A method and system for intelligent optimization of electromagnetic noise in motor controllers

By using multimodal sensor synchronous acquisition and intelligent optimization algorithms, the electromagnetic noise components of the motor system are separated in real time, which solves the problem of lack of specificity and accuracy in noise optimization in traditional methods, realizes adaptive optimization of electromagnetic noise, and improves the noise suppression effect of the motor controller.

CN123122387APending Publication Date: 2026-11-17XUZHOU SCI FICTION ELECTROMECHANICAL EQUIP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611628354.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing electromagnetic noise suppression methods for motor controllers cannot accurately separate and quantify the independent components of electromagnetic noise, resulting in a lack of specificity and precision in noise optimization measures. Furthermore, traditional methods are unable to cope with dynamic changes in motor operating conditions, leading to unstable suppression effects.

Method used

Multimodal sensors are used to synchronously acquire motor system signals. Combined with vibration signal-assisted independent vector analysis and electrical feature matching, electromagnetic noise components are separated in real time. Control parameters are adjusted online through intelligent optimization algorithms to achieve adaptive electromagnetic noise optimization.

Benefits of technology

It achieves high-precision, online separation and quantification of electromagnetic noise in motor systems under complex operating conditions, improves the pertinence and effectiveness of noise optimization, dynamically tracks changes in operating conditions, improves the subjective perception of the human ear, and breaks through the limitations of traditional passive suppression methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122387A_ABST
    Figure CN123122387A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent optimization method and system for electromagnetic noise in motor controllers, relating to the field of motor electromagnetic noise optimization technology. The method simultaneously acquires and preprocesses acoustic, vibration, and electrical signals from the motor system. By combining vibration-assisted independent vector analysis for blind source separation and fine identification based on electrical feature matching, electromagnetic noise components are separated from the mixed acoustic signals. Furthermore, by analyzing the characteristic parameters of the electromagnetic noise components, a joint optimization objective for reducing electromagnetic noise sound pressure level and improving sound quality is generated. Based on this objective, an intelligent optimization algorithm is used to perform online optimization within constraints, generating the optimal control parameter set for achieving electromagnetic noise optimization, which is then sent to the motor controller for execution. This solves the problem that traditional methods cannot accurately isolate independent electromagnetic noise components, leading to a lack of targeted noise optimization measures. It achieves active suppression of electromagnetic noise, improving the efficiency and accuracy of noise optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of electromagnetic noise optimization technology for motors, and in particular to an intelligent optimization method and system for electromagnetic noise of a motor controller. Background Technology

[0002] In high-performance motor drive systems, such as electric vehicles, precision industrial servos, and home appliances, the electromagnetic noise of the motor controller is one of the main factors affecting the overall sound quality of the machine. This noise is mainly caused by the high-frequency pulse width modulation (PWM) harmonics generated by the rapid switching action of power devices, which excite the mechanical structure. Its frequency is usually located in the range of human hearing sensitivity, manifesting as an unpleasant "whistling".

[0003] Existing noise suppression methods focus on passive optimization, such as improving mechanical structures, adding sound insulation materials, or empirically adjusting control parameters (e.g., fixed-frequency PWM, random PWM). However, these methods have significant limitations. Passive noise suppression increases costs and is difficult to handle broadband noise. The noise generated by motor systems during operation is a mixture of various noise sources, including electromagnetic, mechanical, and aerodynamic noise. Traditional methods lack specificity in the noise optimization process. Furthermore, fixed control parameters cannot adapt to the dynamic changes in load, speed, and temperature during motor operation, resulting in unstable electromagnetic noise suppression effects. There is an urgent need for an electromagnetic noise optimization method that can accurately separate electromagnetic noise from the motor controller online and perform adaptive parameter optimization based on the separation results. Summary of the Invention

[0004] This application provides an intelligent optimization method and system for electromagnetic noise of motor controllers, which solves the problem that traditional methods cannot accurately isolate and quantify the independent components of electromagnetic noise, resulting in a lack of specificity and precision in noise optimization measures and low efficiency. The method is based on sound source separation, which decouples the electromagnetic noise component from the mixed noise of the motor system in real time, and drives the control parameters of the motor system to be intelligently optimized online according to this component, thereby achieving accurate and adaptive active suppression of electromagnetic noise of the motor controller.

[0005] This application provides an intelligent optimization method for electromagnetic noise of a motor controller, including: Step S10: Synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller. Step S20: The multimodal sensor signal is preprocessed. By combining vibration signal-assisted independent vector analysis for blind source separation and fine identification based on electrical feature matching, the electromagnetic noise component is separated from the mixed acoustic signal acquired by the microphone array. Step S30: Analyze the characteristic parameters of the electromagnetic noise components, including the main peak frequency, sound pressure level and sharpness, and generate a joint optimization target for reducing the electromagnetic noise sound pressure level and improving sound quality in combination with the current operating conditions. Step S40: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The algorithm searches online within the constraints to generate the optimal control parameter set that achieves electromagnetic noise optimization and sends it to the motor controller for execution.

