An active control method for squeeze film damper based on RBF neural network PID optimization

Abstract

The present invention discloses an active control method for a squeeze film damper based on RBF neural network PID optimization. The method comprises the following steps: obtaining rotor system state parameters through a data acquisition module; inputting the state parameters into an RBF neural network, and utilizing the RBF neural network to perform online optimization of the PID control parameters; and, based on the optimized PID parameters, driving a piezoelectric ceramic actuator via a piezoelectric ceramic controller to adjust the damping force of the squeeze film damper, thereby achieving active control of rotor system vibration. By integrating the nonlinear approximation of an RBF neural network with the stability of PID control, the method addresses the hysteresis and locking issues of traditional passive control, achieves active suppression of rotor system vibration, and significantly improves control accuracy and reliability.

Description

Technical Field

[0001] The present invention belongs to the technical field of vibration control of rotor systems, and in particular relates to an active control method for a squeeze film damper based on RBF neural network PID optimization. Background Art

[0002] Traditional squeeze film dampers (SFDs) employ passive control, suppressing vibrations through the damping effect of oil film squeezing. However, they suffer from inherent drawbacks: Passive SFDs maintain a fixed oil film gap or rely on mechanical adjustment, failing to respond in real time to changes in rotor dynamic loads. This results in a time lag in vibration suppression and makes it difficult to cope with transient vibrations caused by sudden imbalances or base excitations. Existing active control schemes often use linear PID controllers to adjust oil film parameters, but PID parameters rely on precise mathematical models and cannot be adjusted adaptively. This can lead to overshoot and insufficient response speed in nonlinear, time-varying systems. While existing active vibration suppression technologies for aircraft engine rotor systems (electromagnetic bearings and magnetorheological dampers) can improve control accuracy, they face challenges such as high power consumption, complex structures, and reliability. Piezoelectric ceramic actuators offer nanometer-level displacement accuracy and fast response, but their application in dynamic oil film gap adjustment has not yet been deeply integrated with adaptive control algorithms, limiting their potential for active vibration control. Summary of the Invention

[0003] To address these technical issues, the present invention provides an active control method for a squeeze film damper based on RBF neural network PID optimization. By integrating intelligent algorithms with piezoelectric ceramic precision drive technology, this method achieves real-time dynamic adjustment of the oil film gap, effectively suppressing the vibration amplitude of the rotor system under variable speed and load conditions.

[0004] The technical solution of the present invention to solve the above technical problems is:

[0005] An active control method for a squeeze film damper based on RBF neural network PID optimization includes the following steps:

[0006] The data acquisition module is used to obtain the state parameters of the rotor system, which includes a turntable, a rotating shaft, a bearing, a bearing seat, and a squeeze film damper. The turntable is symmetrically installed in the middle of the rotating shaft to simulate the load of the rotating equipment. The two ends of the rotating shaft are supported in the bearing seat by bearings, and the squeeze film damper is installed between the bearing seat and the bearing.

[0007] The state parameters are input into an RBF neural network, and the RBF neural network is used to perform online optimization of the PID control parameters. Specifically, a nonlinear mapping model between oil film thickness and vibration amplitude is established, and the proportional, integral, and differential parameters of the PID controller are optimized online using a gradient descent method. When the vibration amplitude exceeds a preset threshold, a displacement control voltage is generated, and the oil film thickness is adjusted through closed-loop feedback to implement active vibration suppression.

[0008] Based on the optimized PID parameters, the piezoelectric ceramic actuator is driven by the piezoelectric ceramic controller to adjust the damping force of the squeeze film damper to achieve active control of the rotor system vibration.

[0009] Preferably, the state parameters include vibration displacement, vibration velocity and vibration acceleration of the rotor.

[0010] Preferably, the steps of optimizing PID parameters using the RBF neural network are as follows:

[0011] Establish an RBF neural network model, taking the current vibration error value, error change rate and error acceleration of the rotor system as input;

[0012] The RBF neural network is used to adjust the network weights and output the optimized proportional parameters, integral parameters and differential parameters of the PID controller.

