A machine learning based method for reducing noise in electromechanical devices

Abstract

The application discloses a kind of based on machine learning electromechanical equipment noise reduction method, it is related to electromechanical equipment noise reduction field, including: the sound pressure and vibration signal when electromechanical equipment is operated are collected, and the operating state data of electromechanical equipment is collected;Noise characteristics are sparsified reconstruction;And the sound source positioning is carried out to noise characteristics after reconstruction using beam forming algorithm, obtain noise distribution chart;Noise reduction model based on deep learning is constructed, obtain the noise signal after noise reduction;The frequency in the preset frequency range of noise signal is extracted by peak search algorithm;The operating state data collected is analyzed using time-varying linear prediction coding algorithm TVLPC, obtain the noise frequency offset caused by electromechanical equipment operating state exception;Optimal solution is generated using particle swarm algorithm;According to the optimal solution of noise control, extract noise main frequency component to separate, obtain multiple phase inversion noise control signal.The precision of noise reduction control is improved in the application, compared with the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of noise reduction of electromechanical equipment, and in particular to a noise reduction method for electromechanical equipment based on machine learning. BACKGROUND

[0002] In recent years, with the increasing demand for indoor environmental comfort in modern buildings, the noise problems caused by the operation of building electromechanical equipment such as heating, ventilation and air conditioning, elevators, and water pumps have become increasingly prominent. These equipment noises not only reduce the indoor sound environment quality and affect people's work, study and life, but also may cause a decrease in the sound insulation and sound absorption performance of buildings, and even lead to noise nuisance problems. Therefore, how to effectively control the noise of building electromechanical equipment and improve the acoustic performance of buildings has become one of the key problems to be solved in the field of modern building acoustic environment design and control.

[0003] Traditional noise control of building electromechanical equipment mainly adopts passive noise reduction technology, such as installing soundproof covers, sound barriers, and mufflers around the equipment. However, these passive noise reduction measures mainly target medium and high frequency noises, and have limited effect on low frequency noises. In addition, the noise of building electromechanical equipment usually has time-varying and non-stationary characteristics, and its frequency spectrum composition and amplitude will dynamically change with the changes of equipment working conditions and external interference. Traditional passive noise reduction technology is difficult to adapt to such dynamic noise characteristics, and the noise reduction effect is limited.

[0004] In order to overcome the above shortcomings, active noise control (ANC) technology has received more and more attention. ANC technology uses the principle of destructive interference of sound waves, and by applying a secondary sound wave with the same frequency and amplitude but opposite phase as the original noise in the noise field, the original noise can be reduced. Compared with passive noise reduction technology, ANC technology has the advantages of good low-frequency noise reduction effect and adaptability to dynamic noise. However, to achieve efficient and accurate ANC noise reduction, real-time and accurate feature extraction and analysis of equipment noise are required, and the control parameters need to be dynamically adjusted according to the noise characteristics, which puts high requirements on noise signal processing and control algorithms.

[0005] In addition, there are many types of building electromechanical equipment, and the noise spectrum is complex and variable, which poses challenges to the adaptability and robustness of the noise control system. At the same time, the fluctuation of equipment operating conditions also introduces the shift of noise frequency, further increasing the difficulty of ANC system design. Traditional ANC methods are difficult to meet the requirements of high precision, real-time, and adaptive control in complex noise environments. SUMMARY

[0006] In view of the low noise reduction precision of electromechanical equipment in the prior art, the application provides an electromechanical equipment noise reduction method based on machine learning, which generates a spatial distribution map of noise sources and constructs a deep learning model for noise reduction processing, focuses on analyzing the noise frequency offset caused by abnormal operation state of electromechanical equipment, and obtains the optimal control solution through particle swarm optimization, thereby improving the noise reduction control precision.

[0007] The purpose of the application is achieved by the following technical solutions.

[0008] The application provides an electromechanical equipment noise reduction method based on machine learning, which comprises the following steps: collecting sound pressure and vibration signals of electromechanical equipment during operation through acoustic sensors and fiber Bragg grating sensors, and collecting operation state data of the electromechanical equipment; wherein the operation state data includes rotation speed, vibration and temperature; the acoustic sensor includes a microphone; extracting features from the collected sound pressure and vibration signals to obtain noise features; performing sparse reconstruction on the noise features; and using a beam forming algorithm to locate the sound source of the reconstructed noise features to obtain a noise distribution map; constructing a deep learning-based noise reduction model, wherein the noise reduction model takes the noise distribution map as input and outputs a noise-free distribution map; using the trained noise reduction model to perform noise reduction processing on the noise distribution map to obtain a noise-reduced noise signal; performing frequency spectrum analysis on the noise-reduced noise signal, and extracting the frequency within the preset frequency range of the noise signal as the noise main frequency component through a peak value search algorithm; analyzing the collected operation state data using a time-varying linear predictive coding algorithm (TVLPC) to obtain the noise frequency offset caused by abnormal operation state of the electromechanical equipment; taking the noise main frequency component and the noise frequency offset as input, and using a particle swarm algorithm to generate the optimal solution of noise control of the electromechanical equipment; according to the optimal solution of noise control, the noise main frequency component is extracted for separation to obtain multiple phase-inverted noise reduction control signals for noise reduction control.

[0009] The fiber Bragg grating (FBG) sensor is a sensor based on fiber Bragg grating technology. It uses the Bragg grating inside the optical fiber to reflect light of a specific wavelength, and detects the change of the reflected light wavelength to sense the change of external physical quantities. In the application, the fiber Bragg grating sensor is used to collect the vibration signal of the electromechanical equipment during operation. When the equipment vibrates, the optical fiber is subjected to strain, causing the grating period to change, and the reflected light wavelength to shift. By demodulating the wavelength shift of the reflected light, high-sensitivity measurement of the vibration of the electromechanical equipment can be achieved.

[0010] Beamforming is a spatial filtering technique that adjusts the amplitude and phase of signals received by an array of sensors to direct the main lobe of the array towards the target direction and the side lobes towards noise directions, thereby improving the gain of the target signal and suppressing noise interference. In this application, the beamforming algorithm is used for sound source localization of the reconstructed noise features. Using the noise signals received by the acoustic sensor array, an appropriate weight vector is designed to adjust the amplitude and phase of each channel signal, focusing the array response on the direction of the target noise source, thereby achieving spatial localization of the noise source. Common beamforming algorithms include delay-and-sum, Capon, MUSIC, etc.

[0011] Peak search algorithm is an algorithm for finding local maximum points in data sequences. In this application, the peak search algorithm is used to extract the main frequency components from the spectrum of the noise signal after noise reduction. By setting a certain search range and threshold condition, local peak points with amplitude exceeding the threshold are searched in the spectrum data, and the corresponding frequency values are extracted as the main frequency components of the noise. Common peak search algorithms include derivative-based methods, sliding window methods, wavelet transform methods, etc. Peak search can quickly locate important components in the spectrum, providing frequency domain features for subsequent noise control.

[0012] Time-varying linear predictive coding (TVLPC) is an adaptive speech coding algorithm that linearly predicts the short-time stationary characteristics of signals to achieve signal compression and parameter extraction. In this application, the TVLPC algorithm is used to analyze the operating state data of mechanical and electrical equipment and extract the noise frequency offset caused by state abnormalities. By autoregressive modeling of state data such as speed, vibration and temperature, time-varying linear prediction coefficients are obtained, and then the noise frequency offset caused by state abnormalities is estimated according to the changes in the prediction coefficients. The TVLPC algorithm can track the time-varying characteristics of the signal and adaptively extract the state abnormality features to provide compensation for noise control.

[0013] Furthermore, feature extraction is performed on the collected sound pressure and vibration signals to obtain noise features, including: using edge computing, performing time-frequency domain analysis on the sound pressure and vibration signals through short-time Fourier transform to extract statistical features; among which, statistical features include mean, variance, and kurtosis; using edge computing, performing multi-scale decomposition of the sound pressure and vibration signals using a wavelet packet decomposition algorithm to obtain wavelet packet coefficients; calculating the energy of each wavelet packet coefficient to obtain wavelet packet energy features; using edge computing, filtering the spectrum of the sound pressure and vibration signals through a Mel frequency filter bank, taking the logarithm of the filtered spectrum, and then performing a discrete cosine transform to obtain Mel frequency cepstral coefficients (MFCCs), which serve as spectral envelope features; and using a feature-level fusion algorithm to concatenate the statistical features, wavelet packet energy features, and spectral envelope features to obtain noise features.

[0014] Furthermore, to generate a noise distribution map, noise features are first classified. The collected noise feature data is then preprocessed to remove outliers and invalid data, resulting in a standardized noise feature sample set X = {x1, x2, ..., x...}. N}, where x i Let N be the i-th noisy feature sample, and N be the total number of samples. The K-means clustering algorithm is used to classify the noisy feature samples. First, K samples are randomly selected as initial cluster centers. Then, the sum of squared distances from the samples to the cluster centers is iteratively optimized to minimize the sum of squared distances, updating the cluster centers and sample class labels until convergence. The clustering result can be represented as C = {C1, C2, ..., C...} K}, where C k Let C be the k-th noise feature category, where k = 1, 2, ..., K. Each category C... k It contains a set of noise feature samples, denoted as Where N k denoted as the number of samples for the k-th type of noise feature.

[0015] Learn the category dictionary, for each noise feature category C k A complete dictionary D has been constructed. k The initial dictionary can be obtained by randomly selecting N. k Each sample can be used as a dictionary atom, or a basic dictionary such as DCT can be used. The K-SVD algorithm is used to process the dictionary D. k Training and optimization are performed. K-SVD minimizes the reconstruction error through two steps: alternating iterative sparse encoding and dictionary updates. Sparse encoding: a fixed dictionary D. k For each sample x ki Perform sparse representation, i.e., solve Constraints, ||a ki ||0≤T, thus obtaining the sparse coefficient a kiCommonly used sparse representation algorithms include OMP, LARS, etc. Dictionary update: fixed sparse coefficient a ki , update each atom d kj of the dictionary. Decompose the residual matrix by SVD kj , solve d , so that kj , update d kj and a k (j). Repeat sparse coding and dictionary update until the reconstruction error converges or the maximum number of iterations is reached. Get the optimal dictionary D k and sparse coefficient A k1 ={a k2 ,a kNk ,.....,a K} of the kth noise feature.

