A bearingless compound rotor cage asynchronous motor fault-tolerant detection and control method

Abstract

The application relates to the technical field of motor fault-tolerant detection and control, and discloses a bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method, which comprises the following steps: introducing a dynamic kernel width optimized adaptive generalized regression neural network, so that the accuracy and reliability of displacement sensor fault detection are improved. The AGRNN model can adapt to the unevenness of sample distribution, improve the fitting capacity of a nonlinear system, and make the displacement estimation error decrease to a certain extent; the designed four-sensor annular redundant configuration supports displacement reconstruction under the condition of double-sensor fault, significantly improves the fault-tolerant capacity of the system, and enables the bearingless composite rotor cage asynchronous motor to keep stable operation under more severe conditions; the proposed multi-scale anti-interference filtering technology can effectively suppress high-frequency interference, improve the quality of the displacement signal, and improve the anti-interference capacity of the system in a complex industrial environment.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault-tolerant detection and control, and more particularly to a method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor. Background Art

[0002] As a mainstream electromechanical energy conversion device in the industrial sector, asynchronous motors dominate applications such as flywheel energy storage systems, CNC machine tool spindle drives, compressor clusters, and agricultural mechanization. However, the dependence of traditional asynchronous motors on mechanical bearings presents significant technical bottlenecks in high-speed, high-precision, and long-cycle operation. Especially under extreme operating conditions, bearing wear and lubrication failure severely limit equipment reliability and service life. To overcome the bearing limitations of traditional structures, bearingless composite rotor cage asynchronous motors integrate magnetic levitation technology with asynchronous drive principles to achieve rotor levitation and rotation, successfully transitioning from mechanical contact support to magnetic levitation control. However, it is worth noting that the suspension closed-loop feedback control system of bearingless composite rotor cage asynchronous motors relies heavily on accurate detection and real-time feedback from displacement sensors. In complex industrial environments, displacement sensors are susceptible to multiple interferences, such as high-frequency electromagnetic noise, high temperatures and oily environments, and mechanical vibration. If a displacement sensor fails, the closed-loop control system will fail, directly leading to rotor suspension instability, which can lead to serious accidents such as unbalanced suspension forces and rotor-stator collisions.

[0003] Therefore, how the fault-tolerant detection and control method of the bearingless composite rotor cage asynchronous motor can handle the high complexity and meet the real-time response requirements of the high-speed motor, and how the traditional redundant sensor configuration can cope with both single sensor failure scenarios and concurrent multi-sensor failures have become technical problems that need to be solved urgently. Summary of the Invention

[0004] The present invention provides a fault-tolerant detection and control method for a bearingless composite rotor cage-type asynchronous motor, which solves the technical problems in the prior art of how to handle high complexity while meeting the real-time response requirements of high-speed motors, and how to configure traditional redundant sensors to cope with both single-sensor failure scenarios and concurrent multi-sensor failures.

[0005] The present invention provides a fault-tolerant detection and control method for a bearingless composite rotor cage asynchronous motor, comprising:

[0006] Displacement estimation and fault detection using an adaptive generalized regression neural network based on dynamic kernel width optimization;

[0007] Based on displacement estimation and fault detection, high-precision displacement data is obtained by selecting three-phase suspension current, air gap flux and eccentric magnetic pull as neural network input features;

[0008] Based on the adaptive generalized regression neural network, a four-sensor ring redundant configuration is used to achieve multi-sensor fault tolerance. Through online parameter calibration and majority voting displacement reconstruction algorithm, accurate fault tolerance information is obtained.

[0009] Based on the four-sensor ring redundant configuration, multi-scale wavelet filtering is used for anti-interference processing, and the improved SUREShrink threshold rule is adopted to achieve signal reconstruction and noise suppression to obtain fault-tolerant and anti-interference capabilities.

[0010] Furthermore, the adaptive generalized regression neural network based on dynamic kernel width optimization for displacement estimation and fault detection includes:

[0011] Data preprocessing and feature input: three-phase suspension current, air gap flux and eccentric magnetic pull are selected as input features. The data are normalized and then input into AGRNN.

[0012] The AGRNN model is trained by dynamically adjusting the local kernel width parameter and introducing a regularization mechanism;

[0013] The trained AGRNN model is used to estimate the displacement online and generate the residual by subtracting it from the actual sensor value.

[0014] Based on the comparison between the residual error and the dynamic threshold, the sensor failure is determined.

[0015] Furthermore, the adaptive generalized regression neural network includes:

[0016] Utilize residual analysis and dynamic threshold mechanism for fault detection, achieving real-time fault monitoring;

[0017] The adaptive generalized regression neural network structure is designed as a four-layer structure consisting of input layer, pattern layer, summation layer and output layer, in which the summation layer contains two types of neurons: denominator unit neurons and numerator unit neurons.

[0018] Furthermore, the specific steps of the dynamic core width optimization include:

[0019] Introducing the local kernel width parameter σ i , and according to the sample density distribution, it is inversely proportional to the sample density, and its expression is: , where d i Represents the sample point v i Euclidean distance to the nearest neighbor, k d is a hyperparameter;

[0020] The kernel width parameter is iteratively optimized based on the gradient descent method, and its iterative update expression is:

[0021] , to achieve adaptive adjustment of parameters; where RMSE is the root mean square error; η is the learning rate; represents the local kernel width parameter σ i At the tth iteration, the current value of the i-th parameter; is the local kernel width parameter σ i The updated parameter value, that is, the new value of the i-th parameter at the t+1th iteration.

[0022] Introduction The regularization mechanism suppresses overfitting, and the total loss function L is: ; To ensure the stability of the model, where λ is the regularization term and n is the sample size;

[0023] When λ=0.1, the RMSE of the test set is 1.2×10−4mm, which is the lowest value, indicating that moderate regularization can effectively suppress overfitting. When λ=0.0 and λ≥0.5, the RMSE increases significantly, so the regularization coefficient λ is taken as 0.1;

[0024] According to the gradient update formula, it is corrected as follows: ;

[0025] Set the convergence conditions: , where ε is the preset threshold to obtain the convergence of the optimization process.

