing branch.

[0098] Step S205: construct an error fusion layer based on the linear superposition principle.

[0099] Step S205 is specifically based on the first harmonic prediction term and the nonlinear error term, using the linear superposition principle to obtain an error fusion layer and obtain the prediction error. The expression of the prediction error is:

[0100] ;

[0101] in, E pred is the prediction error.

[0102] Step S206: splice the input coding layer, the parallel processing branch and the error fusion layer to obtain an initial circular induction synchronizer error prediction model.

[0103] Step S207: input the training set into the initial circular induction synchronizer error prediction model for training to obtain a trained initial circular induction synchronizer error prediction model.

[0104] Step S208: input the test set into the trained initial circular induction synchronizer error prediction model, adjust the trained initial circular induction synchronizer error prediction model, and obtain a trained initial circular induction synchronizer error prediction model.

[0105] Step S209: Determine the trained initial circular induction synchronizer error prediction model as the circular induction synchronizer error prediction model.

[0106] When doing actual application, construct Figure 2 The circular induction synchronizer error prediction model shown contains the following components:

[0107] Harmonic linear term: first harmonic compensator, the expression is .

[0108] Radial Basis Function Neural Network: Contains Dynamic cluster centers, where K To represent the number of hidden nodes in the radial basis function (RBF) neural network, the activation function is the Gaussian kernel function:

[0109] .

[0110] Fusion output layer: linear superposition .

[0111] Parameter joint optimization uses grid search and cross-validation to optimize hyperparameters. The procedure is as follows:

[0112] param_grid={

[0113] 'n_rbf_centers':[12,15,18],

[0114] 'alpha':[0.1,0.5,1.0]

[0115] };

[0116] grid_search=GridSearchCV(HybridRBFModel(),param_grid,cv=5);

[0117] grid_search.fit(X_train,y_train).

[0118] Optimal parameter combination example: n_rbf_centers=15, α =0.5.

[0119] Then, incremental training is performed. During the deployment phase, new data is continuously received and the model is updated according to the following strategy:

[0120] Dynamic center adjustment: When the MAE of 100 consecutive samples is greater than 0.8, the new radial basis function center is:

[0121] ;

[0122] in, To add the radial basis function center, is the total number of samples, For the current samples, For the The angle-encoded feature vector of each sample.

[0123] Online weight update: Solved using ridge regression, expressed as follows:

[0124] ;

[0125] in, is the regularized solution, is the transposed matrix of the training set matrix, is the training set matrix, is the regularization parameter, is the identity matrix, is the target variable, T is the transpose of the matrix.

[0126] The result of harmonic analysis is the first harmonic coefficient: =-0.3859, =-0.0307; equivalent installation eccentricity: 0.3871; initial phase angle: -175.5°.

[0127] S3. Input the target position-error data into the circular induction synchronizer error prediction model to obtain an error prediction value.

[0128] The prediction flow chart of the circular induction synchronizer error prediction model of the present invention is as follows: Figure 4 As shown, Figure 4 The real-time prediction process includes the following steps: the sensor reads the raw angle θ_raw and transmits it to the MCU; the preprocessing module calculates the sine / cosine features [sinθ, cosθ]; the system executes LUT queries in parallel: the left path calls the harmonic calculation unit to output E_harm, while the right path calculates φ_i using the RBF kernel function and takes a weighted sum to output E_rbf; the fusion layer linearly superimposes E_harm and E_rbf to generate the prediction error E_pred; and finally, the compensation module executes θ_corrected = θ_raw - E_pred to output the corrected angle. The entire process uses a pipeline design, with a single prediction latency of ≤0.2ms. DMA transfer enables parallel processing of computation and data acquisition, ensuring real-time error compensation at a 5kHz sampling rate.

[0129] The real-time compensation test results of the present invention show that the error prediction method based on the hybrid model has a maximum value of 31.5644 code words between the measured data and the predicted data, which can accurately predict the zero-bit error, as shown in Table 1:

[0130] Table 1 Measured error and predicted error

[0131] ;

[0132] The comparison diagram of the measured error and predicted error of the circular induction synchronizer error prediction model of the present invention is as follows: Figure 5 As shown, Figure 5 The coarse and fine zero position deviation after display compensation is less than 50 code words, and the error is less than 0.2°.

[0133] The present invention provides a circular induction synchronizer error prediction system based on a hybrid model, comprising:

[0134] A data acquisition module is used to obtain target position-error data;

[0135] A model construction module is used to construct a circular induction synchronizer error prediction model based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle; the circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch and an error fusion layer;

[0136] The prediction module is used to input the target position-error data into the circular induction synchronizer error prediction model to obtain the error prediction value.

[0137] A hybrid model-based method and system for circular induction synchronizer error compensation primarily includes a circular induction synchronizer, an AD2S1210 digital demodulation module, and a circular induction synchronizer error prediction model. Figure 6 shows the actual system, which includes a circular induction synchronizer, a Renishaw circular encoder, a coupling, a permanent magnet synchronous motor, and a digital demodulation device.