[0013] Preferably, the piezoelectric ceramic actuator adjusts the oil film thickness of the squeeze film damper in real time according to the optimized PID parameters, thereby adjusting the damping force.

[0014] Preferably, the method further includes an offline training step for the RBF neural network: using historical vibration data to train the RBF neural network, determining initial parameters of the network, and improving the efficiency of online optimization.

[0015] Preferably, the piezoelectric ceramic controller includes a 0-150V high-voltage drive circuit and a displacement feedback module.

[0016] Preferably, the number of hidden layer nodes of the RBF neural network is dynamically adjusted according to the number of rotor modes. The specific adjustment rules are as follows: a) the initial number of hidden layer nodes N0 is set to twice the number of dominant modes of the current rotor system; b) the mean square error E between the neural network output and the expected value is calculated in real time. If E continuously exceeds the threshold E th If the control period reaches 3, then add 2 hidden layer nodes; c) If E is continuously lower than 0.5E th If the number of hidden layer nodes reaches 10 control cycles, one hidden layer node is reduced; d) The node number adjustment range is limited to N0±30%; The output layer contains the PID parameter increment △K p , △K i , △Kd and oil film displacement Δh.

[0017] Preferably, the learning rate η of the gradient descent method is set to 0.01-0.05, the momentum factor α is set to 0.85-0.95, and the parameter update period is 10-50ms.

[0018] Preferably, the threshold is dynamically set according to the rotor critical speed, specifically: , where the benchmark value is 20-50μm.

[0019] Compared with the prior art, the present invention has the following technical effects:

[0020] 1) The present invention adopts RBF neural network to construct a nonlinear mapping model of oil film thickness and vibration, and optimizes PID parameters online in real time through gradient descent method, which solves the limitation of traditional PID relying on linear model and can achieve the purpose of controlling rotor system vibration faster, more stably and more accurately.

[0021] 2) This invention uses an integrated piezoelectric ceramic actuator to achieve nanometer-scale displacement and a 0.1-1ms closed-loop response. Compared to traditional mechanical adjustment, the displacement resolution is improved by two orders of magnitude, solving the hysteresis problem of passive SFD.

[0022] 3) The vibration thresholds for starting and shutting down the piezoelectric ceramic actuator in the present invention can be set according to the specific requirements of the rotor operation. After setting, the system can be actively controlled based on this value without human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the active control method of the rotor system using a squeeze film damper according to the present invention.

[0024] Figure 2 This is a diagram of the oil film clearance adjustment mechanism of the present invention.

[0025] Figure 3 This is the flow chart of the RBF-PID control algorithm in the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] like Figure 1 、 2 As shown in , 3, the present invention provides an active control method for a squeeze film damper based on RBF neural network PID optimization, comprising the following steps:

[0028] Step S1: acquiring state parameters of a rotor system through a data acquisition module, wherein the rotor system includes a turntable, a rotating shaft, a bearing, a bearing seat, and a squeeze film damper.

[0029] First, hardware initialization and parameter setting are performed, and the eddy current displacement sensor and piezoelectric ceramic actuator are calibrated to ensure measurement accuracy of ±0.1μm and the actuator is reset to zero. The rotor system comprises a turntable, a rotating shaft, bearings, a bearing seat, and a squeeze film damper. The rotor system consists of a turntable 1, a rotating shaft 2, bearings 3, a bearing seat 4, and a squeeze film damper 12. The turntable 1 is symmetrically mounted in the middle of the rotating shaft 2 to simulate the load of rotating equipment. The rotating shaft 2 is supported by bearings at both ends in the bearing seat 4. The squeeze film damper 12 is installed between the bearing seat 4 and the bearing 3 to provide dynamic damping.

[0030] The piezoelectric ceramic actuator 5 is integrated into the oil film gap adjustment mechanism of the squeeze film damper. Its actuator end connects to the pad 11 through a connecting boss 10, driving the pad 11 to achieve nanometer-scale adjustment of the oil film thickness. The piezoelectric ceramic actuators 5 are evenly distributed along the X and Y axes of the damper housing to ensure uniform dynamic adjustment of the oil film gap. The piezoelectric ceramic controller 6 includes a 0-150V high-voltage drive circuit and a displacement feedback module. Its output voltage resolution is ≤5nm and its control period is 0.1-1ms, meeting the requirements of fast response.