[0016] Joint dictionary sparse representation, splice the K optimal dictionaries into a joint dictionary D = [D1, D2,..., D r ], used to represent newly collected noise features. For newly collected noise features x, sparse representation is performed under the joint dictionary D by OMP algorithm, solving under the constraint condition: ||a||0≤T, to get the sparse coefficient a. Use the sparse coefficient a and the joint dictionary D to reconstruct the noise feature x, to get the reconstructed noise feature x m =Dα.

[0017] Noise source spatial positioning, according to the spatial coordinates (x m ,y m ,z r ) of the M microphones in the acoustic sensor array, m = 1, 2,..., M, construct the element position matrix A and the direction vector of the microphone array, where is the target direction's pitch angle and azimuth angle. Use the geometric information of the microphone array to construct the direction matrix H and the steering vector d of the beamforming algorithm. The direction matrix H is designed according to the expected spatial response, commonly used are delay sum, minimum variance distortion, etc. The steering vector d is set according to the target direction, used to point to the region of interest. Spatial filtering is performed on the reconstructed noise feature x r . Frame x r , multiply each frame of data x H (t) by the direction matrix H and the steering vector d to get the filtered output y(t) = d r(t), t is frame index. y(t) is the noise signal of target direction, which enhances the component of target direction and suppresses the noise of other directions. The spatial filtering output y(t) is subjected to noise source positioning. The generalized cross-correlation function of y(t) between each microphone pair is calculated using GCC algorithm, and the TDOA of each microphone pair is obtained by searching the cross-correlation peak value. Based on the TDOA observation, the positioning equation set of the spatial position of the noise source is constructed. Assuming that the noise source is located at (x s ,y s ,z s ), the TDOA of microphones i and j is τ ij , then it satisfies

[0018] where c is the speed of sound. The TDOA of each microphone pair is substituted to obtain the nonlinear equation set with (x s ,y s ,z s ) as unknown quantities. The least square (LS) algorithm is used to solve the positioning equation set. The nonlinear equation set is converted into a linear equation set, and the LS estimate value of (x s ,y s ,z s ), i.e., the spatial coordinates of the noise source, is obtained by minimizing the error sum of squares.

[0019] The noise distribution map is generated, the noise source coordinates (x s ,y s ,z s ) positioned are matched with the three-dimensional model of the electromechanical equipment, and the position of the noise source is labeled. The spatial interpolation algorithm, such as Kriging interpolation or radial basis function interpolation, is used to map the discrete noise source position coordinates to the continuous space of the three-dimensional model. According to the interpolation result, the isosurface or cloud chart of noise distribution is generated, and the distribution of noise in different regions of the equipment is intuitively displayed.

[0020] Further, the peak search algorithm is used to extract the frequency in the preset frequency range of the noise signal as the noise main frequency component. Specifically, the noise signal after noise reduction is subjected to frame processing. Assuming that the sampling frequency is f s , the frame length is N, and the frame shift is M, the noise signal after frame processing can be expressed as x(n, m), where n = 0, 1,..., N-1 is the time index within the frame, and m = 0, 1,..., M-1 is the frame index. The Hamming window is added to the noise signal after frame processing. The Hamming window function can be expressed as The noise signal after windowing is

[0021] x w(n, m) = x(n, m) x w(n). Fast Fourier transform (FFT) is performed on the windowed noise signal. The FFT point number is N, and the frequency spectrum X(k, m) of the noise signal is obtained, where k = 0, 1,..., N-1 is the frequency index. The frequency resolution is f s / N.

[0022] The peak value is searched, and the main frequency component search range of the noise signal spectrum is set according to the rotating speed f r of the electromechanical device. The multiple frequency of f r is taken as the center frequency of the search range, and the multiple order is L. The center frequency of the lth multiple frequency is

[0023] lf r , l = 1, 2,..., L. With the center frequency lf r as the reference, the search bandwidth is set to B, and the search range of the lth multiple frequency is

[0024] In each search range, a peak value search algorithm is used to detect the local peak value point. Commonly used peak value search algorithms include the first-order difference method, the parabolic interpolation method, etc. The peak value point frequency index searched is set as k

[0025] k l (p), p = 1, 2,..., P0, P0 is the peak value point serial number, and k l (p) corresponds to the peak value amplitude |X(k l (p), m). According to the peak value amplitude size, the first P0 peak value points are selected, and the corresponding frequency values f are extracted as the candidate noise main frequency components.

[0026] The frequencies are merged, the frequency difference between the candidate noise main frequency components is calculated, and the frequency difference matrix D(p, q) = |f l (p)-f l (q), p, q = 1, 2,..., P0 is obtained. When D(p, q) < T, it is considered that the candidate main frequency components corresponding to f l (p) and f l (q) can be merged. For the mergable candidate main frequency components, the frequency average value is taken as the merged main frequency, and the frequency value and the number of the candidate main frequency components are updated. The number of the updated candidate main frequency components is denoted as P1, and the frequency value is denoted as f c (p), p = 1, 2,..., P1.

[0027] The ESPRIT frequency is estimated, and the signal subspace of the ESPRIT algorithm is constructed by using the spectrum peak value corresponding to the updated candidate main frequency components. The dimension of the signal subspace is equal to P1, and the corresponding spectrum peak value vector is

[0028] s = [X(k c (1),m), X(k c (2),m),..., X(k c (P1),m)] T , where k c (p) is the frequency index corresponding to f c (p). The spectrum part except the spectrum peak corresponding to the candidate dominant frequency component is taken as the noise subspace. According to the principle of ESPRIT algorithm, the modified frequency value of the candidate dominant frequency component can be estimated by solving the rotational invariance of the signal subspace. The modified frequency value is denoted as f r (p), p = 1, 2,..., P1.

[0029] The dominant frequency component is selected, the modified frequency value f r (p) is sorted, and the P2 largest frequency values are selected as the final noise dominant frequency components. The value of P2 can be flexibly set according to the spectrum characteristics of the noise signal and the noise reduction requirement, and generally should be less than or equal to P1. The final determined noise dominant frequency component frequency is denoted as f f (p), p = 1, 2,..., P2. At the same time, the amplitude and phase parameters of the corresponding frequency component estimated by the ESPRIT algorithm are output, which are used for subsequent reconstruction of the noise reduction control signal. In this application, the peak search and frequency merging can preliminarily determine the candidate dominant frequency component, reduce the data amount, and reduce the operation complexity. The ESPRIT algorithm further improves the frequency estimation accuracy on the basis of the candidate dominant frequency. Finally, through sorting and screening, the most significant dominant frequency component in the noise signal is obtained.

[0030] Further, the noise frequency offset caused by the abnormal running state of the electromechanical equipment is obtained, including: modeling the collected running state data by using a time-varying linear prediction coding algorithm TVLPC, describing the time-varying characteristics of the state data by using an AR model, estimating the time-varying coefficients of the AR model by using a Levinson-Durbin recursive algorithm, and obtaining a TVLPC model of the state data; wherein the AR model is a short name of Auto-Regressive model, which is a commonly used time series prediction model. In this application, the AR model is used to describe the time-varying characteristics of the running state data of the electromechanical equipment. The AR model assumes that the state data at the current time can be obtained by linear combination of the state data at the previous time and noise, and the mathematical expression is as follows: where x(n) is the state data at the n th time, a i (n) is the i th AR coefficient, p is the model order, and e(n) is the prediction error. The AR model describes the time series correlation and evolution law of the state data by estimating the optimal coefficient, so as to realize the prediction of the future state.

[0031] wherein the Levinson-Durbin recursive algorithm

[0032] The Levinson-Durbin recursive algorithm is an efficient method for estimating the parameters of an AR model. In the present application, the algorithm is used to estimate the coefficients of an AR model that describes the time-varying characteristics of the operating state data of a mechanical and electrical device. The Levinson-Durbin algorithm utilizes the Yule-Walker equations and the recursive structure of the AR model to obtain the optimal estimates of the AR coefficients through iterative calculations. The basic steps are as follows:

[0033] Calculate the autocorrelation function R of the state data xx (i, n) = E[x(n)x(n-i)], i = 0, 1,..., p;

[0034] Initialize the recursive variables: a0(n) = 1, E0(n) = R xx (0, n), k = 1; recursively calculate:

[0035]

[0036] a k (n) = [a k-1,1 (n-1),..., a k-1,k-1 (n-1), 0] - k k (n) [0, a k-1,k-1 (n-1),..., a k-1,1 (n-1)]; update k = k + 1, repeat the steps until k = p + 1.

[0037] The final a p (n) = [a1(n), a2(n),..., ap(n)] obtained is the pth order AR model coefficient of the state data at the nth time. p

[0038] wherein the time-varying characteristic refers to the property that the statistical characteristics of a signal change over time. In the present application, the time-varying characteristic is used to characterize the non-stationarity and dynamic change of the operating state data of a mechanical and electrical device. The statistical parameters of state data with time-varying characteristics, such as mean, variance, autocorrelation function, etc., will change over time, showing a certain trend or periodicity. The AR model describes the time-varying characteristics of the state data by introducing time-varying coefficients a i (n), wherein the time dependence of the coefficients reflects the dynamic correlation and evolution of the state data at different times. By using the time-varying AR model, the time-varying patterns in the state data can be captured, and dynamic monitoring and prediction of the operating state of the device can be achieved.

