Equipment voiceprint noise reduction monitoring method fusing empirical mode decomposition and dual-channel U-Net

Abstract

The invention discloses an equipment voiceprint noise reduction monitoring method fusing empirical mode decomposition and dual-channel U-Net, and relates to the technical field of power equipment state monitoring and voiceprint signal processing. The method comprises the following steps: acquiring an original voiceprint signal when equipment runs, and decomposing the signal into a plurality of intrinsic mode functions by using empirical mode decomposition to realize preliminary separation of noise and useful signals; constructing a dual-channel U-Net network, performing deep extraction on time domain features and frequency domain features of an intrinsic mode function, and introducing an ECA channel attention mechanism to strengthen fusion of effective features; designing a self-adaptive loss function, and optimizing the noise suppression capability of the network on different equipment and under different working conditions; and the noise-reduced voiceprint signal is used for equipment state evaluation to realize early recognition of an abnormal state. The method can be applied to an on-line monitoring system of power equipment, improves the signal-to-noise ratio of the voiceprint signal and the state evaluation accuracy, and has an important value for guaranteeing the stable operation of a power system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment state monitoring and voiceprint signal processing, and particularly relates to a device voiceprint denoising monitoring method fusing empirical mode decomposition and a double-channel U-Net. BACKGROUND

[0002] In a power system, devices such as power transformers, converter valves and water turbines are core components for guaranteeing power production and transmission, and their operating states directly affect the stability and safety of power supply. Voiceprint signals generated during the operation of the devices contain rich state information; for example, mechanical vibration sound and fluid flow sound of a water turbine, and core magnetic strain sound and winding discharge sound of a transformer. Through analysis of these voiceprint signals, device state monitoring and early fault warning can be achieved.

[0003] However, the actual operating environment of the devices is complex, and the voiceprint signals are easily affected by multiple interferences: on the one hand, there are background noises of other devices such as fans and pump bodies in the environment, as well as airflow and electromagnetic radiation interferences; on the other hand, the device voiceprint signals themselves have non-stationary and nonlinear characteristics, and the useful signals are often overwhelmed by noises, resulting in a decrease in the accuracy of subsequent state analysis.

[0004] The existing denoising and monitoring technologies have obvious deficiencies: (1) traditional denoising methods (such as fixed filtering and wavelet transform) rely on preset parameters and are difficult to adapt to complex and variable noise characteristics, and have poor suppression effect on non-stationary and nonlinear noises, which easily leads to distortion of useful signals; (2) a single deep learning model (such as a single-channel U-Net or CNN) can only extract single-dimensional (time domain or frequency domain) features when directly processing original voiceprint signals, and the features are not fully mined, which cannot fully capture device state-related information; (3) most methods are not optimized for the specificity of power device voiceprints, and have weak generalization ability under different devices (such as transformers and water turbines) and different operating conditions (such as full load and light load), which results in unstable denoising precision and affects the accuracy of device state evaluation, and thus it is difficult to meet the demand of the power system for high-reliability monitoring of devices. Therefore, there is an urgent need for a voiceprint denoising monitoring method that can adaptively separate noises and useful signals, deeply mine multi-dimensional features and adapt to multiple devices and multiple operating conditions, so as to improve the reliability and practicality of power device voiceprint monitoring. SUMMARY

[0005] The application provides a device voiceprint noise reduction monitoring method fusing empirical mode decomposition and a double-channel U-Net, which preliminarily separates noise and useful components in a voiceprint signal through the self-adaptive decomposition capability of empirical mode decomposition; extracts time domain and frequency domain features through a double-channel U-Net, strengthens useful feature fusion through an ECA channel attention mechanism, and improves feature utilization efficiency; designs an adaptive loss function to adapt to noise suppression requirements under different devices and working conditions; finally, early identification of device abnormal states is realized through the voiceprint signal after noise reduction, the signal-to-noise ratio of power device voiceprint monitoring is improved, the state evaluation accuracy is improved, and the stable operation of the power system is ensured, thereby solving the problems in the background art.