[0031] The data acquisition module 7 includes an eddy current displacement sensor 9. This sensor monitors the radial vibration amplitude of the turntable 1 in real time, with a sampling frequency of ≥10 kHz and a measurement accuracy of ±0.1 μm. The collected vibration signal is transmitted to the RBF neural network 8 via a data acquisition board.

[0032] In step S2, the state parameters are input into an RBF neural network, and the RBF neural network is used to perform online optimization of the PID control parameters. The specific steps are as follows: a nonlinear mapping model between oil film thickness and vibration amplitude is established, and the proportional, integral, and differential parameters of the PID controller are optimized online using a gradient descent method. When the vibration amplitude exceeds a preset threshold, a displacement control voltage is generated, and the oil film thickness is adjusted through closed-loop feedback to implement active vibration suppression.

[0033] The offline training steps of RBF neural network are as follows:

[0034] Historical vibration data collection and preprocessing: Actual rotor system operating data is used to collect vibration signals covering different operating conditions, including vibration displacement, vibration velocity, vibration acceleration, and corresponding oil film thickness values. A low-pass filter (cutoff frequency = rotor's highest modal frequency × 1.2) is used to filter out high-frequency noise. Input data (vibration error, error change rate, error acceleration) are scaled to the [-1, 1] range, and output data (PID parameters, oil film displacement command) are scaled to the [0, 1] range.

[0035] RBF neural network initialization: K-means clustering algorithm is used to cluster the training data. The number of cluster centers = the number of initial hidden layer nodes N0 (N0 = 2 × the number of dominant modes). The cluster center is used as the center vector of the Gaussian radial basis function. width We take 1.5 times the distance from the jth class sample to the nearest cluster center to ensure that the function effectively covers the input space. We use the Xavier method to avoid gradient vanishing / explosion.

[0036] Training algorithm and optimization target: The loss function uses mean square error (MSE) as the optimization target ,in is the ideal PID parameter and oil film displacement, is the network output. Using the Levenberg-Marquardt algorithm, which converges faster than gradient descent, the validation set error decreases by <0.1% for five consecutive iterations or the maximum number of iterations is reached.

[0037] Parameter preservation: Save the optimal network parameters (center vector ,width , output layer weights), as the initial values ​​for online optimization.

[0038] The construction method of the nonlinear mapping model between oil film thickness and vibration amplitude is as follows:

[0039] Input data definition and preprocessing: The input layer receives the current vibration error value, error change rate and error acceleration, where the vibration error , is the instantaneous deviation between the target and actual vibration amplitude;

[0040] Error change rate Reflects the changing trend of vibration amplitude; error acceleration The second-order differential characterizes the dynamic characteristics of vibration. Data preprocessing: A low-pass filter (with a cutoff frequency of 1.2 times the rotor's highest modal frequency) is used to eliminate high-frequency noise. All input parameters are scaled to the interval [-1, 1] to prevent dimensional differences from affecting model convergence. Time series characteristics of the error, error derivative, and error acceleration are extracted using a 10ms window period.

[0041] Structure of RBF neural network: The input layer of RBF neural network receives the current vibration error value, error change rate and error acceleration. The number of nodes in the hidden layer is dynamically adjusted according to the number of dominant modes of the rotor. The initial number of nodes is twice the number of modes. The increase or decrease of nodes is determined by the real-time mean square error. If the mean square error E exceeds the threshold E for three consecutive control cycles, the node number is adjusted. th , add 2 nodes.

[0042] Adding one node at a time may result in insufficient improvement in model capacity and an inability to quickly respond to sudden vibration modes.

[0043] Adding two nodes can strike a balance between computational efficiency (avoiding frequent adjustments) and model convergence speed.

[0044] If E continues to be lower than 0.5E th Up to 10 control cycles, reducing 1 node.

