An industrial noise identification method and system

Abstract

The application relates to the technical field of audio signal processing, and discloses an industrial noise identification method and system.The method comprises the following steps: acquiring an audio signal and performing time-frequency decomposition to obtain preliminary decomposition data; classifying coupling features of the preliminary decomposition data to obtain a coupling region boundary value; if the coupling region boundary value exceeds a preset boundary threshold value, performing filtering processing to obtain separated noise spectrum data; calculating an overlap ratio of the separated noise spectrum data and a preset target sound; if the overlap ratio leads to extraction deviation, extracting dynamic change characteristics of the separated noise spectrum data and integrating the dynamic change characteristics into a spectrum extraction process to obtain an accurate noise spectrum; calculating an intensity value according to the accurate noise spectrum to determine an intensity evaluation result; comparing the intensity evaluation result with a preset pollution threshold value, calculating a noise pollution degree, and outputting a noise identification result.The method is suitable for industrial environment noise monitoring and can realize accurate extraction of a noise spectrum.

Description

Technical Field

[0001] This invention relates to the field of industrial noise pollution identification technology, and in particular to an industrial noise identification method and system. Background Technology

[0002] Currently, sound recognition is an important technical means for monitoring noise and diagnosing equipment status in industrial settings, and it has significant application value in ensuring production safety and implementing fault early warning.

[0003] In existing technologies, industrial noise identification typically employs a combination of spectral analysis and threshold judgment. For example, after performing time-frequency decomposition of audio signals using Fourier transform or wavelet transform, matching and judgment are performed based on preset noise spectrum templates or energy thresholds to assess noise intensity and pollution levels. However, this method struggles to achieve effective spectral separation when dealing with complex operating conditions where noise spectra are severely coupled and overlapped with useful signals in industrial settings. This results in incomplete noise spectrum extraction, decreased accuracy in intensity assessment, and an inability to reliably identify the degree of noise pollution.

[0004] Therefore, existing technologies have the problem of difficulty in accurately extracting the spectrum of industrial noise. Summary of the Invention

[0005] This invention provides an industrial noise identification method and system to achieve accurate extraction of noise spectrum.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides an industrial noise identification method, comprising:

[0007] Audio signals from an industrial scene are acquired and wavelet basis function matching is performed to generate frequency coefficients; the modulus maxima of the frequency coefficients are calculated, and the singularity locations are determined based on the propagation law of the modulus maxima; the frequency coefficients are separated according to the singularity locations to obtain preliminary decomposed data;

[0008] Based on the preliminary decomposition data, a coupling feature vector is constructed, and a preset support vector machine is used to classify the coupling feature vector to obtain the boundary values ​​of the coupling region;

[0009] If the boundary value of the coupling region exceeds a preset boundary threshold, the coupling frequency band data sequence corresponding to the boundary value of the coupling region is extracted and adaptive iterative decomposition is performed to obtain the intrinsic mode component set; based on the instantaneous frequency fluctuation variance of the intrinsic mode component set, the noise dominant component is screened out; the noise dominant component is subjected to Hilbert transform and stripping processing to obtain the background noise sequence and convert it into frequency domain energy distribution to obtain the separated noise spectrum data;

[0010] Calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and determine whether the overlap ratio causes an extraction deviation;

[0011] When the overlap ratio causes extraction deviation, the time-varying envelope and frequency modulation features of the separated noise spectrum data are extracted and mapped to the frequency domain space, and the amplitude fluctuation trajectory and instantaneous frequency components are calculated; based on the amplitude fluctuation trajectory and the instantaneous frequency components, the correction gain factor and inverse filter coefficient are calculated; based on the correction gain factor and the inverse filter coefficient, a spectrum reconstruction matrix is ​​constructed, and the spectrum reconstruction matrix is ​​analyzed to obtain the accurate noise spectrum; when there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as the accurate noise spectrum.

[0012] The intensity value of the energy is calculated based on the precise noise spectrum, and the intensity assessment result is determined.

[0013] The intensity assessment results are compared with a preset pollution threshold to calculate the noise pollution level and output the noise identification results.

[0014] Secondly, the present invention provides an industrial noise identification system, comprising:

[0015] An audio acquisition module is used to acquire audio signals from an industrial scene and perform wavelet basis function matching to generate frequency coefficients; calculate the modulus maxima of the frequency coefficients, determine the singularity locations based on the propagation law of the modulus maxima; and separate the frequency coefficients according to the singularity locations to obtain preliminary decomposed data.

[0016] The coupling classification module is used to construct coupling feature vectors based on the preliminary decomposition data, and to classify the coupling feature vectors using a preset support vector machine to obtain the boundary values ​​of the coupling region.

[0017] The filtering module is used to extract the coupling frequency band data sequence corresponding to the coupling region boundary value and perform adaptive iterative decomposition if the boundary value of the coupling region exceeds a preset boundary threshold, thereby obtaining a set of intrinsic mode components; based on the instantaneous frequency fluctuation variance of the intrinsic mode component set, the noise dominant component is screened out; the noise dominant component is subjected to Hilbert transform and stripping processing to obtain a background noise sequence and convert it into a frequency domain energy distribution to obtain separated noise spectrum data;

[0018] The overlap determination module is used to calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and to determine whether the overlap ratio causes an extraction deviation.

[0019] The dynamic correction module is used to extract the time-varying envelope and frequency modulation features of the separated noise spectrum data and map them to the frequency domain when the overlap ratio causes extraction deviation; calculate the amplitude fluctuation trajectory and instantaneous frequency components; calculate the correction gain factor and inverse filter coefficient based on the amplitude fluctuation trajectory and instantaneous frequency components; construct a spectrum reconstruction matrix based on the correction gain factor and inverse filter coefficient; and parse the spectrum reconstruction matrix to obtain the accurate noise spectrum. When there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as the accurate noise spectrum.

[0020] The intensity assessment module is used to calculate the intensity value of the energy based on the precise noise spectrum and determine the intensity assessment result;

[0021] The result output module is used to compare the intensity assessment result with the preset pollution threshold, calculate the noise pollution level, and output the noise identification result.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) The present invention classifies the coupling characteristics of noise spectrum and useful signal by using a preset support vector machine, and performs filtering when the coupling strength exceeds the threshold, which can effectively separate noise and useful signal that are seriously overlapping in the frequency domain, and improve the accuracy and completeness of noise spectrum extraction.

[0024] (2) This invention calculates the overlap ratio between the separated noise and the target sound, and dynamically incorporates the time-varying envelope and frequency modulation characteristics of the noise for spectrum reconstruction when it is determined that there is an extraction deviation. This enhances the method's adaptability to non-stationary and dynamically changing noise, thereby achieving stable and reliable noise spectrum extraction in complex time-varying industrial acoustic environments.

[0025] (3) Based on the extraction of accurate noise spectrum, this invention calculates intensity by energy integration and combines it with preset pollution threshold to quantify risk and assess level, thereby achieving a refined and quantitative assessment of the degree of industrial noise pollution. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of an industrial noise identification method provided in the first embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of an industrial noise identification system provided in the second embodiment of the present invention. Detailed Implementation

[0028] 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.