[0039] ​The one-step prediction value and the multi-step prediction value of the state data are calculated by using the TVLPC model, and the prediction values are used as the future trend of the state data; the prediction error of the state data is calculated according to the collected running state data and the future trend obtained by the TVLPC model, and whether the running state of the electromechanical equipment is abnormal is judged by setting a prediction error threshold; for the running state judged as abnormal, the characteristic frequency component reflecting the running abnormality of the electromechanical equipment is obtained by calculating the spectrum and the power spectral density (PSD) of the prediction error sequence; the power spectral density (PSD) is a physical quantity describing the distribution of signal power in the frequency domain. In this application, PSD is used to analyze the frequency domain characteristics of the prediction error sequence when the electromechanical equipment is running abnormally. PSD represents the average power of the prediction error sequence in a unit frequency band, reflecting the contribution of different frequency components to error energy. For a discrete time sequence e(n), n = 1, 2,..., N, its PSD can be estimated by the periodogram method: where, is the discrete Fourier transform of e(n), k = 0, 1,..., N-1 is the normalized frequency. The role of PSD in anomaly detection is to identify significant frequency components related to abnormal states by observing the spectral distribution of the prediction error, and then extract abnormal characteristic frequencies for subsequent fault diagnosis and noise control.

[0040] The characteristic frequency component of the running abnormality is analyzed in association with the extracted noise main frequency component, the frequency offset of the characteristic frequency component of the running abnormality relative to the noise main frequency component is calculated, the mapping relationship between the running state abnormality of the electromechanical equipment and the noise frequency offset is established, and the noise frequency offset caused by the state abnormality of the electromechanical equipment is obtained.

[0041] Further, the noise main frequency component and the noise frequency offset are used as inputs, and a particle swarm algorithm is used to generate the optimal solution of the noise control of the electromechanical equipment, including: establishing an optimization problem of the noise control of the electromechanical equipment, taking the noise reduction effect as the optimization target, constructing a target function of the noise attenuation amount, taking the controller parameters as the optimization variables, establishing the mapping relationship between the optimization variables and the optimization target, and the controller parameters include the control gain, the filter order and the damping coefficient; the extracted noise main frequency component is used as the frequency constraint condition of the target function, and the calculated noise frequency offset is used as the correction coefficient of the target function; the particle swarm optimization algorithm (PSO) is used to solve the optimal solution set of the controller parameters; wherein the controller parameters are used as the position vector of the particle, and the noise attenuation amount is used as the fitness function of the particle, and the optimal solution is searched in the parameter space; according to the noise reduction demand of the user, the solution with the maximum noise attenuation amount is selected from the optimal solution set as the optimal solution of the controller parameters.

[0042] Further, according to the optimal solution of noise control, the noise main frequency components are extracted and separated to obtain a plurality of phase-inverted noise reduction control signals for noise reduction control, including: setting the parameters of the controller according to the optimal solution; tracking the extracted noise main frequency components by using a joint spectrum analysis algorithm, estimating the instantaneous frequency and amplitude of the noise main frequency components by short-time Fourier transform to obtain the dynamic spectrum of the noise main frequency; extracting the parameters of each main frequency component by using a matching pursuit algorithm according to the dynamic spectrum of the noise main frequency, the parameters of the main frequency component including frequency, amplitude and phase; synthesizing control signals with equal amplitude, same frequency and opposite phase of each main frequency component by using frequency modulation and phase inversion algorithm according to the extracted parameters of each main frequency component; inputting the synthesized control signals into the controller with set parameters to perform noise reduction control of the electromechanical equipment.

[0043] The joint spectrum analysis algorithm is an algorithm for analyzing the correlation of two signals in the time-frequency domain. In this application, the joint spectrum analysis algorithm is used to track the time-varying characteristics of the noise main frequency components. The algorithm calculates the short-time Fourier transform of the noise signal and the reference signal to obtain their time-frequency representation, then calculates their cross-power spectrum and cross-correlation coefficient at each time-frequency point to obtain the joint spectrum. The joint spectrum is a two-dimensional complex matrix, whose amplitude reflects the correlation degree of two signals at different time-frequency positions, and whose phase reflects their phase difference. By analyzing the peak value and continuity of the joint spectrum, the trajectory and parameters of the noise main frequency components in the time-frequency domain can be extracted.

[0044] The dynamic spectrum refers to the spectrum that changes with time, reflecting the time-varying characteristics of the frequency components of the signal. In this application, the dynamic spectrum is used to represent the changes of the instantaneous frequency and amplitude of the noise main frequency components. By performing short-time Fourier transform on the noise signal, the frequency spectrum at different times is obtained to form a two-dimensional time-frequency representation, which is the dynamic spectrum of the noise signal. The horizontal axis of the dynamic spectrum is time, the vertical axis is frequency, and the color or gray scale represents the amplitude of the frequency component. By analyzing the energy distribution and peak trajectory in the dynamic spectrum, the variation law of the frequency and amplitude of the noise main frequency components with time can be tracked, providing a basis for subsequent parameter extraction and control signal synthesis.

[0045] The matching pursuit algorithm is an algorithm for tracking the parameter changes of target signals. In this application, the matching pursuit algorithm is used to extract the parameters of each dominant frequency component, including frequency, amplitude and phase, from the dynamic spectrum of noise dominant frequencies. The algorithm realizes continuous tracking of dominant frequency components by searching for the most matching peak points between adjacent time frames of the dynamic spectrum. The specific steps include: searching for local peak points in the dynamic spectrum of the current time frame as candidate dominant frequency components; finding the component closest to the frequency and amplitude of the current candidate component in the dominant frequency component set of the previous time frame as the matching reference; calculating the similarity score between the candidate component and the reference component according to the matching degree of frequency and amplitude; selecting the candidate component with the highest similarity score as the dominant frequency component parameter at the current time. By iteratively executing the above steps, continuous tracking and extraction of the dominant frequency component parameters are realized.

[0046] The phase inversion algorithm is an algorithm for generating a control signal opposite in phase to a target signal. In this application, the phase inversion algorithm is used to synthesize a control signal with the same amplitude, frequency and opposite phase as the dominant frequency component of noise, for active noise reduction control. Specifically, for a sinusoidal noise signal If a control signal with the same amplitude, frequency and opposite phase is generated The two signals will cancel each other out after superposition, achieving noise signal attenuation or elimination. In practical applications, the phase inversion algorithm needs to generate a control signal corresponding to the extracted noise dominant frequency component in real time. The specific steps include: for each dominant frequency component, generate a sinusoidal signal with the same frequency according to its frequency parameter; adjust the amplitude of the sinusoidal signal according to the amplitude parameter so that it is equal to the amplitude of the dominant frequency component; perform a phase shift of π on the sinusoidal signal according to the phase parameter to obtain a control signal component with opposite phase; superimpose the control signals of each dominant frequency component to obtain the total noise reduction control signal. The control signal generated by the phase inversion algorithm can achieve precise interference and cancellation at the noise source, achieving the purpose of noise reduction.

[0047] Further, a deep learning-based noise reduction model is constructed, the noise reduction model taking a noise distribution map as input and a noise-free distribution map as output; the trained noise reduction model is used to reduce noise of the noise distribution map to obtain a noise signal after noise reduction, including: collecting noise signals of the electromechanical equipment under different working conditions, and performing time-frequency analysis and sound source positioning on the collected noise signals to obtain a noise distribution map corresponding to the noise signals; collecting noise-free signals corresponding to the noise signals, and performing time-frequency analysis and sound source positioning on the noise-free signals to obtain a noise-free distribution map corresponding to the noise-free signals; a noise reduction model based on an encoder-decoder structure is constructed, wherein the encoder extracts multi-scale features of the noise distribution map through a convolutional neural network (CNN), and the decoder reconstructs the extracted multi-scale features into a noise-free distribution map through a deconvolutional neural network and a skip connection; a weighted combination of reconstruction loss and adversarial loss is used to construct a loss function of the noise reduction model; the noise distribution map is taken as an input sample of training data, and the corresponding noise-free distribution map is taken as an output sample of training data, the training data is used to train the constructed noise reduction model, the loss function is calculated through forward propagation, and the model parameters are updated through back propagation; the noise distribution map corresponding to the newly collected noise signal is taken as input, the trained noise reduction model is used to extract features of the noise distribution map through the encoder, and the noise-free distribution map is reconstructed through the decoder to obtain a noise distribution map after noise reduction; the noise distribution map after noise reduction is converted into a noise signal after noise reduction through inverse Fourier transform.

[0048] Compared with the prior art, the application has the following advantages:

[0049] A noise feature dimension reduction and reconstruction method based on dictionary learning and sparse representation is proposed, an over-complete dictionary of noise features is adaptively constructed through clustering analysis and K-SVD dictionary training, and low-dimensional representation of noise features is realized by using sparse coding, effectively removing the redundant information of noise features. The reconstructed noise features are subjected to sound source positioning by combining a beamforming algorithm, a spatial distribution map of noise sources is obtained, and the spatial distribution characteristics of device noise are intuitively presented, providing a basis for noise source identification and positioning.

[0050] Aiming at the noise frequency offset problem caused by the change of the running state of the electromechanical equipment, a noise frequency offset estimation method based on time-varying linear predictive coding (TVLPC) is proposed. Through TVLPC modeling of the device state data, the future change trend of the device state is predicted, and the spectral characteristics of the state prediction error are calculated to extract the characteristic frequency components reflecting the device abnormalities. By correlating the noise main frequency and the abnormal characteristic frequency, a mapping relationship between the device state and the noise frequency offset is established, and adaptive estimation of the noise frequency offset is realized, providing a frequency offset compensation basis for subsequent control decisions.