[0045] Node redundancy can lead to overfitting, but removing too many nodes at once may destroy the learned stable modal mapping.

[0046] Gradual reduction (one at a time) ensures a smooth simplification of the model structure, while avoiding accidental deletions through 10 cycles of stability verification (E is continuously below 0.5Eth).

[0047] The total number of nodes is limited to N0 ± 30% to prevent overfitting. The lower limit (N0 - 30%) ensures coverage of basic modalities. The upper limit (N0 + 30%) prevents overfitting and real-time calculation timeouts.

[0048] The hidden layer function uses Gaussian radial basis function , where x is the three-dimensional column vector consisting of the current vibration error value, error change rate and error acceleration, is the center vector of the jth hidden layer node, and its initial value is determined by historical data. is the width parameter, which is dynamically adjusted according to the distance between adjacent nodes to ensure that the function covers the effective area of ​​the input space. The output layer generates the PID parameter increment and the oil film instruction Δh.

[0049] Optimize parameters using gradient descent:

[0050] The loss function is the mean square error As the optimization goal, is the target vibration amplitude, is the actual vibration amplitude.

[0051] Backpropagation calculates the gradient of the loss with respect to the PID parameters using the chain rule: ;in is the learning rate, which is 0.01–0.05, and the momentum factor The value is 0.85–0.95 to accelerate convergence.

[0052] Parameter update: , similarly update and , and limit the parameter variation range to prevent oscillation.

[0053] The dynamic threshold is updated according to the real-time speed. The threshold calculation formula is: , where the baseline value is set to 20–50 μm, representing the maximum allowable vibration amplitude of the rotor in static equilibrium. The threshold value is dynamically adjusted with the speed. If the vibration amplitude exceeds the upper threshold, closed-loop control is initiated and parameters are continuously optimized. If the vibration amplitude falls below the lower threshold, optimization is paused and the current oil film clearance is maintained. The vibration suppression effect is verified by varying the damping parameter C, and the model parameters are iteratively optimized.

[0054] The input layer of the RBF neural network 8 receives the current error value, the rate of error change, and the acceleration of the error change. The number of nodes in the hidden layer is dynamically adjusted based on the number of rotor system modes. The output layer generates PID parameter increments and oil film displacement commands. The RBF neural network 8 uses a gradient descent method to optimize the PID parameters online, with a learning rate η set to 0.01-0.05, a momentum factor α to 0.85-0.95, and a parameter update period of 10-50ms. When the vibration amplitude exceeds a preset threshold, the controller generates a control voltage signal, which is used by the piezoelectric ceramic actuator 5 to adjust the oil film gap, forming a closed-loop feedback control.

[0055] Closed-loop feedback mechanism process:

[0056] Real-time measurement of oil film thickness:

[0057] An integrated capacitive displacement sensor is embedded in the contact surface between the squeeze film damper pad 11 and the boss 10 (see Figure 2 ), directly measure the real-time changes of the oil film gap h.

[0058] Feedback signal processing:

[0059] The sensor signal is converted into digital quantity by signal conditioning and transmitted to the controller with a sampling frequency of ≥10kHz.

[0060] Real-time matching of damping force:

[0061] The damping parameter C is calculated in real time based on the feedback h (formula: C = C0·(h0 / h)²), ensuring that the damping force matches the vibration suppression requirements of the rotor system in real time.

[0062] The preset vibration threshold is set dynamically according to the rotor critical speed, and the calculation formula is:

[0063] Threshold = Reference Value + 0.2 × (Current Speed ​​ / Critical Speed)², where the reference value of 20-50 μm represents the maximum allowable vibration amplitude of the rotor in static equilibrium. When the vibration amplitude exceeds the threshold, RBF neural network 8 initiates closed-loop control, suppressing vibration by adjusting the oil film thickness. When the vibration amplitude falls below the lower limit of the vibration threshold, the actuator stops and the oil film clearance remains at the current value.