[0006] Furthermore, step S10, from the perspective of signal perception, constructs a synchronous observation system covering the entire physical chain of "electrical excitation—structural vibration—acoustic radiation". This system not only collects conventional acoustic signals, but also incorporates electrical quantities and vibration quantities into the scope of synchronous observation, providing multi-dimensional data support for subsequent precise decoupling.

[0007] From the perspective of sensor deployment, acoustic signals are collected using a microphone array arranged around the motor controller housing, with the array geometry designed according to the sound source localization requirements; vibration signals are collected using accelerometers installed on key mechanical structures, including key nodes along the vibration transmission path such as the surface of the motor controller housing, the power module mounting base, and the busbar support; electrical signals are directly collected from the phase current and phase voltage at the output of the motor controller.

[0008] From a synchronization mechanism perspective, the three types of signals mentioned above are synchronously sampled through a multi-channel data acquisition system driven by the same high-precision clock source. All channels share a unified sampling clock and trigger signal, ensuring that the acoustic, vibration, and electrical signals acquired at any given time reflect the system's operating state under the same physical condition. The data is encapsulated and transmitted in a timestamped frame structure.

[0009] Furthermore, step S20 is divided into three levels from a signal processing perspective: preprocessing, blind source separation, and feature recognition, to achieve refined extraction of electromagnetic noise step by step. The preprocessing layer performs standardization processes such as time alignment verification, bandpass filtering, and background noise removal on the synchronously acquired acoustic, vibration, and electrical signals. Short-time Fourier transforms are then performed on the acoustic and vibration signals to convert them to the time-frequency domain, so that the phase information in the complex spectrum can be used for subsequent separation.

[0010] In the blind source separation stage, from a technical perspective of blind source separation algorithms, this step employs the Independent Vector Analysis (IVA) algorithm and innovatively introduces vibration signals as an auxiliary observation channel. The acoustic time-frequency signal from the microphone array is used as the primary observation vector, and the time-frequency signal from the vibration sensor is used as an additional observation channel; the two are concatenated to form an extended observation vector. The IVA algorithm is then used to demix the extended observation vector, yielding multiple initially separated independent sound source components. The introduction of vibration signals leverages the strong correlation between structural radiated noise and vibration response, effectively enhancing the algorithm's ability to separate structural radiated noise and enabling it to distinguish between electromagnetic noise and airborne noise (such as fan aerodynamic noise).

[0011] At the feature recognition level, from the perspective of electrical feature matching, this step utilizes the unique electrical fingerprint of electromagnetic noise to achieve accurate identification. First, the PWM switching frequency, its integer harmonics, and the sideband modulation frequency generated by the fundamental frequency modulation of the motor are extracted from the preprocessed phase current and phase voltage signals to form the electrical feature fingerprint. Then, the amplitude square coherence function of each independent sound source component and the current signal at each characteristic frequency is calculated. If an independent sound source component has high coherence with the current signal at multiple characteristic frequencies, and the main peak of the spectrum is consistent with the main peak of the electrical fingerprint, then the component is identified as the target electromagnetic noise component.

[0012] Furthermore, step S30, from the perspective of sound quality evaluation, focuses not only on the loudness of the noise but also on its subjective perception characteristics. The electromagnetic noise components output in step S20 are quantitatively analyzed to extract acoustic and psychoacoustic parameters such as the main peak frequency, sound pressure level, and sharpness. Sharpness is calculated using the Zwicker model, amplifying the influence of high-frequency components in the noise spectrum on the harshness through a high-frequency weighting function.

[0013] From the perspective of constructing optimization objectives, this step combines the current motor operating conditions (speed, torque, fundamental frequency, DC bus voltage, etc.) and compares the aforementioned characteristic parameters with preset thresholds to determine the optimization priority. For example, when both sound pressure level and sharpness exceed the limit, both are set as joint optimization objectives; when only sharpness exceeds the limit, sharpness is taken as the primary optimization objective. Finally, a structured optimization instruction containing objective values, weighting coefficients, and constraints is generated as input for subsequent intelligent optimization algorithms.

[0014] Furthermore, from the perspective of intelligent optimization algorithms, step S40 models the electromagnetic noise optimization problem as a constrained multi-objective optimization problem. The decision variables are selected as the high-frequency PWM switching frequency of the motor controller, modulation strategy parameters (including random PWM bandwidth, random pulse position probability distribution parameters, and switching angles of specific harmonic elimination PWM, etc.), and dead-zone compensation parameters (including compensation coefficients or lookup tables). The optimization objective is to minimize the weighted acoustic index generated in step S30, and the constraints include system efficiency, torque ripple, current harmonic distortion rate, device temperature rise, switching losses, and variable boundaries.

[0015] From the perspective of online optimization, this step uses the Bayesian optimization algorithm for parameter optimization: First, a Gaussian process surrogate model is established using the current control parameters as the initial sample; then, by maximizing the acquisition function (such as the desired improvement) to balance development and exploration, the next set of control parameters to be evaluated is obtained; the parameters are sent to the motor controller for short-term operation evaluation, and the corresponding target values ​​are calculated using steps S20 and S30 to update the surrogate model; the above process is repeated until convergence, and finally the optimal set of control parameters that satisfies all constraints is output.