[0051] A multi-objective optimization model of noise control of electromechanical equipment is constructed by using particle swarm optimization algorithm, taking noise reduction effect as the target, controller parameters as optimization variables, noise main frequency as frequency constraint, and noise frequency deviation as correction coefficient. By global optimization, the optimal solution set of noise controller parameters is obtained, the parameter self-tuning of noise control system is realized, the blindness and uncertainty of manual parameter adjustment are avoided, and the precision and efficiency of noise control are improved. BRIEF DESCRIPTION OF DRAWINGS

[0052] The present application will be further described in the manner of exemplary embodiments, which will be described in detail through the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same numbers represent the same structures, wherein:

[0053] Figure 1 is an exemplary flowchart of a machine learning-based electromechanical equipment noise reduction method according to some embodiments of the present application;

[0054] Figure 2 is an exemplary flowchart of generating noise features according to some embodiments of the present application;

[0055] Figure 3 is an exemplary flowchart of generating noise distribution map according to some embodiments of the present application;

[0056] Figure 4 is an exemplary flowchart of obtaining P2 noise main frequency components according to some embodiments of the present application. DETAILED DESCRIPTION

[0057] The method and system provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0058] As Figure 1As shown, the sound pressure and vibration signals of the electromechanical equipment during operation are collected by the acoustic sensor and the fiber Bragg grating sensor, and the operating state data of the electromechanical equipment is collected; wherein the operating state data includes rotation speed, vibration and temperature; the acoustic sensor includes a microphone; the collected sound pressure and vibration signals are subjected to feature extraction to obtain noise features; the noise features are subjected to sparse reconstruction; and the beam forming algorithm is used to locate the sound source of the reconstructed noise features to obtain a noise distribution map; a deep learning-based noise reduction model is constructed, the noise reduction model takes the noise distribution map as input and the noise-free distribution map as output; the trained noise reduction model is used to reduce the noise of the noise distribution map to obtain the noise signal after noise reduction; the noise signal after noise reduction is subjected to frequency spectrum analysis, and the peak search algorithm is used to extract the frequency in the preset frequency range of the noise signal as the noise main frequency component; the time-varying linear predictive coding algorithm TVLPC is used to analyze the collected operating state data to obtain the noise frequency offset caused by the abnormal operating state of the electromechanical equipment; the noise main frequency component and the noise frequency offset are taken as input, and the particle swarm algorithm is used to generate the optimal solution of the noise control of the electromechanical equipment; according to the optimal solution of the noise control, the noise main frequency component is extracted for separation to obtain multiple phase-inverted noise reduction control signals for noise reduction control.

[0059] As shown in Figure 2 The acoustic sensor array and the fiber Bragg grating (FBG) sensor array are arranged, wherein the acoustic sensor array is composed of multiple microphones for collecting sound pressure signals of the electromechanical equipment during operation; the FBG sensor array is composed of multiple FBG sensors for collecting vibration signals of the electromechanical equipment during operation. At the same time, rotation speed sensors, vibration sensors and temperature sensors are installed at key positions of the electromechanical equipment for collecting operating state data of the electromechanical equipment, including rotation speed, vibration and temperature. The sound pressure signals output by the acoustic sensor array and the vibration signals output by the FBG sensor array are collected in real time by using a data acquisition card and data acquisition software, and the collected analog signals are converted into digital signals. For the sound pressure signal, a 24-bit analog-to-digital converter (ADC) is used to ensure sufficient dynamic range and signal-to-noise ratio; for the vibration signal, a 16-bit ADC is used to meet the accuracy requirements of the vibration signal collection. According to the noise frequency range and vibration frequency range of the electromechanical equipment, a suitable sampling frequency is set, such as 44.1 kHz or 48 kHz, to avoid frequency aliasing and information loss.

[0060] The collected sound pressure and vibration signals are subjected to feature extraction to obtain noise features, including: using an edge computing node, performing short-time Fourier transform (STFT) on the collected sound pressure signals and vibration signals to convert time-domain signals into time-frequency domain representations. The time-domain signals are frame segmented using overlapping sliding windows, and each frame of signal is multiplied by a Hamming window or a Hanning window to reduce the influence of spectral leakage. According to the frequency range of the noise of the electromechanical equipment, a suitable window length and overlap ratio are selected, such as a window length of 1024 sampling points and an overlap ratio of 50%. Fast Fourier transform (FFT) is performed on each frame of signal to obtain a short-time spectrum, i.e. a time-frequency domain representation. Statistical features are extracted from the short-time spectrum, including mean, variance and kurtosis. The mean reflects the energy distribution of the spectrum, the variance reflects the dispersion degree of the spectrum, and the kurtosis reflects the peak degree of the spectrum. By calculating the amplitude mean, variance and kurtosis of each frequency point, a statistical feature vector of the spectrum is obtained. The statistical feature vectors of all frames are spliced to obtain a statistical feature sequence of the sound pressure signals and the vibration signals.

[0061] The collected sound pressure signals and vibration signals are subjected to wavelet packet decomposition (WPD) using an edge computing node. A suitable wavelet basis function, such as db4 or sym4, is selected for multi-scale decomposition of the signals. At each scale, the wavelet coefficients are further divided into two parts to obtain wavelet packet coefficients. According to the frequency range of the noise of the electromechanical equipment and the time resolution requirement, a suitable number of decomposition layers is selected, such as 4 layers or 5 layers. The energy feature is calculated for the coefficients of each wavelet packet node. The energy of the wavelet packet coefficients is measured using a two-norm or a variance to obtain the energy value of the wavelet packet node. The energy values of all wavelet packet nodes are combined to form a feature vector representing the energy distribution of the sound pressure signals and the vibration signals in different frequency bands and time scales, i.e. the wavelet packet energy feature.

[0062] The obtained short-time spectrum is subjected to Mel frequency filtering using an edge computing node. According to the auditory characteristics of the human ear, a Mel frequency filter bank is designed to filter and weight the spectrum. The center frequencies and bandwidths of the filter bank are distributed according to the Mel frequency scale, with dense filters in the low frequency part and sparse filters in the high frequency part. The filtered spectrum is taken as a logarithm to obtain a Mel frequency log power spectrum. Discrete cosine transform (DCT) is performed on the Mel frequency log power spectrum to obtain Mel frequency cepstral coefficients (MFCC). DCT concentrates the energy of the spectrum on the low-order coefficients, which is beneficial for extracting the envelope features of the spectrum. The first 12 to 20 order MFCCs are selected as the spectral envelope features, representing the spectral shape and dynamic changes of the sound pressure signals and the vibration signals. A feature-level fusion algorithm is used to splice the statistical features, wavelet packet energy features and spectral envelope features to obtain a noise feature vector of the sound pressure signals and the vibration signals. The feature vector contains multi-scale and multi-dimensional information in the time domain, frequency domain and time-frequency domain, and fully characterizes the characteristics of the noise of the electromechanical equipment.

[0063] The application fully excavates the time-frequency joint features of sound pressure signals and vibration signals by using time-frequency analysis methods such as STFT and WPD, and overcomes the limitations of traditional time-domain or frequency-domain analysis; multi-domain features such as statistical features, wavelet packet energy features and spectral envelope features are used to describe noise characteristics from multiple angles such as energy distribution, frequency composition and time scale, thereby improving the representation ability and discrimination of noise features; through a feature-level fusion algorithm, the multi-domain features are optimally combined, the deficiencies of single features are overcome, the complementarity of different features is fully utilized, and a joint representation of noise features is formed.

[0064] As shown in Figure 3 , the noise features are sparsely reconstructed; and a beamforming algorithm is used to locate the sound sources of the reconstructed noise features to obtain a noise distribution map, including: based on the extracted noise features, a K-means clustering algorithm is used to classify the noise features. The noise feature vector is taken as the input of clustering, and through iterative optimization, the intra-class distance is minimized and the inter-class distance is maximized to obtain K noise feature categories. The value of K is determined according to the complexity of the mechanical and electrical equipment noise and prior knowledge, and is usually 3 to 10. For each noise feature category, an overcomplete dictionary is constructed. The number of atoms of the overcomplete dictionary is greater than the dimension of the noise feature, and has redundancy and sparsity. The K-SVD algorithm is used to iteratively train the overcomplete dictionary, and the sparse coding and dictionary updating are alternately performed to minimize the reconstruction error, thereby obtaining the optimal dictionary corresponding to the noise feature category. The number of atoms of the dictionary and the number of iterations are balanced according to the reconstruction accuracy and computational efficiency.

[0065] A joint dictionary is constructed, and for K noise feature categories, the respective optimal dictionaries D k ,k=1,2,.....,K are obtained through the K-SVD dictionary learning algorithm, wherein the number of columns of D k is the number of dictionary atoms, and the number of rows is the dimension of the noise feature. The K optimal dictionaries are spliced by columns to form a joint dictionary D=[D1,D2,......,D K ], and the number of columns of the joint dictionary is the sum of the number of all dictionary atoms, and the number of rows is the same as that of a single dictionary. The joint dictionary is used for sparse representation, and for a newly collected noise feature y, the joint dictionary D is used for sparse representation to obtain sparse coefficients a, such that y≈Da and the number of non-zero elements of a is as small as possible. The orthogonal matching pursuit (OMP) algorithm is used to solve the sparse representation problem: initialize the residual vector r=y, the sparse coefficient vector a=0, the iteration number t=0, and the support set When t<T (T is the maximum number of iterations), ||r|| 2 >ε (ε is the reconstruction error threshold): calculate the inner product of the residual r and each column of the joint dictionary D, select the column index with the maximum absolute value of the inner product Add it to the support set Λ = Λ U {j}. Solve the non-zero elements of a on the dictionary columns corresponding to the support set Λ using least squares, such that Minimize. Update the residual r = y - D Λ a Λ , and the iteration number t = t + 1. Output the sparse coefficient a, where the index of the non-zero element is the element in the support set Λ, and the value of the non-zero element is the least squares solution. The number of non-zero elements of the sparse coefficient a is much smaller than the dimension of the noise feature y, realizing the low-dimensional representation of the noise feature.

[0066] Reconstruct the noise feature using the sparse coefficient. Reconstruct the noise feature y using the joint dictionary D and the sparse coefficient a, to obtain the reconstructed noise feature y rec . The reconstruction formula is: y rec = Da, that is, multiply the sparse coefficient a and the joint dictionary D to obtain the reconstructed noise feature y rec . The reconstructed noise feature y rec removes the redundant and noise components in the original noise feature y, and improves the quality and robustness of the feature. The present application uses K-means clustering and K-SVD dictionary learning to classify and sparsely reconstruct the noise feature, removes the redundant and noise components in the noise feature, extracts the essential mode of the noise feature, and improves the quality of the noise feature. Specifically, first, the K-SVD dictionary learning algorithm is used to learn the optimal dictionary for each noise feature category, and then all the optimal dictionaries are concatenated by columns to form a joint dictionary. For newly collected noise features, use the joint dictionary for sparse representation, use the OMP algorithm to solve the sparse representation problem, obtain the sparse coefficient of the noise feature under the joint dictionary, and realize the low-dimensional representation of the noise feature. Finally, the sparse coefficient and the joint dictionary are used to reconstruct the noise feature, remove the redundant and noise components in the noise feature, and improve the feature quality and robustness. The present application effectively mines the internal structure and mode of the noise feature through dictionary learning and sparse representation, and reduces the dimension of the noise feature.