[0026] Furthermore, the displacement dx is estimated online e ,dy e and compared with the value ds measured by the sensor and filtered x_denoised ,ds y_denoised Subtraction generates residual e x , e y .

[0027] ;

[0028] The dynamic threshold is set as: ;

[0029] in, is the mean of the residuals in the stable suspension stage, is the variance of the residual in the stable suspension stage;

[0030] The fault judgment rule is set as follows: or , it is determined that the corresponding direction sensor is faulty.

[0031] Furthermore, the four-sensor ring redundant configuration includes:

[0032] By providing an x-axis direction sensor S1 and a y-axis direction sensor S2, the redundant sensor S3 is arranged along the direction with an angle of 0°<α<90° to the positive direction of the x-axis, and the redundant sensor S4 is arranged along the direction with an angle of 90°<β<180° to the positive direction of the x-axis, forming a four-sensor ring layout.

[0033] Furthermore, the four-sensor ring redundant configuration further includes:

[0034] The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds3 of the three displacement sensors S1, S2, and S3 are then synchronously recorded, and the objective function J(α) is constructed based on the least squares method: ;

[0035] The optimal solution of the actual installation angle α is iteratively solved through the gradient descent optimization algorithm; the calibrated α value is updated to the control system parameter table to complete the dynamic correction;

[0036] The calibration process for β is similar to that for α. The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds4 of the three displacement sensors S1, S2, and S4 are then synchronously recorded. The objective function J(β) is constructed using the least squares method: ;

[0037] The optimal solution of the actual installation angle β is iteratively solved through the gradient descent optimization algorithm; the calibrated β value is updated to the control system parameter table to complete the dynamic correction;

[0038] Within the linear measurement range of the displacement sensor, the change in the measured displacement value is linearly related to the change in the output voltage k: Where Δu and Δds represent the absolute value of the change in the displacement sensor output voltage and the absolute value of the change in the rotor displacement, respectively, and the data is unified by making the gains of all displacement sensors the same.

[0039] Further, the analysis of various sensor failures includes:

[0040] When sensor S1, which is used as the x-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the x-axis displacement sensor, and the x-axis reconstructed displacement is obtained as: ;

[0041] When sensor S2, which is used as the y-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the y-axis displacement sensor, and the y-axis reconstructed displacement is obtained as: ;

[0042] When sensors S1 and S2, which are used as sensors for the x-axis and y-axis directions, fail at the same time, sensors S3 and S4 are used to reconstruct the x- and y-axis displacements. The reconstructed x- and y-axis displacements are obtained as follows:

[0043] ;

[0044] The designed ring-based redundant sensors and majority voting mechanism support displacement reconstruction under dual sensor failures to achieve fault-tolerant operation under dual sensor failures.

[0045] Furthermore, the multi-sensor fault-tolerant strategy of the ring redundant layout includes:

[0046] Multi-sensor fault-tolerant design and multi-scale anti-interference filtering;

[0047] The multi-scale anti-interference filtering includes: performing three-layer wavelet decomposition on the original displacement signal, selecting the db4 mother wavelet to obtain high-frequency detail coefficients D1, D2, D3 and low-frequency approximation coefficient A3; achieving high-precision displacement extraction under interference through layered signal processing; ds signal decomposition expression: ;

[0048] An improved SUREShrink threshold rule is used for high-frequency detail coefficients to suppress high-frequency noise:

[0049] Where, D j is the coefficient of the original signal at the jth position, is the coefficient after denoising, λ th is the threshold adjustment coefficient, sign(D j ) means taking D j The sign of , which means positive is 1, negative is -1, and zero is 0; Indicates threshold shrinkage of the absolute value. When the absolute value is less than the threshold, it is set to zero, otherwise the threshold is subtracted. For the derived threshold adjustment item, the threshold is adaptively adjusted according to the data scale;

[0050] Reconstruct the denoised signal ds_ using the thresholded coefficients denoised : ;in, is the approximate coefficient of the third level decomposition.

[0051] The application has the beneficial effects that: the application introduces the adaptive generalized regression neural network (AGRNN) with dynamic kernel width optimization, improves the accuracy and reliability of displacement sensor fault detection. The AGRNN model can adapt to the unevenness of sample distribution, improve the fitting ability of the nonlinear system, so that the displacement estimation error is reduced, and the determination coefficient is improved to 0.985; the designed four-sensor ring redundancy configuration supports displacement reconstruction under the condition of double-sensor fault, significantly improves the fault tolerance of the system, and enables the CCR-BIM to operate stably under more severe conditions; the proposed multi-scale anti-interference filtering technology can effectively suppress high-frequency interference, improve the quality of the displacement signal, and improve the anti-interference ability of the system in complex industrial environments; the method has strong engineering practicability, is easy to implement, and can be widely applied to flywheel energy storage systems, numerical control machine tool spindles, compressors and other scenes with high requirements for bearingless motor suspension control reliability. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method flow diagram provided in an embodiment of the application;

[0053] Figure 2 is an RMSE convergence curve diagram of the bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method under different learning rates provided in an embodiment of the application;

[0054] Figure 3 is an AGRNN structure diagram of the bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method provided in an embodiment of the application;

[0055] Figure 4 is an AGRNN fault detection flow diagram of the bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method provided in an embodiment of the application;

[0056] Figure 5 is a four-sensor ring redundancy configuration diagram of the bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method provided in an embodiment of the application;

[0057] Figure 6 is a control diagram of the bearingless composite rotor cage asynchronous motor fault-tolerant detection and control method provided in an embodiment of the application. DETAILED DESCRIPTION

[0058] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0059] At least one embodiment of the present invention discloses a method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor, such as Figure 1 Shown, including:

[0060] Step 1: Displacement estimation and fault detection are performed based on an adaptive generalized regression neural network with dynamic kernel width optimization.