[0016] From an engineering execution perspective, after obtaining the optimal parameter set, it is encapsulated into a command frame recognizable by the controller and sent to the motor controller via a real-time communication link. The controller performs a smooth switch in the next PWM cycle or at the synchronous interruption point to avoid current surges caused by parameter abrupt changes. Simultaneously, a safety monitoring and rollback mechanism is set: if efficiency, temperature rise, torque, or other indicators exceed safety thresholds after switching, the system immediately rolls back to the previous safe parameter configuration.

[0017] This application also provides an intelligent optimization system for electromagnetic noise of a motor controller, including: Multi-sensor synchronous sensing module: used to synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller. Hybrid noise decoupling module: used to preprocess multimodal sensor signals, and separate electromagnetic noise components from the hybrid acoustic signals acquired by the microphone array by combining vibration signal-assisted independent vector analysis blind source separation and fine identification based on electrical feature matching; Electromagnetic noise characteristic analysis module: used to analyze the characteristic parameters of electromagnetic noise components, including peak frequency, sound pressure level and sharpness, and generate a joint optimization target to reduce electromagnetic noise sound pressure level and improve sound quality in combination with the current operating conditions; Intelligent optimization decision module: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The module seeks optimization online within the constraints, generates the optimal control parameter set to achieve electromagnetic noise optimization, and sends it to the motor controller for execution.

[0018] This application discloses the following technical effects: This application provides an intelligent optimization method and system for electromagnetic noise of a motor controller. Through multi-sensor fusion and sound source separation algorithms, it achieves high-precision, online separation and quantification of electromagnetic noise components in motor systems under complex operating conditions. This transforms the optimization target of motor noise from "total noise" to "specific electromagnetic noise," significantly improving the targeting and effectiveness of the optimization. The proposed method combines online electromagnetic noise analysis with online tuning of the motor controller, forming a complete active suppression closed loop. It dynamically tracks changes in operating conditions, achieving adaptive optimization and overcoming the limitations of traditional passive and open-loop suppression methods. Furthermore, the intelligent optimization target in this method not only includes sound pressure level but also introduces sharpness, a psychoacoustic index, to improve the subjective perception of the human ear. The algorithm can find the optimal solution under multiple objectives and constraints, achieving a comprehensive balance in noise optimization. The noise optimization framework of this method is not dependent on specific motor and controller types and can be adapted to various electromagnetic synchronous motors, induction motors, and other systems through model adaptation. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating an intelligent optimization method for electromagnetic noise of a motor controller, provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the electromagnetic noise separation process provided in an embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the structure of an intelligent optimization system for electromagnetic noise of a motor controller, provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1: This application provides an intelligent optimization method for electromagnetic noise of a motor controller, such as... Figure 1 As shown, the method includes: Step S10: Synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller.

[0024] In this embodiment, the multimodal sensor consists of an acoustic signal acquisition unit, a vibration signal acquisition unit, and an electrical signal acquisition unit. The signals of the multimodal sensor are completely synchronized in the time domain, accurately capturing the instantaneous motor operating state covering the complete process of electrical excitation, mechanical vibration, and sound radiation. Microphone arrays capture spatial sound field information for the separation and localization of mixed motor noise. Vibration sensors are installed at key nodes in the force-sound transmission path to directly capture structural sound sources and assist in the identification of structural radiated noise. Phase current and phase voltage signals obtain excitation characteristics from the motor system for coherence calculation and identification of electromagnetic noise components. The acoustic signal acquisition unit uses four microphones with a frequency range of 20Hz-40kHz, which are arranged around the motor controller housing and the adjacent motor body. The four microphones are arranged in a square array directly above the motor controller housing, about 20cm away from the housing surface. The geometric center of the array is aligned with the heat dissipation center of the controller power module, covering the main acoustic radiation surface of motor noise. The microphones are connected to the data acquisition card through a digital interface to ensure the anti-interference of signal transmission. When arranging the microphones, they should be avoided from being directly facing the air outlet of the cooling fan to reduce airflow interference.

[0025] The vibration signal acquisition unit employs three IEPE piezoelectric accelerometers, whose frequency response range covers the main excitation frequency band of the motor controller's electromagnetic noise. The sensors are mounted on key mechanical structures. The first sensor is mounted on the heatsink surface directly below the IGBT substrate of the motor controller's power module, used to directly acquire structural vibrations caused by the power module's switching excitation. The second sensor is mounted on the axial center of the motor housing near the stator core, used to acquire motor body vibrations. The third sensor is mounted on the DC bus capacitor bracket, used to acquire capacitor vibrations. All sensors are rigidly bonded to the measured surface using high-strength cyanoacrylate adhesive. The mounting surfaces are polished and cleaned to ensure that the coupling stiffness meets the high-frequency response requirements. In the electrical signal acquisition unit, the phase current signal uses two Hall current sensors, which are connected in series to the U-phase and V-phase output terminals of the motor controller, respectively, with a measurement range of 0 to ±150A; the phase voltage signal uses a high-voltage differential voltage probe with a bandwidth of 100MHz and an attenuation ratio of 50:1, which is connected across the UV and VW terminals of the controller, respectively; the W-phase current can be calculated using Kirchhoff's current law; all electrical sensors are connected to the data acquisition system via shielded cables to suppress electromagnetic interference.