[0029] Reference Figure 1 The first embodiment of the present invention provides an industrial noise identification method, comprising the following steps:

[0030] S11, acquire the audio signal from the industrial scene, and perform time-frequency decomposition on the audio signal through wavelet transform to obtain preliminary decomposition data;

[0031] S12, construct a coupling feature vector based on the preliminary decomposition data, and classify the coupling feature vector using a preset support vector machine to obtain the boundary values ​​of the coupling region;

[0032] S13, If the boundary value of the coupling region exceeds the preset boundary threshold, then the boundary value of the coupling region is filtered to obtain the separated noise spectrum data;

[0033] S14, calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and determine whether the overlap ratio causes an extraction deviation;

[0034] S15, when the overlap ratio causes extraction deviation, the dynamic change characteristics of the separated noise spectrum data are extracted and incorporated into the spectrum extraction process to obtain an accurate noise spectrum; when there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as an accurate noise spectrum.

[0035] S16, Calculate the intensity value of the energy based on the precise noise spectrum, and determine the intensity assessment result;

[0036] S17, compare the intensity assessment result with the preset pollution threshold, calculate the noise pollution level and output the noise identification result.

[0037] In step S11, an audio signal from an industrial scene is acquired, and the audio signal is decomposed into time-frequency components using wavelet transform to obtain preliminary decomposed data, including:

[0038] S111 acquires audio signals from industrial scenes and converts them into digital audio sequences;

[0039] S112, Match wavelet basis functions according to the spectral energy distribution of the digital audio sequence to generate frequency coefficients;

[0040] S113, Calculate the modulus maxima of the frequency coefficients, and determine the location of the singular point based on the propagation law of the modulus maxima;

[0041] S114, Based on the location of the singularity point, the frequency coefficients are separated to obtain preliminary decomposition data.

[0042] In practical applications, to acquire audio signals from industrial scenarios, a successive approximation ADC (analog-to-digital converter) can be used with a fixed sampling frequency. For example, the audio signal is sampled at equal intervals at 44.1kHz, and the analog voltage value of each sampling point is quantized into a 16-bit digital value to generate a discrete-time digital sequence. , which is a digital audio sequence, where n is the sample point number.

[0043] First, the digital audio sequence Perform a short-time Fourier transform using a Hamming window with a window length of 1024 points and an overlap rate of 75%, and calculate its spectral energy distribution. If the frequency of transient events, where the amplitude change exceeds 30% of the total signal amplitude within a 0.5ms time window, is higher than 100 times per second, a Haar wavelet with a short support set is used for decomposition. If the signal's spectral energy is widely distributed in the mid-to-high frequency band (1000Hz to 8000Hz), and its accumulated energy in this band accounts for more than 60% of the total energy of the entire band, a Sym8 wavelet basis function with good symmetry is used for decomposition to maintain good phase characteristics. It should be noted that the method for calculating the frequency of transient events is as follows: first, the digital audio sequence... Non-overlapping frame segmentation is performed with a window length of 0.5ms, and each frame has 1000 pixels. Calculate the amplitude change of the signal within each frame. ,in The global maximum amplitude of the signal. This represents the signal's global minimum amplitude; if this change exceeds the signal's global maximum amplitude... If a frame is identified as a transient event frame, then the frame is determined to be a transient event frame. The total number of transient event frames is then counted across all frames. The total signal duration is (s), where N is the total number of sampling points, Sampling frequency. Frequency of transient events. , calculated as The unit is times per second. If If the signal contains a large number of transient impulse components, then the Haar wavelet of the short support set is selected for subsequent decomposition.

[0044] After filtering, downsampling is performed. The downsampling factor can be set to 2, meaning only convolution results with even (or odd) indices are retained. This ultimately generates a set of results including high-frequency coefficients. With low frequency coefficient The frequency coefficient, where j is the number of decomposition levels ( ), Typically, the number of layers is set to 5-8 based on the sampling frequency and signal characteristics; this embodiment uses 5 layers, where k is the coefficient index. After transformation, the high-frequency coefficients... It mainly reflects the details and noise components of the signal, and the low-frequency coefficient. This preserves the overall trend and main energy structure of the signal.

[0045] Next, the high-frequency coefficients were analyzed separately. and low frequency coefficient Calculate its modulus and Next, the maximum value of the modulus is found in each coefficient sequence. For any position k, if the following condition is met... and Then this position is the point of maximum modulus. For low-frequency coefficients... The same rules are used for maximum detection.

[0046] It is worth noting that the propagation law of modulo maxima is as follows: if singularities exist in the signal, such as impulses, transients, or abrupt changes, then at different decomposition scales j, the modulo maxima corresponding to the same singular event will form a propagation chain whose amplitude gradually decreases with increasing scale while maintaining alignment. By tracing the correspondence of these maxima across different scales, a propagation path from fine to coarse scales can be established. The time-domain location corresponding to the intersection of these paths is the singularity location. A set of singularity locations is obtained by statistically analyzing all locations. If a maximum point appears only at a single scale and has no corresponding propagation, it is considered noise and is excluded. Finally, the locations of all singularities verified by the propagation law and their corresponding scale-amplitude characteristics are output as the structural basis for subsequent signal separation.

[0047] Based on the set of singularity locations At each decomposition scale The above high frequency coefficient Perform separation for each singularity. Expand the scale-adaptive protection zone to both sides from it as the center. ,in , Let be the time-domain support length of the wavelet basis used at the current scale j. All waveslet basis bases falling within any guard interval are considered. The high-frequency coefficients within are denoted as transient high-frequency components. High-frequency coefficients that do not fall within any protected area are denoted as stationary high-frequency components. Low-frequency coefficient The overall trend component is retained as low frequency. The above three types of components are reorganized according to scale to form preliminary decomposed data. Including transient high-frequency components Stable high-frequency components and low-frequency trend components Each component retains its temporal location, scale, and coefficient value information.

[0048] In step S12, a coupling feature vector is constructed based on the preliminary decomposition data, and a preset support vector machine is used to classify the coupling feature vector to obtain the boundary values ​​of the coupling region, including:

[0049] S121, Extract the spectral energy density data of the preliminary decomposition data, and construct a coupling feature vector based on the spectral energy density data;

[0050] S122, Input the coupled feature vector into a preset support vector machine, establish the hyperplane equation, and calculate the geometric distance between sample points;

[0051] S123, determine the classification interval based on the geometric distance of the sample points, solve the decision function for the classification interval, and obtain the boundary of the coupling region;

[0052] S124, extract the spectral energy value corresponding to the boundary of the coupling region, and confirm the spectral energy value as the boundary value of the coupling region.