[0067] According to the spatial position coordinates of each microphone in the acoustic sensor array, the array signal processing algorithm is used to calculate the array element position matrix and the direction vector of the microphone array. The array element position matrix represents the geometric arrangement of the microphone array, and the direction vector represents the directivity of the microphone array. Using the geometric information of the microphone array, the direction matrix and the steering vector of the beamforming algorithm are constructed for spatial filtering and sound source positioning. The reconstructed noise features are used as the input of the beamforming algorithm, and the direction matrix and the steering vector are used for spatial filtering of the noise features. By adjusting the direction matrix and the steering vector, the main lobe of the microphone array is aligned to the target direction, the interference in other directions is suppressed, the noise signal in the target direction is extracted, and the spatial filtering result is obtained. Spatial filtering enhances the target noise signal and provides high-quality input for sound source positioning. The spatial filtering of the reconstructed noise features using the beamforming algorithm enhances the noise signal in the target direction, suppresses the interference in other directions, and improves the accuracy and reliability of sound source positioning. Spatial filtering combines the spatial correlation and directivity of the noise signal, fully utilizes the spatial sampling advantage of the acoustic sensor array, and provides high-quality input for noise source positioning.

[0068] For the spatially filtered noise signal, the generalized cross-correlation (GCC) algorithm is used to calculate the cross-correlation function of each channel noise signal in the acoustic sensor array. By finding the maximum value of the cross-correlation function, the time difference of arrival (TDOA) of each channel noise signal is estimated. TDOA reflects the time difference of sound source arrival at different microphones and contains information about the spatial position of the sound source. TDOA is used as an observation, combined with the direction matrix of the microphone array, to construct a positioning equation set for the spatial position of the noise source. The positioning equation set establishes a mathematical relationship between TDOA and the spatial coordinates of the sound source, and is a nonlinear equation set. The least squares (LS) algorithm is used to solve the positioning equation set. By minimizing the sum of squared errors between the observed and estimated values of TDOA, the spatial position coordinates of the noise source are obtained. The LS algorithm has the characteristics of fast convergence and global optimization, and is suitable for real-time sound source positioning. Through the GCC algorithm and the LS algorithm, the TDOA of the noise signal is estimated, the spatial position coordinates of the noise source are solved, and accurate positioning of the noise source is achieved. Sound source positioning reveals the spatial distribution of the noise source, providing a basis for targeted noise control. Combined with the three-dimensional model of the mechanical and electrical equipment, a noise distribution map is generated, which visually displays the position and intensity of the noise source. According to the spatial position coordinates of the noise source, combined with the three-dimensional model of the mechanical and electrical equipment, spatial interpolation and mapping algorithms are used to generate a noise distribution map. The noise distribution map displays the position and intensity distribution of the noise source in the mechanical and electrical equipment space in the form of color or contour lines.

[0069] A deep learning-based denoising model is constructed, which takes a noise distribution map as input and outputs a noise-free distribution map. The trained denoising model is used to process the noise distribution map to obtain a denoised noise signal. The process includes collecting noise signals of mechanical and electrical equipment under different working conditions, including different rotating speeds, loads, and fault types. Time-frequency analysis and sound source localization are performed on the collected noise signals to obtain the corresponding noise distribution map. The noise distribution map represents the joint distribution characteristics of the noise signal in the time-frequency domain and the spatial domain, providing input samples for the training of the denoising model. Under the same working conditions, noise-free signals corresponding to the noise signals are collected as the target for denoising. Time-frequency analysis and sound source localization are performed on the noise-free signals to obtain the corresponding noise-free distribution map. The noise-free distribution map represents the time-frequency domain and spatial domain characteristics of the mechanical and electrical equipment under normal operation, providing output samples for the training of the denoising model.

[0070] A denoising model based on an encoder-decoder structure is constructed. The encoder uses a convolutional neural network (CNN) to extract multi-scale features of the noise distribution map through multiple convolution and pooling operations. The decoder uses an inverse convolutional neural network to gradually restore the extracted multi-scale features to a noise-free distribution map through multiple inverse convolution and upsampling operations. A skip connection is introduced between the encoder and the decoder to concatenate the shallow features of the encoder with the deep features of the decoder, preserving the detailed information of the noise distribution map and improving the reconstruction quality. An end-to-end denoising model is constructed to directly establish the mapping relationship from the noise distribution map to the noise-free distribution map, avoiding the complex noise modeling and decomposition process in traditional methods, simplifying the denoising process, and improving the denoising efficiency. The encoder-decoder structure uses convolutional neural networks to extract multi-scale features of the noise distribution map, fully exploiting the joint features of the noise signal in the time-frequency domain and the spatial domain, capturing local and global patterns of the noise, and fusing the shallow features of the encoder with the deep features of the decoder to preserve the detailed information of the noise distribution map and improve the fidelity of the denoised signal, avoiding information loss caused by excessive smoothing.

[0071] A weighted combination of reconstruction loss and adversarial loss is used to construct the loss function of the denoising model. The reconstruction loss measures the pixel-level difference between the denoised noise distribution map and the noise-free distribution map, calculated using mean square error (MSE) or mean absolute error (MAE). The adversarial loss introduces a discriminator network to distinguish between the generated noise-free distribution map and the real noise-free distribution map, calculated using binary cross-entropy. By adjusting the weights of the reconstruction loss and the adversarial loss, the fidelity and realism of the denoising are balanced. The reconstruction loss ensures that the denoised noise distribution map is close to the noise-free distribution map at the pixel level, maintaining the detailed features of the signal. The adversarial loss uses the discriminator network to identify the differences between the generated noise-free distribution map and the real noise-free distribution map, prompting the generated noise-free distribution map to be more realistic and have better visual quality.

[0072] The noise distribution map is taken as the input sample of the training data, and the corresponding noise-free distribution map is taken as the output sample of the training data to construct a pair of training data sets. The constructed denoising model is trained end-to-end by using the training data. In the forward propagation process, the noise distribution map is input into the encoder to extract multi-scale features, and then the noise-free distribution map is reconstructed through the decoder to calculate the reconstruction loss and the adversarial loss. In the backward propagation process, the parameters of the model are updated by using an optimization algorithm (such as Adam) according to the gradient of the loss function to iteratively optimize the denoising performance. The noise distribution map corresponding to the newly collected noise signal is taken as the input, and the trained denoising model is used for inference. The noise distribution map is extracted through the encoder to extract multi-scale features, and then the noise-free distribution map is reconstructed through the decoder to obtain the denoised noise distribution map. The denoised noise distribution map retains the key components of the noise signal and removes the noise interference, thereby achieving effective suppression of the noise.

[0073] The denoised noise distribution map is converted into a time-domain signal by inverse Fourier transform or inverse wavelet transform to obtain the denoised noise signal. Compared with the original noise signal, the noise component of the denoised noise signal is significantly weakened, and the signal quality is significantly improved, thereby providing a high-quality data basis for subsequent feature extraction and fault diagnosis.

[0074] The denoised noise signal is subjected to frequency spectrum analysis, and the frequency in the preset frequency range of the noise signal is extracted by a peak search algorithm as the noise main frequency component, including: the obtained denoised noise signal is subjected to frequency spectrum analysis. A Hamming window with a frame length of N is used to frame and window the noise signal to reduce the influence of spectrum leakage. The windowed noise signal is subjected to fast Fourier transform (FFT) to convert the time-domain signal into a frequency-domain signal to obtain a spectrum representation of the noise signal. The spectrum represents the amplitude distribution of the noise signal at different frequencies.

[0075] As Figure 4As shown, the main frequency component search range of the noise signal spectrum is set according to the rotating speed information of the electromechanical equipment. Specifically, the rotating speed information of the electromechanical equipment is obtained. The rotating speed data of the electromechanical equipment is collected in real time through a rotating speed sensor installed on the electromechanical equipment, such as an encoder or a Hall sensor. The rotating speed data can be expressed in revolutions per minute (RPM) or revolutions per second (Hz). According to the type and noise characteristics of the electromechanical equipment, the frequency multiplication relationship between the main frequency component and the rotating frequency is determined. Generally, the noise main frequency component of the electromechanical equipment appears at an integer multiple of the rotating frequency, such as 1 times, 2 times, 3 times, etc. Different types of electromechanical equipment may have different frequency multiplication characteristics, which need to be determined according to the structure, noise mechanism and experience of the equipment. The base frequency value corresponding to the rotating frequency is calculated. According to the obtained rotating speed data, it is converted into a frequency value as the base frequency of the noise signal spectrum. For example, if the rotating speed is 1500 RPM, the base frequency is 1500 / 60 = 25 Hz. The frequency multiplication range of the noise main frequency component is determined. According to the determined frequency multiplication relationship, the frequency multiplication range of the noise main frequency component is set, usually between 1 times and 10 times. For example, if the frequency multiplication relationship is 1, 2, 3 times, the frequency multiplication range of the noise main frequency component is [1, 2, 3]. The frequency range of the noise main frequency component is calculated. Multiply the base frequency value by the frequency multiplication range to get the frequency range of the noise main frequency component. For example, if the base frequency is 25 Hz and the frequency multiplication range is [1, 2, 3], the frequency range of the noise main frequency component is [25, 50, 75] Hz. Consider the uncertainty of the frequency range. Due to the fluctuation of the actual rotating speed, there is a certain uncertainty in the estimation of the base frequency and the main frequency component. In order to improve the robustness of the main frequency component search, a certain frequency tolerance can be introduced on the basis of the frequency range. For example, the frequency range can be expanded to [(1-ε)×f, (1+ε)×f], where f is the original frequency value and ε is the frequency tolerance coefficient, usually 0.05 to 0.1. The calculated frequency range of the noise main frequency component is taken as the target region for subsequent spectrum peak search. The spectrum peak search algorithm will perform local peak point detection in this frequency range to extract the main frequency component of the noise signal.

