A noise feedback-based cooling fan speed smoothing control method

Abstract

The application relates to the technical field of automatic control, and particularly discloses a heat dissipation fan rotating speed smooth control method based on noise feedback, which comprises the following steps: based on the transient temperature response of a thermocouple array, extracting a thermal wave dispersion relationship and predicting the spatiotemporal evolution of vortex structure, outputting a heat flux density field and a vortex structure time sequence marker; in response to the time sequence marker, collecting a pulsating vortex field and aerodynamic noise, establishing a sound-vortex coupling transfer function, identifying a matching window, and generating a modulation spectrum instruction; acquiring motor multi-physical field data, inputting the heat flux density field, a steady-state reference value and matching window parameters into an electromagnetic-thermal-mechanical multi-frequency response model, solving an optimal broadband rotating speed track and driving a fan; constructing a reward signal according to actual operation deviation, optimizing a strategy through deep reinforcement learning, and reversely injecting each model to realize self-evolution. Through four-order progression of thermodynamic reconstruction-sound-vortex modulation-multi-field optimization-self-evolution, the application realizes predictive smooth control of the heat dissipation fan.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and specifically to a method for smooth control of cooling fan speed based on noise feedback. Background Technology

[0002] With the rapid development of artificial intelligence, high-performance computing, and mobile terminal devices, the heat flux density of electronic chips is increasing exponentially. As a core component of active cooling systems, the control strategy of the cooling fan directly determines the device's heat dissipation efficiency, power consumption, and user acoustic experience. Currently, mainstream cooling fan control technologies are mainly divided into two categories: The first type is closed-loop control technology based on temperature feedback. This technology collects real-time temperature data using temperature sensors placed on or near the heating element, and combines this data with a PID control algorithm or a preset temperature-speed curve to generate a PWM duty cycle command to adjust the fan speed. This type of technology is mature and reliable, but it is essentially a lag control—the control system only begins to respond after heat is generated and conducted to the temperature sensor. This lag leads to two problems: first, temperature overshoot and sudden speed changes coexist; when a sudden high load occurs, the fan needs to rapidly increase from low to high speed, producing noticeable whistling noise and mechanical shock; second, it cannot perceive the microscopic dynamics of convection heat transfer within the thermal boundary layer, resulting in coarse speed regulation and difficulty in achieving refined thermal management.

[0003] The second type is resonance frequency avoidance technology based on NVH testing. Before the fan leaves the factory, frequency sweep testing identifies the speed points that cause structural resonance or aerodynamic noise peaks, stores them in a blacklist or applies notch filtering, and the control algorithm skips these dangerous speed ranges during operation. Although this type of technology improves acoustic quality to some extent, it has inherent defects: First, the acoustic characteristics of fans are time-varying. After long-term operation, bearing wear, dust accumulation leading to dynamic imbalance deterioration, and lubricant aging can cause the factory-preset quiet speed point to drift. The originally quiet speed gradually produces abnormal noise, and existing technology cannot detect and adapt to this aging drift. Second, this method only passively avoids resonance points and fails to reveal the intrinsic coupling relationship between the vortex field in the fan wake region and aerodynamic noise at the mechanistic level. Therefore, it cannot suppress noise generation at its source.

[0004] In recent years, a few studies have begun to explore the introduction of acoustic sensors into fan control, using noise sound pressure level monitoring as feedback for auxiliary adjustment. However, these improvements remain at the simple feedback level of a single sound pressure level threshold, and have four fundamental limitations: First, they only focus on steady-state sound pressure levels, completely ignoring the destructive impact of millisecond-level transient noise pulses on user experience; second, they cannot identify the microscopic precursors of abnormal noises, thus only passively compensating after the noise occurs, rather than preventing it in advance; third, the control objective is singular, unable to simultaneously consider multiple conflicting physical field objectives such as heat dissipation efficiency, acoustic experience, electromagnetic thermal stability, and mechanical reliability; and fourth, the control strategy is statically preset and calibrated offline, fixed once shipped from the factory, unable to continuously optimize as equipment ages or operating conditions change, lacking self-evolution capabilities.

[0005] In summary, existing technologies are limited by the mindset of delayed control, passive avoidance, single-objective optimization, and static preset, and have not yet formed an intelligent cooling fan control method that can simultaneously achieve advanced thermal field prediction, noise source suppression, multi-field collaborative optimization, and self-evolution of control strategies. Summary of the Invention

[0006] The purpose of this invention is to provide a method for smooth control of cooling fan speed based on noise feedback, so as to solve the problems in existing cooling fan control technology, such as sudden speed changes caused by thermal feedback lag, inability to suppress aerodynamic noise at the source due to ignoring the acoustic-vortex coupling mechanism, and inability to adapt to aging drift and achieve multi-field collaborative optimization due to static preset control strategy.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for smoothing the speed control of a cooling fan based on noise feedback includes the following steps: S1. Obtain the transient temperature sequence of the heat sink, reconstruct the non-equilibrium entropy yield field, predict the evolution trajectory of the thermal boundary layer, and output the expected heat flux density vector field and vortex structure time sequence markers; S2. In response to the vortex structure timing mark, acquire the wake vorticity field and far-field noise pressure, establish the acoustic vortex transfer function, identify the matching window, and generate a rotational speed modulation spectrum command based on the matching window and the expected heat flux density vector field. S3. Obtain the multi-physics field of the motor, input the expected heat flux density vector field and speed modulation spectrum command into the multi-frequency response model, solve the optimal broadband speed trajectory, and generate the drive signal; S4. Collect actual operational deviations to form reward signals, optimize strategies through reinforcement learning, and update the model in reverse to achieve self-evolution.

[0008] As a preferred embodiment of the present invention, S1 specifically includes: S11. Collect transient temperature response sequences at each measuring point, divide the heat sink into multiple local control volumes, calculate the entropy yield density distribution of each local control volume, solve the Onsager coupling transport coefficient matrix, and extract the effective thermal diffusion tensor and thermal wave dispersion relation under the action of thermal-mass-fluid coupling. S12. Based on the thermal wave dispersion relation, calculate the functional relationship between the thermal disturbance phase velocity and the frequency, solve the generation frequency of the flow vortex and spanwise vortex induced by the micro-groove through the vortex transport equation, and predict the spatiotemporal evolution trajectory of the vortex structure in the thermal boundary layer from generation, development to breakup. S13. Input the spatiotemporal evolution trajectory into the electromechanical coupling response model, and combine the rotational inertia of the fan rotor at the current speed with the electromagnetic torque response time to calculate the advance compensation time required from issuing the control command to the vortex structure reaching the corresponding position on the heat sink surface; at the advance compensation time, output the expected heat flux density vector field distribution of each spatial point on the heat sink surface, as well as the time sequence mark of each vortex structure reaching the heat sink surface.

[0009] As a preferred embodiment of the present invention, S11 specifically includes: S111. Based on a micro thermocouple array embedded in the heat sink substrate with a non-uniform topological distribution, the transient temperature response sequence of each measurement point is synchronously acquired at a sampling frequency of not less than 1kHz, and the transient temperature response sequence is subjected to wavelet denoising processing to eliminate the high-frequency measurement noise introduced by the thermocouple contact thermal resistance. S112. Based on the geometric structure and material thermophysical parameters of the heat sink substrate, the heat sink is divided into multiple non-overlapping local control volumes. The positions of each thermocouple measuring point are used as control volume nodes. The heat flux density and temperature gradient of each control volume boundary are discretized and solved by the finite volume method. The entropy production rate distribution of each local control volume is calculated according to the entropy production rate calculation formula in irreversible thermodynamics. S113. Taking the minimum entropy production density of each local control volume as the objective function and the linear phenomenological relationship between heat flux density and temperature gradient and pressure gradient as the constraint, the Onsager coupled transport coefficient matrix is ​​solved by variational method to invert the non-equilibrium thermodynamic force-fluid relationship and extract the effective heat diffusion tensor under the action of thermo-mass-fluid coupling. S114. Based on the effective thermal diffusion tensor, and combined with the coupled boundary conditions of the energy equation and the momentum equation, the characteristic equation of thermal wave propagation is solved by dispersion analysis, and the thermal wave dispersion relation of the thermal disturbance phase velocity with frequency is extracted as the input parameter for predicting the thermal disturbance propagation characteristics in subsequent steps.

[0010] As a preferred embodiment of the present invention, S12 specifically includes: S121. Based on the thermal wave dispersion relation, solve the functional relationship between the thermal disturbance phase velocity and the angular frequency, identify the critical frequency point where dispersion characteristics appear during the thermal wave propagation process, and divide the low-frequency thermal diffusion dominant region and the high-frequency thermal wave propagation dominant region. S122. Obtain the geometric parameters of the microgroove structure on the heat sink surface, including groove depth, groove spacing, groove inclination angle and groove arrangement density. Combined with the mainstream surface velocity formed by fan blowing, calculate the flow vortex generation frequency and spanwise vortex generation frequency induced by the microgroove structure by solving the vortex transport equation. S123. Using the flow vortex generation frequency and spanwise vortex generation frequency as vorticity disturbance source terms, and taking the critical frequency point as the boundary, establish the coupled wave equation of thermal disturbance and vorticity disturbance in the high-frequency thermal wave propagation dominance region, and solve the cross-spectral density function of vorticity fluctuation and temperature fluctuation in the thermal boundary layer. S124. Based on the spatiotemporal distribution characteristics of the cross-spectral density function and combined with the boundary layer stability theory, predict the complete life cycle of the vortex structure from initial generation, linear growth, nonlinear saturation to final breakage, and generate the spatiotemporal evolution trajectory of the vortex structure within the thermal boundary layer. The spatiotemporal evolution trajectory includes the position coordinates, vorticity intensity, characteristic scale and corresponding timestamp of the vortex structure.

[0011] As a preferred embodiment of the present invention, S2 specifically includes: S21. In response to the advance compensation moment, collect the pulsating vorticity field data and far-field aerodynamic noise pressure data in the downstream wake region of the fan, perform intrinsic orthogonal decomposition on the pulsating vorticity field data, extract the spatial basis functions and time coefficients, and establish the acoustic-vortex coupling transfer function using the system identification method. S22. Perform phase-locking analysis on the time sequence markers output by S1 and the time coefficients extracted by S21, and identify the matching window between the vortex structure shedding phase and the sound pressure pulsation dissipation valley value through cross-correlation function or phase synchronization algorithm. S23. Generate a rotational speed modulation spectrum instruction based on the matching window, the rotational speed modulation spectrum instruction including applying a small rotational speed fluctuation during the matching window to enhance acoustic vortex energy dissipation, and calculating a steady-state rotational speed reference value outside the matching window based on the expected heat flux density vector field distribution.