[0006] To solve the above technical problems, the application is implemented through the following technical solutions:

[0007] The device voiceprint noise reduction monitoring method fusing empirical mode decomposition and a double-channel U-Net of the application comprises the following steps:

[0008] S1, collecting an original voiceprint signal during device operation, specifically, arranging sensors in the operation area of power transformers, converter valves, water turbines and other devices, and collecting the original voiceprint signal generated during device operation; using empirical mode decomposition to decompose the original voiceprint signal into multiple intrinsic mode functions, and preliminarily separating noise and useful signals according to the correlation coefficient and frequency characteristics of the intrinsic mode functions;

[0009] The S1 step specifically comprises the following steps:

[0010] S11, selecting a piezoelectric acoustic sensor with an accuracy of ±1 dB and a frequency response range of 20 Hz-20 kHz, arranging it in the device operation area at an interval of 5 m-10 m, and collecting the original voiceprint signal at a sampling frequency and 16 bit quantization bit number that meet the Nyquist sampling theorem;

[0011] S12, performing empirical mode decomposition on the original voiceprint signal: identifying local maximum points and minimum points of the signal, fitting upper and lower envelope lines through cubic spline interpolation, calculating envelope mean values and obtaining IMF candidate components by subtracting the mean values from the original signal, and iterating until the candidate components meet the conditions that the number of extreme points and zero-crossing points is equal or the difference is ≤1 and the envelope mean value at any time is 0, to determine the first valid IMF; taking the remaining signal as a new original signal and repeating the decomposition until the remaining component is a monotonic function or a constant, to obtain multiple IMFs;

[0012] S13, calculating the correlation coefficients of each IMF and the original signal, and retaining the IMFs with a correlation coefficient >0.3; applying an attenuation weight to high-frequency IMFs and maintaining the original amplitude of medium and low-frequency IMFs to complete the preliminary separation of noise and useful signals;

[0013] S2, empirical mode decomposition, performing an empirical mode decomposition operation on the collected original voiceprint signal to decompose it into a plurality of intrinsic mode functions, realizing preliminary separation of noise and useful signal, and screening intrinsic mode functions containing effective information; double-channel feature extraction;

[0014] A double-channel U-Net network is constructed, which includes a time domain branch and a frequency domain branch, and the time domain features and the frequency domain features of the preliminarily separated intrinsic mode functions are extracted respectively; feature fusion optimization is performed, an ECA channel attention mechanism is introduced, useful feature fusion is strengthened, noise-related redundant features are suppressed, and information critical to equipment condition monitoring is highlighted;

[0015] The S2 step specifically comprises the following steps:

[0016] S21, a double-channel U-Net network architecture is built: the time domain branch adopts a sliding window blocking method with a window length of 256 and a step length of 128, converts the IMF into a two-dimensional time domain feature map with a dimension of [window number x 256]; the frequency domain branch performs short-time Fourier transform on the IMF using a Hanning window with a window length of 512 and an overlap rate of 50%, obtains a frequency spectrum and transposes it into a frequency domain feature map with a dimension of [window number x frequency point number], ensuring that the time dimension is aligned with the time domain; the two types of feature maps are respectively input into U-Net with the same structure, and the output dimension is [window number x frequency point number x 64] time domain and frequency domain features;

[0017] S22, ECA channel attention feature fusion is performed: element-level averaging is performed on the time domain and frequency domain features to obtain aggregated features, which are compressed into channel descriptors through global average pooling; one-dimensional convolution is used to nonlinearly map the descriptors, the convolution kernel size is calculated according to the formula k=floor(log2(C)+1), C is the number of channels 64, and the attention weight is obtained through Sigmoid activation; the unified features are weighted using the weight, the residual connection is introduced to add the fused features and the aggregated features, and the final output features are obtained;

[0018] S23, the output features are mapped to a single-channel feature map through 1x1 convolution with one convolution kernel, the overlapping part is removed and spliced according to the 50% overlap rate, and a denoised signal with the same dimension as the original signal is generated;

[0019] S3, a self-adaptive loss function is designed, which is a weighted sum of reconstruction loss, signal-to-noise ratio loss and feature loss, and the weights of each loss term are dynamically adjusted according to the device type and working condition to optimize the network's ability to suppress noise in different scenarios;

[0020] The EMD operation is performed on the collected original voiceprint signal to decompose the original voiceprint signal into a plurality of intrinsic mode functions, to realize preliminary separation of noise and useful signal, and to screen out intrinsic mode functions containing effective information; and the double-channel feature extraction is performed.