[0053] First, the preliminary decomposition data Spectral energy density estimation is performed on each component. For transient high-frequency components... and stable high-frequency components Perform short-time Fourier transforms on each component using a Hanning window with a window length set to twice the wavelet support length at the corresponding scale and an overlap rate of 50%, to obtain the components in each frequency band. Energy density on and For low-frequency trend components Calculate its power spectral density directly The power spectral density uses the Hanning window as the window function.

[0054] Next, based on the aforementioned spectral energy density data, a coupling feature vector characterizing the degree of coupling between noise and useful signal is constructed. For each time-frequency unit... Extract the following features: energy ratio feature Frequency band energy ratio Trend coupling Transient steady energy difference The features mentioned above in the same time-frequency unit are concatenated in sequence to obtain the feature sub-vector of that unit; then all frequency bands within that time period are concatenated. The feature vectors are connected in frequency band order to form the coupled feature vector corresponding to that time period.

[0055] It should be noted that the set of coupled feature vectors constructed in the above steps is represented as... Its corresponding category label is represented as Both are used as input data to a pre-defined support vector machine (SVM) classifier. This indicates that the sample belongs to a strongly coupled region. This indicates a weakly coupled or uncoupled region. The label data comes from a historical industrial audio database, from which historical samples are used to train the SVM. Each historical sample contains the original noisy frequency, the corresponding background noise reference signal, and the target sound reference signal. The background noise reference signal is collected during a stable period when the equipment is stopped, and the target sound reference signal is collected at close range using a high-sensitivity microphone in an environment where the equipment is operating alone and there are no other interfering sound sources. During labeling, preliminary decomposition data is extracted from each historical sample, and the transient energy percentage of each frequency band is calculated. and the energy proportion of this frequency band If both conditions are met... and Then all time-frequency units corresponding to the sample in the frequency band are marked as strongly coupled ( Otherwise, it is marked as weakly coupled. Furthermore, the annotation results can be randomly checked and verified by professional acoustic engineers to ensure consistency of no less than 90%.

[0056] It should be noted that the training process of SVM is as follows: first, the set of labeled feature vectors is... Z-score normalization was performed. A radial basis function was selected. and initialize parameters initial values ​​of slack variables Regularization parameters The sequential minimum optimization algorithm is used to solve the convex optimization problem. The constraints are The optimal solution is obtained by iteratively updating the Lagrange multipliers. , And support vectors, thereby establishing the optimal classification hyperplane. For any eigenvector X, its geometric distance to the hyperplane is calculated as follows: After training, the hold-out method is used for validation, requiring the accuracy on the test set to be no less than 85%. Otherwise, the parameters are adjusted and the model is retrained. Finally, the model parameters are solidified for online coupled classification.

[0057] It is worth noting that, based on the geometric distance of the sample points Determine the boundary range of the classification margin. Include all positive class samples in the training set ( The minimum geometric distance of ) is denoted as All negative class samples ( The maximum geometric distance is denoted as The classification margin is defined as follows: This interval reflects the range of geometric distances between the nearest samples on either side of the decision boundary, i.e., the classification margin width is... .

[0058] To obtain the boundary of the coupling region, a decision threshold needs to be determined within the classification interval. .set up That is, take the geometric distance value corresponding to the midpoint of the classification margin. Convert to decision function value:

[0059] ,

[0060] in For any positive class sample, geometric distance is usually chosen. The largest positive class sample. For a symbolic function, take symbols, If the value is greater than or equal to 0, add 1; otherwise, subtract 1. Let be the norm of the normal vector of the optimal hyperplane.

[0061] Finally, the boundary of the coupled region is defined by the following equation:

[0062] .

[0063] The BFGS method, a quasi-Newton method, is used to solve the problem. The iteration stops when the equation is satisfied. All X values ​​satisfying this equation constitute the boundary surface in the feature space, which is the decision boundary for dividing the strongly coupled and weakly coupled regions. In the high-dimensional feature space, this boundary corresponds to a hypersurface; after mapping back to the time-frequency domain, the coupling region boundaries of each time-frequency unit with varying coupling strength are obtained.

[0064] The boundary equation of the coupling region Map back to the time-frequency domain. For each time-frequency unit... , and its corresponding coupling feature vector Substitute into the boundary equations and calculate The value of . If (in For boundary tolerance, it is usually taken as If the time-frequency unit is located on the boundary of the coupling region, then it is considered to be located on the boundary of the coupling region. Taking the maximum value multiplied by 5% as the tolerance is based on the distribution characteristics of samples near the boundary in the training set. In one example, in SVM classification, approximately 90% of the samples near the boundary have a geometric distance of within the margin width. The tolerance is calculated by multiplying the maximum value within 10% by 5%, based on the distribution characteristics of samples near the boundary in the training set. In SVM classification, approximately 90% of the samples near the boundary, i.e., those with a geometric distance within the margin width... Within 10% The value should not exceed 2-3 times this threshold. Therefore, a 5% ratio can filter out samples that clearly belong to one side while preserving areas with blurred boundaries.

[0065] For all time-frequency cells falling on the boundary, extract their corresponding spectral energy values. This value is the energy density calculated in step S121. and The sum is obtained, Then, all boundary units... By frequency band Statistical analysis was performed, and the median energy at the boundary of each frequency band was taken as the boundary value of the coupling region for that frequency band. .in, The time-domain index set for boundary cells, containing all that satisfy... The time point t.

[0066] Finally, the boundary values ​​of all frequency bands Combined into vectors This vector is the boundary value of the coupling region, representing the energy threshold for dividing the coupling state between noise and useful signal in the frequency domain. In practical identification, if the energy of a certain time-frequency unit... If the condition is met, the unit is determined to be a strongly coupled region; otherwise, it is a weakly coupled region.

[0067] In step S13, if the boundary value of the coupling region exceeds a preset boundary threshold, filtering is performed on the boundary value of the coupling region to obtain separated noise spectrum data, including:

[0068] S131, if the boundary value of the coupling region exceeds a preset boundary threshold, extract the coupling frequency band data sequence corresponding to the boundary value of the coupling region;

[0069] S132, perform adaptive iterative decomposition on the coupled frequency band data sequence to obtain the set of intrinsic mode components;

[0070] S133, based on the instantaneous frequency fluctuation variance of the intrinsic mode component set, the noise-dominant component is selected;

[0071] S134, perform Hilbert transform and stripping on the dominant noise component to obtain a background noise sequence, and convert the background noise sequence into a frequency domain energy distribution to obtain separated noise spectrum data.

[0072] First, With preset boundary threshold One by one, comparisons are made. The preset boundary threshold is set based on historical data statistics, typically 1.5 to 2 times the average energy of the background noise, for example, 1.8. This value achieved a recall rate of 92% and a precision rate of 88% for identifying the coupling region in the test set. The background noise is the average energy of the stationary background component. The average energy of the background noise is calculated during a period without target sound, containing only ambient background noise. This period is usually extracted separately before or after the start of audio acquisition, lasting 10 seconds. The calculation covers the entire frequency band. ,in To analyze the upper limit of the frequency band, half of the sampling rate is usually taken, such as... The energy of each frequency band is obtained by integrating the power spectral density of the background noise signal within the corresponding frequency band during that time period. It should be noted that the historical data originates from an industrial noise multi-scenario benchmark database constructed based on this invention. This database contains a total of 10,000 labeled audio samples from typical industrial scenarios, such as stamping workshops, fan rooms, pump stations, and assembly lines. Each sample includes the original mixed audio with a sampling rate of 44.1 kHz and a duration of 10 seconds; synchronously acquired pure target sound, i.e., the sound of equipment running without load; and a background noise reference signal, which was acquired during equipment downtime.