[0012] As a preferred embodiment of the present invention, S21 specifically includes: S211. In response to the advance compensation moment, a particle image velocimeter or hot-wire anemometer arranged in the downstream wake region of the fan is used to collect pulsating velocity field data at a sampling frequency of not less than 5kHz, and pulsating vorticity field data is obtained through the vorticity calculation formula; at the same time, far-field aerodynamic noise pressure data is collected synchronously through a microphone array arranged in the far field. S212. Perform eigenorthogonal decomposition on the pulsating vorticity field data, construct a spatial two-point correlation matrix, and extract the top N modes with the largest eigenvalues ​​as the dominant flow modes by solving the eigenvalue problem of the correlation matrix, thereby obtaining the corresponding spatial basis functions and time coefficients that reflect the evolution of mode energy over time. S213. Using the time coefficient as the input signal and the far-field aerodynamic noise pressure data as the output signal, a subspace-based system identification method is adopted to establish the transfer function relationship between the two in the frequency domain, and obtain the amplitude frequency characteristic curve and phase frequency characteristic curve of the acoustic-vortex coupling transfer function. S214. The acoustic vortex coupling transfer function is verified by cross-correlation analysis. The coherence function between the output predicted by the input signal through the transfer function and the actual output is calculated. When the coherence function is greater than a preset threshold in the main frequency band, the validity of the acoustic vortex coupling transfer function is confirmed, and it is used as the basis for identifying the matching window in subsequent steps.

[0013] As a preferred embodiment of the present invention, S22 specifically includes: S221. Take the timing mark of the vortex structure reaching the heat sink surface output by S1 as the reference time axis, take the time coefficient obtained by S212 as the signal to be analyzed, perform Hilbert transformation on the time coefficient, extract its instantaneous phase information, and obtain the curve of the instantaneous phase change with time during the vortex structure shedding process. S222. Perform wavelet packet decomposition on the synchronously acquired far-field aerodynamic noise pressure data, extract the energy envelope in each frequency band, identify the time point when the sound pressure pulsation energy reaches a local minimum as the dissipation valley time, and record the instantaneous phase corresponding to each dissipation valley time. S223. Perform cross-correlation analysis on the instantaneous phase curve of the vortex structure shedding and the instantaneous phase corresponding to the dissipation valley value, calculate the phase difference between the two, and determine the phase offset when the cross-correlation function reaches its maximum value through cyclic shift search as the optimal matching phase between the vortex structure shedding phase and the sound pressure pulsation dissipation valley value. S224. Taking the optimal matching phase as the center, and based on the complete life cycle of the vortex structure from generation to dissipation, set the allowable phase fluctuation range and construct a time window as the matching window. The start and end times of the matching window correspond to specific phase intervals within the vortex structure shedding cycle.

[0014] As a preferred embodiment of the present invention, S3 specifically includes: S31. The transient Joule heat distribution data of the stator winding is obtained by a distributed fiber optic grating sensor embedded in the stator slot of the motor; the eddy current loss data of the permanent magnet is obtained by a thin film thermocouple attached to the rotor surface; and the contact stress fluctuation data of the bearing rolling element is obtained by an acceleration sensor installed on the outer ring of the bearing combined with a dynamic model inversion. S32. Input multiple data into the multi-frequency response model, take the time sequence mark as the reference time axis, and use a multi-objective optimization algorithm to solve the optimal broadband speed trajectory. The broadband speed trajectory includes a fundamental frequency component for meeting heat dissipation requirements, a perturbation component for acoustic vortex dissipation, and a notch component for avoiding structural resonance. S33. Generate a three-phase duty cycle sequence of a space vector pulse width modulation signal based on the optimal wideband speed trajectory, and output the three-phase duty cycle sequence to the motor driver to drive the cooling fan to run according to the optimal wideband speed trajectory.

[0015] As a preferred embodiment of the present invention, S32 specifically includes: S321. Construct an electromagnetic-thermal-mechanical coupled multi-frequency response model, which includes a three-dimensional thermal network sub-model describing the Joule thermal dynamics of the stator winding, an electromagnetic field sub-model describing the eddy current loss of the rotor permanent magnet, and a dynamic sub-model describing the contact stress fluctuation of the bearing rolling elements. Each sub-model achieves bidirectional data exchange through thermal-mechanical coupling boundary conditions. S322. The expected heat flux density vector field distribution output by S1, the steady-state speed reference value calculated by S23, the small speed fluctuation parameters corresponding to the matching window, the transient Joule heat distribution data, eddy current loss data and contact stress fluctuation data obtained by S31 are loaded into the multi-frequency response model as input parameters. S323. Using the timing mark output by S13 as the reference time axis, set an optimization time window, and solve the multi-frequency response model in the time domain within the optimization time window to obtain the heat sink entropy yield fluctuation time history curve, sound power level fluctuation time history curve, stator temperature rise fluctuation time history curve and bearing contact fatigue index time history curve. S324. Taking the minimum weighted integral of the four factors—root mean square of heat sink entropy yield fluctuation, root mean square of sound power level fluctuation, peak value of stator temperature rise fluctuation, and cumulative value of bearing contact fatigue index—as the objective function, a multi-objective particle swarm optimization algorithm is used to search for the optimal solution in a wide bandwidth, and the output is the optimal wideband speed trajectory containing fundamental frequency component, perturbation component and notch component. S325. Perform Fourier analysis on the optimal broadband rotational speed trajectory to verify whether the notch component is sufficiently separated from the system structure resonance frequency. If the separation degree is lower than the preset threshold, adjust the penalty function coefficient in the multi-objective optimization algorithm and re-execute step S324 until the requirements are met.

[0016] As a preferred embodiment of the present invention, S4 specifically includes: S41. Collect the heat sink entropy productivity field evolution trajectory, sound power level temporal fluctuation, stator temperature rise response curve, and bearing vibration characteristic data after the actual operation of the optimal broadband speed trajectory; calculate the deviation between the heat sink entropy productivity field evolution trajectory and the predicted expected heat flux density vector field distribution in S1, the deviation between the sound power level temporal fluctuation and the sound pressure pulsation dissipation valley value identified in S2, and the deviation between the stator temperature rise response curve and the stator temperature rise fluctuation target value in the optimal solution obtained in S3, and use the abnormal impact peak value exceeding the preset threshold in the bearing vibration characteristic data as a penalty factor to jointly constitute a multi-dimensional reward signal; S42. Input the multidimensional reward signal into the deep reinforcement learning network; the deep reinforcement learning network uses the expected heat flux density vector field distribution output by S1, the amplitude-frequency and phase-frequency characteristics of the acoustic-vortex coupling transfer function identified by S2, and the sub-model parameters of the multi-frequency response model solved by S3 as the state space, and the frequency and amplitude of the perturbation component, the center frequency and notch depth of the notch component, and the rate of change of the fundamental frequency component in the speed modulation spectrum command as the action space, and iteratively optimizes the control strategy through the policy gradient method, and outputs the optimized control strategy parameters; S43. The optimized control strategy parameters are back-injected into the thermal wave dispersion relation prediction model of S1, the matching window identification algorithm of S2, and the multi-frequency response model of S3. The boundary conditions of the thermal wave dispersion relation are dynamically updated, the identification threshold of the matching window is corrected, and the coupling weight of each sub-model in the multi-frequency response model is adjusted to realize the online self-evolution of the entire control system.

[0017] Compared with the prior art, the present invention has the following advantages: 1. Breaking through the traditional temperature feedback lag, the thermo-mass flow coupling relationship is inverted through the minimum entropy production rate principle, the thermal wave dispersion relationship is extracted and the vortex structure evolution trajectory is predicted, and the advance compensation duration is generated, upgrading the ex-post compensation to ex-ante prediction, thus avoiding sudden changes in rotational speed from the root.

[0018] 2. By identifying the phase matching window through the acoustic vortex transfer function, a small rotational speed fluctuation is applied to convert the vortex kinetic energy into heat energy dissipation, thereby achieving active noise reduction by vortex silencing; at the same time, a multi-frequency response model with the minimum weighted integral of four objectives is constructed to solve the optimal rotational speed trajectory containing the fundamental frequency, perturbation, and notch components, thereby achieving smooth control of the four fields of thermo-acoustic-electrical-mechanical coordination.

[0019] 3. The actual operating deviation is constructed as a reward signal. The control strategy is iteratively optimized through reinforcement learning. The thermal wave dispersion model, matching window algorithm and multi-frequency response model are injected in reverse to realize the two-way interaction of positive prediction and reverse correction. This enables the control system to continuously evolve with aging and operating conditions, and complete the intelligent leap from static preset to dynamic evolution. Attached Figure Description

[0020] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the method described in Embodiment 1 of the present invention.

[0022] Figure 2 This is a framework diagram of the system described in Embodiment 2 of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments. Example 1

[0025] like Figure 1 As shown, the present invention provides a method for smooth control of cooling fan speed based on noise feedback, comprising the following steps: S1. Obtain the transient temperature sequence of the heat sink, reconstruct the non-equilibrium entropy yield field, predict the evolution trajectory of the thermal boundary layer, and output the expected heat flux density vector field and vortex structure time series markers; specifically including: S11. Entropy yield field reconstruction and extraction of thermo-mass flow coupling parameters, specifically: S111. Transient temperature response sequence acquisition and denoising, specifically: A micro thermocouple array with a non-uniform topological distribution is embedded inside the heat sink substrate. This non-uniform topological distribution is optimized based on the geometric characteristics of the heat sink substrate and the expected heat flux density gradient. Measurement points are densely arranged in areas where the heat flux density changes drastically, and sparsely arranged in areas where the heat flux density is gentle.

[0026] The transient temperature response sequence at each measuring point is synchronously acquired at a sampling frequency of not less than 1 kHz. This sampling frequency ensures that high-frequency thermal disturbance information during the heat wave propagation process can be captured.

[0027] The acquired transient temperature response sequence was subjected to wavelet denoising. The Daubechies wavelet basis function was used to decompose the signal into multiple scales, identify and threshold filter out the high-frequency measurement noise introduced by the thermocouple contact thermal resistance, and retain the characteristic frequency components of heat wave propagation to obtain the denoised transient temperature response sequence. This denoised transient temperature response sequence is used as the basis data for subsequent entropy yield calculation.

[0028] S112. Local control volume partitioning and entropy yield density distribution calculation, specifically: Based on the geometric structure and thermal properties of the heat sink substrate, including thermal conductivity, density, specific heat capacity, and thermal diffusivity, the heat sink is divided into multiple non-overlapping local control volumes. The volumetric dimensions of each local control volume are adaptively adjusted according to the density of thermocouple measuring points and the rate of change of the temperature gradient. Using the location of each thermocouple measuring point as the node center of the local control volume, a heat conduction network topology is established between the control volumes.

[0029] The heat flux density vector and temperature gradient vector of each local control volume boundary are discretized using the finite volume method. A second-order upwind scheme is used during the discretization process to ensure computational accuracy, and a fully implicit time-progression scheme is used to ensure numerical stability.