[0064] The adjustment of oil film thickness directly affects the damping parameters of the squeeze film damper, and the relationship is: C = C0 (h0 / h) 2 , where h0 is the initial oil film thickness, C0 is the initial damping parameter, and h0 is the initial oil film thickness. By dynamically adjusting the oil film gap, the system can match the vibration suppression requirements of the rotor system in real time.

[0065] In step S3, based on the optimized PID parameters, the piezoelectric ceramic controller drives the piezoelectric ceramic actuator to adjust the damping force of the squeeze film damper to achieve active control of the vibration of the rotor system.

[0066] Specifically, based on optimized PID parameters, the piezoelectric ceramic controller executes the displacement command. The displacement feedback module verifies accuracy, and changes in the oil film gap are fed back to the damping parameters in real time, accelerating vibration attenuation. The system continuously monitors the vibration amplitude at a 0.1-1ms cycle. If the vibration returns to below the lower threshold and remains stable for more than 10 cycles, optimization stops and the current oil film gap is maintained. This closed-loop process is cyclically executed to ensure stable vibration amplitude control of the rotor system under variable speed and load conditions, ultimately achieving active vibration suppression with millisecond-level response and nanometer-level precision.

Claims

1. An active control method for a squeeze film damper based on RBF neural network PID optimization, characterized in that: The following steps are involved: The data acquisition module is used to obtain the state parameters of the rotor system, which includes a turntable, a rotating shaft, a bearing, a bearing seat, and a squeeze film damper. The turntable is symmetrically installed in the middle of the rotating shaft to simulate the load of the rotating equipment. The two ends of the rotating shaft are supported in the bearing seat by bearings, and the squeeze film damper is installed between the bearing seat and the bearing. The state parameters are input into an RBF neural network, and the RBF neural network is used to perform online optimization of the PID control parameters. Specifically, a nonlinear mapping model between oil film thickness and vibration amplitude is established, and the proportional, integral, and differential parameters of the PID controller are optimized online using a gradient descent method. When the vibration amplitude exceeds a preset threshold, a displacement control voltage is generated, and the oil film thickness is adjusted through closed-loop feedback to implement active vibration suppression. The number of hidden layer nodes of the RBF neural network is dynamically adjusted according to the number of rotor modes. The specific adjustment rules are as follows: a) The initial number of hidden layer nodes N0 is set to twice the number of dominant modes of the current rotor system; b) The mean square error E between the neural network output and the expected value is calculated in real time. If E exceeds the threshold Eth for three control cycles, two hidden layer nodes are added; c) If E is continuously lower than 0.5Eth for 10 control cycles, one hidden layer node is reduced; d) The node number adjustment range is limited to N0±30%; the output layer contains the PID parameter increment △K p , △K i , △K d and oil film displacement Δh; Based on the optimized PID parameters, the piezoelectric ceramic actuator is driven by the piezoelectric ceramic controller to adjust the damping force of the squeeze film damper to achieve active control of the rotor system vibration.

2. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The state parameters include vibration displacement, vibration velocity and vibration acceleration of the rotor.

3. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The steps of optimizing PID parameters using the RBF neural network are as follows: Establish an RBF neural network model, taking the current vibration error value, error change rate and error acceleration of the rotor system as input; The RBF neural network is used to adjust the network weights and output the optimized proportional parameters, integral parameters and differential parameters of the PID controller.

4. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The piezoelectric ceramic actuator adjusts the oil film thickness of the squeeze film damper in real time according to the optimized PID parameters, thereby adjusting the damping force.

5. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: It also includes the offline training steps of RBF neural network: using historical vibration data to train RBF neural network and determine the initial parameters of the network.

6. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The piezoelectric ceramic controller includes a 0-150V high-voltage drive circuit and a displacement feedback module.

7. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The learning rate η of the gradient descent method is set to 0.01-0.05, the momentum factor α is set to 0.85-0.95, and the parameter update period is 10-50ms.

8. The active control method for a squeeze film damper based on RBF neural network PID optimization according to claim 1, characterized in that: The threshold is set dynamically according to the critical speed of the rotor, specifically: threshold = reference value + 0.2×( )², where the benchmark value is 20-50μm.

Images