[0138] Circular synchronizer: The circular synchronizer is a dual-channel absolute circular induction synchronizer with a speed ratio of 1:180 and a diameter of 120mm.

[0139] AD2S120 digital demodulation module, the structural block diagram is as follows Figure 7 As shown: The AD2S120 digital demodulation module includes a power supply module, an AD2S1210 demodulation module, an STM32 driver module, an OPA548T differential power amplifier module, and a UAF42 weak signal processing module.

[0140] The power modules include the AMS1117-5.0, AMS1117-3.3V, and TPS5430 dual-channel positive and negative switching power supplies, which output +5V, +3.3V, and ±12V power supply voltages.

[0141] The AD2S1210 demodulation module uses single-phase excitation and dual-phase output amplitude detection to perform position conversion. It receives the SPI drive signal from the STM32 and outputs a 10 kHz sinusoidal excitation signal to the OPA548T differential power amplifier module. It receives the four-channel coarse and fine channel sensing signals from the UAF42 and outputs dual 14-bit coarse and fine channel positions to the STM32 driver module.

[0142] The STM32 driver module is used to drive the AD2S1210 internal configuration registers, control registers, position registers, and operating modes, receive dual 14-bit coarse and fine channel positions, and deploy a hybrid error compensation model.

[0143] The OPA548T differential power amplifier module is used to differentially amplify the sinusoidal excitation signal from the AD2S1210. Differential power amplification can reduce planar current fluctuations and improve the circuit's current carrying capacity. The amplified excitation signal current can reach 1A.

[0144] The UAF42 module receives weak coarse and fine sensing signals from the circular induction synchronizer and performs two-stage amplification. The first-stage common-phase amplifier has a gain of 40 times and a bandwidth of 25KHz. The second-stage common-phase amplifier has a gain of 37.5 times and a bandwidth of 26.7KHz, with a total gain of 1500 times.

[0145] Circular induction synchronizer error prediction model: The circular induction synchronizer error prediction model includes an input coding layer, parallel processing branches and an error fusion layer.

[0146] The input coding layer converts the 14-bit angle value (0-16384) output by the circular induction synchronizer into sine and cosine components through normalization, eliminates the periodic mutation of the angle, and generates the input feature vectors sinθ and cosθ.

[0147] The parallel processing branch includes a first harmonic linear term, a radial basis function neural network term, and a dynamic parameter optimization mechanism.

[0148] The first harmonic linear term is used to construct the first harmonic compensation term to correct the fundamental frequency error caused by installation eccentricity, where the coefficient and Dynamic fitting by least squares method.

[0149] The radial basis function neural network item includes a radial basis function network for designing dynamic clustering, the hidden layer center is adaptively generated by the K-means algorithm, the activation function is a Gaussian kernel function, and the output is a nonlinear error term.

[0150] The dynamic optimization mechanism uses grid search to jointly optimize hyperparameters such as the number of radial basis function centers (10-20) and the regularization coefficient (0.1-1.0), and combines 5-fold cross-validation to screen the optimal combination to avoid overfitting.

[0151] The dynamic optimization mechanism introduces ridge regression to solve the radial basis function output weights, constrains the model complexity through L2 regularization, and improves the robustness in noisy environments.

[0152] The error fusion layer outputs the final prediction error through linear superposition to achieve coordinated compensation of linear and nonlinear errors.

[0153] The oscilloscope measurement diagram of the induction signal of the circular induction synchronizer error prediction system of the present invention, that is, the amplitude of the single-ended signal of the coarse channel SIN of the dual-channel circular induction synchronizer mentioned in this application is about 5mVp-p. After filtering, amplification and processing by the signal processing circuit, the amplitude of the output signal is as follows: Figure 8 As shown, the amplitude is about 3.16Vp-p, the waveform is smooth, the sine characteristics are obvious, and the phase offset with the excitation signal is less than 44 degrees. Figure 9 As shown in the figure, the TPS5430 uses positive and negative power outputs, and the actual measured efficiency is about 88%. The negative voltage conversion rate is lower than the positive voltage.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A circular induction synchronizer error prediction method based on a hybrid model, characterized in that: The following steps are involved: S1. Obtain target position-error data; S2. Based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle, a circular induction synchronizer error prediction model is constructed; The circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch and an error fusion layer; S3. Input the target position-error data into the circular induction synchronizer error prediction model to obtain an error prediction value; In S2, a circular induction synchronizer error prediction model is constructed based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism, and the linear superposition principle. Specifically, it includes: S201, obtaining a historical position-error data set; the historical position-error data set includes: historical position-error data and historical position-error data prediction results; S202, constructing a training set and a test set based on the historical position-error dataset; S203. Based on the normalization principle, construct an input coding layer; S204, constructing a parallel processing branch based on the first harmonic linear term, the radial basis function neural network term, and the dynamic parameter optimization mechanism; S205. Based on the linear superposition principle, construct an error fusion layer; S206, splicing the input coding layer, the parallel processing branch and the error fusion layer to obtain an initial circular induction synchronizer error prediction model; S207, inputting the training set into the initial circular induction synchronizer error prediction model for training, to obtain a trained initial circular induction synchronizer error prediction model; S208, inputting the test set into the trained initial circular induction synchronizer error prediction model, adjusting the trained initial circular induction synchronizer error prediction model, and obtaining a trained initial circular induction synchronizer error prediction model; S209: Determine the trained initial circular induction synchronizer error prediction model as the circular induction synchronizer error prediction model.