[0030] According to the entropy production rate calculation formula in irreversible thermodynamics, that is, the local entropy production rate density is equal to the dot product of the heat flux density vector and the temperature gradient vector divided by the negative value of the square of the temperature, the entropy production rate density distribution of each local control volume is calculated. This entropy production rate density distribution characterizes the spatial distribution characteristics of irreversible loss in the heat transfer process inside the heat sink.

[0031] S113. Onsager's solution to the coupled transport coefficient matrix and extraction of the effective thermal diffusivity tensor, specifically: Based on the entropy productivity density distribution of each local control volume calculated by S112, a global entropy productivity minimization objective function is constructed. This objective function is the integral of the product of the entropy productivity density of each local control volume and the volume of the control volume.

[0032] A linear phenomenological relationship between heat flux density and temperature and pressure gradients is established as a constraint condition. This linear phenomenological relationship is described by the Onsager reciprocal relation and includes multiple transport mechanisms such as heat conduction, heat diffusion, heat penetration and thermomechanical coupling.

[0033] The Onsager coupled transport coefficient matrix satisfying the constraints is solved using the variational method. This matrix is ​​a symmetric positive definite matrix, and its elements characterize the coupling strength between different thermodynamic forces and thermodynamic flows. Based on the obtained Onsager coupled transport coefficient matrix, the non-equilibrium thermodynamic force-flow relationship is inverted, and the cross-transport effect under thermo-mass flow coupling is identified.

[0034] The effective thermal diffusion tensor under the action of thermo-mass-fluid coupling is extracted from the Onsager coupling transport coefficient matrix. This effective thermal diffusion tensor is a second-order symmetric tensor. Its principal components represent the equivalent thermal diffusion capacity along different directions, and its cross components represent the anisotropic thermal transport characteristics caused by thermo-mass-fluid coupling.

[0035] S114. Extraction of thermal wave dispersion relation, specifically: Based on the effective thermal diffusion tensor extracted from S113, coupled boundary conditions for the energy and momentum equations are established. These coupled boundary conditions consider convective heat transfer between the heat sink substrate and the cooling fluid, the contact thermal resistance between the heat sink substrate and the heat source, and the volumetric heat source effect within the heat sink substrate. Substituting these coupled boundary conditions into the governing equations for heat wave propagation, the characteristic equation for heat wave propagation is solved through dispersion analysis. This characteristic equation is a transcendental equation concerning the complex wave number.

[0036] Solving the characteristic equation yields the dispersion curve of the thermal disturbance phase velocity as a function of frequency, identifying the normal and anomalous dispersion regions in the heat wave propagation process. The thermal wave dispersion relation of the thermal disturbance phase velocity as a function of frequency is extracted; this relation includes the phase velocity-frequency function and the group velocity-frequency function, serving as input parameters for predicting the thermal disturbance propagation characteristics and delineating the dominant thermal diffusion and heat wave propagation regions in subsequent step S12.

[0037] S12. Spatiotemporal evolution prediction of vortex structures driven by thermal wave dispersion, specifically: S121. Analysis of thermal wave dispersion characteristics and identification of critical frequencies, specifically: The thermal wave dispersion relation output from S114 is received, and the phase velocity-frequency function and group velocity-frequency function contained therein are analyzed to establish a continuous functional relationship between the thermal disturbance phase velocity and angular frequency. The first derivative of this functional relationship is solved to calculate the rate of change of group velocity relative to frequency, and the extreme points and inflection points in the group velocity curve are identified.

[0038] Based on the threshold criterion of group velocity change rate, a critical frequency point is determined where significant dispersion characteristics appear during heat wave propagation. This critical frequency point divides the heat wave dispersion curve into two characteristic regions: in the low-frequency region below the critical frequency point, heat disturbance propagation is dominated by thermal diffusion, and the phase velocity decreases monotonically with increasing frequency, exhibiting normal dispersion characteristics; in the high-frequency region above the critical frequency point, heat disturbance propagation is dominated by thermal wave mechanisms, and the phase velocity increases with frequency rather than changing monotonically, exhibiting anomalous dispersion characteristics. The numerical value of the critical frequency point and the frequency boundaries between the low-frequency thermal diffusion-dominated region and the high-frequency heat wave propagation-dominated region are output as the frequency partitioning basis for subsequent eddy-thermal coupling analysis.

[0039] S122. Calculation of vorticity generation frequency for microgroove structures, specifically: The geometric parameters of the microgroove structure on the heat sink surface are obtained, including groove depth, groove spacing, groove inclination angle, and groove density. These parameters are acquired through 3D optical scanning or digital extraction from design drawings. The mainstream surface velocity formed by fan blowing is measured or estimated; this mainstream surface velocity is a time-averaged velocity reflecting the macroscopic flow intensity of the cooling fluid on the heat sink surface. The ratio of groove depth to groove spacing is used as a dimensionless roughness parameter. Combined with the groove inclination angle and density, an equivalent roughness model of the microgroove structure is established.

[0040] Based on the equivalent roughness model and the mainstream surface velocity, the flow-oriented vortex generation frequency induced by the microgroove structure is calculated by solving the vortex transport equation. This flow-oriented vortex generation frequency is related to the groove spacing and the mainstream surface velocity, reflecting the shedding frequency of vortex trains arranged along the mainstream direction. Simultaneously, the spanwise vortex generation frequency induced by the microgroove structure is calculated. This spanwise vortex generation frequency is related to the groove depth and the local shear velocity, reflecting the vortex roll-up frequency perpendicular to the mainstream direction. The numerical values ​​of the flow-oriented and spanwise vortex generation frequencies, along with the corresponding dimensionless Strouhal number, are output as the perturbation source term input for the subsequent vortex-thermal coupled wave equation.

[0041] S123. Solving the eddy-thermal coupled wave equation and calculating the cross-spectral density, specifically: Using the flow-direction and spanwise vortex generation frequencies calculated in S122 as vortex disturbance source terms, a harmonic excitation model of vortex pulsation is established. Taking the critical frequency point identified in S121 as the boundary, a coupled wave equation for thermal disturbance and vortex disturbance is established within the high-frequency thermal wave propagation dominance region. This coupled wave equation considers the thermal buoyancy effect of temperature pulsation on vortex generation and the enhancement effect of vortex pulsation on heat transport, reflecting the two-way coupling mechanism of heat and mass flow.

[0042] The coupled wave equations are transformed into a set of algebraic equations about frequency and wave vector by performing a time-space-frequency domain transformation. The amplitude-frequency and phase-frequency characteristics of vorticity fluctuations and temperature fluctuations in the thermal boundary layer are solved by an iterative method.

[0043] Based on the obtained amplitude-frequency and phase-frequency characteristics, the cross-spectral density function of vorticity fluctuations and temperature fluctuations is calculated. This cross-spectral density function is a complex-valued function, whose magnitude characterizes the covariance intensity of vortex-thermal disturbances, and whose phase characterizes the time lag relationship between vorticity fluctuations and temperature fluctuations. The distribution matrices of the cross-spectral density function in the normal and spanwise directions of the thermal boundary layer are output as frequency domain feature inputs for subsequent vortex structure lifecycle prediction.

[0044] S124. Prediction of the life cycle of vortex structures and generation of spatiotemporal evolution trajectories, specifically: The cross-spectral density function output from S123 is received, and its peak position and full width at half maximum (FWHM) of the modulus distribution are extracted to identify the coherent structural scale of vortex fluctuations and temperature fluctuations within the thermal boundary layer. Combining the Orr-Sommerfeld equation from boundary layer stability theory, the spatial growth rate of the cross-spectral density function at different frequencies is analyzed to determine the linear stability characteristics of the vortex structure.

[0045] Based on the results of linear stability analysis, a weak nonlinear interaction model is introduced to predict the complete life cycle of the vortex structure from the initial linear growth stage, to the vortex pairing and merging stage in the nonlinear saturation stage, and finally to the small-scale turbulence stage in the final breaking stage. The duration and characteristic scale evolution of each stage are calculated.

[0046] By integrating linear stability analysis, nonlinear evolution models, and dispersion relation corrections, the spatiotemporal evolution trajectory of the vortex structure within the thermal boundary layer is generated and stored in the form of a discrete time series. This spatiotemporal evolution trajectory is output, including the three-dimensional position coordinates of the vortex structure at each time point, the vorticity intensity tensor, the characteristic scale vector, and the corresponding timestamp, serving as the input parameters for the electromechanical coupling advance compensation calculation in step S13.

[0047] S13. Electromechanical coupling advanced compensation and expected heat flow field output, specifically: S131. Receive the spatiotemporal evolution trajectory output by S124, analyze the vortex structure position coordinates, vortex intensity, characteristic scale, and timestamp information contained therein, and establish the spatiotemporal mapping relationship of the vortex structure reaching the heat sink surface. Input this spatiotemporal evolution trajectory into the electromechanical coupled response model, which includes a mechanical inertia sub-model of the fan rotor and an electromagnetic torque response sub-model of the motor. The mechanical inertia sub-model uses the rotor's moment of inertia and current speed as state variables, while the electromagnetic torque response sub-model uses the stator current and rotor magnetic field as state variables.

[0048] S132. Based on the moment of inertia of the fan rotor at the current speed, calculate the mechanical time constant required for the rotor to accelerate from rest to the target speed or adjust from the current speed to the new speed; at the same time, based on the control bandwidth of the motor driver and the electrical time constant of the stator winding, calculate the electromagnetic response time from the issuance of the command to the actual establishment of the electromagnetic torque; superimpose the mechanical time constant and the electromagnetic response time to obtain the total response delay time of the fan system.

[0049] S133. Based on the timestamp information in the spatiotemporal evolution trajectory of the vortex structure, calculate the propagation time required for each vortex structure to develop from its current position and reach the corresponding position on the heat sink surface. Subtract the total response delay time of the fan system from this propagation time to obtain the advance compensation time. If the advance compensation time is positive, it indicates that control commands need to be issued in advance to achieve synchronization between the vortex structure and the control action. If the advance compensation time is negative, it indicates that the propagation speed of the vortex structure is faster than the system response capability, and it is necessary to activate the emergency acceleration mode or adjust the control target.

[0050] S134. At the calculated advance compensation time, based on the characteristic scale and vortex intensity of each vortex structure in the spatiotemporal evolution trajectory, and combined with the convective heat transfer coefficient distribution of the heat sink surface, calculate the expected heat flux density vector field distribution at each spatial point on the heat sink surface. This expected heat flux density vector field distribution contains information on the magnitude and direction of the heat flux density. Simultaneously, output the time sequence marker of each vortex structure arriving at the heat sink surface. This time sequence marker is a relative time series with respect to the advance compensation time. This expected heat flux density vector field distribution and time sequence marker serve as input parameters for step S2, used for the speed modulation spectrum optimization of the acoustic-vortex coupling dissipation mechanism.