[0026] During the multimodal sensor signal acquisition process, a uniform sampling rate is set for all sensor channels. The key operating event of the motor controller, namely the speed reaching the set value, is used as a global hardware trigger signal to start the data acquisition process, and acoustic, vibration and electrical signal acquisition is started synchronously. Sampling is performed according to the same clock beat to ensure that each sampling point corresponds to the state at the same moment, and a uniform timestamp is added to all data. The acquired multimodal sensor signals are packaged into data frames, and the timestamp, microphone array data, vibration acceleration data, three-phase current and three-phase voltage data are recorded.

[0027] Step S20: The multimodal sensor signal is preprocessed. By combining vibration signal-assisted independent vector analysis for blind source separation and fine identification based on electrical feature matching, the electromagnetic noise component is separated from the mixed acoustic signal acquired by the microphone array.

[0028] In this embodiment, as Figure 2 As shown, this step converts the raw synchronous multimodal sensor signal acquired by S10 into an independent and analyzable electromagnetic noise component, including signal preprocessing, initial separation of blind sources, and feature-guided fine identification. Step S21 involves preprocessing the multimodal sensor signals to transform the original synchronization signals into time-accurate, frequency-band-uniform, noise-free, and standard data suitable for the analysis domain, providing input for subsequent blind source separation and feature recognition. This includes the following sub-steps: Time alignment verification and data segmentation: Check the start point of the timestamps of acoustic, vibration, and electrical signals. When the motor system is initially powered on, inject a synchronous pulse signal that can be detected by all sensors. Locate the rising edge of this pulse in all sensor signals and calculate its relative time. If the time difference is less than one sampling interval, the time alignment is confirmed to be effective; otherwise, interpolation correction is performed. Divide long-term, continuous sensor signals into a series of analysis frames to process non-stationary signals and meet the requirements of subsequent time-frequency analysis. The frame length is set according to the time resolution and frequency resolution to ensure that it includes multiple PWM cycles and motor electrical cycles. The frame overlap rate is 50%. Use a Hanning window to weight each frame of data to reduce spectral leakage caused by segmentation.

[0029] Background noise subtraction and bandpass filtering: Obtain background noise data before the motor system is powered on, and calculate the average power spectral density of the background noise. For each analysis frame, the power spectral density of the acoustic and vibration signals is calculated. ,implement The corrected acoustic and vibration signals are obtained, among which As a subtraction factor, it is used to compensate for minor environmental changes. Through noise floor subtraction, the influence of fixed background noise such as fans and ambient noise is directly reduced in the frequency domain. In addition, for each frame of time-domain signal in the acoustic and vibration channels, an identical digital bandpass filter is applied with a high-pass cutoff frequency of 200 Hz to filter out low-frequency vibrations (such as chassis resonance and unbalanced forces) and acoustic noise that are unrelated to electromagnetic noise. The low-pass cutoff frequency is set to be slightly higher than the highest electromagnetic noise frequency of the target analysis to filter out high-frequency electronic noise and aliasing components.

[0030] Short-Time Fourier Transform and Complex Time-Spectrum Generation: A Short-Time Fourier Transform (STFT) is performed on the filtered acoustic and vibration signals of each frame. The window length is the same as the frame length to obtain sufficient frequency resolution to distinguish close harmonics. The number of overlapping points is consistent with the overlap rate when dividing the frames. For each analysis frame of each sensor, the STFT outputs a complex matrix containing the amplitude and phase information of each time-frequency point. Finally, the complex time-spectrum set of the microphone array acoustic signal and the complex time-spectrum set of the vibration signal of each vibration acceleration sensor are obtained.

[0031] Step S22: Using the preprocessed acoustic signal and the vibration signal as an auxiliary reference, a blind source separation algorithm based on independent vector analysis is used to initially separate several independent sound source components from the mixed acoustic signal. These components include multiple independent components related to the electromagnetic noise of the motor controller and structural vibration. The detailed steps are as follows: Using the pre-processed acoustic signal from the microphone array as the primary observation signal, an acoustic observation vector is generated. The time-frequency domain data from the vibration sensor is used as an additional observation channel and concatenated with the acoustic observation vector. For each time-frequency point, the complex spectral values ​​from the four microphones and the complex spectral values ​​from the three vibration sensors are concatenated to form a seven-dimensional complex extended observation vector. ,in and These represent the frequency and time frame index, respectively; this vector simultaneously contains information about sound field propagation and structural vibration. In practice, to balance the differences in dimensions and amplitudes among different sensors, the spectral amplitude of each channel can be normalized to have a mean of 0 and a variance of 1.