[0073] For each frequency band If satisfied , If the preset boundary threshold is met, the frequency band is determined to be a strongly coupled band; otherwise, it is determined to be an uncoupled or weakly coupled band and no further processing is performed. All strongly coupled bands constitute a set:

[0074]

[0075] Next, from the original audio sequence Extract the time-domain data sequence corresponding to the strongly coupled frequency band. The specific steps are as follows: Design an FIR bandpass filter bank, where the passband of each filter corresponds to a strongly coupled frequency band. The transition bandwidth is set to 20% of the bandwidth, and the stopband attenuation is not less than 40dB; Each signal passes through a filter to obtain a set of sub-band signals. Align and superimpose all sub-band signals in the time domain to obtain the coupled frequency band data sequence. .like If the set is empty, the filtering process is skipped, and the original signal is directly output as noise spectrum data.

[0076] It should be noted that the coupled frequency band data sequence As input, an adaptive noise-complete set empirical mode decomposition method is used for iterative decomposition. The core of this step is to improve the mode aliasing problem of traditional empirical mode decomposition in strongly coupled signals by introducing adaptive white noise assistance.

[0077] The specific steps involve generating an adaptive white noise sequence that matches the shape of the signal spectrum. The original signal is superimposed with each noise sequence to obtain a noise-assisted signal set. For each Perform empirical mode decomposition and extract intrinsic mode components through an iterative selection process: for the current residual signal First, extract all its extreme points, then use cubic spline interpolation to fit the upper and lower envelopes respectively. and Calculate the mean envelope and use Update candidate components; repeat this process until... The conditions for IMF compliance are met: the number of extreme points and the number of zero-crossing points differ by no more than 1, and the mean envelope is close to zero throughout the entire time period, meaning the instantaneous amplitude of the mean envelope does not exceed 5% of the maximum amplitudes of the upper and lower envelopes. (The sentence ends abruptly here, likely due to an incomplete or corrupted source.) This is denoted as an IMF component. Update residuals The decomposition continues until the residuals exhibit a monotonic trend or the energy falls below a preset energy threshold, such as 1% of the original signal energy. After performing the above decomposition on each of the M noise-assisted signals, the corresponding IMFs of the same order are integrated and averaged to obtain the final set of intrinsic mode components. , where K is the number of IMF layers obtained by decomposition, each IMF component reflects the oscillation characteristics of the signal at different time scales, and the components are approximately orthogonal to each other.

[0078] It should be noted that the method for generating white noise sequences includes: firstly, calculating the strongly coupled signal. power spectral density ,in for Fourier transform. A linear-phase FIR filter is constructed based on this power spectrum. Its amplitude-frequency response is set to This makes the filter amplitude profile proportional to the signal spectral envelope. A Gaussian white noise sequence with zero mean and 1 variance is generated. Frequency shaping is performed using this filter to obtain... Finally, The amplitude is scaled down to within 10% of the original signal's maximum amplitude, i.e. .

[0079] It is worth noting that, for the aforementioned set of intrinsic mode components... Each component in Perform a Hilbert transform to obtain its analytic signal. ,in This represents the Hilbert transform. Next, the instantaneous frequency of each component is calculated. .in Represents an imaginary number, squared as 1. To analyze the phase angle of the signal, the rate of change of the instantaneous frequency is calculated using a differential approximation, with a fixed step size of one sampling interval. Then, the variance of the instantaneous frequency fluctuation for each component is calculated. Finally, based on the fact that noise components typically exhibit high frequency fluctuation characteristics, a fluctuation variance threshold is set. in This is the adjustment factor, typically taken as 1.5. If the instantaneous frequency fluctuation variance of a certain component satisfies... If the component is determined to be a dominant noise component, it will be included in the noise component set. Otherwise, it is determined to be the dominant component of the signal.

[0080] It should be noted that the set of dominant noise components... Perform a Hilbert transform to construct the analytic signals for each component. in Let j denote the Hilbert transform operator, where j is the imaginary unit. Next, the instantaneous envelope of each dominant noise component is calculated. With instantaneous phase , The noise component is stripped by phase modulation and amplitude adjustment to remove any potentially residual signal-related components. Specifically, each component is phase-randomized and its envelope structure is preserved by amplitude rescaling to obtain the stripped noise component. in It is an independent and identically distributed random phase sequence, which follows the... Uniformly distributed.

[0081] Then, the real part of the stripped and analyzed signal is taken to obtain the background noise component. . For all dominant noise components Superimposing these sequences in the time domain yields the background noise sequence:

[0082] ,

[0083] Finally, for Perform a discrete Fourier transform and calculate its power spectral density. As the separated noise spectrum data, it represents the energy distribution of the noise components extracted from the original signal in the frequency domain.

[0084] In step S14, the overlap ratio between the separated noise spectrum data and the preset target sound is calculated, and it is determined whether the overlap ratio causes an extraction deviation, including:

[0085] S141, the separated noise spectrum data is mapped to the frequency domain of the preset target sound to construct a spectrum energy density comparison matrix;

[0086] S142, Analyze the frequency overlap interval in the spectral energy density comparison matrix, calculate the amplitude masking effect value, and perform frequency domain weighted integration on the amplitude masking effect value to obtain the overlap ratio;

[0087] S143, if the overlap ratio is higher than the preset deviation threshold, it is determined that the overlap ratio causes an extraction deviation.

[0088] It should be noted that the reference spectrum of the preset target sound is obtained. The spectrum is the power spectral density obtained by Fourier transforming the normal operating sound of equipment or standard speech signal collected in a quiet environment. The separated noise spectrum data... and Alignment is performed in the frequency domain, with both using the same frequency resolution. And frequency band division, among which, , With a sampling frequency of 44.1kHz, N FFT The number of points in the Fourier transform is 4096. Next, a spectral energy density comparison matrix is ​​constructed. Its dimensions are ,in This represents the total number of frequency bands. The b-th row of the matrix corresponds to the frequency band. The two columns are as follows:

[0089]

[0090] The unit of energy density is To facilitate subsequent calculations, the global maximum value is usually normalized to obtain the result. The normalized spectral energy density comparison matrix is ​​denoted as... Its element value range is It intuitively reflects the relative intensity distribution of target sound and noise energy in each frequency band.