[0021] The S3 step specifically comprises the following sub-steps:

[0022] S31, constructing a total loss function: the designed total loss function is a weighted sum of a reconstruction loss L rec , a signal-to-noise ratio loss L snr , and a feature loss Lf eat , that is, L total =α·L rec +β·L snr +γ·Lf eat , wherein α+β+γ=1, α is a reconstruction loss weight, β is a signal-to-noise ratio loss weight, and γ is a feature loss weight;

[0023] S32, the reconstruction loss L rec uses the overall error of measuring the network output noise reduction signal and the clean label signal (the device voiceprint signal without noise) to ensure that the noise reduction signal is consistent with the useful signal in the waveform, and the formula is:

[0024]

[0025] wherein N is the number of sampling points, is the amplitude of the i th sampling point of the output signal, y i is the amplitude of the i th sampling point of the clean label, and the basic weight α=0.5 is used as the core benchmark of the loss function;

[0026] S33, the signal-to-noise ratio loss L snr is calculated by optimizing the signal-to-noise ratio of the output signal; when the signal-to-noise ratio of the original signal is <10dB, β is adjusted from 0.2 to 0.4, otherwise, β remains 0.2; γ is determined by α and β, that is, γ=1-α-β.

[0027] S34, the feature loss Lf eat uses the cosine similarity calculation of the output of the full connection layer of the pre-trained VGG16 network, and the formula is:

[0028]

[0029] wherein VGG(.) is the output of the full connection layer of the pre-trained VGG16 network (extracting deep features), and cos(.) is the cosine similarity function.

[0030] Network training and optimization: input IMF of S1 step, clean label as output, minimize total loss function with Adam optimizer (learning rate 0.001, batch size 32), iterative training until loss converges (e.g. loss fluctuation <0.001 after 500 iterations), get optimized denoising model;

[0031] S4, use the denoised voiceprint signal for equipment state evaluation, extract the time domain key features and frequency domain key features of the denoised signal, and compare them with the voiceprint feature benchmark library in the normal operation of the equipment, to realize early identification of abnormal state, assist online monitoring and fault warning of power equipment;

[0032] S4 step specifically includes the following steps:

[0033] S41, extract the time domain key features including peak factor, kurtosis, root mean square value and the frequency domain key features including main frequency of the denoised signal;

[0034] S42, collect normal denoised voiceprint signals under rated operating conditions of the equipment, extract the above key features and calculate the mean mu and standard deviation sigma, and construct the feature benchmark range [mu-3sigma, mu+3sigma];

[0035] S43, compare the key features of the actual denoised signal with the benchmark range, if it is out of range, trigger an abnormal warning, realize early identification of abnormal state.

[0036] The present application has the following beneficial effects compared with the prior art:

[0037] (1) High denoising precision: combining adaptive decomposition of EMD and multi-dimensional feature extraction of double-channel U-Net, non-stationary and nonlinear noise can be effectively separated, and the signal-to-noise ratio of the denoised signal is improved by 10-15dB, which is much better than traditional filtering methods;

[0038] (2) Strong generalization ability: the adaptive loss function can dynamically adjust the weight according to the type of equipment (transformer, water turbine, etc.) and operating conditions (full load, light load, etc.), and can maintain stable denoising precision in different scenarios;

[0039] (3) Accurate state evaluation: based on the high-quality voiceprint signal after denoising, the key features are extracted, and compared with the normal benchmark, the early identification of abnormal state can be realized, the false alarm rate is less than 5%, and reliable basis is provided for power equipment fault warning;

[0040] (4) Strong engineering practicability: the method flow is clear, the sensor arrangement and network training are easy to land, and it can be directly integrated into the existing power equipment online monitoring system without large-scale hardware modification, reducing the application cost.

[0041] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above at the same time. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Fig. 1 This is a simplified flowchart of the device acoustic noise reduction monitoring method that integrates empirical mode decomposition and dual-channel U-Net according to the present invention;

[0044] Fig. 2 This is a complete flowchart of the device acoustic noise reduction monitoring method that integrates empirical mode decomposition and dual-channel U-Net according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] This invention leverages the adaptive signal decomposition capability of Empirical Mode Decomposition (EMD) to accurately and initially separate noise and useful components from acoustic signature signals. It employs a dual-channel U-Net network to deeply extract time-domain and frequency-domain features, and combines this with an ECA channel attention mechanism to enhance useful feature fusion and improve feature utilization efficiency. An adaptive loss function is designed to dynamically adapt to the noise suppression requirements of different devices and operating conditions, balancing noise removal and useful signal retention. Based on the high-quality acoustic signature signal after noise reduction, early identification of abnormal equipment conditions is achieved, improving the signal-to-noise ratio and accuracy of status assessment in power equipment acoustic signature monitoring, and ensuring the stable operation of the power system.