[0032] Suppose there exists Each statistically independent source signal, in this embodiment =4, corresponding to possible main sound sources including electromagnetic noise, mechanical bearing noise, cooling fan aerodynamic noise, and inverter capacitive noise; the extended observation vector is modeled as:

[0033] in, Representing frequency point The corresponding unknown mixing matrix, Represents the time-frequency vector of the source signal; Independent vector analysis aims to estimate a set of frequency-related unmixing matrices. This makes the output components :

[0034] And satisfy for each frequency The output components are independent of each other, and for each source signal, its components at different frequencies have vector correlation. This embodiment uses a natural gradient solution based on minimizing mutual information. In the separate solution process, the identity matrix is ​​first defined for each frequency point. Initialize the demixing matrix That is, the initial assumption is that the first four components of the acoustic signal are the source signal; then for each source signal... It is decomposed into multiple frequency components along the frequency range, forming the source vector. Next, a generalized Gaussian model is used to calculate the nonlinear evaluation function based on the statistical characteristics of the source vector. Update using the natural gradient algorithm :

[0035] in, This represents the update transformation amount of the unmixing matrix in the current iteration. This indicates a direct proportion, and is multiplied by the learning rate during actual updates; Represents the identity matrix. This represents the expectation operation. Represents the conjugate transpose, used to calculate correlation; nonlinear evaluation function. It reflects the statistical properties of the source vector, and

[0036] The guided algorithm finds the correct independent components; to avoid trivial solutions, after each iteration, the algorithm... Perform symmetric orthogonalization; when The separation iteration process ends when the value is less than a preset threshold, and the converged result is obtained. ; Using the converged unmixing matrix, the estimated time-frequency distribution of the source signal is calculated. Each row corresponds to the time spectrum of each isolated independent source; an inverse short-time Fourier transform is performed on the time spectrum of each isolated source to reconstruct the corresponding time-domain waveform; Finally, a set was obtained The time-domain signal components are used as candidate independent sound source components. These components are initially separated based on statistical independence and physical correlation. They include one or more components that are highly correlated with structural vibration, involving electromagnetic, mechanical and aerodynamic noise. The target electromagnetic noise component (caused by electromagnetic excitation and radiated through the structure) is implied in it, but it has not yet been clearly identified at this time.

[0037] Vibration signals are selected as an additional observation channel that is strongly correlated with structural radiated noise, providing a physical constraint for the separation algorithm. When searching for independent components, the algorithm tends to assign those components that show strong correlation in both acoustic and vibration signals to the same source vector. This allows noise propagated and radiated through structural vibration (such as shell radiated noise caused by electromagnetic excitation and mechanical impact noise) to be separated from the mixed sound more accurately and robustly, while noise propagated purely through air paths (such as far-field fan aerodynamic noise and ambient background noise) is separated into other components, assisting in the accurate identification of electromagnetic noise.

[0038] Step S23: Extract the PWM switching frequency and its harmonics and sideband modulation frequencies from the preprocessed phase current and phase voltage signals as electrical characteristics of the motor controller's electromagnetic noise. Perform coherence analysis and matching between the spectrum of each sound source component initially separated in step S22 and the electrical characteristics to identify independent sound source components with a coherence coefficient greater than a set threshold as target electromagnetic noise components. Then, use adaptive filtering to enhance the signal, obtaining a clean electromagnetic noise time-frequency domain signal and its characteristic parameters, including: High-resolution spectrum analysis is performed on the preprocessed phase current and phase voltage signals to identify frequency peaks in the power spectral density. The frequency with the highest amplitude is recorded as the PWM switching frequency. Integer multiples of the motor's fundamental frequency and the sideband frequencies generated by its harmonic modulation are also recorded. All the identified frequencies and their corresponding amplitudes and signal-to-noise ratios are used to construct an electrical feature vector of electromagnetic noise. For each candidate independent sound source component initially separated in step S22, calculate its power spectral density; for each characteristic frequency in the electrical eigenvector... Calculate candidate independent sound source components motor phase current signal The squared amplitude coherence function at this frequency point yields the coherence coefficient:

[0039] in, Indicates at the characteristic frequency point Candidate independent sound source components The coherence coefficient between the sound source component and the electrical characteristics ranges from [0,1]. The closer the value is to 1, the stronger the linear causal relationship between the sound source component and the electromagnetic noise. For cross power spectral density, and The power spectral density is the self-power spectral density; Set a coherence threshold Based on experience Statistical analysis of each candidate independent sound source component The number of frequency points with a coherence coefficient greater than a threshold across all characteristic frequencies. ,like satisfy , If the total number of characteristic frequencies is given, and the peak frequency of the candidate source matches the peak frequency of the electrical characteristic, then... Identified as a target electromagnetic noise component. Based on experience.

[0040] Step S30: Analyze the characteristic parameters of the electromagnetic noise components, including the main peak frequency, sound pressure level, and sharpness, and generate a joint optimization objective to reduce the electromagnetic noise sound pressure level and improve sound quality in combination with the current operating conditions.