[0061] Step 2: Based on displacement estimation and fault detection, three-phase suspension current, air gap flux, and eccentric magnetic pull are selected as neural network input features to obtain high-precision displacement data;

[0062] Step 3: Based on the adaptive generalized regression neural network, a four-sensor ring redundant configuration is used to achieve multi-sensor fault tolerance. Accurate fault tolerance information is obtained through online parameter calibration and majority voting displacement reconstruction algorithm.

[0063] In step 4, based on the four-sensor ring redundant configuration, multi-scale wavelet filtering is used for anti-interference processing, and the improved SUREShrink threshold rule is used to achieve signal reconstruction and noise suppression to obtain fault-tolerant and anti-interference capabilities.

[0064] Example 1

[0065] Working Principle of the CCR-BIM: The electromagnetic structure and control mechanism of the bearingless composite cage rotor bearingless induction motor (CCR-BIM) are based on the deep integration of magnetic levitation technology and asynchronous motors. Its stator design utilizes a dual-winding integrated architecture. In addition to the traditional asynchronous motor's torque winding, an independent three-phase suspension force winding is embedded. These two windings share the same core but have different pole-pair numbers. Precisely controlling the current phase and amplitude of these two windings enables coordinated control of rotor rotation and radial suspension. The rotor utilizes an inner rotor and an outer rotor structure: the inner rotor is a specially designed fixed-pole rotor, while the outer rotor is a solid core. Currents I1 and I2 flow through the torque winding and suspension winding, respectively. When a three-phase symmetrical current with a frequency of ω1 and a pole-pair number of p1 flows through the torque winding, it generates a magnetic field rotating at the synchronous speed in the air gap, driving the rotor. When a current with a frequency of ω2 and a pole-pair number of p2 flows through the suspension winding, a controllable radial suspension magnetic field is generated.

[0066] To achieve stable levitation force output, the pole pairs and frequency of the two sets of windings must meet the following conditions: P2=P1±1, ω2=ω1. Under this condition, the torque magnetic field and the levitation magnetic field are coupled through the magnetic field modulation effect, forming an asymmetric air gap magnetic flux distribution. By adjusting the initial angle of the two windings, a controllable levitation force F can be generated on the rotor. x With F y The suspension force can be dynamically balanced by adjusting the current vector of the suspension winding, ensuring that the rotor is stably suspended in the center position without mechanical contact.

[0067] The expression of the controllable radial suspension force of CCR-BIM in the rotating two-phase coordinate system is:

[0068] ;

[0069] Where, F x and F y are the suspension forces on the inner and outer rotors along the x-axis and y-axis, respectively. B1 and B2 are the magnetic induction intensities of the torque and suspension windings, r and r' are the diameters of the inner and outer rotors, l is the effective length of the rotor, μ0 is the air gap permeability, and θ s is the initial angle difference between the torque winding and the suspension force winding, μ Am and μ Fe is the magnetic permeability of the outer rotor and the inner rotor.

[0070] When the rotor and outer rotor are both made of silicon steel, , then the incremental suspension force caused by the outer rotor can be expressed as:

[0071] ;

[0072] Therefore, the CCR-BIM suspension force expression can be simplified as:

[0073] ;

[0074] Perform coordinate transformation, and the expression of suspension force in dq rotating coordinate system is:

[0075] ;

[0076] Where K m is the suspension force coefficient, i 2sd and i 2sq are the d-axis and q-axis components of the suspension winding current, ψ 1rd and ψ 1rq are the d-axis and q-axis magnetic flux components of the torque winding, respectively.

[0077] The rotor motion equation is expressed as:

[0078] ;

[0079] Where, F zx and F zy Indicates the external disturbance force of the rotor, F sx and F sy represents the eccentric magnetic pull of the rotor, m is the mass of the motor rotor, and represent the quadratic differential terms of x and y respectively.

[0080] The eccentric magnetic pull is expressed as: ;

[0081] Where e is the rotor eccentricity value.

[0082] Example 2

[0083] Fault detection of displacement sensors: Design and fault detection of adaptive generalized regression neural network (AGRNN).

[0084] Theoretical basis and dynamic kernel width optimization of AGRNN: A generalized regression neural network (GRNN) directly approximates the joint probability density function using sample data, thereby calculating the conditional expectation as the regression output. A generalized regression neural network (GRNN) is a nonparametric regression method based on probability density function estimation and is a variant of the radial basis function (RBF) neural network.

[0085] Let V and U be two random variables, f(v,u) be their joint probability density, and v be an observation value of the random variable V. Then the conditional expectation of the random variable U under the observation value v is as follows:

[0086] , the conditional expectation E(U|v) is called the regression of the random variable U on v. Among them, u(v) is the regression function, f V (v) is the marginal probability density of random variable V. The joint probability density f(v,u) and marginal probability density f(v) can be obtained by training samples (v i ,u i )(i=1…n) and calculated using the non-parametric estimator.

[0087] The traditional GRNN kernel width parameter σ is a fixed value. However, in practical applications, the large difference in sample density between different regions may cause the model to overfit or underfit the prediction of sparse areas. By adaptively adjusting the kernel width, the local adaptability of the model can be improved. The output of the traditional GRNN with fixed kernel width is expressed as:

[0088] In practical applications, large differences in sample density across different regions can lead to overfitting or underfitting of model predictions in sparse regions. Adaptively adjusting the kernel width can improve the model's local adaptability. To this end, this paper proposes an adaptive generalized regression neural network (AGRNN) control algorithm and introduces an adaptive kernel width optimization strategy. This algorithm improves the model's adaptability to sample distribution by dynamically adjusting the local kernel width.