[0047] Please see Figs. 1-2 As shown, the device acoustic noise reduction monitoring method of the present invention, which integrates empirical mode decomposition and dual-channel U-Net, achieves accurate suppression of complex noise by combining the adaptive signal separation capability of empirical mode decomposition with the feature learning advantages of deep learning. The following is a detailed description of each step:

[0048] Step S1: Acquire the raw acoustic signature signal during device operation, and use empirical mode decomposition to decompose the signal into multiple intrinsic mode functions, achieving preliminary separation of noise and useful signal. The process includes:

[0049] Step S11: original voiceprint signal collection. A high-sensitivity acoustic sensor is selected, such as a piezoelectric sensor with a precision of ±1 dB and a frequency response range covering 20 Hz-20 kHz. In a certain hydroelectric station turbine plant, four sensors are arranged at an interval of 5 m along the circumferential direction of the water turbine to ensure that the weak and complex voiceprint signals of the water turbine, power transformer, and converter valve during operation can be effectively captured. An appropriate sampling frequency is set to meet the Nyquist sampling theorem, ensuring complete acquisition of the frequency components of the device voiceprint signal; the sampling frequency meets the Nyquist sampling theorem, such as when the device voiceprint main frequency is ≤10 kHz, the sampling frequency is set to 25 kHz, the quantization bit number is set to 16 bits, and the signal amplitude accuracy is ensured; the continuous collection time is adjusted according to the device working condition; the collection time is adjusted according to the stability of the working condition; for example, 5-10 minutes for steady-state working condition and more than 30 minutes for dynamic working condition.

[0050] Step S12: empirical mode decomposition (EMD) processing and signal preliminary separation. The EMD decomposition process is performed as follows: traverse the signal x(t), identify all local maximum points and local minimum points, and use cubic spline interpolation to fit the maximum points to obtain the upper envelope line u(t) and the minimum points to obtain the lower envelope line l(t); calculate the mean values of the upper and lower envelope lines, formula: m1(t) = (u(t) + l(t)) / 2;

[0051] Subtract the envelope mean from the original signal to obtain the first intrinsic mode function (IMF) candidate component:

[0052] h1(t) = x(t) - m1(t)

[0053] Repeat steps 1-2 for the candidate component h1(t) (the upper limit of the iteration number is set to 100 times) until it meets the two core conditions of IMF: ① the number of extreme points in the entire signal is equal to or differs by no more than 1 from the number of zero-crossing points; ② the mean values of the upper and lower envelope lines at any time are 0, at which time h1(t) is determined as the first effective IMF component;

[0054] Calculate the residual signal r1(t) = x(t) - h1(t), and repeat the above decomposition process with r1(t) as the new "original signal" to obtain h2(t), h3(t),..., h n (t) in turn until the residual component r n (t) is a monotonic function or a constant, and the decomposition stops.

[0055] Step S13: IMF component screening and noise separation: calculate the correlation coefficient corr(h i ,x) of each IMF component h i (t) and the original signal x(t), formula:

[0056]

[0057] wherein and are the h i (t) and x(t) are the mean values. T is the signal length; the IMF components with correlation coefficient corr(h i , x) > 0.3 are reserved, which are considered to contain more useful signals related to equipment operation; the frequency characteristics of the IMF components are analyzed: the high-frequency IMF is mainly composed of environmental noise and sensor background noise, and a decay weight is applied thereto; the low-frequency IMF contains characteristic sound prints of equipment mechanical vibration and fluid flow, and the original amplitude is maintained to realize preliminary separation of noise and useful signals.

[0058] Step S2: a dual-channel U-Net network is constructed to extract the time domain characteristics and frequency domain characteristics of the intrinsic mode function respectively, and the ECA channel attention mechanism is introduced after channel unification to strengthen the fusion of useful features. The process includes:

[0059] Step S21: dual-channel U-Net network architecture construction. The overall structure of the network: the network contains two parallel modules of time domain branch and frequency domain branch, each branch is composed of "encoder (downsampling)-decoder (upsampling)", the fusion of shallow features and deep features is realized through jump connection, and finally the unified dimension fusion features are output through the feature fusion layer.