[0041] In this embodiment, the peak frequency of the electromagnetic noise component is the PWM switching frequency; Calculate the sound pressure level of the electromagnetic noise component within the set frequency range of 150Hz-20kHz. :

[0042] in, Summing all Fourier transform spectral lines that cover electromagnetic noise. Indicates the weighting coefficient. Represents frequency The corresponding sound pressure mean square value, The reference sound pressure level is 20. ; The sharpness calculation method for electromagnetic noise components adopts the widely accepted Zwicker sharpness model, which, based on psychoacoustic principles, quantifies the subjective perception of the harshness of high-frequency noise components by the human ear. First, the 1 / 3 octave band spectrum of the electromagnetic noise is converted into sound intensity level, and then the specific sharpness corresponding to each frequency band is calculated. The formula is expressed as:

[0043] The unit of sharpness is: , which represents the sharpness of narrowband noise at a center frequency of 1 Bark at 1 kHz and 60 dB. Bark is a psychoacoustic scale that simulates the frequency filtering characteristics of the basilar membrane in the human ear. 0~24 Bark corresponds to the audio range of 0 Hz to 15 kHz. Indicates electromagnetic noise in the critical frequency band The instantaneous sound intensity is obtained by mapping the 1 / 3 octave spectrum of the noise signal onto the Bark scale; This represents a high-frequency weighting function, emphasizing the contribution of high-frequency components to the perceived harshness. This represents the total instantaneous sound intensity of electromagnetic noise. The normalization constant is used to make the sharpness of narrowband noise at 1kHz and 60dB exactly 1. The calibration coefficients are introduced.

[0044] In this embodiment, the current operating status, including the motor speed, is read from the motor controller. =3000rpm, load torque =50Nm, DC bus voltage =310V, motor fundamental frequency =200Hz. It was observed that the main peak frequency is exactly equal to the PWM switching frequency of 10kHz, and the sideband energy at 10kHz±200Hz is relatively strong, indicating that there is a large current ripple modulation; the main reason for the high sharpness is that the 20kHz and 30kHz harmonic components have high energy; therefore, it is determined that the optimization direction should include adjusting the switching frequency to move the position of the main peak and disperse the harmonic energy, and optimizing the dead zone compensation to suppress low-order harmonics.

[0045] Using the reduction of sharpness as an evaluation index for improving sound quality, the higher the sharpness, the more high-frequency harmonics are contained in the electromagnetic noise, which will produce an unpleasant subjective feeling even if the total sound pressure level is not high. Based on the reduction of electromagnetic noise sound pressure level and sharpness, a joint optimization objective for electromagnetic noise is established, forming a minimization problem of the value function.

[0046] Step S40: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The algorithm searches online within the constraints to generate the optimal control parameter set that achieves electromagnetic noise optimization and sends it to the motor controller for execution.

[0047] In this embodiment, a Bayesian optimization method is selected, and the weighted calculation formula of the sound pressure level and sharpness of the electromagnetic noise component is used as the objective value function. The decision variables consist of the PWM switching frequency, modulation strategy parameters, and dead-time compensation parameters. The modulation strategy parameters include the modulation index, random PWM parameters, and switching angles for specific harmonic elimination. The dead-time compensation parameters are a set of compensation coefficients related to the current polarity and amplitude, used to reduce voltage distortion caused by dead time. The constraints include performance, electrical, loss, and variable boundary constraints to maintain a balance between system efficiency and loss.

[0048] Using the current control parameters as initial samples, four additional samples are generated using Latin hypercube sampling, resulting in a total of five sets of parameters. Each set of parameters is sequentially sent to the controller, which is then allowed to run stably for 2 seconds. The corresponding parameters are then calculated using steps S20 and S30. and The actual values ​​of the constraints are recorded. Then, a Gaussian process surrogate model is built for each objective, using the Matérn5 / 2 kernel function, and regularization noise of 0.1 is added.

[0049] An improved acquisition function is employed to search for candidate points in the parameter space that maximize the function value. In each iteration, the L-BFGS-B optimizer searches from multiple random starting points to avoid getting trapped in local optima. The found candidate parameter combinations are then sent to the controller for evaluation. After obtaining the true target value, the sample set and Gaussian process model are updated. In this embodiment, the maximum number of iterations is set to 30.

[0050] Since there are multiple optimization objectives, this embodiment uses a linear weighted combination as a single objective, with the objective function... for:

[0051] in and The reference sound pressure level and sharpness are used to eliminate dimensional differences. For the constraints, the external point penalty function method is used, which adds the constraint violation as a square term to the objective function. In this way, the algorithm will move in the direction that satisfies the constraints during the search.

[0052] During the iterative optimization loop, based on the current surrogate model, the sampling function samples in the optimal region predicted by the model and the region with high uncertainty, respectively. The combination of control parameters to be tested is obtained through numerical optimization and sent to the motor controller. The evaluation window is run to calculate the new target value function value, which forms a new data pair and is added to the sample set. The surrogate model is trained with the updated sample set to optimize and generate new control parameters. The above steps are repeated until the maximum number of iterations is reached. The control parameters that satisfy all constraints and have the smallest target value function value are found in the sample set and are taken as the optimal control parameter set.

[0053] The optimal control parameter set, including the specific optimal PWM switching frequency, modulation strategy parameters, and dead-time compensation parameter values, is encapsulated into a recognizable instruction frame for the motor controller and sent to the controller. After receiving the instruction, the controller switches the relevant parameters to the optimal parameter set in the next PWM cycle.