[0089] Assume that any sample point v i The corresponding local kernel width is σ i and make the local kernel width σ i Inversely proportional to the sample density. Local kernel width σ i The expression is: , where d i Indicates v i to its kth d Euclidean distance between neighbors, k d is a hyperparameter, the output of the proposed AGRNN is expressed as:

[0090] , whose goal is to minimize the root mean square error of the prediction error, which can be expressed as: , the gradient descent method is used for update, and the gradient calculation is expressed as: ;

[0091] in: ;

[0092] Implement σ through iterative operations iUpdate, the iterative update expression is:

[0093] , where η is the learning rate, until convergence.

[0094] In order to obtain the best value of η, a numerical test was conducted on η, and the obtained curve is as follows: Figure 2 As shown in Figure 2. When η = 0.1, RMSE oscillates violently and cannot converge stably. When η = 0.01, RMSE decreases smoothly and converges to 1.2×10 after about 300 iterations. − 4 mm. When η=0.001, the convergence speed is too slow and more than 500 iterations are required. Therefore, according to experimental tests, when η=0.01, the RMSE decreases smoothly and converges to 1.2×10 after about 300 iterations. −4 mm, in order to balance the convergence speed and stability, the learning rate η is set to 0.01.

[0095] To prevent overfitting and excessive adjustment of the kernel width, the regularization term λ is defined based on the above formula and introduced The regularization mechanism suppresses overfitting, and the total loss function L is: In order to reasonably select the regularization coefficient λ, it is necessary to verify it through cross-validation. The verification results are shown in Table 1.

[0096] Table 1: Effect of regularization coefficient on RMSE

[0097] As shown in Table 1, when λ=0.1, the RMSE of the test set is 1.2×10 −4 mm is at its lowest value, indicating that moderate regularization can effectively suppress overfitting. However, when λ=0.0 and λ≥0.5, RMSE increases significantly, so the regularization coefficient λ is preferably 0.1.

[0098] The gradient update formula is corrected to: ;

[0099] The convergence conditions are set as: , where ε is the preset threshold, usually set to 1×10 -6 When the kernel width changes for five consecutive iterations are all less than the threshold, the algorithm is considered to have converged and the optimization process is terminated.

[0100] like Figure 3 As shown in Figure 3, the proposed AGRNN structure contains 4 layers: input layer, pattern layer, summation layer and output layer.

[0101] The input layer is used to receive input samples x i (i=1...n), and pass the sample to the pattern layer, whose number of neurons is determined by the dimension of the input vector.

[0102] Each neuron x in the pattern layer i Corresponding to a local kernel width σ i , whose value is adaptively adjusted by the sample density.

[0103] The summation layer consists of two types of neurons: denominator unit neurons and numerator unit neurons. There is only one denominator unit neuron in the entire adaptive GRNN, and its connection weight with all pattern layer neurons is 1, while the numerator unit neuron performs weighted summation of the outputs of all pattern layer neurons.

[0104] The output layer outputs the estimated value vector of AGRNN, and the result is obtained by dividing the numerator unit neuron by the denominator unit neuron. , but the weight is determined by the dynamic kernel width parameter σ i Decide.

[0105] like Figure 4 As shown in the figure, the fault detection process and implementation of AGRNN are:

[0106] Data preprocessing and feature input: AGRNN is suitable for CCR-BIM rotor displacement estimation due to its strong nonlinear fitting ability and fast training characteristics. This paper combines the adaptive kernel width optimization strategy to propose a robust displacement sensor fault detection method. The specific steps are as follows:

[0107] Select the three-phase suspension current i during normal operation of CCR-BIM 2sa 、i 2sb 、i 2sc , air gap flux ψ 1α , ψ 1β and eccentric magnetic pull F sx 、F sy A total of 7-dimensional vectors are used as input features. The data is normalized and then input into AGRNN to eliminate dimensional differences. A total of 1000 samples are selected, and each sample vector contains a total of 7 elements, forming a 7-dimensional feature vector.

[0108] AGRNN model training and optimization, the model training process includes the following steps:

[0109] Initialize kernel width parameters: for each sample point v i , calculate it to the kth d The distance d between neighboring points i , and set the initial kernel width σ i , hyperparameter k d Set to 5;

[0110] Set the learning rate η=0.01, the regularization coefficient λ=0.1, and the maximum number of iterations N max =500, convergence threshold ε=1×10 -6 mm;

[0111] The kernel width is optimized using the gradient descent method, and the objective function is:

[0112] ;

[0113] Implement σ through iterative operations i Update, the iterative update expression is:

[0114] ; Check convergence conditions: when Or when the maximum number of iterations is reached, the optimization is terminated;

[0115] Save the optimized kernel width parameter σ i and the corresponding model weights.

[0116] The training set, validation set, and test set were divided into a ratio of 7:1.5:1.5, respectively. Early stopping was used during training to prevent overfitting. When the RMSE on the validation set did not improve for five consecutive iterations, training was terminated and the optimal model was regressed. The final RMSE of the test set was 1.2×10 -4 mm, MAE is 0.98×10 -4 mm, and R² was 0.985, indicating that the model had good fitting performance.

[0117] Residual generation and dynamic threshold decision: Fault detection is based on the residual analysis between the AGRNN model estimate and the sensor measured value. The specific steps are as follows:

[0118] Online estimation of displacement dx e ,dy e and compared with the value ds measured by the sensor and filtered x_denoised ,ds y_denoised Subtraction generates residual e x , e y .

[0119] ;

[0120] The dynamic threshold is set as: ;

[0121] in, is the mean of the residuals in the stable suspension stage, is the variance of the residual in the stable suspension stage.

[0122] The fault judgment rule is set as follows: or , the corresponding direction sensor is determined to be faulty, triggering fault alarm and fault-tolerant control; the dynamic threshold is adaptively adjusted according to the system operating status, which improves the accuracy of fault detection.