[0091] It should be noted that the spectral energy density comparison matrix... Analyze each frequency band line by line to identify overlapping frequency ranges. If within a certain frequency band... The above simultaneously satisfy:

[0092]

[0093] in The energy activation threshold, typically set to 0.1, indicates that the frequency band belongs to the frequency overlap region, denoted as . For each overlapping frequency band Calculate its amplitude masking effect numerical value. :

[0094]

[0095] in This is a small constant used to prevent division by zero. This represents the masking strength of noise on the target sound in that frequency band; a positive value indicates that the noise energy is higher than the target sound, and the masking effect is significant. Let... The ratio of noise energy to target sound energy is represented by the decibel value, which is obtained by converting the masking effect back to a linear proportion.

[0096] Next, the amplitude masking effect values ​​are subjected to frequency domain weighted integration to calculate the overlap ratio:

[0097] ,

[0098] Among them, weight This is used to emphasize the influence of higher energy frequency bands. The weight is proportional to the total energy of the frequency band (target sound + noise), reflecting that the higher the energy of the frequency band, the greater its contribution to the overall masking effect. , is the normalization factor. It is a dimensionless quantity with a range of values. This value reflects the overall degree to which noise masks the target sound in the frequency spectrum. The closer this value is to 1, the stronger the masking effect caused by overlap, and the greater the deviation that may be introduced during the noise extraction process.

[0099] It should be noted that if the aforementioned overlap ratio , If the preset deviation threshold is used, it is determined that the frequency domain overlap between the noise and the target sound has a significant impact on the separation quality.

[0100] It should be noted that setting a preset deviation threshold Statistical analysis was conducted based on 10,000 samples from the aforementioned industrial noise multi-scenario benchmark database. Specifically, ROC curve analysis showed a classification accuracy of 85%, with a 95% confidence interval of [0.28, 0.32]. This threshold ensures that Type I errors (false alarms) are kept below 5% while maximizing Type II error (false alarm) control. The Bootstrap method was used to resample 1000 times, and the threshold distribution was calculated to ultimately determine the appropriate threshold. .

[0101] In step S15, when the overlap ratio causes an extraction deviation, the dynamic change characteristics of the separated noise spectrum data are extracted and incorporated into the spectrum extraction process to obtain an accurate noise spectrum; when there is no overlap ratio causing an extraction deviation, the separated noise spectrum data is directly output as an accurate noise spectrum.

[0102] In one implementation, when the overlap ratio causes extraction deviation, the dynamic change characteristics of the separated noise spectrum data are extracted and incorporated into the spectrum extraction process to obtain an accurate noise spectrum, including:

[0103] S151, if the overlap ratio causes an extraction deviation, extract the time-varying envelope and frequency modulation features of the separated noise spectrum data;

[0104] S152, map the time-varying envelope and the frequency modulation feature to the frequency domain space, and calculate the amplitude fluctuation trajectory and instantaneous frequency component;

[0105] S153, Calculate the correction gain factor and inverse filter coefficient based on the amplitude fluctuation trajectory and the instantaneous frequency component;

[0106] S154. Based on the modified gain factor and the inverse filter coefficient, construct a spectrum reconstruction matrix, analyze the spectrum reconstruction matrix, and obtain the accurate noise spectrum.

[0107] It should be noted that if the overlap ratio is determined to cause extraction bias, then the separated noise spectrum data... Time-frequency feature extraction is performed. Because... The frequency domain static spectrum needs to be converted back to its time domain representation first. The noise time domain sequence is obtained through inverse discrete Fourier transform:

[0108] ,

[0109] The phase spectrum is generated using random phases, which are uniformly distributed to ensure the reconstructed signal retains the random characteristics of the noise. Next, [the following steps are performed]... Perform Hilbert transform to construct its analytic signal. Extract the following dynamic features from the analytical signal: time-varying envelope features This sequence characterizes the time-domain variation of noise amplitude, reflecting the dynamic characteristics of energy fluctuations. Instantaneous frequency characteristics. The first-order difference approximation calculation is performed using the central difference method, with a step size of 2 sampling points. The starting point... Using forward difference, for the endpoint Backward differential modulation is used, with a step size of one sampling point interval. Frequency modulation depth. ,in The noise center frequency is the median of the instantaneous frequencies. This parameter quantifies the relative intensity of the frequency modulation. Envelope modulation index. ,in and These are the standard deviation and mean of the envelope sequence, respectively, reflecting the degree of amplitude modulation. Finally, these dynamic features are integrated into a dynamic change feature set. It should be noted that the time-varying envelope sequence extracted in step S151... and instantaneous frequency sequence Preprocessing was performed, including removing the trend using a linear trend and Z-score normalization. , .

[0110] Next, the time-domain dynamic features are mapped to the frequency domain space using short-time Fourier transform, and the normalized envelope sequence is then processed. Perform STFT envelope spectrum calculation The window function used is the Hamming window function. The window length is typically twice the main period of the envelope signal, for example, 50ms. It reflects the time-varying energy distribution of the envelope wave across different frequency bands (i.e., modulation frequencies). Instantaneous frequency wave spectrum By analyzing the normalized instantaneous frequency sequence The STFT is obtained by performing the same method as calculating the envelope spectrum. This spectrum characterizes the frequency domain pattern of instantaneous frequency changes and reveals the main components of frequency modulation.

[0111] Based on the above time-frequency representation, key dynamic components are extracted, and amplitude fluctuation trajectories are extracted. In the envelope spectrum Extracting energy The amplitude trajectory corresponding to the strongest modulation frequency component:

[0112]

[0113] This trajectory reflects the main time-varying pattern of the envelope fluctuations. Instantaneous frequency components. In the instantaneous frequency fluctuation spectrum Extract the frequency modulation component with the strongest energy:

[0114]

[0115] This component represents the intensity variation of the dominant frequency modulation mode.

[0116] The final output is the amplitude fluctuation trajectory. and instantaneous frequency components Together, these two constitute the time-frequency characterization of noise dynamics.

[0117] It should be noted that the amplitude fluctuation trajectory mentioned above... With instantaneous frequency components Smoothing and normalization were performed, and a low-pass filter using locally weighted regression (Loess) with a window size of 20 was applied. A quadratic polynomial was used to extract the amplitude fluctuation trajectory. With instantaneous frequency components The trend component is obtained by normalizing the trend component by its maximum value. , Correction gain factor It is used to dynamically compensate for noise energy loss or distortion caused by overlap, and its calculation incorporates the normalized trajectory of amplitude and frequency modulation:

[0118] ,

[0119] Among them, the weighting coefficient , Based on historical data, these parameters reflect the contributions of amplitude fluctuations and frequency modulation to energy loss. Restricted to Within a range, usually This is to ensure the numerical stability of the correction process.