[0054] Example 2: The intelligent optimization system for electromagnetic noise of a motor controller provided in this embodiment of the invention can execute the intelligent optimization method for electromagnetic noise of a motor controller provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method, such as... Figure 3 As shown, it includes the following modules: Multi-sensor synchronous sensing module: used to synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller. Hybrid noise decoupling module: used to preprocess multimodal sensor signals, and separate electromagnetic noise components from the hybrid acoustic signals acquired by the microphone array by combining vibration signal-assisted independent vector analysis blind source separation and fine identification based on electrical feature matching; Electromagnetic noise characteristic analysis module: used to analyze the characteristic parameters of electromagnetic noise components, including peak frequency, sound pressure level and sharpness, and generate a joint optimization target to reduce electromagnetic noise sound pressure level and improve sound quality in combination with the current operating conditions; Intelligent optimization decision module: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The module seeks optimization online within the constraints, generates the optimal control parameter set to achieve electromagnetic noise optimization, and sends it to the motor controller for execution.

[0055] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent optimization of electromagnetic noise of a motor controller, characterized in that, The method includes: Step S10: Synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller. Step S20: The multimodal sensor signal is preprocessed. By combining vibration signal-assisted independent vector analysis for blind source separation and fine identification based on electrical feature matching, the electromagnetic noise component is separated from the mixed acoustic signal acquired by the microphone array. Step S30: Analyze the characteristic parameters of the electromagnetic noise components, including the main peak frequency, sound pressure level and sharpness, and generate a joint optimization target for reducing the electromagnetic noise sound pressure level and improving sound quality in combination with the current operating conditions. Step S40: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The algorithm searches online within the constraints to generate the optimal control parameter set that achieves electromagnetic noise optimization and sends it to the motor controller for execution.

2. The method of claim 1, wherein the method further comprises: In step S10, the multimodal sensor consists of an acoustic signal acquisition unit, a vibration signal acquisition unit, and an electrical signal acquisition unit. The signals from the multimodal sensor are completely synchronized in the time domain, accurately capturing the instantaneous motor operating state covering the entire process of electrical excitation, mechanical vibration, and sound radiation. Microphone arrays capture spatial sound field information for the separation and localization of mixed motor noise. Vibration sensors are installed at key nodes in the force-sound transmission path to directly capture structural sound sources and assist in the identification of structural radiated noise. Phase current and phase voltage signals obtain excitation characteristics from the motor system for coherence calculation and identification of electromagnetic noise components.

3. The method of claim 1, wherein the method further comprises: Step S20, based on the acquired synchronous multimodal sensor signals, performs hybrid noise decoupling, specifically including the following steps: Step S21: Preprocess the multimodal sensor signal to transform the original synchronization signal into standard data that is time-accurate, has a uniform frequency band, is free from noise interference, and has been converted to a suitable analysis domain, providing input for subsequent blind source separation and feature recognition; Step S22: Using the preprocessed acoustic signal and the vibration signal as an auxiliary reference, a number of independent sound source components are initially separated from the mixed acoustic signal through a blind source separation algorithm based on independent vector analysis. These components include multiple independent components related to the electromagnetic noise of the motor controller and structural vibration. Step S23: Extract the PWM switching frequency and its harmonics and sideband modulation frequencies from the preprocessed phase current and phase voltage signals as electrical characteristics of the electromagnetic noise of the motor controller. Perform coherence analysis and matching between the spectrum of each sound source component initially separated in step S22 and the electrical characteristics to identify independent sound source components with a coherence coefficient greater than a set threshold as target electromagnetic noise components. Then, use adaptive filtering to enhance the signal and obtain a clean electromagnetic noise time-frequency domain signal and its characteristic parameters.

4. The method of claim 3, wherein the method further comprises: The blind source separation algorithm based on independent vector analysis includes: Using the preprocessed acoustic signal as the main observation signal, an acoustic observation vector is generated. The time-frequency domain data of the vibration sensor is used as an additional observation channel and concatenated with the acoustic observation vector to form an extended observation vector containing the complex spectrum of the acoustic signal and vibration signal at each time-frequency point. Suppose there are N statistically independent source signals. Independent vector analysis aims to estimate a set of frequency-dependent unmixing matrices such that the output components are represented as the product of the unmixing matrix and the extended observation vector, and satisfy that for each frequency, the output components are independent of each other, and for each source signal, its components at different frequencies are vector-dependent.

5. The method of claim 4, wherein the method further comprises: The classification solution process in the blind source separation algorithm includes: First, an identity matrix is ​​defined, and the unmixing matrix is ​​initialized for each frequency point. Second, for each source signal, it is decomposed into multiple frequency components along the frequency to form a source vector. Then, a generalized Gaussian model is used to calculate a nonlinear evaluation function based on the statistical characteristics of the source vector, and the unmixing matrix is ​​updated in combination with the natural gradient algorithm. To avoid trivial solutions, the unmixing matrix is ​​symmetrically orthogonalized after each iteration. When the updated value of the unmixing matrix is ​​less than a preset threshold, the separation iteration process ends, and the converged unmixing matrix is ​​obtained. Using the converged unmixing matrix, the estimated time-frequency distribution of the source signal is calculated to obtain the output components. Each row of the output components corresponds to the time spectrum of each separated independent source. An inverse short-time Fourier transform is performed on the time spectrum of each separated source to reconstruct the corresponding time-domain waveform, and finally a set of N time-domain signal components are obtained as candidate independent sound source components.