[0060] Construction of time domain input I t : for each IMF component (t) screened in S12, a sliding window blocking method is used to convert it into a two-dimensional feature map: let the window length W = 256 and the window step s = 128, then the window number H = [L / S], where L is the total length of the IMF component h i (t), and finally a two-dimensional time domain feature map I t with a dimension of HxW is obtained.

[0061] Construction of frequency domain input I f : each IMF component h i (t) screened in S12 is converted to the frequency domain by short-time Fourier transform (STFT): a Hanning window (window length = 512, overlap rate = 50%) is set, and a spectrum S with a dimension of FxH is calculated, where F = 256 is the number of frequency points. H is the number of time frames, which is consistent with the window number of the time domain branch); the spectrum is transposed to HxF to ensure that the time dimension is aligned with the time domain feature map.

[0062] The time domain input I t and the frequency domain input I fThe input is fed into two U-Net models with identical structures, and the two models output temporal features F respectively. t and frequency domain features F f Both have the same dimensions, C×H×W, where C is the number of channels, which can be set according to the actual situation, and is generally set to 64.

[0063] Step S22: ECA Channel Attention Feature Fusion. ECA Channel Attention Weight Calculation: The importance weights of each channel feature are adaptively learned through the ECA mechanism. Specific steps:

[0064] Calculate the aggregated feature: Perform an element-wise averaging operation on the sum to obtain the aggregated feature, as shown in the formula:

[0065]

[0066] Global Information Compression: Global Average Pooling (GAP) is performed on the aggregated features to compress the two-dimensional features of each channel into a single value, resulting in the channel descriptor z. c The formula is:

[0067]

[0068] Where z c Let c = 1, 2, ..., 64 be the global descriptor for the c-th channel.

[0069] Nonlinear mapping and weight normalization: One-dimensional convolution is used to normalize the channel descriptor z = [z c For each c = 1, 2, ..., 64, a non-linear mapping is performed, and the kernel size is calculated using the following formula:

[0070] k = floor(log2(C) + 1);

[0071] After mapping and normalization using the Sigmoid activation function, the attention for each channel is obtained:

[0072] w = Sigmoid(Conv1(z,k))

[0073] The unified time-domain and frequency-domain features are weighted using attention weight w, as shown in the formula:

[0074] F fusion =w·F t +(1-w)·F f

[0075] To avoid feature information loss during the fusion process, residual connections are introduced to add the fused features to the aggregated features, resulting in the final output features:

[0076] F out =F fusion +Fagg

[0077] F out The dimensions are 64×H×W, which are used for subsequent network training and noise reduction output.

[0078] Step S23: Reconstruct the network output features to generate a single-channel feature map of the denoised signal with the same dimension as the original signal.

[0079] First, the multi-channel feature maps are passed through a 1×1 convolution (with only one convolution kernel) in the output layer, thus mapping the multi-channel features into single-channel feature maps:

[0080] F final =Conv 1×1 (F out (k=1)

[0081] Because the original signal was segmented using an overlap strategy with a window length W = 256 and a step size s = 128 (overlap rate 50%), the overlapping portion needs to be removed during splicing to restore the original length: F final For each window (H in total), take its center s = 128 sampling points (non-overlapping region); stitch all non-overlapping regions together in the original window order to obtain a signal with a total length of H×s;

[0082]

[0083] In the above formula, ":" is a wildcard, meaning "all", that is, taking all elements from column 1 to column W in row h. h represents the row, the h-th row, h = 1, 2, ... H; S is the block step size.

[0084] Step S3: Design an adaptive loss function to optimize the network's ability to suppress noise under different devices and operating conditions.

[0085] This step involves constructing an adaptive loss function with multiple sub-loss terms and dynamically adjusting the weights of each loss term. This allows the network to balance noise suppression and useful signal retention under different devices and operating conditions, improving the model's generalization ability and noise reduction accuracy. The process includes:

[0086] Step S31: Constructing the adaptive loss function. The designed total loss function is the reconstruction loss L. rec Signal-to-noise ratio loss L snr Feature loss L feat Weighted sum:

[0087] L total =α·L rec +β·L snr +γ·L feat

[0088] Where α, β, and γ are dynamically adjusted weight coefficients, satisfying α + β + γ = 1, and each sub-loss term is defined as follows:

[0089] Reconstruction loss: Mean square error (MSE) is used to measure the overall error between the network's output denoised signal and the clean labeled signal, ensuring that the denoised signal is consistent with the useful signal in waveform. The formula is:

[0090]

[0091] Where N is the number of signal sampling points. Let y be the amplitude of the i-th sampling point of the output signal. i The amplitude of the i-th sampling point of the clean label; the base weight α = 0.5, which serves as the core benchmark of the loss function.