[0051] S2. In response to the vortex structure timing marker, acquire the wake vorticity field and far-field noise pressure, establish the acoustic vortex transfer function, identify the matching window, and generate a rotational speed modulation spectrum command based on the matching window and the expected heat flux density vector field; specifically including: S21. Establishment of the acoustic-vortex coupling transfer function, specifically as follows: S211. Synchronous acquisition of pulsating vorticity field and far-field aerodynamic noise pressure, specifically: In response to the lead compensation moment of S13 output, the high-speed data acquisition system is triggered, enabling the particle image velocimeter or hot-wire anemometer located in the downstream wake region of the fan and the microphone array located in the far field to enter synchronous acquisition mode. The particle image velocimeter acquires pulsating velocity field data in the downstream wake region of the fan at an equivalent sampling frequency of not less than 5 kHz, or the hot-wire anemometer acquires data at a real-time sampling frequency of not less than 5 kHz. This sampling frequency ensures the resolution of the high-frequency dynamic characteristics of vortex structure shedding, capturing complete spectral information from low-frequency large-scale structures to high-frequency small-scale turbulence. Simultaneously, the microphone array located in the far field synchronously acquires far-field aerodynamic noise pressure data at the same time reference. This microphone array employs a helical or random geometric arrangement to cover wideband directivity. The line connecting the array center and the fan rotation center is perpendicular to the mainstream direction, and the array radius is determined based on the measurement frequency range and sound wave wavelength.

[0052] The acquired fluctuating velocity field data undergoes quality checks, eliminating outlier data points caused by insufficient particle concentration or thermal filament breakage. Fluctuating vorticity field data is then obtained through antisymmetric component calculation of the velocity gradient tensor. This data includes three-dimensional components: flow vorticity, normal vorticity, and spanwise vorticity. The fluctuating vorticity field data is then aligned with far-field aerodynamic noise pressure data according to timestamps to establish synchronized data pairs. These synchronized data pairs serve as the initial input for subsequent intrinsic orthogonal decomposition and system identification.

[0053] S212. Intrinsic orthogonal decomposition and dominant flow mode extraction, specifically: The system receives fluctuating vorticity field data from the S211 output and constructs a three-dimensional vorticity field snapshot matrix within the analysis time window. Each column of this matrix represents the vorticity field distribution in the entire spatial domain at a given moment. A spatial correlation matrix is ​​calculated, obtained by multiplying the vorticity field snapshot matrix by its transpose. Its elements characterize the statistical correlation between vorticity fluctuations at different spatial locations, reflecting the coherent spatial scale of the vortex structure.

[0054] The correlation matrix is ​​decomposed using eigenvalues, and the eigenvalues ​​are sorted in descending order. The top N modes with the largest eigenvalues ​​are extracted as the dominant flow modes. The value of N is determined based on the cumulative energy percentage threshold, typically the smallest mode with a cumulative energy percentage exceeding 90%. The spatial basis functions corresponding to each dominant flow mode are obtained. These spatial basis functions are orthogonally normalized spatial distribution patterns, representing the spatial morphological characteristics of the vortex structure. Simultaneously, time coefficients reflecting the energy evolution of each mode over time are obtained. These time coefficients are the inner product of the spatial basis functions and the original vorticity field snapshot, representing the transient energy amplitude and phase evolution of the dominant flow modes. The spatial basis functions and corresponding time coefficients of the top N dominant flow modes are output. These time coefficients serve as the input signals for subsequently establishing the acoustic-vortex coupling transfer function.

[0055] S213. Identification of acoustic-vortex coupled transfer function systems, specifically: The system receives the time coefficient output from S212 as the input signal and the far-field aerodynamic noise pressure data synchronously acquired from S211 as the output signal, establishing a multi-input single-output system identification framework. The input dimension is N, representing the number of dominant flow modes, and the output dimension is 1, representing the far-field sound pressure. A subspace-based system identification method is employed, constructing the Hankel matrix of the input and output data. The observable and controllable subspaces of the system are extracted through QR decomposition and singular value decomposition, and the state-space realization of the system is estimated.

[0056] The estimated state-space model is transformed to the frequency domain, and the frequency response functions from each time coefficient to the far-field sound pressure are calculated. Integrating these frequency response functions yields the amplitude-frequency characteristic curve and phase-frequency characteristic curve of the acoustic-vortex coupling transfer function. The amplitude-frequency characteristic curve characterizes the energy transfer efficiency from vortex fluctuations to sound pressure fluctuations at different frequencies, while the phase-frequency characteristic curve characterizes the phase lag relationship at different frequencies, reflecting the causal delay characteristics in the vortex-acoustic conversion process. The stability of the acoustic-vortex coupling transfer function is checked to ensure that all poles lie within the unit circle or the left half-plane, guaranteeing the physical realizability of the transfer function.

[0057] S214. Verification of the acoustic-vortex coupling transfer function, specifically: The acoustic-vortex coupling transfer function established by S213 is received, and this transfer function is used to perform forward prediction on the input signal, i.e., the time coefficient, to calculate the predicted far-field sound pressure output. The predicted output is compared with the actual far-field aerodynamic noise pressure data collected by S211, and the cross-spectral density and their respective autospectral density are calculated. Then, the coherence function is calculated, which characterizes the degree of linear correlation between the predicted output and the actual output in the frequency domain, and its value ranges from 0 to 1.

[0058] A coherence function threshold is set, typically 0.8 or higher, to verify whether the coherence function exceeds this threshold within the vortex shedding frequency and its harmonic frequency bands. If the coherence function exceeds the threshold within the main frequency band, the validity of the acoustic-vortex coupling transfer function is confirmed, indicating that the established transfer function can accurately describe the causal relationship between vorticity fluctuations and sound pressure fluctuations. If the coherence function is below the threshold within the main frequency band, return to S212 to adjust the number of dominant flow modes N or return to S213 to change the system identification algorithm parameters and re-identify the model. The verified valid acoustic-vortex coupling transfer function, including its amplitude-frequency response curve and phase-frequency response curve, serves as the basis for identifying the matching window between the vortex structure shedding phase and the sound pressure fluctuation dissipation valley in the subsequent step S22.

[0059] S22. Identification of the matching window between the phase of vortex structure shedding and the valley value of sound pressure dissipation, specifically: S221. Instantaneous phase extraction of vortex structure shedding, specifically: The timing marker of the vortex structure arriving at the heat sink surface, output from S13, is received and used as a reference time axis to establish an absolute time reference system. This reference time axis covers the entire time interval from the lead compensation moment to the predicted complete dissipation of the vortex structure. The time coefficient obtained from S212, which characterizes the transient energy amplitude of the dominant flow mode, is received and used as the signal to be analyzed. A Hilbert transform is performed on this time coefficient, and an analytical signal is constructed through convolution operations. The real part of this analytical signal is the original time coefficient, and the imaginary part is the orthogonal component after the Hilbert transform.

[0060] Instantaneous amplitude and instantaneous phase are calculated based on the analytic signal. The instantaneous amplitude is the magnitude of the analytic signal, reflecting the transient energy intensity of the dominant flow mode. The instantaneous phase is the argument of the analytic signal, calculated using the arctangent function and expanded to continuous phase, yielding the instantaneous phase variation curve over time during the vortex structure shedding process. This instantaneous phase curve reflects the periodic phase evolution of the dominant flow mode, with the phase change rate corresponding to the instantaneous frequency. The output vortex structure shedding instantaneous phase curve serves as the reference signal for subsequent cross-correlation analysis with the phase of the sound pressure dissipation valley.

[0061] S222. Identification of the valley moment of sound pressure pulsation dissipation, specifically: The system receives far-field aerodynamic noise pressure data synchronously acquired by the S211. Using Morlet or Daubechies wavelets, matching the dominant frequency of vortex shedding, as basis functions, it performs wavelet packet decomposition on the far-field aerodynamic noise pressure data, decomposing the signal into different frequency bands and obtaining wavelet coefficients for each band. The wavelet coefficients in each band are squared and smoothed along the time axis to extract the energy envelope of each band. This energy envelope reflects the evolution trend of sound pressure fluctuation energy over time.

[0062] A local minimum search algorithm is applied to the energy envelope to identify the time point when the sound pressure fluctuation energy reaches a local minimum as the dissipation trough moment. This dissipation trough moment corresponds to the peak energy dissipation during the acoustic vortex energy conversion process. The instantaneous phase corresponding to each dissipation trough moment is recorded. This instantaneous phase is obtained by interpolation of the vortex structure shedding instantaneous phase curve obtained by S221 at the corresponding moment. The dissipation trough moment sequence and its corresponding instantaneous phase sequence are output as the target signal for subsequent cross-correlation analysis.

[0063] S223. The optimal matching phase is determined, specifically: The instantaneous phase curve of vortex structure shedding output from S221 and the instantaneous phase sequence corresponding to the dissipation valley moment output from S222 are obtained. Cross-correlation analysis is performed on the two sets of phase signals to calculate the cross-correlation coefficient under different time delays. This cross-correlation coefficient characterizes the degree of linear correlation between the two sets of signals at a specific delay.

[0064] By employing a cyclic shift search, all possible phase shifts are traversed within the vortex structure shedding cycle at a fixed step size, and the cross-correlation function value is calculated for each phase shift. The phase shift that maximizes the cross-correlation function is determined as the optimal matching phase. This optimal matching phase corresponds to the best time alignment between the vortex structure shedding phase and the valley of sound pressure fluctuation dissipation, indicating that the energy conversion efficiency of the vortex structure to acoustic energy is highest under this phase relationship. The value of the optimal matching phase and its corresponding peak cross-correlation function are output as the central reference for subsequently constructing the matching window.

[0065] S224. Matching window construction, specifically: The system receives the optimal matching phase determined by S223, and simultaneously receives the complete lifecycle duration of the vortex structure from generation to dissipation predicted by S124. Centered on the optimal matching phase, and based on the tolerance of the complete lifecycle duration of the vortex structure and the acoustic-vortex coupling energy conversion efficiency, an allowable phase fluctuation range is set. This phase fluctuation range is typically taken as 10% to 30% of the lifecycle duration to ensure that effective acoustic-vortex coupling is maintained at the window boundary.

[0066] The phase fluctuation range is converted into a time window, and the start and end times of the matching window are calculated. The start time corresponds to the phase interval within the vortex structure shedding cycle, which is the optimal matching phase minus half the window width. The end time corresponds to the phase interval within the optimal matching phase plus half the window width. The start and end times of the matching window correspond to a specific phase interval within the vortex structure shedding cycle, within which the vortex structure shedding and sound pressure dissipation exhibit optimal coupling characteristics. The start time, end time, center phase, and window width of the matching window are output as the time boundary conditions for generating the rotational speed modulation spectrum command in step S23.

[0067] S23. Speed ​​modulation spectrum command generation, specifically: S231. Receive the matching window output by S224, parse the start time, end time, center phase and window width information contained in the matching window, and establish window boundary conditions in the time domain; at the same time, receive the expected heat flux density vector field distribution output by S13, parse the spatial point coordinates, heat flux density magnitude and direction information contained in the expected heat flux density vector field distribution, and calculate the total heat load demand of the heat sink surface.