[0120] Inverse filter coefficients This is used to selectively enhance the noise spectrum in the frequency domain to recover frequency band details lost due to the masking effect. First, a time-frequency masking template is constructed. The template is dominated by amplitude fluctuations in the modulation frequency. With instantaneous frequency dominating modulation frequency It is formed by superimposing Gaussian weighted projections:

[0121] ,

[0122] in, and Take 1 / 3 octave bandwidth of the corresponding frequency band, with the reference frequency being the instantaneous frequency of the strongest modulation component in the envelope fluctuation, extracted in step S152, with an attenuation coefficient e = 2.0. The inverse filter coefficient is calculated as follows:

[0123]

[0124] Among them, the enhancement strength factor =0.4. Final output correction gain factor. With inverse filter coefficients .

[0125] It should be noted that the time-varying correction gain factor calculated based on step S153... and time-frequency domain inverse filter coefficients Construct a time-frequency domain spectral reconstruction matrix This matrix is ​​used for the estimation of the original noise spectrum. This is used to generate the corrected noise spectrum. Reconstruction matrix in This is the time-frequency window function matrix, typically taken as the time-frequency distribution of the Hanning window, used to control the time-frequency resolution balance during the reconstruction process. Next, the original noise spectrum is estimated... Multiplying the noise spectrum element-wise with the reconstruction matrix yields the preliminary reconstructed noise spectrum. ,in, This represents the Hadamard product (element-by-element multiplication). To ensure energy conservation and physical plausibility of the reconstructed spectrum, [the following is omitted as it is not explicitly stated in the original text]. Perform constraint optimization:

[0126]

[0127] in It is the Frobenius norm. This is a time-frequency total variation regularization term used to suppress spurious fluctuations introduced by reconstruction. Here, is the regularization coefficient, typically taken as 0.1. This convex optimization problem can be solved using the gradient projection method, where the initial step size is set to 1.0, the reduction factor to 0.5, the maximum number of iterations to 200, and Armijo constant. The convergence condition also requires the relative gradient norm. And relative solution changes ,in The optimized noise spectrum is obtained by solving the problem. Then, by integrating along the time dimension, the static, precise noise spectrum is obtained. , where T is the total duration of the signal.

[0128] In step S16, the intensity value of the energy is calculated based on the precise noise spectrum, and the intensity assessment result is determined, including:

[0129] S161, Analyze the precise noise spectrum to obtain discrete frequency point data and calculate the spectral density function;

[0130] S162, Integrate the spectral density function to obtain the interval frequency band energy value, and construct a full-band energy accumulation sequence based on the interval frequency band energy value;

[0131] S163, perform smoothing calculation on the full-band energy accumulation sequence to generate an instantaneous intensity fluctuation curve, and extract dynamic intensity feature quantities based on the instantaneous intensity fluctuation curve;

[0132] S164, perform strength evaluation on the dynamic strength characteristic quantity to obtain the strength evaluation result.

[0133] It should be noted that the precise noise spectrum is... Stored in the form of discrete frequency points, frequency resolution From the original signal sampling rate With the number of spectral analysis points N FFT Decide, If the frequency range covered by the spectrum is ,in Then the discrete frequency point sequence is:

[0134]

[0135] The corresponding spectral amplitude, i.e., energy density, is In this embodiment, a uniform logarithmic power spectral density of dB / Hz is used. To facilitate subsequent energy integration, the spectral amplitude is converted into a linear power spectral density function. After conversion The unit is .

[0136] Discrete data points are obtained through cubic spline interpolation. Fit to a continuous spectral density function .

[0137] It should be noted that the entire frequency band range It is divided into L consecutive frequency bands, each with a bandwidth of [missing value]. The frequency band is divided using equal intervals. For example, if the frequency range... Frequency bandwidth is taken ,but For the first frequency band Its frequency band energy value By analyzing the spectral density function The integral is obtained by integrating over this interval. The trapezoidal rule is used for integration to ensure computational accuracy. Trapezoid integral formula. in, The spectral density function is the interval between adjacent discrete frequency points. The values ​​at these discrete points Next, based on the energy values ​​of each frequency band... The cumulative energy values ​​are calculated in ascending order of frequency to construct a full-band energy accumulation sequence. , The sequence satisfies , That is, the total noise energy. This indicates the frequency from the lowest to the highest. The cumulative energy at the end of each frequency band reflects the cumulative distribution characteristics of noise energy in the frequency domain. The final output is a full-band energy accumulation sequence. This reflects the band-by-band cumulative distribution of noise energy in the frequency domain.

[0138] It should be noted that the full-band energy accumulation sequence... Smoothing is performed to suppress high-frequency fluctuations introduced by frequency band division and numerical integration. A Savitzky-Golay filter is used for convolutional smoothing, with a window length of [value missing]. The unit is frequency bands. Based on historical data testing, the window length is within the range of 3-7 frequency bands. While suppressing the random fluctuations (mainly high-frequency components) caused by numerical integration, it avoids over-smoothing the true energy accumulation trend (low-frequency components). The polynomial order is 2, resulting in a smoothed accumulation sequence. The instantaneous intensity fluctuation curve is obtained by performing a first-order difference on the smoothed sequence. in For the first The width of each frequency band.

[0139] in, It reflects the rate of change of noise energy per unit frequency band in the frequency domain, that is, the gradient of noise intensity distribution along the frequency range. Based on Extract the following dynamic intensity features: maximum fluctuation amplitude. It characterizes the maximum fluctuation intensity of noise intensity across the entire frequency band; fluctuation energy concentration. The calculation formula is: The value is kurtosis minus 3, which reflects the sharpness of the fluctuation distribution. A positive value indicates that the fluctuation is concentrated in a few frequency bands, while a negative value indicates that the fluctuation distribution is flat. The dominant frequency band represents the arithmetic mean of the instantaneous intensity fluctuation curve over the entire frequency band. Through calculation The short-time Fourier transform (STFT) or wavelet transform is used to extract the frequency band where the energy is most concentrated, and the corresponding wave frequency is obtained. Fluctuation attenuation coefficient Through the In the high-frequency band, such as Perform exponential fitting The optimal fit is obtained and This reflects the attenuation rate of noise energy fluctuations in the high-frequency band; it should be added that... The parameters in the exponential fitting are defined as amplitude coefficients, representing the initial amplitude value of the exponential curve at the starting point of the frequency band index; these features are combined into a dynamic intensity feature vector. .

[0140] It should be noted that the dynamic intensity feature vector is input into the pre-trained intensity assessment model. This model is a regression model based on Gradient Boosting Decision Tree (GBDT). The training data comes from three types of labeled samples: first, measured acoustic data, accounting for approximately 60%, obtained by simultaneously measuring noise signals using a sound level meter in a standard acoustic environment to obtain the A-weighted equivalent sound level. The noise intensity is categorized into three sets of data: 1) Real-world intensity labels in dB; 2) Simulated synthetic data, accounting for approximately 25%, generated based on a noise source power simulation model, labeled as theoretical power level in dB; and 3) Subjective rating data, accounting for approximately 15%, from multiple listeners who subjectively rated the noise intensity from 0 to 100 points, with the average score used as the label. To eliminate dimensional differences, all intensity labels are standardized to dimensionless intensity scores. A linear mapping is used for the sound pressure level / sound power level (dB). ,in , This represents the actual possible range. For subjective ratings, it is directly used as... .