[0092] Signal-to-noise ratio loss: By optimizing the signal-to-noise ratio (SNR) of the output signal, the network's ability to suppress noise is enhanced. The formula is as follows:

[0093]

[0094] in, This represents the signal-to-noise ratio of the output signal.

[0095] When the original input signal In high-noise scenarios, adjust β from the base value of 0.2 to 0.4 to prioritize noise suppression; when the original SNR is ≥10dB, keep β = 0.2.

[0096] Feature loss: Protecting key features of the equipment's acoustic signature (such as the dominant frequency of mechanical vibration and fluid pulse characteristics) to avoid excessive noise reduction leading to the loss of useful information. The formula is:

[0097]

[0098] Where VGG(.) is the output of the fully connected layer of the pre-trained VGG16 network (extracting deep features), and cos(.) is the cosine similarity function;

[0099] Network training and optimization: Using the IMF from step S1 as input and clean labels as output, the total loss function is minimized using the Adam optimizer (learning rate 0.001, batch size 32), and the training is iterated until the loss converges (e.g., loss fluctuation <0.001 after 500 iterations) to obtain the optimized denoising model.

[0100] Step S4: Use the denoised voiceprint signal for equipment status assessment to achieve early identification of abnormal states based on the high-quality denoised voiceprint signal. The process includes:

[0101] Step S41: Key Feature Extraction. For the denoised signal, this patent extracts key features in the time domain (peak factor, kurtosis, root mean square value) and key features in the frequency domain (dominant frequency).

[0102] Step S42: Equipment Status Assessment and Anomaly Identification. Feature Baseline Construction: Collect 100-200 sets of normal noise-reduced acoustic signature signals from the equipment under rated operating conditions (rated load, rated voltage), extract the aforementioned key features, calculate the mean μ and standard deviation σ of each feature, and construct a baseline range [μ-3σ, μ+3σ]. Anomaly Identification: Collect the equipment acoustic signature signal in real time, extract key features after noise reduction in steps S1-S3, and compare them with the baseline range; if any feature exceeds the range, trigger an anomaly warning and mark the anomaly type (e.g., a main frequency shift may correspond to mechanical imbalance, increased kurtosis may correspond to a local impact fault).

[0103] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A device acoustic signature noise reduction monitoring method integrating empirical mode decomposition and dual-channel U-Net, characterized in that, Includes the following steps: S1. Collect the original voiceprint signal during the operation of the acquisition device, and use empirical mode decomposition to decompose the original voiceprint signal into multiple intrinsic mode functions. Based on the correlation coefficient and frequency characteristics of the intrinsic mode functions, the noise and useful signal are initially separated. S2. Construct a dual-channel U-Net network, which includes a time-domain branch and a frequency-domain branch. Extract time-domain features and frequency-domain features from the initially separated intrinsic mode functions, respectively. After channel unification, introduce the ECA channel attention mechanism to enhance the fusion of useful features and suppress noise redundancy features. S3. Design an adaptive loss function, which is a weighted sum of reconstruction loss, signal-to-noise ratio loss and feature loss. The weights of each loss term are dynamically adjusted according to the equipment type and working conditions to optimize the network's ability to suppress noise in different scenarios. S4. Use the noise-reduced voiceprint signal for equipment status assessment, extract the key features in the time domain and frequency domain of the noise-reduced signal, and compare it with the voiceprint feature benchmark library when the equipment is running normally to achieve early identification of abnormal conditions.

2. The device acoustic signature noise reduction monitoring method that integrates empirical mode decomposition and dual-channel U-Net according to claim 1, characterized in that: Step S1 specifically includes the following sub-steps: S11. Select piezoelectric acoustic sensors with an accuracy of ±1dB and a frequency response range of 20Hz-20kHz, and arrange them at intervals of 5m-10m in the equipment operating area to meet the sampling frequency of the Nyquist sampling theorem and the 16-bit quantization bit depth to acquire the original acoustic signature signal. S12. Perform empirical mode decomposition on the original voiceprint signal: identify local maxima and minima of the signal, fit the upper and lower envelopes by cubic spline interpolation, calculate the envelope mean, and subtract the mean from the original signal to obtain IMF candidate components. Iterate until the candidate components satisfy: the number of extreme points and zero-crossing points are equal or differ by ≤1 and the envelope mean is 0 at any time, and determine the first valid IMF; repeatedly decompose the remaining signal as a new original signal until the remaining components are monotonic functions or constants, and obtain multiple IMFs. S13. Calculate the correlation coefficient between each IMF and the original signal, and retain the IMFs with a correlation coefficient > 0.3; for the retained IMFs, apply attenuation weights to the high-frequency IMFs and maintain the original amplitude of the mid- and low-frequency IMFs to complete the initial separation of noise and useful signal.