[0076] In the main frequency component search range, a peak search algorithm is used to perform local peak point detection. By sliding a window on the spectrum, the spectrum amplitudes in the window are compared to find the local maximum value point, which is the candidate noise main frequency component. Select the P0 local peak points with the largest spectrum amplitudes, extract the corresponding frequency values as the noise main frequency components screened out initially.

[0077] Frequency merging operation, calculate the frequency difference of the candidate noise main frequency component. For the P0 candidate noise main frequency components obtained, the frequency difference between them is calculated. Let the frequency value of the candidate main frequency component be The calculation formula of the frequency difference ΔF is: ΔF(i,j) = |f i -fj |,i,j = 1,2,......,P0, and i≠j; set a frequency merging threshold ε. According to the type and noise characteristics of the electromechanical device, set a frequency merging threshold ε to determine whether two candidate principal frequency components belong to the same principal frequency component. The selection of threshold ε needs to balance the frequency resolution and the frequency estimation error, and is usually 1 to 2 times the spectral resolution. Determine whether the frequency difference is less than the threshold. For each pair of candidate principal frequency components, determine whether their frequency difference ΔF(i,j) is less than the frequency merging threshold ε. If ΔF(i,j) < ε, it is considered that the two candidate principal frequency components are caused by the same principal frequency component and need to be merged. Merge the candidate principal frequency components. For the candidate principal frequency components that meet the merging condition, take the average of their frequency values as the frequency of the merged principal frequency component. Let the frequency values of the candidate principal frequency components that need to be merged be f , and the frequency of the merged principal frequency component be f merge , the calculation formula is:

[0078] Update the candidate principal frequency component set. Add the frequency of the merged principal frequency component to the candidate principal frequency component set, and delete the original merged candidate principal frequency component. Repeat until all candidate principal frequency components that meet the merging condition are processed. The frequency merging operation determines whether the candidate principal frequency components belong to the same principal frequency component by calculating the frequency difference between them, and averages the candidate principal frequency components that meet the merging condition to obtain the frequency of the merged principal frequency component. The frequency merging operation can eliminate the frequency estimation deviation of the principal frequency component caused by the spectral resolution limitation, and improve the accuracy of the principal frequency component frequency estimation.

[0079] The ESPRIT algorithm extracts the frequency of the noise principal frequency component. For the obtained noise signal spectrum, the autocorrelation matrix R of the signal is constructed. Let the noise signal spectrum be X(k), k = 0, 1,......, N-1, where N is the number of spectral points, and the calculation formula of the autocorrelation matrix R is: M is the order of the autocorrelation matrix, and * represents the complex conjugate. Perform eigenvalue decomposition on the autocorrelation matrix R to obtain eigenvalues λ1, λ2,......, λ M and corresponding eigenvectors v1, v2,......, v M . Arrange the eigenvalues in descending order, and arrange the corresponding eigenvectors accordingly. According to the size of the eigenvalues, select the eigenvectors corresponding to the first P largest eigenvalues to form the signal subspace Us. The value of P is usually the number of noise principal frequency components, which can be determined according to prior knowledge or by threshold judgment. Use the rotational invariance of the signal subspace Us to construct the frequency estimation matrix Φ. Let the dimension of Us be MxP, and the calculation formula of the frequency estimation matrix Φ is: where U s1and U s2 Us, and U denotes pseudo-inverse, Pseudo-Inverse, usually refers to Moore-Penrose Pseudoinverse, is a concept for processing non-square matrix (or singular matrix) matrix. Solve the eigenvalues μ1, μ2,..., μ P of the frequency estimation matrix Φ. i There is a relationship between the frequency f i of the noise dominant frequency component and the noise dominant frequency component. Where angle(*) represents the phase angle of a complex number, in radians. The frequency estimation value f i obtained is taken as the frequency estimation result of the noise dominant frequency component. The ESPRIT algorithm realizes high-resolution frequency estimation by constructing the signal subspace and using the subspace rotation invariance, and can accurately extract the frequency value of the noise dominant frequency component.

[0080] The ESPRIT algorithm constructs the autocorrelation matrix of the signal and performs eigenvalue decomposition, selects the signal subspace, constructs the frequency estimation matrix using the rotation invariance of the subspace, solves the eigenvalues of the matrix, and obtains the frequency estimation value of the noise dominant frequency component. The ESPRIT algorithm has high frequency resolution and noise interference resistance, and can accurately estimate the frequency value of the noise dominant frequency component, providing important frequency domain feature information for subsequent noise analysis and fault diagnosis. Sort the frequency values estimated by the ESPRIT algorithm, and select the P1 largest frequency values as the final noise dominant frequency components. The value of P1 can be set according to the type of electromechanical equipment and experience, usually 3 to 5 main frequency components can cover most of the noise energy. The selected noise dominant frequency components represent the main frequency characteristics of the noise signal after noise reduction.

[0081] To obtain the noise frequency offset caused by abnormal running state of electromechanical equipment, first model the running state data, preprocess the collected electromechanical equipment running state data, including denoising, normalization, etc., to obtain the normalized state data sequence x(n), n = 1, 2,..., N, where N is the data length. The time-varying linear prediction coding algorithm (TVLPC) is used to model the state data x(n). TVLPC describes the time-varying characteristics of the data through an autoregressive (AR) model, with model order p, expressed as where a i (n) is the time-varying AR coefficient, and e(n) is the prediction error.

[0082] The time-varying AR coefficients of the TVLPC model are estimated by using the Levinson-Durbin recursive algorithm. The algorithm iteratively computes the AR coefficients by minimizing the mean square value of the prediction error. Let R xx (i,n) be the autocorrelation function of the state data,

[0083] R xx (i,n) = E[x(n)x(n-i)], i = 0, 1,..., p, is solved by the Levinson-Durbin algorithm as follows:

[0084] According to the estimated time-varying AR coefficients a i (n), the TVLPC model of the state data is established as follows:

[0085] Specifically, the TVLPC algorithm proposes to describe the time-varying characteristics of the state data by using the time-varying AR model, in which the key parameter is the time-varying AR coefficient a i (n). The Levinson-Durbin algorithm is an effective method for estimating the time-varying AR coefficient a i (n) in the TVLPC model, which obtains the optimal AR coefficient estimate by recursively solving the Yule-Walker equation. By substituting the AR coefficient a i (n) estimated by the Levinson-Durbin algorithm into the TVLPC model, the time-varying AR model representation of the state data is obtained as follows: The TVLPC algorithm provides the framework of the time-varying AR model, and the Levinson-Durbin algorithm is an effective tool for solving the parameters of the model.

[0086] Predicting the running state, the one-step prediction value and the multi-step prediction value of the state data are calculated by using the established TVLPC model. m is the prediction step. The one-step prediction formula is The multi-step prediction is realized by iterative computation as follows: where x (n+m-i) represents the prediction value at the time n+m-i.

[0087] 2.2 Taking the prediction values and as the change trend of the state data at future time, the prediction values reflect the expected evolution law of the running state of the electromechanical equipment.

[0088] Detecting the running abnormality, the prediction error sequence of the state data is calculated as follows: The prediction error reflects the deviation between the actual collected state data and the predicted value of the TVLPC model. A prediction error threshold δ is set, and when |e(n)|>δ, the running state at the nth moment is judged to be abnormal; otherwise, it is judged to be normal. The threshold δ can be determined according to experience or statistical rules, such as δ=3σ, where σ is the standard deviation of the prediction error. The prediction error data corresponding to the running state moment judged to be abnormal is extracted, denoted as the abnormal prediction error sequence ea(n), n=1,2,.....,N a . a , where N a is the number of abnormal states.

[0089] The abnormal feature frequency is extracted, and the frequency domain analysis is performed on the abnormal prediction error sequence ea(n). The fast Fourier transform (FFT) is used to calculate the frequency spectrum E a (f) of ea(n). The frequency spectrum reflects the amplitude distribution of different frequency components in the prediction error sequence. The power spectral density PSD of the abnormal prediction error sequence is calculated.

[0090] PSD describes the power distribution of the signal at each frequency point, which can be estimated by the periodogram method:

[0091] According to the frequency spectrum and PSD, the significant frequency components of the abnormal prediction error sequence are extracted as the feature frequency reflecting the abnormal running of the mechanical and electrical equipment. By setting an amplitude threshold or a statistical test method, the frequency points with concentrated energy and related to the abnormal state are screened out, denoted as the abnormal feature frequency set F a ={f1,f2,.....,f M}, where M is the number of abnormal feature frequencies.

[0092] The noise frequency offset is calculated, and the extracted noise main frequency components are denoted as F0={F 01 ,F 02 ,.....,F 0K}, where K is the number of main frequency components. The correlation between each frequency in the abnormal feature frequency set F a and the noise main frequency component F0 is calculated using a correlation analysis method, such as the cross-correlation function or the coherence function. For each abnormal feature frequency f a in F i , i=1,2,....,M, find the noise main frequency F 0j with the largest correlation, j=1,2,.....,K. Calculate the frequency offset Δf i of f 0j relative to F i . i 0j ​, to obtain the offset of the i-th abnormal characteristic frequency relative to the noise main frequency. A mapping relationship table between the abnormal state of the electromechanical equipment and the noise frequency offset is established, containing abnormal characteristic frequencies f i , corresponding noise main frequencies F 0j , and frequency offsets Δf i . The mapping relationship links the abnormal state to the noise spectrum, providing a basis for subsequent noise control.

[0093] In this application, the time-varying linear prediction coding algorithm TVLPC is used to analyze the collected operating state data, and the noise frequency offset caused by the abnormal operating state of the electromechanical equipment is obtained. The TVLPC algorithm describes the time-varying characteristics of the state data through an autoregressive model, estimates the time-varying AR coefficients using the Levinson-Durbin recursive algorithm, and constructs a dynamic prediction model for the state data. Based on the TVLPC model, the predicted value of the state data is calculated, and the prediction error is used to determine whether the operating state of the electromechanical equipment is abnormal. For abnormal states, the frequency domain characteristics of the prediction error sequence are analyzed, and the characteristic frequency components reflecting the operating abnormality are extracted. The abnormal characteristic frequency and the noise main frequency component are correlated and analyzed, and the frequency offset is calculated, establishing a mapping relationship between the abnormal state of the electromechanical equipment and the noise frequency offset. Using this mapping relationship, the noise frequency offset can be estimated based on the operating state prediction error, providing a basis for subsequent fault diagnosis and predictive maintenance.