[0068] S232. For the matching window period, based on the amplitude-frequency characteristic curve of the acoustic vortex coupling transfer function established in S213, identify the frequency component with the highest energy transfer efficiency as the pulsation reference frequency; according to the gain peak value of the amplitude-frequency characteristic curve at this frequency, combined with the fan speed adjustment range and mechanical strength constraints, calculate the small speed pulsation amplitude that can enhance acoustic vortex energy dissipation without causing structural resonance; determine the phase compensation amount according to the phase-frequency characteristic curve, and select a sinusoidal waveform or an optimized non-sinusoidal waveform as the pulsation waveform. The optimized non-sinusoidal waveform is solved by a genetic algorithm or gradient descent method, with maximizing the acoustic vortex energy dissipation efficiency as the objective function; integrate the above parameters to form the small speed pulsation parameters, which include the pulsation frequency, pulsation amplitude, pulsation waveform, and phase compensation amount.

[0069] S233. For the time interval outside the matching window, calculate the total heat flux flux of the heat sink surface based on the spatial integral value of the expected heat flux density vector field distribution; combine the fan's heat dissipation performance curve with the goal of minimizing the heat sink entropy productivity to establish a steady-state speed optimization model; solve the optimization model to obtain the steady-state speed reference value that meets the heat dissipation requirements and has the lowest operating energy consumption. The steady-state speed reference value is a speed command that is either time-invariant or slowly time-varying.

[0070] S234. The minute speed fluctuation parameters during the matching window are sequentially concatenated with the steady-state speed reference value outside the matching window. A smooth transition function is used at the window boundary to eliminate speed jumps, generating a complete speed modulation spectrum command. This speed modulation spectrum command is a time-varying speed command sequence, covering all time nodes from the current moment to the next control cycle. The speed modulation spectrum command is output, which includes timing markers, speed values, and corresponding status indicators. It serves as the input parameter for broadband speed trajectory synthesis in step S3, and is used for coupling and solving with the electromagnetic-thermal-mechanical multi-frequency response model.

[0071] S3. Obtain the multiphysics field of the motor, input the expected heat flux density vector field and speed modulation spectrum command into the multi-frequency response model, solve for the optimal broadband speed trajectory, and generate the drive signal; specifically including: S31. Motor multi-source loss data acquisition, specifically: S311. By using a distributed fiber optic grating sensor embedded in the stator slot of the motor, the transient Joule heat distribution data of the stator winding is obtained with sampling parameters of spatial resolution not less than 1 cm and temporal resolution not less than 100 Hz. The distributed fiber optic grating sensor utilizes the temperature sensitivity of the Bragg wavelength to measure the temperature field distribution of each slot of the stator winding in real time, and then inverts the Joule heat generation rate based on the winding resistance-temperature characteristics.

[0072] S312. By using a thin-film thermocouple array mounted on the rotor surface, the eddy current loss data of the permanent magnet is obtained at a sampling frequency of not less than 1kHz. The thin-film thermocouple array is arranged at key positions on the surface of the permanent magnet to measure the local temperature rise caused by eddy current loss, and the transient eddy current loss power is calculated in combination with the heat capacity parameters of the permanent magnet.

[0073] S313. By using an accelerometer installed on the outer ring of the bearing, the bearing vibration acceleration signal is collected at a sampling frequency of not less than 10kHz. Combined with the bearing dynamic model and the frequency characteristics of the rolling elements, the contact stress fluctuation data of the bearing rolling elements is obtained by inversion. This contact stress fluctuation data reflects the dynamic load distribution and fatigue damage accumulation trend inside the bearing.

[0074] S314. Align the transient Joule thermal distribution data, eddy current loss data, and contact stress fluctuation data according to the timestamps to establish a synchronized multi-source loss dataset, which will serve as the input parameters for the subsequent multi-frequency response model.

[0075] S32. Construction of multi-frequency response model and solution of optimal wideband speed trajectory, specifically: S321. Construct an electromagnetic-thermal-mechanical coupled multi-frequency response model, which consists of three mutually coupled sub-models, specifically: A three-dimensional thermal network sub-model describing the Joule thermal dynamics of the stator winding is proposed. This sub-model discretizes the stator winding into lumped heat capacity nodes connected by a thermal resistance network, considering copper losses, iron losses, and convective heat transfer from cooling airflow. An electromagnetic field sub-model describing the eddy current losses of the rotor permanent magnet is proposed. This sub-model solves the eddy current field equations in the rotating coordinate system based on the time-stepping finite element method, calculating the eddy current density distribution and power loss inside the permanent magnet. A dynamic sub-model describing the contact stress fluctuations of the bearing rolling elements is proposed. This sub-model considers the Hertzian contact between the rolling elements and the inner and outer rings, cage collisions, and the dynamic characteristics of the lubricating film, solving for the time-varying contact force and stress distribution of the rolling elements.

[0076] Thermo-mechanical coupling boundary conditions are established between the sub-models: the temperature rise of the stator winding affects the winding resistance and Joule heat generation, forming a thermo-electric positive feedback; the thermal deformation caused by rotor eddy current losses alters the air gap magnetic field distribution, affecting electromagnetic torque fluctuations; the frictional heat caused by bearing contact stress fluctuations alters the lubricant viscosity, affecting the bearing dynamic characteristics. Iterative solutions enable bidirectional data exchange between the sub-models, ensuring that the multi-frequency response model accurately describes the coupled dynamic behavior of the electromagnetic, thermal, and mechanical domains.

[0077] S322. Input parameter loading, specifically: The system receives the expected heat flux density vector field distribution output by S13, analyzes its spatial distribution characteristics and temporal evolution trend, and uses it as the boundary heat flux input condition for the 3D thermal network sub-model. It also receives the steady-state rotational speed reference value calculated by S23, which serves as the steady-state operating point reference for the multi-frequency response model. Furthermore, it receives the minute rotational speed pulsation parameters corresponding to the matching window generated by S23, including pulsation frequency, pulsation amplitude, pulsation waveform, and phase compensation, which serve as the time-varying rotational speed excitation for the electromagnetic field sub-model. The system receives transient Joule heat distribution data acquired by S31, which serves as the initial temperature field and heat source distribution verification data for the 3D thermal network sub-model. It also receives eddy current loss data, which serves as the benchmark for comparing the eddy current loss calculation results of the electromagnetic field sub-model. Finally, it receives contact stress fluctuation data, which serves as the dynamic load verification input for the dynamic sub-model. All the above input parameters are then loaded into the corresponding sub-models of the multi-frequency response model according to their physical domain classification, and the solution time step and convergence tolerance are set to prepare for time-domain solution.

[0078] S323. Time-domain solution and performance index calculation, specifically: Receive the timing mark output by S13, use the timing mark as the reference time axis, and establish an optimized time window from the current moment to the end of the predictive control cycle. This optimized time window covers the complete life cycle of the vortex structure from generation, propagation to dissipation.

[0079] The multi-frequency response model is solved in the time domain within the optimization time window. An implicit integration algorithm is used to ensure numerical stability. The coupling equations of electromagnetic field, temperature field and mechanical field are solved iteratively in each time step to obtain the transient response of state variables such as stator winding temperature distribution, permanent magnet eddy current loss power and bearing contact stress.

[0080] Based on the solution results, four key performance indicators are calculated: the time history curve of heat sink entropy yield fluctuation is calculated based on the heat sink surface temperature gradient and heat flux density, characterizing the irreversible loss fluctuation in the heat transfer process; the time history curve of sound power level fluctuation is calculated based on fan speed pulsation and acoustic vortex coupling transfer function, characterizing the aerodynamic noise radiation intensity fluctuation; the time history curve of stator temperature rise fluctuation is calculated based on the difference between the highest stator winding temperature and the ambient temperature, characterizing the motor thermal load state; and the time history curve of bearing contact fatigue index is calculated based on bearing contact stress amplitude and cycle number, characterizing the bearing life consumption rate. The time history curves of the four performance indicators are output as the evaluation basis for the multi-objective optimization algorithm.

[0081] S324. Multi-objective optimization solution, specifically: Using the weighted integral of the four factors obtained from S323—root mean square of heat sink entropy yield fluctuation, root mean square of sound power level fluctuation, peak value of stator temperature rise fluctuation, and cumulative value of bearing contact fatigue index—as the objective function, a multi-objective optimization problem is established. The weight coefficients of each performance index are dynamically adjusted according to heat dissipation priority, noise limitation, motor safety, and bearing life requirements.

[0082] The multi-objective particle swarm optimization algorithm is used to search for the optimal solution in a wide bandwidth, which covers the entire frequency band from DC to the bearing pass frequency. Each particle in the particle swarm represents a set of wideband rotational speed trajectory parameters, including the amplitude and rate of change of the fundamental frequency component, the frequency and amplitude of the perturbation component, and the center frequency and notch depth of the notch component.

[0083] For each particle's corresponding broadband rotation trajectory, a multi-frequency response model is invoked for rapid solution, calculating the objective function value. Particle velocity and position are updated using non-dominated sorting and crowding calculations, iteratively searching for the Pareto optimal front. From the Pareto optimal solution set, a compromise optimal solution is selected based on decision preferences, outputting the optimal broadband rotation trajectory containing fundamental frequency components, perturbation components, and notch components. This optimal broadband rotation trajectory is stored in a time series format based on time-series labels.

[0084] S325. Verification and optimization adjustment of notch components, specifically: Fourier analysis was performed on the optimal broadband speed trajectory output by S324 to calculate the speed spectrum distribution and identify the center frequency and bandwidth of the notch component. The system structural resonant frequencies were obtained, including the passing frequency of the fan blades, the natural frequency of the motor stator, and the critical speed of the bearing-rotor system. The separation degree between the center frequency of the notch component and the resonant frequencies of each structure was calculated. This separation degree is defined as the ratio of the absolute value of the frequency difference to the half-bandwidth of the resonant frequency.

[0085] The separation degree is compared with a preset threshold, typically set to 2 to 3, to ensure that the notch component effectively avoids the resonance region. If the separation degree is lower than the preset threshold, it indicates that the notch component is too close to the structural resonance frequency, which may lead to resonance amplification risk. In this case, the penalty function coefficient in the multi-objective optimization algorithm is adjusted to increase the penalty weight of solutions close to the resonance frequency, and step S324 is re-executed for optimization. This process continues until the separation degree between the notch component and all structural resonance frequencies meets the preset threshold requirement, confirming the validity of the optimal broadband rotational speed trajectory, and outputting the final verified optimal broadband rotational speed trajectory.

[0086] S33. Drive signal generation and execution, specifically: Receive the optimal wideband speed trajectory verified by S325, analyze the temporal superposition relationship of the fundamental frequency component, perturbation component and notch filter component contained in the optimal wideband speed trajectory, and calculate the instantaneous speed target value at each moment.