3. The device acoustic signature noise reduction monitoring method that integrates empirical mode decomposition and dual-channel U-Net according to claim 1, characterized in that: Step S2 specifically includes the following sub-steps: S21. Construct a dual-channel U-Net network architecture: The time-domain branch uses a sliding window block method with a window length of 256 and a step size of 128 to convert the IMF into a two-dimensional time-domain feature map with a dimension of [number of windows × 256]; the frequency-domain branch uses a Hanning window with a window length of 512 and an overlap rate of 50% to perform a short-time Fourier transform on the IMF, obtain a spectrogram, and transpose it into a frequency-domain feature map with a dimension of [number of windows × number of frequency points], ensuring that the time dimension is aligned with the time domain; input the two types of feature maps into a U-Net with the same structure, and output time-domain and frequency-domain features with a dimension of [number of windows × number of frequency points × 64]; S22. Perform ECA channel attention feature fusion: Element-wise averaging is performed on the time-domain and frequency-domain features to obtain aggregated features, which are then compressed into channel descriptors via global average pooling. One-dimensional convolution is used to perform non-linear mapping on the descriptors, with the kernel size calculated according to the formula k = floor(log2(C) + 1), where C is the number of channels (64). Attention weights are obtained by Sigmoid activation. The unified features are weighted using these weights, and residual connections are introduced to add the fused features to the aggregated features to obtain the final output features. S23. Map the output features to a single-channel feature map using a 1×1 convolution kernel, remove the overlapping parts with a 50% overlap rate, and then stitch them together to generate a denoised signal with the same dimension as the original signal.

4. The device acoustic signature noise reduction monitoring method integrating empirical mode decomposition and dual-channel U-Net as described in claim 1, characterized in that, The S3 step specifically includes the following sub-steps: S31. Constructing the total loss function: The designed total loss function is the reconstruction loss L. rec Signal-to-noise ratio loss L snr Feature loss Lf eat Weighted sum: L total =α·L rec +β·L snr +γ·Lf eat , where α+β+γ=1, α is the reconstruction loss weight, β is the signal-to-noise ratio loss weight, and γ is the feature loss weight; S32, Reconstruction Loss L rec The overall error between the network's output denoised signal and the clean label signal is measured to ensure that the waveform of the denoised signal is consistent with the useful signal. The formula is as follows: Where N is the number of sampling points, Let y be the amplitude of the i-th sampling point of the output signal. i The amplitude of the i-th sampling point of the clean label; the base weight α = 0.5, serving as the core benchmark of the loss function; S33, Signal-to-noise ratio loss L snr By optimizing the output signal-to-noise ratio calculation, when the original signal-to-noise ratio is <10dB, β is adjusted from 0.2 to 0.4, otherwise it remains at 0.2; γ is determined by inverse calculation using α and β. S34, Feature Loss Lf eat The cosine similarity is calculated using the outputs of the fully connected layers of a pre-trained VGG16 network, using the following formula: Where VGG(.) is the output of the fully connected layer of the pre-trained VGG16 network (extracting deep features), and cos(.) is the cosine similarity function.

5. The device acoustic signature noise reduction monitoring method integrating empirical mode decomposition and dual-channel U-Net as described in claim 1, characterized in that, The S4 step specifically includes the following sub-steps: S41. Extract key time-domain features of the denoised signal, including peak factor, kurtosis, and root mean square value, as well as key frequency-domain features including the dominant frequency. S42. Collect the normal noise reduction ripple signal under the rated operating conditions of the equipment, extract the above key features and calculate the mean μ and standard deviation σ, and construct the feature reference range [μ-3σ,μ+3σ]. S43. Compare the key features of the actual noise-reduced signal with the reference range. If the signal exceeds the range, trigger an abnormal warning to achieve early identification of abnormal conditions.

Images