[0094] Taking the noise main frequency component and the noise frequency offset as inputs, the particle swarm algorithm is used to generate the optimal solution for electromechanical equipment noise control, including: establishing an optimization problem for electromechanical equipment noise control, taking the noise reduction effect as the optimization objective, and defining the objective function J(x) of the noise attenuation amount: where A n is the amplitude of the original noise signal, A c is the amplitude of the controlled noise signal, and x is the controller parameter vector. The controller parameters are selected as optimization variables, including control gains K p , K i , K d , filter order N, and damping coefficient ξ, forming the optimization variable vector x = [K p , K i , K d , N, ξ]. The mapping relationship between the optimization variable x and the optimization objective J(x) is established, usually through the cascade of the controller transfer function and the noise transfer function.

[0095] The noise main frequency component and the frequency offset are included in the constraint conditions of the optimization problem, and the noise main frequency component f is extracted to construct the frequency constraint condition: min(f i ) ≤ f ≤ max(f i .), where i = 1, 2,..., P1, f is the dominant component of the control post-noise signal. According to the calculated noise frequency offset Δf1, Δf2,..., Δf M , the modified term of the objective function is constructed: where i = 1, 2,..., M, w i is the weight coefficient of the frequency offset.

[0096] The particle swarm optimization (PSO) algorithm is used to solve the optimal solution set of the controller parameters. The particle swarm is initialized, and the position vector x i and the velocity vector v i of N particles are randomly generated, where i = 1, 2,..., N. The position vector x i represents the value of the controller parameters, and the velocity vector v i represents the search direction and step size of the parameters. The fitness value of each particle, i.e., the value of the objective function , is calculated, which represents the noise attenuation amount corresponding to the particle position. The individual optimal position p i and the global optimal position p g of each particle are updated. The individual optimal position p i represents the optimal solution found by the i-th particle in the search history, and the global optimal position p g represents the optimal solution found by the entire particle swarm in the search history. According to the individual optimal position and the global optimal position, the velocity vector and the position vector of each particle are updated:

[0097] where t is the iteration number, w is the inertia weight, c1 and c2 are the learning factors, and r1 and r2 are random numbers between 0 and 1. Repeat until the termination condition is met, such as reaching the maximum iteration number or the change in the fitness value being less than a threshold value. The optimal solution set obtained by the particle swarm search is output, including the optimal value of the controller parameters and the corresponding noise attenuation amount.

[0098] According to the noise reduction demand of the user, the optimal solution is selected from the optimal solution set. The solutions in the optimal solution set are sorted in descending order of noise attenuation. According to the user-specified noise reduction target, such as the target value of noise attenuation or the noise limit, the solution that meets the requirement is selected from the optimal solution set as the optimal solution of the controller parameters. If the user does not specify a specific noise reduction target, the solution with the largest noise attenuation is selected as the optimal solution by default. The present application uses the particle swarm algorithm to generate the optimal solution of the noise control of the electromechanical equipment by taking the noise main frequency component and the noise frequency offset as inputs. By establishing an optimization problem of noise control, taking the noise reduction effect as the target and the controller parameters as the optimization variables, taking the noise main frequency component as the frequency constraint condition and taking the noise frequency offset as the correction term of the objective function, a mathematical model of noise control is constructed. The particle swarm optimization algorithm PSO is used to search for the optimal solution set of the controller parameters in the parameter space through the position update and velocity update of the particles, and the parameter optimization of noise control is realized. According to the noise reduction demand of the user, the solution with the largest noise attenuation is selected from the optimal solution set as the final controller parameters, and the effective control of the noise of the electromechanical equipment is realized. The present application comprehensively utilizes the frequency domain characteristics of the noise signal and the time domain characteristics of the running state of the electromechanical equipment, and finds the optimal parameters of noise control through an intelligent optimization algorithm.

[0099] According to the optimal solution of noise control, the noise main frequency component is extracted for separation, and the electromechanical equipment noise reduction control is performed. First, the controller parameters are set. According to the optimal solution of noise control, the parameters of the controller are set. The optimal solution contains parameter information such as the optimal gain, order, and filter type of the controller. The parameter values in the optimal solution are assigned to the controller to adjust the structure and coefficients of the controller so that it achieves the optimal control performance. The type of the controller can be an adaptive filter, a feedforward compensator, a feedback controller, etc., which is selected according to the specific control strategy.

[0100] The noise main frequency component tracking is performed using the joint spectrum analysis algorithm. The extracted noise main frequency component signal x(t) and the reference signal y(t) are preprocessed. The reference signal can be a delayed version of the noise signal, an approximate model of the noise signal, or other related signals. The x(t) and y(t) are centered and normalized to remove the direct current component and amplitude influence of the signals, and zero-mean unit-variance signals x'(t) and y'(t) are obtained. The preprocessed signals x'(t) and y'(t) are divided into overlapping frames with a length of N, and the overlapping length between frames is L. Each frame is represented as x i '(n) and y i '(n), i is the frame number, and n is the time index within the frame, n=0,1,.....,N-1.

[0101] The time frame is calculated. For each frame signal xi '(n) and y i '(n) are subjected to short-time Fourier transform (STFT) to obtain their time-frequency representations X i (f,t) and Y i (f,t). The calculation formula of STFT is: where f is the frequency variable, t is the time variable, and w(n) is an analysis window function such as Hamming window or Gaussian window. According to the STFT results, the power spectrum P xx,i (f,t) and P yy,i (f,t) of each frame are calculated. xy,i (f,t). P xx,i (f,t) = |X i (f,t)| 2 , P yy,i (f,t) = |Y i (f,t)| 2 , P xy,i (f,t) = X i (f,t)Y i *(f,t), where * represents complex conjugate operation.

[0102] The joint spectrum C xy (f,t) of x(t) and y(t) is estimated using the power spectrum and cross power spectrum. where M is the total number of frames. The joint spectrum C xy (f,t) is a complex matrix, whose amplitude |C xy (f,t)| reflects the correlation degree of x(t) and y(t) at frequency f and time t, and the phase ∠C xy (f,t) reflects the phase difference of them at frequency f and time t.

[0103] The dominant frequency of the noise is tracked according to the amplitude |C xy (f,t)| of the joint spectrum. The local peak value of |C xy (f,t)| corresponds to the area with high correlation of x(t) and y(t) in the time-frequency domain, representing the time-frequency track of the dominant frequency component of the noise. Peak value search and threshold judgment are performed on |C xy (f,t)| to extract the significant dominant frequency component of the noise. Set the amplitude threshold T, when |C xy (f,t)| > T, the time-frequency position of the dominant frequency component of the noise is determined.(f, t) is considered to exist a significant dominant frequency component. The extracted noise dominant frequency components are analyzed for continuity and tracked. Since the noise dominant frequency components have continuity in time, the most likely trajectory connection is searched between adjacent time frames of the joint spectrum to form a complete dominant frequency trajectory. The tracked noise dominant frequency trajectory is smoothed to remove isolated points and short-time mutations to obtain a stable dominant frequency component variation curve.

[0104] According to the dynamic spectrum S k (t), the matching pursuit algorithm is used to extract the parameters of each dominant frequency component. The matching pursuit realizes the continuous tracking and parameter estimation of the spectrum peak by measuring the similarity between the spectrum peak at the current time and the spectrum peak at the previous time. For the kth dominant frequency component, the local peak point is searched in the dynamic spectrum S k (t), and the corresponding frequency f k (t), amplitude A k (t) and phase are extracted. The frequency parameter f k (t) is the peak frequency of the dynamic spectrum at time t; the amplitude parameter A k (t) is the amplitude value at the peak frequency; and the phase parameter can be estimated by continuous wavelet transform or Hilbert transform. The extracted parameter sequence is smoothed to remove noise and mutation points to obtain a continuous and smooth dominant frequency component parameter trajectory.

[0105] The control signal is synthesized. According to the extracted parameters of each dominant frequency component, the frequency modulation and phase inversion algorithm are used to synthesize a control signal with the same amplitude, the same frequency and the opposite phase as the dominant frequency component. For the kth dominant frequency component, the control signal corresponding to it is synthesized as The frequency modulation takes the instantaneous frequency of the dominant frequency component as the modulation frequency of the control signal, and the phase inversion introduces a phase shift of π based on the phase of the dominant frequency component to realize vibration cancellation. The synthesized dominant frequency control signals are superimposed to obtain the total noise reduction control signal The control signal is synchronized with the original noise dominant frequency component in frequency, equal in amplitude, and opposite in phase, thereby realizing accurate noise reduction control.

[0106] The output control signal is input to the controller with the set parameters. The controller generates a secondary vibration or sound wave that matches the noise dominant frequency component according to the input control signal to cancel and weaken the original noise. Through the action of the controller, active noise reduction control of the electromechanical equipment is realized, the influence of the noise dominant frequency component is reduced, and the vibration and noise performance of the equipment is improved.

Claims

1. A noise reduction method for electromechanical equipment based on machine learning, characterized in that, include: Acoustic sensors and fiber Bragg grating sensors are used to collect sound pressure and vibration signals during the operation of electromechanical equipment, and to collect operating status data of the electromechanical equipment; the operating status data includes speed, vibration and temperature; the acoustic sensors include microphones; Noise features are obtained by extracting features from the collected sound pressure and vibration signals; The noise features are sparsified and reconstructed; then, a beamforming algorithm is used to locate the source of the reconstructed noise features to obtain a noise distribution map. A deep learning-based denoising model is constructed, which takes a noise distribution map as input and a noise-free distribution map as output. The trained denoising model is then used to denoise the noise distribution map to obtain the denoised noise signal. The noise signal after noise reduction is subjected to spectral analysis, and the frequencies within a preset frequency range of the noise signal are extracted by the peak search algorithm as the main frequency components of the noise. The time-varying linear predictive coding algorithm (TVLPC) is used to analyze the collected operating status data to obtain the noise frequency shift caused by abnormal operating status of electromechanical equipment. The optimal solution for noise control of electromechanical equipment is generated by using the dominant noise frequency component and noise frequency offset as inputs and the particle swarm optimization algorithm. Based on the optimal solution for noise control, the dominant frequency component of the noise is extracted and separated to obtain multiple phase-reversed noise reduction control signals for noise reduction control.