[0087] Based on the number of motor pole pairs and the instantaneous speed target value, the motor power supply frequency is calculated; based on the voltage-frequency coordinated control strategy, the corresponding power supply voltage amplitude is calculated; using the space vector pulse width modulation algorithm, the power supply frequency and voltage amplitude are converted into the switching state of the three-phase inverter, generating a three-phase duty cycle sequence of the space vector pulse width modulation signal. This three-phase duty cycle sequence has a carrier period as the time resolution and covers the entire control cycle.

[0088] The three-phase duty cycle sequence is output to the motor driver via a high-speed communication interface. The motor driver controls the switching of power switching devices according to the duty cycle sequence, generating three-phase AC power with adjustable frequency and amplitude. This three-phase AC power drives the permanent magnet synchronous motor of the cooling fan, enabling the fan speed to track the optimal wideband speed trajectory, achieving multi-objective synergistic optimization of heat sink heat dissipation, eddy current dissipation, and mechanical reliability.

[0089] S4. Collect actual operational deviations to form reward signals, optimize the strategy through reinforcement learning, and update the model in reverse to achieve self-evolution; specifically including: S41. Construction of multidimensional reward signals, specifically: S411. Acquire multi-physics response data after actual operation of the optimal broadband speed trajectory. Using a miniature thermocouple array in S111, continuously acquire the transient temperature response sequence of the heat sink at a sampling frequency of no less than 1kHz. Based on the entropy yield density calculation method in S112, reconstruct the evolution trajectory of the heat sink entropy yield field after actual operation. Using a microphone array in S211, synchronously acquire far-field aerodynamic noise pressure data. Based on the sound power level calculation formula, obtain the temporal fluctuation of the sound power level. Using a distributed fiber optic grating sensor in S31, acquire the transient temperature distribution of the stator winding to obtain the stator temperature rise response curve. Using a bearing outer ring accelerometer in S31, acquire the bearing vibration acceleration signal. After feature extraction, obtain bearing vibration characteristic data, including vibration amplitude, spectral distribution, and impact indicators.

[0090] S412. Calculate the deviation between the actual response and the predicted target for each physical field: For the evolution trajectory of the heat sink entropy yield field, it is compared with the expected heat flux density vector field distribution output by S13 in time and space. The difference in entropy yield density between the two at the same spatial point and time is calculated. The degree of deviation is quantified by the root mean square error or weighted average absolute error. This deviation reflects the prediction accuracy of the heat wave dispersion relation prediction model in S1.

[0091] For the temporal fluctuation of sound power level, the phase is compared with the time of the sound pressure pulsation dissipation valley identified by S222. The time offset between the actual minimum time of sound power level and the predicted dissipation valley time is calculated, as well as the difference between the amplitude of sound power level fluctuation and the expected dissipation depth. This deviation reflects the modeling accuracy of the acoustic vortex coupling transfer function in S2.

[0092] For the stator temperature rise response curve, its amplitude and phase are compared with the stator temperature rise fluctuation target value solved by S323. The relative error and time lag between the actual temperature rise peak and the target peak are calculated. This deviation reflects the thermal prediction accuracy of the multi-frequency response model in S3.

[0093] S413. Identify abnormal bearing conditions and generate a penalty factor. A threshold is determined for the bearing vibration characteristic data. This threshold is set based on the statistical characteristics of historical bearing operating data, typically the mean plus three standard deviations. Abnormal impact peaks exceeding the preset threshold are identified. These abnormal impact peaks characterize rolling element defects, lubrication failures, or misalignment faults. The amplitude, frequency, and duration of the abnormal impact peaks are comprehensively quantified into a penalty factor. This penalty factor is negative, and its amplitude is positively correlated with the degree of abnormality.

[0094] S414. Integrate the above deviations and penalty factors to form a multi-dimensional reward signal. This multi-dimensional reward signal is in vector form, with each component corresponding to the heat sink entropy yield prediction deviation, sound power level dissipation matching deviation, stator temperature rise tracking deviation, and bearing anomaly penalty factor, respectively. The weight of each component is dynamically adjusted according to the control priority. For example, the weight of the entropy yield deviation is increased during the critical heat dissipation period, and the weight of the sound power deviation is increased during the noise-sensitive period. This multi-dimensional reward signal serves as the environmental feedback for the deep reinforcement learning network in step S42, and is used to evaluate the overall performance of the current control strategy.

[0095] S42. Optimization of deep reinforcement learning strategies, specifically: S421. Construct the state space of the deep reinforcement learning network. This state space includes the expected heat flux density vector field distribution output in S13, which represents the predicted heat load state of the heat sink; it includes the amplitude-frequency characteristic curve and phase-frequency characteristic curve of the acoustic-vortex coupling transfer function established in S213, which represents the frequency domain characteristics of acoustic-vortex energy conversion; and it includes the parameters of each sub-model of the multi-frequency response model constructed in S321, including the thermal resistance and thermal capacity parameters of the three-dimensional thermal network sub-model, the material electromagnetic parameters of the electromagnetic field sub-model, and the contact stiffness and damping parameters of the dynamic sub-model, which represent the physical characteristics and current operating state of the motor system.

[0096] S422. Construct the action space of the deep reinforcement learning network. This action space uses the adjustable parameters of the rotational speed modulation spectrum command generated in S23 as variables, including the frequency and amplitude of the perturbation component, which determines the intensity of acoustic vortex dissipation within the matching window; the center frequency and notch depth of the notch component, which determines the structural resonance avoidance effect; and the rate of change of the fundamental frequency component, which determines the agility of the heat dissipation response. The value range of each action variable is set according to the mechanical constraints and thermal safety constraints of the motor to ensure the executability of the action.

[0097] S423. The multidimensional reward signal constructed in S41 is input into the deep reinforcement learning network as an environmental feedback signal. A policy gradient method, such as proximal policy optimization or the soft actor-critic algorithm, is used to iteratively optimize the control policy. This policy gradient method approximates the policy function and value function through a neural network, calculates the policy gradient based on the reward signal, and updates the network parameters to maximize the cumulative expected reward. During the iteration process, the agent selects an action based on the current state, observes the new state and reward after executing the action, stores experience samples in the replay buffer, and breaks data correlation through random sampling to improve learning efficiency.

[0098] S424. Output the optimized control strategy parameters. These control strategy parameters are the neural network weights and biases, or the specific parameterized form of the policy function, representing the optimal mapping relationship from the state space to the action space. These optimized control strategy parameters serve as the basis for model updates in step S43, guiding the adaptive adjustment of the prediction model and control algorithm in S1-S3.

[0099] S43. The control system is online and self-evolving, specifically: S431. The optimized control strategy parameters output from S42 are back-injected into the thermal wave dispersion relation prediction model of S1. Based on the learning results of the control strategy on the thermal wave propagation characteristics, the boundary conditions of the thermal wave dispersion relation are dynamically updated, including the equivalent convective heat transfer coefficient between the heat sink substrate and the cooling fluid, and the spatiotemporal correction factor of the heat source distribution inside the heat sink. This update is achieved through online parameter identification or transfer learning, so that the thermal wave dispersion relation extracted in S114 more accurately reflects the actual heat-mass flow coupling characteristics.

[0100] S432. The optimized control strategy parameters are back-injected into the matching window identification algorithm of S2. Based on the learning results of the control strategy on the acoustic-vortex coupling phase relationship, the identification threshold of the matching window constructed in S224 is corrected, including the allowable fluctuation range of the optimal matching phase and the adaptive adjustment coefficient of the window width. This correction enables the matching window identified by S22 to more accurately capture the optimal coupling time between the vortex shedding phase and the valley of acoustic pressure dissipation, thereby improving the acoustic-vortex energy dissipation efficiency.

[0101] S433. The optimized control strategy parameters are back-injected into the multi-frequency response model of S3. Based on the learning results of the control strategy on the electromagnetic-thermal-mechanical coupling characteristics, the coupling weights of each sub-model in the multi-frequency response model constructed in S321 are adjusted, including the thermo-electric feedback strength between the three-dimensional thermal network sub-model and the electromagnetic field sub-model, and the frictional-thermal coupling coefficient between the dynamic sub-model and the thermal network sub-model. This adjustment makes the time-domain solution of S323 more accurate in predicting the multi-physics dynamic response and improves the solution quality of the multi-objective optimization in S324.

[0102] S434. Through the reverse injection of the above three aspects, the entire control system achieves online self-evolution. This online self-evolution enables the thermal wave dispersion relationship prediction model of S1, the matching window identification algorithm of S2, and the multi-frequency response model of S3 to be continuously optimized according to actual operating data, adapting to the characteristic drift caused by heat sink aging, environmental changes, and operating condition migration, forming a closed-loop adaptive control architecture from perception, decision-making to execution, ensuring the optimal performance and reliability of the cooling fan speed smooth control method in long-term operation. Example 2

[0103] like Figure 2 As shown, a noise feedback-based cooling fan speed smoothing control system is used to implement a noise feedback-based cooling fan speed smoothing control method, including: A. Thermodynamic reconstruction prediction module; specifically including: The transient temperature acquisition unit is used to synchronously acquire the transient temperature response sequence of each measuring point of the miniature thermocouple array at a sampling frequency of not less than 1kHz, and perform wavelet denoising processing. The entropy productivity calculation unit is connected to the transient temperature acquisition unit. It is used to divide the heat sink into multiple local control volumes, solve the heat flux density and temperature gradient of each control volume boundary using the finite volume method, and calculate the entropy productivity density distribution of each local control volume. The force-fluid relationship inversion unit is connected to the entropy yield calculation unit. It is used to solve the Onsager coupled transport coefficient matrix by variational method with the objective function of minimizing the local entropy yield density, invert the non-equilibrium thermodynamic force-fluid relationship, and extract the effective thermal diffusion tensor. The thermal wave dispersion analysis unit, connected to the force-flow relationship inversion unit, is used to solve the characteristic equation of thermal wave propagation based on the effective thermal diffusion tensor and extract the thermal wave dispersion relationship of thermal disturbance phase velocity as a function of frequency. The vortex structure prediction unit, connected to the thermal wave dispersion analysis unit, is used to combine the geometric parameters of the microgroove structure on the heat sink surface with the mainstream velocity, calculate the vortex generation frequency induced by the microgroove by solving the vortex transport equation, and predict the spatiotemporal evolution trajectory of the vortex structure in the thermal boundary layer. The advance compensation calculation unit, connected to the vortex structure prediction unit, is used to input the spatiotemporal evolution trajectory into the electromechanical coupling response model, combine the fan rotor rotational inertia and electromagnetic torque response time to calculate the advance compensation duration, and output the expected heat flux density vector field distribution and vortex structure arrival time marker at the advance compensation moment.