2. The circular induction synchronizer error prediction method based on a hybrid model according to claim 1, characterized in that: In S203, based on the normalization principle, an input coding layer is constructed, which specifically includes: The historical position-error data is normalized to obtain the sine feature component and the cosine feature component, and the input coding layer is obtained.

3. The circular induction synchronizer error prediction method based on a hybrid model according to claim 2, characterized in that: In S204, a parallel processing branch is constructed based on the first harmonic linear term, the radial basis function neural network term, and the dynamic parameter optimization mechanism, specifically including: Based on the sine and cosine characteristic components, the first harmonic prediction term is constructed through the least squares dynamic fitting; According to the K-means algorithm and Gaussian kernel function, the radial basis function neural network term is constructed and the nonlinear error term is obtained; Grid search and ridge regression algorithms are used to optimize radial basis function neural network terms to obtain optimized radial basis function neural network terms; Calculate the nonlinear error according to the optimized radial basis function neural network term to obtain the nonlinear error term; The first harmonic prediction term and the optimized radial basis function neural network term are spliced ​​together to obtain a parallel processing branch.

4. The method for predicting circular induction synchronizer error based on a hybrid model according to claim 3, characterized in that: The grid search and ridge regression algorithms are used to optimize the radial basis function neural network terms to obtain the optimized radial basis function neural network terms, including: Using grid search to optimize hyperparameters in the radial basis function neural network term to obtain an initial optimized radial basis function neural network term; the hyperparameters include: the number of radial basis function centers and the regularization coefficient; The ridge regression algorithm is used to constrain the complexity of the initial optimized radial basis function neural network term and obtain the optimized radial basis function neural network term.

5. The circular induction synchronizer error prediction method based on hybrid model according to claim 3 is characterized in that: The expression of the first harmonic prediction term is: ； in, is the first harmonic prediction, is the sinusoidal characteristic component, is the cosine eigencomponent, is the sinusoidal eigencomponent coefficient, is the cosine eigencomponent coefficient.

6. The circular induction synchronizer error prediction method based on hybrid model according to claim 3, characterized in that: The expression of the Gaussian kernel function is: ； in, is the input of the network, is the i-th node, is the Gaussian kernel function, for e The exponential function with base , is the center vector of the ith node in the network, is the base width parameter of the node; The expression of the nonlinear error term is: ； in, is the nonlinear error, For the summation symbol, is the RBF output weight, is the Gaussian kernel function.

7. The method for predicting circular induction synchronizer error based on hybrid model according to claim 3, characterized in that: In S205, based on the linear superposition principle, an error fusion layer is constructed, which specifically includes: Based on the first harmonic prediction term and the nonlinear error term, the error fusion layer is obtained by adopting the linear superposition principle, and the prediction error is obtained.

8. The method for predicting circular induction synchronizer error based on hybrid model according to claim 7, characterized in that: The expression of the prediction error is: ； in, is the prediction error.

9. A circular induction synchronizer error prediction system based on a hybrid model, characterized in that: include: A data acquisition module is used to obtain target position-error data; A model building module is used to construct a circular induction synchronizer error prediction model based on the normalization principle, the first harmonic linear term, the radial basis function neural network term, the dynamic parameter optimization mechanism and the linear superposition principle; The circular induction synchronizer error prediction model includes: an input coding layer, a parallel processing branch and an error fusion layer, specifically including: Acquire a historical position-error data set; the historical position-error data set includes: historical position-error data and historical position-error data prediction results; Construct training and test sets based on the historical position-error dataset; Based on the normalization principle, construct the input encoding layer; Based on the first harmonic linear term, radial basis function neural network term and dynamic parameter optimization mechanism, a parallel processing branch is constructed; Based on the principle of linear superposition, an error fusion layer is constructed; The input coding layer, parallel processing branch and error fusion layer are spliced ​​together to obtain the initial circular induction synchronizer error prediction model; Inputting the training set into the initial circular induction synchronizer error prediction model for training, thereby obtaining a trained initial circular induction synchronizer error prediction model; Inputting the test set into the trained initial circular induction synchronizer error prediction model, adjusting the trained initial circular induction synchronizer error prediction model, and obtaining a trained initial circular induction synchronizer error prediction model; determining the trained initial circular induction synchronizer error prediction model as the circular induction synchronizer error prediction model; The prediction module is used to input the target position-error data into the circular induction synchronizer error prediction model to obtain the error prediction value.

Images