[0123] AGRNN fault detection process: when a sensor fault is detected, the system automatically switches to a ring redundancy fault-tolerant mode, reconstructs the displacement using the remaining sensor information, and ensures continuous operation of the system.

[0124] Performance verification of displacement estimation of AGRNN: To evaluate the fitting performance of the AGRNN model, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are selected as evaluation indicators, and are compared with the traditional fixed kernel width GRNN model. As shown in Table 1, the RMSE of AGRNN on the test set is 1.2×10 -4 mm, which is lower than the 1.56×10 -4 mm of GRNN. The MAE of AGRNN on the test set is 0.98×10 -4 mm, which is lower than the 1.32×10 -4 mm of traditional GRNN. The R² value of AGRNN is improved compared to the 0.942 of traditional GRNN. These results show that the AGRNN model has stronger fitting ability for the nonlinear relationship of displacement, as shown in Table 2.

[0125] Table 2: Performance comparison of AGRNN and traditional GRNN

[0126] To further quantitatively evaluate the fitting effect of the model, the above quantitative analysis results are mutually verified. The residual magnitude of AGRNN is concentrated in the ±1.0×10 -4 mm interval, and shows an approximately symmetric normal distribution characteristic, indicating that the system deviation of displacement estimation is effectively controlled. In contrast, the residual distribution of traditional GRNN spans ±5.6×10 -4 mm, and there are significant outliers, verifying the superiority of the AGRNN optimization strategy in error control.

[0127] Convergence analysis through the training process reveals the optimization mechanism of the proposed algorithm. Under the same conditions of an initial learning rate η = 0.01, AGRNN exhibits significant convergence advantages, with an exponential decay characteristic in the RMSE curve, reaching a stable state only after 182 iterations, with improved convergence speed compared to the traditional GRNN model. In contrast, the fixed kernel width GRNN lacks a parameter self-adaptive adjustment mechanism, and its optimization process is trapped in a local extremum, with oscillation phenomena in the later convergence period, verifying the effectiveness of the gradient adaptive learning strategy.

[0128] Example 3

[0129] Ring redundant sensor configuration design and displacement reconstruction.

[0130] To address the limitation of existing cross-redundancy configurations, which only support single-sensor failures, the present invention proposes a multi-sensor fault-tolerant strategy based on a ring-shaped redundant layout. The existing sensor fault detection device is equipped with an x-axis sensor S1, a y-axis sensor S2, and a redundant sensor S3. S3 is installed at a 45° angle to the positive direction of the x-axis to supplement displacement measurement information. When a single sensor (such as S1) fails, the system reconstructs the x-axis displacement using the measurements of S2 and S3. However, when two sensors fail simultaneously, the data from the remaining sensors cannot meet the dimensional requirements for displacement reconstruction, resulting in displacement detection failure.

[0131] like Figure 5 As shown, four sensors, S1, S2, S3, and S4, form a ring-shaped redundant sensor configuration. Specifically, the x-axis sensor S1 and the y-axis sensor S2 are arranged. Redundant sensor S3 is arranged at an angle of 0° < α < 90° to the positive x-axis, and redundant sensor S4 is arranged at an angle of 90° < β < 180° to the positive x-axis, forming a four-sensor ring layout. The main advantages of this configuration are:

[0132] (1) When any two sensors fail, the position information of the remaining two sensors can still reconstruct the complete xy plane displacement data;

[0133] (2) The ring layout structure enables each sensor to have the same signal-to-noise ratio and signal quality, avoiding the degradation of system reliability caused by the degradation of some sensor signals in the traditional configuration.

[0134] During the system startup phase, the sensor installation angle calibration is required to eliminate the impact of installation deviation on displacement reconstruction accuracy. The calibration process is as follows:

[0135] The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds3 of the three displacement sensors S1, S2, and S3 are then synchronously recorded, and the objective function J(α) is constructed based on the least squares method: ;

[0136] The optimal solution of the actual installation angle α is iteratively solved through the gradient descent optimization algorithm; the calibrated α value is updated to the control system parameter table to complete the dynamic correction;

[0137] The calibration process for β is similar to that for α. The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds4 of the three displacement sensors S1, S2, and S4 are then synchronously recorded. The objective function J(β) is constructed using the least squares method: ;

[0138] The optimal solution of the actual installation angle β is iteratively solved through the gradient descent optimization algorithm; the calibrated β value is updated to the control system parameter table to complete the dynamic correction;

[0139] This process effectively eliminates sensor installation deviation and ensures the accuracy of coordinate transformation through periodic oscillation excitation and data fitting.

[0140] Within the linear measurement range of the displacement sensor, the change in the measured displacement value is linearly related to the change in the output voltage: Where Δu and Δds represent the absolute value of the change in the displacement sensor output voltage and the absolute value of the change in the rotor displacement, respectively. To facilitate subsequent calculations and unify the formulas, the gains of all displacement sensors are assumed to be the same.

[0141] Displacement reconstruction algorithm: According to different sensor failure conditions, the present invention designs corresponding displacement reconstruction strategies:

[0142] Analyze various sensor failures separately.

[0143] When sensor S1, which is used as the x-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the x-axis displacement sensor, and the x-axis reconstructed displacement is obtained as: ;

[0144] When sensor S2, which is used as the y-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the y-axis displacement sensor, and the y-axis reconstructed displacement is obtained as: ;

[0145] When sensors S1 and S2, which are used as sensors for the x-axis and y-axis directions, fail at the same time, sensors S3 and S4 are used to reconstruct the x- and y-axis displacements. The reconstructed x- and y-axis displacements are obtained as follows:

[0146] ;

[0147] The improved ring redundant sensor design and majority voting mechanism support displacement reconstruction under dual sensor failures and fault-tolerant operation under dual sensor failures, thus improving system reliability.