6. The method of claim 3, wherein the method further comprises: The coherence analysis and matching includes the following detailed steps: Spectral analysis is performed on the preprocessed phase current and phase voltage signals to identify frequency peaks in the power spectral density. The frequency with the highest amplitude is recorded as the PWM switching frequency. Integer multiples of the motor's fundamental frequency and the sideband frequencies generated by its harmonic modulation are also recorded. All the identified frequencies and their corresponding amplitudes and signal-to-noise ratios are used to construct an electrical feature vector of electromagnetic noise. For each candidate independent sound source component initially separated in step S22, calculate its power spectral density; for each characteristic frequency in the electrical feature vector, calculate the amplitude square coherence function of the candidate independent sound source component and the motor phase current signal at that frequency point based on the power spectral density, and obtain the coherence coefficient. Set a coherence threshold and count the number of frequency points where the coherence coefficient of each candidate independent sound source component is greater than the threshold at all characteristic frequencies. If the proportion of the number of frequency points greater than the threshold to the total number of frequency points is greater than the set threshold, and the main peak frequency of the candidate independent sound source component is consistent with the main peak frequency of the electrical characteristics, then it is identified as the target electromagnetic noise component.

7. The method of claim 1, wherein the method further comprises: In step S30, the sound pressure level of the electromagnetic noise component is calculated in a set frequency range : wherein summing all Fourier transform lines of the electromagnetic noise, denotes a weighting coefficient, denotes a frequency corresponding to the sound pressure squared, denotes that the reference sound pressure is 20 ; The formula for calculating the sharpness of electromagnetic noise components is as follows: where the unit of sharpness is , represents the sharpness of a 1 kHz, 60 dB narrowband noise with a center frequency of 1 Bark, Bark is a psychoacoustic scale that mimics the basilar membrane frequency filtering of the human ear, and 0-24 Bark corresponds to an audio range of 0 Hz to 15 kHz; represents the instantaneous sound intensity of the electromagnetic noise in the critical band , mapped onto the Bark scale from the 1 / 3 octave spectrum of the noise signal; represents a high-frequency weighting function that emphasizes the contribution of high-frequency components to the harshness perception; represents the total instantaneous sound intensity of the electromagnetic noise; Using the reduction of sharpness as an evaluation index for improving sound quality, a joint optimization objective for electromagnetic noise is established based on the reduction of electromagnetic noise sound pressure level and sharpness, forming a minimization problem of the value function.

8. The method of claim 1, wherein the method further comprises: In step S40, a Bayesian optimization method is selected, and the weighted calculation formula of the sound pressure level and sharpness of the electromagnetic noise component is used as the objective value function. The decision variables consist of the PWM switching frequency, modulation strategy parameters, and dead-time compensation parameters. The modulation strategy parameters include the modulation index, random PWM parameters, and switching angles for specific harmonic elimination. The dead-time compensation parameters are a set of compensation coefficients related to the current polarity and amplitude, used to reduce voltage distortion caused by dead time. The constraints include performance, electrical, loss, and variable boundary constraints to maintain a balance between system efficiency and loss.

9. The method of claim 8, wherein the method further comprises: The Bayesian optimization uses the control parameters (i.e., decision variables) and the corresponding electromagnetic noise sound pressure level and sharpness as a sample set, and establishes a Gaussian process regression model as a surrogate model for the joint optimization objective to predict the target value function value under any control parameters. During the iterative optimization loop, based on the current surrogate model, a sampling function is defined to sample in the optimal region predicted by the model and in the region with high uncertainty. The combination of control parameters to be tested is obtained through numerical optimization and sent to the motor controller. The evaluation window is run to calculate the new target value function value, which forms a new data pair and is added to the sample set. The surrogate model is trained with the updated sample set to optimize and generate new control parameters. The above steps are repeated until the maximum number of iterations is reached. The control parameters that satisfy all constraints and have the smallest target value function value are found in the sample set and are taken as the optimal control parameter set.

10. An intelligent optimization system for electromagnetic noise of a motor controller, characterized in that, The system is used to implement the intelligent optimization method for electromagnetic noise of a motor controller according to any one of claims 1-9, and the system comprises: Multi-sensor synchronous sensing module: used to synchronously acquire multi-modal sensor signals of the motor system, including acoustic signals acquired by a microphone array arranged around the motor controller housing, vibration signals acquired by vibration sensors installed on key mechanical structures, and phase current and phase voltage signals output by the motor controller. Hybrid noise decoupling module: used to preprocess multimodal sensor signals, and separate electromagnetic noise components from the hybrid acoustic signals acquired by the microphone array by combining vibration signal-assisted independent vector analysis blind source separation and fine identification based on electrical feature matching; Electromagnetic noise characteristic analysis module: used to analyze the characteristic parameters of electromagnetic noise components, including peak frequency, sound pressure level and sharpness, and generate a joint optimization target to reduce electromagnetic noise sound pressure level and improve sound quality in combination with the current operating conditions; Intelligent optimization decision module: Based on the joint optimization objective, an intelligent optimization algorithm is adopted. The high-frequency pulse width modulation (PWM) switching frequency, modulation strategy parameters, and dead-zone compensation parameters of the motor controller are used as decision variables. The module seeks optimization online within the constraints, generates the optimal control parameter set to achieve electromagnetic noise optimization, and sends it to the motor controller for execution.