2. The machine learning-based noise reduction method for electromechanical equipment according to claim 1, characterized in that: Feature extraction is performed on the collected sound pressure and vibration signals to obtain noise features, including: By utilizing edge computing, time-frequency domain analysis of sound pressure and vibration signals is performed through short-time Fourier transform to extract statistical features, including mean, variance, and kurtosis. Using edge computing, wavelet packet decomposition algorithm is used to decompose sound pressure and vibration signals at multiple scales to obtain wavelet packet coefficients; the energy of each wavelet packet coefficient is calculated to obtain wavelet packet energy characteristics; Using edge computing, the spectra of sound pressure and vibration signals are filtered through a Mel frequency filter bank. The logarithm of the filtered spectrum is then taken, followed by discrete cosine transform to obtain the Mel frequency cepstral coefficients (MFCCs), which serve as the spectral envelope features. A feature-level fusion algorithm is used to concatenate statistical features, wavelet packet energy features, and spectral envelope features to obtain noise features.

3. The machine learning-based noise reduction method for electromechanical equipment according to claim 2, characterized in that: The noise distribution map is obtained, including: Based on noise characteristics, a clustering analysis algorithm is used to classify the noise characteristics, resulting in K noise characteristic categories; For each noise feature category, an overcomplete dictionary is constructed, and the K-SVD algorithm is used to iteratively train the overcomplete dictionary to obtain the optimal dictionary for the corresponding noise feature. Using a joint dictionary composed of K optimal dictionaries, the newly acquired noise features are sparsely represented. The sparse representation problem is solved by the orthogonal matching pursuit (OMP) algorithm to obtain the sparse coefficients of the noise features under the joint dictionary. By utilizing the sparse coefficients of the noise features, the noise features are reconstructed using a joint dictionary to obtain the reconstructed noise features. Based on the spatial coordinates of each microphone in the acoustic sensor array, the array signal processing algorithm is used to calculate the element position matrix and direction vector of the microphone array as geometric information; using the geometric information of the acoustic sensor array, the direction matrix and steering vector of the beamforming algorithm are constructed. The reconstructed noise features are spatially filtered using the direction matrix and steering vector to extract the noise signal in the target direction. The spatial filtering result is then used as the input for noise source localization. The cross-correlation function of the noise signals in each channel of the acoustic sensor array is calculated using the generalized cross-correlation (GCC) algorithm. By obtaining the maximum value of the cross-correlation function, the time difference of arrival (TDOA) of the noise signals in each channel is calculated. Using the time difference of arrival (TDOA) as an observation, a set of localization equations for the spatial location of the noise source is constructed based on the direction matrix. The least squares (LS) algorithm is used to solve the localization equations to obtain the spatial coordinates of the noise source. Based on the spatial coordinates of the noise source and the three-dimensional model of the electromechanical equipment, a noise distribution map is generated through spatial interpolation and mapping algorithms.

4. The machine learning-based noise reduction method for electromechanical equipment according to claim 3, characterized in that: The clustering analysis algorithm is the K-means clustering algorithm.

5. The machine learning-based noise reduction method for electromechanical equipment according to any one of claims 1 to 4, characterized in that: The noise signal is extracted within a preset frequency range using a peak search algorithm, and these frequencies are used as the dominant noise frequency components, including: Based on the noise signal after noise reduction, a Hamming window with a frame length of N is used to divide and window the noise signal, and a Fast Fourier Transform (FFT) is performed on the windowed noise signal to obtain the noise signal spectrum. Based on the rotational speed of the electromechanical equipment, the search range of the main frequency component of the noise signal spectrum is set, and the harmonic of the rotational speed frequency is used as the center frequency of the search range; Within the search range of the main frequency components, a peak search algorithm is used to detect local peak points in the spectrum of the noise signal, obtain the P0 local peak points with the largest spectral amplitude, and extract the corresponding frequency values ​​as candidate noise main frequency components; The signal subspace of the ESPRIT algorithm is constructed using the P0 spectral peaks corresponding to the candidate noise main frequency components, and the spectral part other than the spectral peaks corresponding to the candidate noise main frequency components is used as the noise subspace. The order of the ESPRIT algorithm is determined based on the number P0 of the candidate noise dominant frequency components, and the order is equal to P0. The rotation invariance of the signal subspace is solved by the ESPRIT algorithm, and the corrected frequency values ​​of P1 candidate noise main frequency components are obtained. The calculated P1 corrected frequency values ​​are sorted, and the P2 frequencies with the largest corrected frequency values ​​are selected as the final noise dominant frequency components.

6. The machine learning-based noise reduction method for electromechanical equipment according to claim 5, characterized in that: Before calculating the frequency values ​​of the dominant noise component, the following steps are also included: Calculate the frequency difference between each pair of candidate noise dominant frequency components; When the frequency difference is less than the set threshold, the corresponding two candidate main frequency components are merged, and the average frequency value of the two candidate main frequency components is taken as the merged main frequency.

7. The machine learning-based noise reduction method for electromechanical equipment according to claim 6, characterized in that: The noise frequency shift caused by abnormal operating conditions of electromechanical equipment is obtained, including: The time-varying linear predictive coding algorithm (TVLPC) is used to model the collected operating status data. The AR model is used to describe the time-varying characteristics of the status data. The time-varying coefficients of the AR model are estimated by the Levinson-Durbin recursive algorithm to obtain the TVLPC model of the status data. Using the TVLPC model, one-step and multi-step predicted values ​​of state data are calculated, and the predicted values ​​are used as the future trend of state data. Based on the collected operational status data and the future trend obtained from the TVLPC model, the prediction error of the status data is calculated. By setting a prediction error threshold, it is determined whether the operating status of the electromechanical equipment is abnormal. For operating states judged to be abnormal, the characteristic frequency components reflecting the abnormal operation of electromechanical equipment are obtained by calculating the spectrum and power spectral density (PSD) of the prediction error sequence. The characteristic frequency components of the abnormal operation are correlated with the extracted noise main frequency components. The frequency offset of the characteristic frequency components of the abnormal operation relative to the noise main frequency components is calculated. The mapping relationship between the abnormal operation status of the electromechanical equipment and the noise frequency offset is established, and the noise frequency offset caused by the abnormal operation status of the electromechanical equipment is obtained.

8. The machine learning-based noise reduction method for electromechanical equipment according to claim 1, characterized in that: The optimal solution for noise control of electromechanical equipment is generated using the particle swarm optimization algorithm, including: An optimization problem for noise control of electromechanical equipment is established, with noise reduction effect as the optimization objective. An objective function for noise attenuation is constructed, and controller parameters are used as optimization variables. A mapping relationship between optimization variables and optimization objective is established. The controller parameters include control gain, filter order, and damping coefficient. The extracted noise main frequency component is used as the frequency constraint condition of the objective function, and the calculated noise frequency offset is used as the correction coefficient of the objective function. The optimal solution set of the controller parameters is obtained by using the particle swarm optimization algorithm (PSO). The controller parameters are used as the position vectors of the particles, and the noise attenuation is used as the fitness function of the particles. The optimal solution is searched in the parameter space. Based on the user's noise reduction requirements, the solution with the largest noise attenuation is selected from the optimal solution set as the optimal solution for the controller parameters.

9. The machine learning-based noise reduction method for electromechanical equipment according to claim 8, characterized in that: Based on the optimal solution for noise control, the dominant frequency components of the noise are extracted and separated, including: Based on the optimal solution, set the controller parameters; The extracted noise main frequency components are tracked using a joint spectrum analysis algorithm. The instantaneous frequency and amplitude of the noise main frequency components are estimated by short-time Fourier transform to obtain the dynamic spectrum of the noise main frequency. Based on the dynamic spectrum of the noise dominant frequency, the parameters of each dominant frequency component are extracted using a matching pursuit algorithm. The parameters of the dominant frequency component include frequency, amplitude, and phase. Based on the parameters of each extracted main frequency component, a control signal with the same amplitude, frequency, and phase opposite to each main frequency component is synthesized using frequency modulation and phase inversion algorithms. The synthesized control signal is input into the controller after setting the parameters to perform noise reduction control of electromechanical equipment.

10. The machine learning-based noise reduction method for electromechanical equipment according to claim 9, characterized in that: Constructing a deep learning-based noise reduction model includes: Noise signals from electromechanical equipment under different operating conditions are collected, and time-frequency analysis and sound source localization are performed on the collected noise signals to obtain noise distribution maps corresponding to the noise signals. Collect noise-free signals corresponding to noise signals, perform time-frequency analysis and sound source localization on the noise-free signals, and obtain noise-free distribution maps corresponding to the noise-free signals; A noise reduction model based on an encoder-decoder structure is constructed, in which the encoder extracts multi-scale features of the noise distribution map through a convolutional neural network (CNN), and the decoder reconstructs the extracted multi-scale features into a noise-free distribution map through a deconvolutional neural network and skip connections. We construct the loss function of the denoising model by using a weighted combination of reconstruction loss and adversarial loss; The noise distribution map is used as the input sample of the training data, and the corresponding noise-free distribution map is used as the output sample of the training data. The noise reduction model is trained using the training data, the loss function is calculated through forward propagation, and the model parameters are updated through back propagation. The noise distribution map corresponding to the newly acquired noise signal is used as input. The trained noise reduction model is used to extract the features of the noise distribution map through the encoder and the decoder to reconstruct the noise-free distribution map, thus obtaining the noise reduction noise distribution map. The noise distribution map after noise reduction is converted into a noise signal after noise reduction by inverse Fourier transform.

Images