[0141] The model maps feature vectors to intensity scores through multi-layer decision tree ensemble. :

[0142]

[0143] in, For the first The prediction output of each decision tree is calculated, where v is the learning rate (typically 0.1) and U is the number of trees (typically 100). Simultaneously, a baseline intensity value is calculated based on the overall characteristics of the frequency band energy accumulation sequence. ,in, The total noise energy, As the reference sound pressure level, For signal duration, Sampling frequency, The sampling sensitivity of the microphone is a parameter inherent to the device itself, measured in V / Pa. Ultimately, the intensity assessment result is a weighted average of the intensity score and the baseline intensity value. Among them, the weight weight = 0.7, which focuses more on the intensity change characteristics reflected by dynamic features.final This is a dimensionless scalar, typically ranging from 0 to 100; a higher value indicates a higher noise intensity. Additionally, an output intensity level label is provided; if I... final <30, marked as "low intensity"; if 30≤I final <60, marked as "medium intensity"; if I final ≥60 is marked as "high strength". The strength assessment result format is as follows: .

[0144] It should be noted that the weights were set based on the optimization results of grid search cross-validation. The weights were determined when the risk prediction error was minimized on the test set through grid search and cross-validation. On a dataset containing 5000 historical noise samples, 5-fold cross-validation was performed on the candidate weights weight ∈ {0.5, 0.6, 0.7, 0.8, 0.9}, using the mean absolute error (MAE) and Pearson correlation coefficient between the intensity assessment results and the true sound pressure level labels as indicators. Experimental results show that when weight = 0.7, the MAE reaches its minimum of 43 dB, and the correlation coefficient R reaches its maximum of 0.912, indicating that the combination of dynamic features and baseline energy information is optimal at this value.

[0145] In step S17, the intensity assessment result is compared with a preset pollution threshold to calculate the noise pollution level and output the noise identification result, including:

[0146] S171, Based on the intensity assessment results, index the standard pollution threshold from the preset pollution threshold matrix;

[0147] S172, calculate the Euclidean distance between the intensity assessment result and the standard pollution threshold, generate a difference distribution vector, and input the difference distribution vector into a preset risk quantification function to obtain the impact index;

[0148] S173, Determine the noise pollution level based on the impact index, and use a nonlinear correction factor to perform a weighted calculation on the intensity assessment results to obtain a corrected intensity value;

[0149] S174, the modified intensity value is associated with the noise pollution level and encapsulated to output the noise identification result.

[0150] It should be noted that a contamination threshold matrix T is pre-stored. hazard This matrix is ​​constructed based on different industrial scenarios, equipment types, and exposure times, and its dimensions are: ,in Indicates the number of noise types / scene categories. The matrix data, representing the number of intensity level segments, is sourced from a multi-scenario industrial noise benchmark database. Statistical analysis yielded the intensity thresholds for pollution occurrence in each scenario. These raw data were then compared with the intensity assessment results.final The same normalized mapping is converted to a dimensionless threshold and organized into a matrix by scene category (rows) and intensity level (columns). Matrix elements Indicates the first In similar scenarios, the intensity level is The corresponding standard pollution threshold, in units of I based on the intensity assessment result. final Consistent dimensionless scalar.

[0151] The indexing process is as follows: First, determine the scene category r. Then, determine the intensity level segment c, based on the intensity assessment result I. final The pre-defined segmented interval is determined. For example, for Level 1: 0 ≤ I final <20, Level 2: 20≤I final <40, Level 3: 40≤I final <60, Level 4: 60≤I final <80, Level 5: I final ≥80, let I final It falls into the c-th level segment. The index standard pollution threshold is determined from matrix T. hazard Extract the corresponding standard pollution threshold. .

[0152] For example, if the current scenario is a "stamping workshop" (r=3), the strength assessment result I final =58 falls into level 3, then the standard pollution threshold T std =T hazard (3,3), assuming the value in the matrix is ​​55. The final output index yields the standard contamination threshold T. std This is used for subsequent difference calculations and risk quantification.

[0153] It should be noted that the method for determining scene categories includes maintaining a scene feature database. Each record contains equipment acoustic features, i.e., noise spectrum templates of typical equipment, such as fans, pumps, and cutting machines; environmental acoustic tags, i.e., background noise spectrum features, reverberation time, environmental type, such as workshop, outdoor, and machine room; and job type codes, i.e., process codes defined according to standard job classifications, such as ISO 10628. During matching, the spectral envelope features of the input audio signal are first extracted and dynamically time-warped (DTW) distance is calculated between it and the equipment acoustic templates in the database. The equipment type with the smallest distance is selected as the candidate. The equipment acoustic templates are obtained by extracting MFCC features of the noise of various types of equipment under standard operating conditions. Simultaneously, sound field analysis, such as the ratio of direct sound to reverberation sound energy, is used to match the environmental acoustic tags, and the job type code is matched in conjunction with the user-input job description. The three factors are combined using a weighted voting strategy: equipment matching weight 0.5, environment matching weight 0.3, and job matching weight 0.2. If the overall matching degree is higher than 0.7, the corresponding scenario category r is determined; otherwise, it is classified as a general industrial scenario (r=0).

[0154] It is worth noting that the standard pollution threshold T obtained from the above steps std The strength assessment result I obtained in step S16 final Compare them and calculate their absolute difference. Since noise pollution depends not only on intensity differences but also on factors such as frequency distribution and exposure time, a multidimensional difference distribution vector is constructed. This vector contains the following four dimensions of difference measures: intensity difference... (Directly reflects the degree of intensity deviation); differences in the proportion of high-frequency frequencies ,in H represents the energy percentage of the noise signal in the high-frequency band (>4000Hz). std This is a reference value for the high-frequency proportion of standard noise in this scenario; fluctuation stability differences. ,in Instantaneous intensity fluctuation curve standard deviation The standard deviation of fluctuation is a reference value under standard scenarios; exposure time coefficient. ,That The actual estimated exposure time is expressed in minutes. This represents the upper limit of the allowed continuous exposure time in this scenario. The difference distribution vector is... Next, the Euclidean distance D is calculated as a measure of overall variance. .

[0155] The difference distribution vector d is input into a preset risk quantification function, which is a nonlinear mapping based on the radial basis function (RBF):

[0156] ,

[0157] in, This serves as a reference difference vector (usually the zero vector, indicating complete compliance with the standard). The scale parameter controls the sensitivity, and is usually set to a value that is... The setting is based on the analysis of the distribution of the difference vector norm ‖d‖ in historical data. It was found that its 75th percentile is about 0.7-0.8 times the maximum value. This setting allows the influence index to enter the rapid rise zone when the difference reaches the 75th percentile, thus maintaining good discrimination under common working conditions. The vector of greatest historical difference; The normalization factor is the historical maximum Euclidean distance. The final output impact index... For range The dimensionless value of , the closer it is to 1, the higher the risk of pollution.