[0148] Majority voting mechanism: When three of the four sensors are functioning properly, the system enhances stability by constructing possible displacement estimation combinations and calculating the Euclidean distances between them. The combination with the smallest mean distance is selected as the final displacement value. This majority voting mechanism significantly improves the reliability of displacement reconstruction and can generate accurate displacement information even in the presence of interference or signal distortion.

[0149] Example 4

[0150] Multi-scale anti-interference filter design.

[0151] Wavelet threshold denoising principle and implementation: To improve the anti-interference performance of the CCR-BIM displacement measurement system in complex industrial environments, it is necessary to suppress contamination of the displacement signal by high-frequency noise and non-stationary interference. Traditional filtering methods (such as low-pass and bandpass filtering) can eliminate noise in specific frequency bands, but their ability to handle non-stationary interference is limited and can easily cause loss of useful signals. This paper designs a wavelet multi-scale threshold denoising filtering strategy that, through layered signal processing, achieves high-precision displacement extraction in the presence of interference.

[0152] Wavelet transform has good time-frequency localization characteristics, can analyze signals at multiple scales, and is suitable for processing non-stationary signals and transient interference. The wavelet threshold denoising process of the present invention includes the following three main steps:

[0153] Wavelet decomposition: Perform a three-layer wavelet decomposition on the original displacement signal ds(t), using the db4 mother wavelet as the basis function. The db4 wavelet has good orthogonality, tight support, and high-order moment vanishing properties, making it suitable for processing mechanical vibration signals. Decomposition yields high-frequency detail coefficients D1, D2, and D3, and low-frequency approximation coefficient A3. The ds signal decomposition expression is:

[0154] ;

[0155] The frequency range corresponding to each frequency band is:

[0156] D1: 1kHz-10kHz, mainly includes electromagnetic interference and sensor high-frequency noise;

[0157] D2: 500Hz-1kHz, including medium-frequency mechanical vibration interference;

[0158] D3: 100Hz-500Hz, including low-frequency vibration components;

[0159] A3: 0-100Hz, contains useful displacement information.

[0160] Threshold processing: An improved SUREShrink threshold rule is used for high-frequency coefficients to suppress high-frequency noise. Traditional hard thresholding and soft thresholding methods often lead to over-smoothing or discontinuity problems.

[0161] Use a larger threshold coefficient ; To strongly suppress noise; for D containing some useful information j , a smaller threshold coefficient is used to retain signal details, where D j is the coefficient of the original signal at the jth position, is the coefficient after denoising, λ th is the threshold adjustment coefficient, σ j is the estimated value of the noise standard deviation of the jth layer, and N is the sample size.

[0162] Signal reconstruction: Use the thresholded wavelet coefficients to reconstruct the denoised signal and perform phase correction to eliminate the phase delay caused by filtering. Use the thresholded coefficients to reconstruct the denoised signal ds _denoised : ,The reconstructed signal is restored to its original sampling rate through quadratic spline interpolation, and the group delay caused by wavelet transform is eliminated through delay compensation to ensure the phase matching of the signal with the original control loop.

[0163] AGRNN model training and verification:

[0164] Select the three-phase suspension current i during normal operation of CCR-BIM 2sa 、i 2sb 、i 2sc 、air gap flux α-β axis component ψ 1α , ψ 1β and eccentric magnetic pull F sx 、F sy As the input feature parameters of the neural network, a 7-dimensional feature vector was constructed, and a total of 1000 samples were selected. The data was normalized and input into the AGRNN, with the training set, validation set, and test set divided into a ratio of 7:1.5:1.5.

[0165] Early stopping was used during training to prevent overfitting. When the RMSE on the validation set did not improve after five consecutive iterations, training was terminated and the optimal model was reverted. The training convergence curve showed that AGRNN converged 2.7 times faster than traditional GRNN and achieved higher final accuracy, validating the effectiveness of the dynamic kernel width optimization strategy.

[0166] In the fault detection performance test, three fault types, namely S1 and S2 sensor signal drift, open circuit and short circuit, were simulated respectively. The fault detection accuracy of AGRNN was improved to a certain extent, and the average detection time was delayed, meeting the real-time control requirements.

[0167] Ring redundant sensor fault tolerance verification: Sensor fault tolerance testing is carried out on the CCR-BIM experimental platform.

[0168] The test is divided into two scenarios: single sensor failure and double sensor failure:

[0169] Single sensor fault test:

[0170] When the S1 sensor fails, the system automatically activates the fault-tolerance mechanism and reconstructs the x-axis displacement using S3 and S4. After the failure, the x-axis displacement residual, rx, temporarily exceeds the threshold T. However, the fault-tolerance control brings the residual back to normal within a certain period of time, stabilizing the rotor position near the set point with a deviation within ±5 μm. This validates the effectiveness of the fault-tolerance strategy for single-sensor failures.

[0171] Dual sensor fault test:

[0172] Simulating simultaneous failures of S1 and S2, the system reconstructed the x- and y-axis displacements using S3 and S4. Test results showed that even in the extreme case of simultaneous failure of both primary sensors, the system maintained stable rotor suspension, maintaining position control accuracy within ±15μm. While this accuracy was somewhat lower than normal operating conditions, it still met safe operation requirements and could sustain long-term operation, validating the proposed method's robust fault tolerance in dual-sensor failure scenarios.

[0173] Calibration test:

[0174] To address the issue of sensor installation angle deviation, an online calibration experiment was conducted. Under conventional sensor installation conditions, the displacement reconstruction error is large without calibration. However, using the proposed calibration method, the reconstruction error is significantly reduced, validating the effectiveness of the online parameter calibration mechanism.

[0175] Anti-interference performance test: To verify the effectiveness of the multi-scale anti-interference filter design, the filter performance test was conducted in a high-frequency interference environment. The test conditions include: high-frequency electromagnetic interference; high interference frequency; high interference intensity; interference direction: perpendicular to the displacement sensor measurement direction;

[0176] Test results: The displacement fluctuation amplitude without filtering is higher than that after traditional low-pass filtering;

[0177] The fluctuation amplitude is significantly reduced after using the proposed multi-scale filtering.