[0104] B. Acoustic-vortex coupling modulation module, connected to the thermodynamic reconstruction prediction module; specifically including: The flow field and sound field acquisition unit is used to acquire pulsating velocity field data and calculate pulsating vorticity field data through a particle image velocimeter or a hot-wire anemometer, and simultaneously acquire far-field aerodynamic noise pressure data through a microphone array. The intrinsic orthogonal decomposition unit, connected to the flow field and acoustic field acquisition unit, is used to perform intrinsic orthogonal decomposition on the pulsating vorticity field data and extract the spatial basis functions and corresponding time coefficients of the dominant flow modes. The transfer function identification unit, connected to the intrinsic orthogonal decomposition unit, is used to establish the amplitude-frequency and phase-frequency characteristics of the acoustic-vortex coupling transfer function using the system identification method with the time coefficient as input and the aerodynamic noise pressure data as output. The phase matching analysis unit connects the transfer function identification unit and the thermodynamic reconstruction prediction module. It is used to perform phase-locking analysis on the time series marker and time coefficient, extract the instantaneous phase through Hilbert transform, and identify the matching window between the vortex structure shedding phase and the sound pressure pulsation dissipation valley. The modulation spectrum generation unit, connected to the phase matching analysis unit and the thermodynamic reconstruction prediction module, is used to generate a speed modulation spectrum command based on the matching window. The speed modulation spectrum command includes a small speed fluctuation parameter applied during the matching window and a steady-state speed reference value calculated outside the matching window based on the expected heat flux density vector field distribution.

[0105] C. A multi-frequency response synthesis module, connected to both the thermodynamic reconstruction prediction module and the acoustic-vortex coupling modulation module; specifically including: The multi-physics acquisition unit is used to acquire transient Joule heat distribution data through fiber optic grating sensors embedded in the stator slots, acquire eddy current loss data through thin-film thermocouples attached to the rotor surface, and acquire contact stress fluctuation data through acceleration sensors mounted on the outer ring of the bearing combined with dynamic model inversion. The multi-frequency response model unit connects the multi-physics acquisition unit, the thermodynamic reconstruction prediction module, and the acoustic-vortex coupling modulation module. It is used to construct an electromagnetic-thermal-mechanical coupled multi-frequency response model that includes a thermal network sub-model, an electromagnetic field sub-model, and a dynamic sub-model. The input parameters are loaded with time-domain solutions using time-series markers as the reference time axis. The multi-objective optimization unit, connected to the multi-frequency response model unit, is used to minimize the weighted integral of four factors: root mean square of heat sink entropy yield fluctuation, root mean square of sound power level fluctuation, peak value of stator temperature rise fluctuation, and cumulative value of bearing contact fatigue index. The multi-objective optimization algorithm is used to solve the optimal broadband speed trajectory containing fundamental frequency component, perturbation component and notch filter component, and outputs the target value of stator temperature rise fluctuation. The drive signal generation unit, connected to the multi-objective optimization unit, is used to generate a three-phase duty cycle sequence of space vector pulse width modulation signal based on the optimal wideband speed trajectory, and output it to the motor driver.

[0106] D. Self-evolutionary learning module, which connects to the thermodynamic reconstruction prediction module, the acoustic-vortex coupling modulation module, and the multi-frequency response synthesis module; specifically including: The deviation calculation unit is connected to the thermodynamic reconstruction prediction module, the acoustic vortex coupling modulation module and the multi-frequency response synthesis module respectively. It is used to collect the heat sink entropy yield field evolution trajectory, acoustic power level time-series fluctuation, stator temperature rise response curve and bearing vibration characteristic data after actual operation, calculate the deviation between each response and the corresponding prediction target, and use the abnormal impact peak of bearing vibration as a penalty factor to jointly form a multi-dimensional reward / penalty signal. The reinforcement learning network unit connects to the deviation calculation unit. It is used to optimize the control strategy iteratively through the policy gradient method, using the expected heat flux density vector field distribution, the amplitude and phase frequency characteristics of the acoustic eddy coupling transfer function, and the parameters of each sub-model of the multi-frequency response model as the state space, and the frequency and amplitude of the perturbation component, the center frequency and notch depth of the notch component, and the rate of change of the fundamental frequency component as the action space, and output the optimized control strategy parameters. The back-injection unit, connected to the reinforcement learning network unit, thermodynamic reconstruction prediction module, acoustic-vortex coupling modulation module, and multi-frequency response synthesis module, is used to back-inject the optimized control strategy parameters, dynamically update the boundary conditions of the thermal wave dispersion relation, correct the identification threshold of the matching window, and adjust the coupling weights of each sub-model in the multi-frequency response model, thereby realizing the online self-evolution of the control system.

[0107] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: This invention overcomes the hysteresis limitation of traditional temperature feedback control by acquiring transient temperature response sequences through a micro thermocouple array. Based on the minimum entropy yield principle, it inverts the non-equilibrium thermodynamic force-fluid relationship, introducing the second law of thermodynamics into the field of fan control. Unlike traditional PID control, which only senses the current temperature value, this invention extracts the effective heat diffusion tensor and the thermal wave dispersion relationship to quantitatively describe the propagation speed and dispersion characteristics of thermal disturbances within the heat sink. Combined with the eddy transport effect induced by the microgroove structure, it predicts in advance the generation time, evolution trajectory, and precise timing of eddy structures within the thermal boundary layer reaching the heat sink surface. This allows the control system to generate advance compensation time before the heat flux peak arrives, outputting the expected heat flux density vector field distribution. This upgrades the traditional post-control compensation to pre-control prediction, fundamentally avoiding sudden speed changes caused by thermal hysteresis and providing high-fidelity prior physical field information for subsequent precise modulation.

[0108] This invention extracts the dominant modes of the pulsating vorticity field through intrinsic orthogonal decomposition, establishes a quantitative mapping relationship between the vorticity field and aerodynamic noise pressure—the acoustic-vortex coupling transfer function—and identifies the matching window between the vortex structure shedding phase and the valley value of sound pressure pulsation dissipation. Applying minute rotational speed fluctuations based on this window can directionally convert vortex kinetic energy into heat dissipation, achieving an active noise reduction mechanism through vortex silencing, fundamentally suppressing the source of aerodynamic noise rather than passively isolating it. Simultaneously, this invention constructs a multi-frequency response model of electromagnetic-thermal-mechanical coupling, unifying four conflicting physical field objectives—heat sink entropy yield fluctuation, sound power level fluctuation, stator temperature rise fluctuation, and bearing contact fatigue index—into a weighted integral minimization optimization framework, solving for the optimal broadband rotational speed trajectory including fundamental frequency components, perturbation components, and notch filter components. This trajectory satisfies dynamic heat dissipation requirements while enhancing acoustic vortex dissipation through perturbation components and avoiding structural resonance through notch components. It achieves a dynamic balance between heat dissipation efficiency, acoustic experience, electromagnetic thermal stability, and mechanical reliability. The speed change is smooth and imperceptible, completely eliminating the noise abrupt changes and mechanical shocks caused by single-objective optimization in traditional control.

[0109] This invention overcomes the technical bottlenecks of traditional control strategies involving offline calibration and fixed parameters. It constructs a multi-dimensional reward signal from the evolution trajectory of the heat sink entropy yield field, the temporal fluctuation of the sound power level, the stator temperature rise response curve, and the bearing vibration characteristic data after actual operation, and inputs this signal into a deep reinforcement learning network. Using the expected heat flux density field, acoustic-vortex coupling transfer function characteristics, and multi-frequency response model parameters as the state space, and the key parameters of the perturbation component, notch component, and fundamental frequency component as the action space, the control strategy is iteratively optimized using the strategy gradient method. More importantly, this invention achieves a bidirectional interactive mechanism of forward prediction and reverse correction. The optimized control strategy parameters are injected back into the thermal wave dispersion relation prediction model, the matching window recognition algorithm, and the multi-frequency response model, dynamically updating the boundary conditions of the thermal wave dispersion relation, correcting the recognition threshold of the matching window, and adjusting the coupling weights of each sub-model. This self-evolving architecture enables the control system to continuously optimize as the fan ages, accumulates dust, and changes operating conditions. The operating experience of each device can be used to improve its own control model, achieving a fundamental leap from being fixed at the factory to becoming increasingly intelligent with use, opening a new path for the intelligent control of cooling fans.

[0110] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the content of the present invention, but they should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0111] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for smoothing the speed control of a cooling fan based on noise feedback, characterized in that, include: Obtain the transient temperature sequence of the heat sink, reconstruct the non-equilibrium entropy yield field, predict the evolution trajectory of the thermal boundary layer, and output the expected heat flux density vector field and vortex structure time sequence markers; In response to the vortex structure timing mark, the wake vorticity field and far-field noise pressure are collected, the acoustic vortex transfer function is established, the matching window is identified, and the rotational speed modulation spectrum command is generated based on the matching window and the expected heat flux density vector field. The multiphysics field of the motor is obtained, and the expected heat flux density vector field and speed modulation spectrum command are input into the multi-frequency response model to solve the optimal broadband speed trajectory and generate the drive signal. The actual operational deviations are collected to form reward signals. The strategy is optimized through reinforcement learning, and the model is updated in reverse to achieve self-evolution.

2. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 1, characterized in that, The process of acquiring the transient temperature sequence of the heat sink, reconstructing the non-equilibrium entropy yield field, predicting the thermal boundary layer evolution trajectory, and outputting the expected heat flux density vector field and vortex structure time series markers specifically includes: The transient temperature response sequence at each measuring point is collected, the heat sink is divided into multiple local control volumes, the entropy yield density distribution of each local control volume is calculated, and the Onsager coupling transport coefficient matrix is ​​solved to extract the effective thermal diffusion tensor and thermal wave dispersion relation under the action of thermal-mass-fluid coupling. Based on the thermal wave dispersion relation, the functional relationship between the thermal disturbance phase velocity and frequency is calculated. The generation frequencies of the flow vortices and spanwise vortices induced by the microgrooves are solved by the vortex transport equation, and the spatiotemporal evolution trajectory of the vortex structure in the thermal boundary layer is predicted. The spatiotemporal evolution trajectory is input into the electromechanical coupling response model. Combining the rotational inertia of the fan rotor at the current speed with the electromagnetic torque response time, the advance compensation time required from issuing the control command to the vortex structure reaching the corresponding position on the heat sink surface is calculated. At the advance compensation time, the expected heat flux density vector field distribution of each spatial point on the heat sink surface and the time sequence mark of each vortex structure reaching the heat sink surface are output.

3. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 2, characterized in that, The transient temperature response sequences at each measuring point are collected, the heat sink is divided into multiple local control volumes, the entropy yield density distribution of each local control volume is calculated, and the Onsager coupling transport coefficient matrix is ​​solved to extract the effective thermal diffusion tensor and thermal wave dispersion relation under the thermal-mass-fluid coupling effect. Specifically, this includes: Based on a micro thermocouple array embedded in the heat sink substrate with a non-uniform topological distribution, transient temperature response sequences at each measuring point are acquired synchronously, and wavelet denoising processing is performed on the transient temperature response sequences. Based on the geometric structure and material thermophysical parameters of the heat sink substrate, the heat sink is divided into multiple non-overlapping local control volumes. The positions of each thermocouple measuring point are used as control volume nodes. The heat flux density and temperature gradient of each control volume boundary are discretized and solved by the finite volume method. The entropy production rate distribution of each local control volume is calculated according to the entropy production rate calculation formula in irreversible thermodynamics. The objective function is to minimize the entropy production density of each local control volume, and the linear phenomenological relationship between heat flux density and temperature and pressure gradients is used as a constraint. The Onsager coupled transport coefficient matrix is ​​solved by variational method to invert the non-equilibrium thermodynamic force-fluid relationship and extract the effective heat diffusion tensor under the action of thermo-mass-fluid coupling. Based on the effective thermal diffusion tensor, and combined with the coupled boundary conditions of the energy equation and the momentum equation, the characteristic equation of thermal wave propagation is solved by dispersion analysis, and the thermal wave dispersion relation of thermal disturbance phase velocity with frequency is extracted.

4. The method for smooth control of cooling fan speed based on noise feedback according to claim 3, characterized in that, Based on the aforementioned thermal wave dispersion relation, the functional relationship between the thermal disturbance phase velocity and frequency is calculated. The generation frequencies of the streamwise and spanwise vortices induced by the microgrooves are solved using the vortex transport equation, and the spatiotemporal evolution trajectory of the vortex structure within the thermal boundary layer is predicted. Specifically, this includes: Based on the aforementioned thermal wave dispersion relation, the functional relationship between the thermal disturbance phase velocity and angular frequency is solved, the critical frequency point where dispersion characteristics appear during thermal wave propagation is identified, and the low-frequency thermal diffusion-dominant region and the high-frequency thermal wave propagation-dominant region are divided. The geometric parameters of the microgroove structure on the heat sink surface are obtained. Combined with the mainstream surface velocity formed by the fan blowing, the flow vortex generation frequency and spanwise vortex generation frequency induced by the microgroove structure are calculated by solving the vortex transport equation. Using the flow vortex generation frequency and spanwise vortex generation frequency as vortex disturbance source terms, and taking the critical frequency point as the boundary, a coupled wave equation of thermal disturbance and vortex disturbance is established in the high-frequency thermal wave propagation dominance region, and the cross-spectral density function of vortex pulsation and temperature pulsation in the thermal boundary layer is solved. Based on the spatiotemporal distribution characteristics of the cross-spectral density function and combined with the boundary layer stability theory, the complete life cycle of the vortex structure from initial generation, linear growth, nonlinear saturation to final breakage is predicted, and the spatiotemporal evolution trajectory of the vortex structure within the thermal boundary layer is generated.

5. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 4, characterized in that, In response to the vortex structure timing marker, the wake vorticity field and far-field noise pressure are acquired, an acoustic vortex transfer function is established, a matching window is identified, and a rotational speed modulation spectrum command is generated based on the matching window and the expected heat flux density vector field, specifically including: In response to the advance compensation moment, the pulsating vorticity field data and far-field aerodynamic noise pressure data of the downstream wake region of the fan are collected. The pulsating vorticity field data are subjected to intrinsic orthogonal decomposition to extract spatial basis functions and time coefficients. The acoustic-vortex coupling transfer function is established by using the system identification method. Phase-locking analysis is performed on the time sequence markers and time coefficients, and the matching window between the vortex structure shedding phase and the sound pressure pulsation dissipation valley value is identified by cross-correlation function or phase synchronization algorithm. A speed modulation spectrum instruction is generated based on the matching window, which includes applying a small speed fluctuation during the matching window to enhance acoustic vortex energy dissipation, and calculating a steady-state speed reference value outside the matching window based on the expected heat flux density vector field distribution.

6. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 5, characterized in that, In response to the aforementioned lead compensation moment, fluctuating vorticity field data and far-field aerodynamic noise pressure data in the downstream wake region of the fan are collected. The fluctuating vorticity field data undergoes intrinsic orthogonal decomposition to extract spatial basis functions and time coefficients. An acoustic-vortex coupling transfer function is established using a system identification method, specifically including: In response to the aforementioned advance compensation moment, pulsating velocity field data is collected by a particle image velocimeter or hot-wire anemometer arranged in the downstream wake region of the fan, and pulsating vorticity field data is obtained by the vorticity calculation formula; at the same time, far-field aerodynamic noise pressure data is collected synchronously by a microphone array arranged in the far field. The pulsating vorticity field data is subjected to intrinsic orthogonal decomposition to construct a spatial two-point correlation matrix. By solving the eigenvalue problem of the correlation matrix, the top N modes with the largest eigenvalues ​​are extracted as the dominant flow modes, and the corresponding spatial basis functions and time coefficients reflecting the evolution of mode energy over time are obtained. Using the time coefficient as the input signal and the far-field aerodynamic noise pressure data as the output signal, a subspace-based system identification method is adopted to establish the transfer function relationship between the two in the frequency domain, and obtain the amplitude frequency characteristic curve and phase frequency characteristic curve of the acoustic-vortex coupling transfer function. The acoustic vortex coupling transfer function is verified by cross-correlation analysis. The coherence function between the output predicted by the input signal and the actual output is calculated. When the coherence function is greater than a preset threshold in the main frequency band, the effectiveness of the acoustic vortex coupling transfer function is confirmed.

7. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 6, characterized in that, Phase-locked analysis is performed on the time series markers and time coefficients. The matching window between the vortex structure shedding phase and the acoustic pressure pulsation dissipation valley is identified using a cross-correlation function or phase synchronization algorithm. Specifically, this includes: The timing of the vortex structure reaching the heat sink surface is used as the reference time axis, and the time coefficient is used as the signal to be analyzed. The time coefficient is subjected to Hilbert transformation to extract its instantaneous phase information, and the instantaneous phase change curve of the vortex structure shedding process over time is obtained. Wavelet packet decomposition was performed on the synchronously acquired far-field aerodynamic noise pressure data to extract the energy envelope in each frequency band. The time point when the sound pressure pulsation energy reached a local minimum was identified as the dissipation valley moment, and the instantaneous phase corresponding to each dissipation valley moment was recorded. Cross-correlation analysis is performed on the instantaneous phase curve of the vortex structure shedding and the instantaneous phase corresponding to the dissipation valley. The phase difference between the two is calculated. The phase shift when the cross-correlation function reaches its maximum value is determined by cyclic shift search. This phase shift is the optimal matching phase between the vortex structure shedding phase and the sound pressure pulsation dissipation valley. Centered on the optimal matching phase, and based on the complete life cycle of the vortex structure from generation to dissipation, an allowable phase fluctuation range is set, and a time window is constructed as the matching window.

8. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 7, characterized in that, The process of acquiring the multiphysics field of the motor involves inputting the expected heat flux density vector field and speed modulation spectrum command into a multi-frequency response model to solve for the optimal broadband speed trajectory and generate a drive signal. Specifically, this includes: Transient Joule heat distribution data of the stator winding is obtained by a distributed fiber optic grating sensor embedded in the stator slot of the motor; eddy current loss data of the permanent magnet is obtained by a thin-film thermocouple attached to the rotor surface; and contact stress fluctuation data of the bearing rolling elements is obtained by an acceleration sensor installed on the outer ring of the bearing combined with a dynamic model inversion. Multiple data are input into the multi-frequency response model, and the time sequence mark is used as the reference time axis. The optimal wideband speed trajectory is solved by a multi-objective optimization algorithm. A three-phase duty cycle sequence of a space vector pulse width modulation signal is generated based on the optimal wideband speed trajectory. The three-phase duty cycle sequence is then output to the motor driver to drive the cooling fan to run according to the optimal wideband speed trajectory.

9. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 8, characterized in that, The process of inputting multiple data into a multi-frequency response model, using the time series marker as the reference time axis, and employing a multi-objective optimization algorithm to solve for the optimal broadband rotational speed trajectory specifically includes: A multi-frequency response model with electromagnetic-thermal-mechanical coupling is constructed. The multi-frequency response model includes a three-dimensional thermal network sub-model describing the Joule thermal dynamics of the stator winding, an electromagnetic field sub-model describing the eddy current loss of the rotor permanent magnet, and a dynamic sub-model describing the contact stress fluctuation of the bearing rolling elements. Each sub-model achieves bidirectional data exchange through thermal-mechanical coupling boundary conditions. The expected heat flux density vector field distribution, steady-state speed reference value, small speed fluctuation parameters corresponding to the matching window, transient Joule heat distribution data, eddy current loss data, and contact stress fluctuation data are loaded as input parameters into the multi-frequency response model. Using the time sequence marker as the reference time axis, an optimization time window is set, and the multi-frequency response model is solved in the time domain within the optimization time window to obtain the heat sink entropy yield fluctuation time history curve, the sound power level fluctuation time history curve, the stator temperature rise fluctuation time history curve, and the bearing contact fatigue index time history curve. Using the minimum weighted integral of the four factors—root mean square of heat sink entropy yield fluctuation, root mean square of sound power level fluctuation, peak value of stator temperature rise fluctuation, and cumulative value of bearing contact fatigue index—as the objective function, a multi-objective particle swarm optimization algorithm is used to search for the optimal solution in a wide bandwidth, and the output is the optimal broadband speed trajectory containing fundamental frequency components, perturbation components, and notch filter components.

10. The method for smoothing the speed control of a cooling fan based on noise feedback according to claim 9, characterized in that, The collected actual operational deviations constitute a reward signal, which is then used to optimize the model through reinforcement learning and back-update the model to achieve self-evolution. Specifically, this includes: The heat sink entropy productivity field evolution trajectory, sound power level temporal fluctuation, stator temperature rise response curve, and bearing vibration characteristic data are collected after the actual operation of the optimal broadband speed trajectory. The deviations between the heat sink entropy productivity field evolution trajectory and the expected heat flux density vector field distribution, the deviations between the sound power level temporal fluctuation and the sound pressure pulsation dissipation valley, and the deviations between the stator temperature rise response curve and the stator temperature rise fluctuation target value in the optimal solution are calculated respectively. Abnormal impact peaks exceeding a preset threshold in the bearing vibration characteristic data are used as penalty factors to jointly constitute a multi-dimensional reward signal. The multidimensional reward signal is input into a deep reinforcement learning network. The deep reinforcement learning network uses the expected heat flux density vector field distribution, the amplitude and phase frequency characteristics of the acoustic vortex coupling transfer function, and the parameters of each sub-model of the multi-frequency response model as the state space, and the frequency and amplitude of the perturbation component in the speed modulation spectrum command, the center frequency and notch depth of the notch component, and the rate of change of the fundamental frequency component as the action space. The control strategy is iteratively optimized through the policy gradient method, and the optimized control strategy parameters are output. The optimized control strategy parameters are back-injected into the thermal wave dispersion relation prediction model, the matching window recognition algorithm, and the multi-frequency response model to dynamically update the boundary conditions of the thermal wave dispersion relation, correct the recognition threshold of the matching window, and adjust the coupling weights of each sub-model in the multi-frequency response model.

Images