[0178] At the same time, the filtering performance under non-stationary interference (simulated mechanical shock) conditions was tested. Compared with the traditional method, the recovery time of the proposed multi-scale filtering method was shortened, indicating that it has excellent processing capabilities for non-stationary interference and verifies the effectiveness of the proposed multi-scale filtering strategy.

[0179] By introducing the improved SUREShrink threshold rule and using adaptive threshold coefficients for different decomposition levels, the signal detail retention is improved compared with the traditional hard threshold method, and the noise suppression effect is also enhanced, which reflects the advantages of the proposed method in anti-interference filtering.

[0180] like Figure 6 As shown in the figure, the bearingless composite rotor cage asynchronous motor is divided into two parts: suspension force control and speed control.

[0181] The suspension force control process is as follows:

[0182] Displacement Detection: Four sensors acquire voltage signals based on distance and, after a k-voltage / displacement transformation, convert them into displacement signals: ds1, ds2, ds3, and ds4. Under normal operating conditions, only ds1 and ds2 signals are used. In fault mode, a residual trigger mechanism switches to a multi-sensor fusion reconstruction strategy.

[0183] Multi-scale anti-interference filtering: Perform multi-scale anti-interference filtering on the detected ds1, ds2, ds3, and ds4 filter signals and output ds1_ denoised 、ds2_ denoised 、ds3_ denoised 、ds4_ denoised Filter the signal.

[0184] AGRNN fault detection: using ψ 1α , ψ 1β 、F sx 、F sy and multi-scale anti-interference filtering branch filtering signal ds1_ denoised 、ds2_ denoised 、ds3_ denoised 、ds4_ denoised And the suspension force winding drive current i output by current pulse width modulation 2sa 、i 2sb and i 2sc As input parameter conditions, the residual is calculated in real time through the adaptive gradient neural network (AGRNN). If the residual exceeds the threshold, it is judged as a fault and the displacement reconstruction is triggered.

[0185] Fault-tolerant control and displacement reconstruction: Fault-free mode directly adopts ds1_ denoised ,ds2_ denoised As effective displacement and Output signal; Fault mode: Displacement reconstruction is performed according to the following fault types:

[0186] When sensor S1, which is used as the x-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the x-axis displacement sensor, and the x-axis reconstructed displacement is obtained as: ;

[0187] When sensor S2, which is used as the y-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the y-axis displacement sensor, and the y-axis reconstructed displacement is obtained as: ;

[0188] When sensors S1 and S2, which are used as sensors for the x-axis and y-axis directions, fail at the same time, sensors S3 and S4 are used to reconstruct the x- and y-axis displacements. The reconstructed x- and y-axis displacements are obtained as follows:

[0189] ;

[0190] Combine multiple data decision and displacement reconstruction output and The output signals are respectively different from the initial setting displacement signal and (usually 0mm) to produce an error signal. The error signal is converted into a suspension force F through a PID controller. x and F y signal, and output the dq coordinate system through the force / current exchange link and Current signal, then and The current signal is transformed into a three-phase suspension force driving current signal through 2 / 3 coordinate transformation. 、 and ,Will 、 and The current signal is adjusted through the current pulse width modulation link, and the output is a controllable suspension force driving current signal 、 and , where the current signals, and are part of the input features of AGRNN fault detection. 、 and Drive the rotor suspension of the bearingless composite rotor cage asynchronous motor to achieve a complete suspension force closed-loop feedback control loop.

[0191] The speed control process is as follows:

[0192] Detect the rotation speed of the bearingless composite rotor cage asynchronous motor With the set value torque (usually the set target speed) to produce an error signal. The error signal is converted into torque through the PI controller signal, the torque Signal and set air gap flux signal As input signal, it is input to the air gap field oriented controller.

[0193] Observe air gap flux and As the input signal, it is input into the force / current conversion part of the suspension force control link.

[0194] Air gap magnetic field oriented control: The air gap magnetic field oriented controller outputs the dq coordinate system , Current signal, then and The current signal is transformed into a three-phase rotating drive current signal through 2 / 3 coordinate transformation. 、 and .

[0195] Current pulse width modulation: current signal 、 and Through the pulse width modulation module, the current size is adjusted through the current pulse width modulation link, and the output is a controllable rotation drive current signal 、 and , current signal 、 and Drive the rotor of the bearingless composite rotor cage asynchronous motor to rotate and realize a complete speed closed-loop feedback control loop.

[0196] The entire system works together through multiple subsystems and processes to achieve precise control of suspension force and speed.

[0197] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor, characterized in that: include: Displacement estimation and fault detection using an adaptive generalized regression neural network based on dynamic kernel width optimization; Based on displacement estimation and fault detection, high-precision displacement data is obtained by selecting three-phase suspension current, air gap flux and eccentric magnetic pull as neural network input features; Based on the adaptive generalized regression neural network, a four-sensor ring redundant configuration is used to achieve multi-sensor fault tolerance. Through online parameter calibration and majority voting displacement reconstruction algorithm, accurate fault tolerance information is obtained. Based on a four-sensor ring redundant configuration, multi-scale wavelet filtering is used for anti-interference processing, and an improved SUREShrink threshold rule is used to achieve signal reconstruction and noise suppression to obtain fault-tolerant and anti-interference capabilities. Among them, the adaptive generalized regression neural network based on dynamic kernel width optimization for displacement estimation and fault detection includes: Data preprocessing and feature input: three-phase suspension current, air gap flux and eccentric magnetic pull are selected as input features. The data are normalized and then input into AGRNN. The AGRNN model is trained by dynamically adjusting the local kernel width parameter and introducing a regularization mechanism; The trained AGRNN model is used to estimate the displacement online and generate the residual by subtracting it from the actual sensor value. Based on the comparison between the residual error and the dynamic threshold, the sensor failure is determined.

2. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 1, characterized in that: The adaptive generalized regression neural network includes: Utilize residual analysis and dynamic threshold mechanism for fault detection, achieving real-time fault monitoring; The adaptive generalized regression neural network structure is designed as a four-layer structure consisting of input layer, pattern layer, summation layer and output layer, in which the summation layer contains two types of neurons: denominator unit neurons and numerator unit neurons.

3. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 2, characterized in that: The specific steps of the dynamic core width optimization include: Introducing the local kernel width parameter σ i , and according to the sample density distribution, it is inversely proportional to the sample density, and its expression is: , where d i Represents the sample point v i Euclidean distance to the nearest neighbor, k d is a hyperparameter; The kernel width parameter is iteratively optimized based on the gradient descent method, and its iterative update expression is: , to achieve adaptive adjustment of parameters; where RMSE is the root mean square error, η is the learning rate, represents the local kernel width parameter σ i At the tth iteration, the current value of the i-th parameter, is the local kernel width parameter σ i The updated parameter value, that is, the new value of the i-th parameter at the t+1th iteration; Introduction The regularization mechanism suppresses overfitting, and the total loss function L is: , to ensure the stability of the model, where λ is the regularization term and n is the sample size; When λ=0.1, the test set RMSE is 1.2×10 −4 mm is at its lowest value, indicating that moderate regularization can effectively suppress overfitting, while RMSE increases significantly when λ=0.0 and λ≥0.5, so the regularization coefficient λ is taken as 0.1; According to the gradient update formula, it is corrected as follows: ; Set the convergence conditions: , where ε is the preset threshold to obtain the convergence of the optimization process.

4. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 1, characterized in that: Online estimation of displacement dx e ,dy e and compared with the value ds measured by the sensor and filtered x_denoised ,ds y_denoised Subtraction generates residual e x , e y ; ； The dynamic threshold is set as: ; in, is the mean of the residuals in the stable suspension stage, is the variance of the residual in the stable suspension stage; The fault judgment rule is set as follows: or , it is determined that the corresponding direction sensor is faulty.

5. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 4, characterized in that: The four-sensor ring redundant configuration includes: By providing an x-axis direction sensor S1 and a y-axis direction sensor S2, the redundant sensor S3 is arranged along the direction with an angle of 0°<α<90° to the positive direction of the x-axis, and the redundant sensor S4 is arranged along the direction with an angle of 90°<β<180° to the positive direction of the x-axis, forming a four-sensor ring layout.

6. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 4, characterized in that: The four-sensor ring redundant configuration further includes: The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds3 of the three displacement sensors S1, S2, and S3 are then synchronously recorded, and the objective function J(α) is constructed based on the least squares method: ; The optimal solution of the actual installation angle α is iteratively solved through the gradient descent optimization algorithm; the calibrated α value is updated to the control system parameter table to complete the dynamic correction; The calibration process for β is similar to that for α. The rotor is controlled to generate a sinusoidal oscillation with an amplitude of ±10 μm near its center as a micro-amplitude excitation signal. The real-time measurement values ​​ds1, ds2, and ds4 of the three displacement sensors S1, S2, and S4 are then synchronously recorded. The objective function J(β) is constructed using the least squares method: ; The optimal solution of the actual installation angle β is iteratively solved through the gradient descent optimization algorithm; the calibrated β value is updated to the control system parameter table to complete the dynamic correction; Through periodic oscillation excitation and data fitting, the sensor installation deviation is effectively eliminated and the accuracy of coordinate transformation is ensured.

7. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 6, characterized in that: Within the linear measurement range of the displacement sensor, the change in the measured displacement value is linearly related to the change in the output voltage k: Where Δu and Δds represent the absolute value of the change in the displacement sensor output voltage and the absolute value of the change in the rotor displacement, respectively, and the data is unified by making the gains of all displacement sensors the same.

8. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 1, characterized in that: Analysis of various sensor failures includes: When sensor S1, which is used as the x-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the x-axis displacement sensor, and the x-axis reconstructed displacement is obtained as: ; When sensor S2, which is used as the y-axis direction sensor, fails, sensors S3 and S4 are used to reconstruct the y-axis displacement sensor, and the y-axis reconstructed displacement is obtained as: ; When sensors S1 and S2, which are used as sensors for the x-axis and y-axis directions, fail at the same time, sensors S3 and S4 are used to reconstruct the x- and y-axis displacements. The reconstructed x- and y-axis displacements are obtained as follows: ； The designed ring-based redundant sensors and majority voting mechanism support displacement reconstruction under dual sensor failures to achieve fault-tolerant operation under dual sensor failures.

9. The method for fault-tolerant detection and control of a bearingless composite rotor cage asynchronous motor according to claim 1, characterized in that: The multi-sensor fault-tolerance strategy for the ring redundant layout includes: Multi-sensor fault-tolerant design and multi-scale anti-interference filtering; The multi-scale anti-interference filtering includes: performing three-layer wavelet decomposition on the original displacement signal, selecting the db4 mother wavelet to obtain high-frequency detail coefficients D1, D2, D3 and low-frequency approximation coefficient A3; achieving high-precision displacement extraction under interference through layered signal processing; ds signal decomposition expression: ; An improved SUREShrink threshold rule is used for high-frequency detail coefficients to suppress high-frequency noise: Where, D j is the coefficient of the original signal at the jth position, is the coefficient after denoising, λ th is the threshold adjustment coefficient, sign(D j ) means taking D j The sign of , which means positive is 1, negative is -1, and zero is 0; Indicates threshold shrinkage of the absolute value. When the absolute value is less than the threshold, it is set to zero, otherwise the threshold is subtracted. For the derived threshold adjustment item, the threshold is adaptively adjusted according to the data scale; Reconstruct the denoised signal ds using the thresholded coefficients _denoised : ;in, is the approximate coefficient of the third level decomposition.

Images