[0158] It should be noted that the standard high-frequency ratio H std The H value is obtained by statistically analyzing the high-frequency energy percentage of all compliant noise samples in this scenario from a historical standard noise database, i.e., samples whose intensity does not exceed the safety threshold. Specifically, for each compliant sample, the proportion of its energy in the high-frequency band (>4000Hz) relative to the total energy is calculated, and then the median of all sample proportions is taken as H. std Standard deviation of standard fluctuation The acquisition method is similar; the median of the standard deviation sequence of the instantaneous intensity fluctuation curve is extracted from the same batch of compliant samples as... .

[0159] It should be noted that, according to the aforementioned impact index This is mapped to a preset noise pollution level. A segmented threshold is used for division; if... Then it is classified as Level I (low pollution). It is classified as Level II (medium pollution). It is classified as Level III (high pollution). This is classified as Level IV (severe pollution). Each level corresponds to a weighting coefficient. Level I Level II Level III Level IV The weighting was based on historical accident statistical analysis. Retrospective analysis of 5000 noise-related accidents showed that for each increase in pollution level, the probability of the accident increased by approximately 1.0, 1.2, 1.5, and 2.0 times, respectively. Next, a nonlinear correction factor was constructed based on the nonlinear characteristics of the impact index. Used for evaluating the original strength results I finalPollution-adaptive weighting was applied. The correction factor was designed as an S-curve.

[0160]

[0161] Among them, parameters Control the maximum correction range. Adjust the curve steepness. , where is the offset, causing the correction factor to begin to change significantly when the influence index is of moderate magnitude, and tanh is the hyperbolic tangent function. This function in When it approaches 0.6, When the value approaches 1.4, the pollution risk is smoothly modulated to the intensity value.

[0162] Finally, the grade weighting coefficients and nonlinear correction factors are combined to perform a weighted calculation on the original intensity assessment results, resulting in the corrected intensity value. .

[0163] It should be noted that the corrected intensity value The noise pollution level is structurally correlated to generate the following noise identification result. For example, the format can be {original intensity value: I} final Corrected intensity value: Pollution levels: Level I (Low Pollution) / Level II (Medium Pollution) / Level III (High Pollution) / Level IV (Severe Pollution), Level weight: Correction factor: Impact Index: Difference vector: Reference threshold: T std Calculate the timestamp: }

[0164] In summary, this invention discloses an industrial noise identification method that achieves accurate extraction of the noise spectrum.

[0165] Reference Figure 2 The second embodiment of the present invention provides an industrial noise identification system, comprising:

[0166] The audio acquisition module is used to acquire audio signals from industrial scenarios and perform time-frequency decomposition on the audio signals using wavelet transform to obtain preliminary decomposed data.

[0167] The coupling classification module is used to construct coupling feature vectors based on the preliminary decomposition data, and to classify the coupling feature vectors using a preset support vector machine to obtain the boundary values ​​of the coupling region.

[0168] The filtering module is used to perform filtering on the boundary value of the coupling region if the boundary value of the coupling region exceeds a preset boundary threshold, so as to obtain the separated noise spectrum data.

[0169] The overlap determination module is used to calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and to determine whether the overlap ratio causes an extraction deviation.

[0170] The dynamic correction module is used to extract the dynamic change characteristics of the separated noise spectrum data and incorporate them into the spectrum extraction process when the overlap ratio causes extraction deviation, so as to obtain an accurate noise spectrum; when there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as an accurate noise spectrum.

[0171] The intensity assessment module is used to calculate the intensity value of the energy based on the precise noise spectrum and determine the intensity assessment result;

[0172] The result output module is used to compare the intensity assessment result with the preset pollution threshold, calculate the noise pollution level, and output the noise identification result.

[0173] It should be noted that the industrial noise identification system provided in this embodiment of the invention is used to execute all the process steps of the industrial noise identification method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0174] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0175] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. An industrial noise recognition method characterized by, include: Acquire audio signals from industrial scenes and perform wavelet basis function matching to generate frequency coefficients; Calculate the modulus maxima of the frequency coefficients, determine the singularity locations based on the propagation law of the modulus maxima, and separate the frequency coefficients according to the singularity locations to obtain preliminary decomposition data; Based on the preliminary decomposition data, a coupling feature vector is constructed, and a preset support vector machine is used to classify the coupling feature vector to obtain the boundary values ​​of the coupling region; If the boundary value of the coupling region exceeds a preset boundary threshold, the coupling frequency band data sequence corresponding to the boundary value of the coupling region is extracted and adaptive iterative decomposition is performed to obtain the intrinsic mode component set; based on the instantaneous frequency fluctuation variance of the intrinsic mode component set, the noise-dominant component is selected. The dominant noise component is subjected to Hilbert transform and stripping to obtain a background noise sequence, which is then converted into a frequency domain energy distribution to obtain separated noise spectrum data. Calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and determine whether the overlap ratio causes an extraction deviation; When the overlap ratio causes extraction deviation, the time-varying envelope and frequency modulation features of the separated noise spectrum data are extracted and mapped to the frequency domain space, and the amplitude fluctuation trajectory and instantaneous frequency components are calculated; based on the amplitude fluctuation trajectory and the instantaneous frequency components, the correction gain factor and inverse filter coefficient are calculated; based on the correction gain factor and the inverse filter coefficient, a spectrum reconstruction matrix is ​​constructed, and the spectrum reconstruction matrix is ​​analyzed to obtain the accurate noise spectrum; when there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as the accurate noise spectrum. The intensity value of the energy is calculated based on the precise noise spectrum, and the intensity assessment result is determined. The intensity assessment results are compared with a preset pollution threshold to calculate the noise pollution level and output the noise identification results.

2. The industrial noise identification method according to claim 1, characterized in that, The process of acquiring audio signals from an industrial scene and performing wavelet basis function matching to generate frequency coefficients includes: Acquire audio signals from industrial scenes and convert them into digital audio sequences; Frequency coefficients are generated by matching wavelet basis functions to the spectral energy distribution of the digital audio sequence.

3. The industrial noise identification method according to claim 1, characterized in that, The step of constructing a coupling feature vector based on the preliminary decomposition data, and classifying the coupling feature vector using a preset support vector machine to obtain the boundary values ​​of the coupling region includes: Extract the spectral energy density data from the preliminary decomposition data, and construct a coupled feature vector based on the spectral energy density data; The coupled feature vectors are input into a preset support vector machine to establish the hyperplane equation and calculate the geometric distance between sample points. Based on the geometric distance of the sample points, the classification interval is determined, and the decision function is solved for the classification interval to obtain the boundary of the coupling region. Extract the spectral energy value corresponding to the boundary of the coupling region, and confirm the spectral energy value as the boundary value of the coupling region.

4. The industrial noise identification method according to claim 1, characterized in that, The step of calculating the overlap ratio between the separated noise spectrum data and the preset target sound, and determining whether the overlap ratio causes extraction deviation, includes: The separated noise spectrum data is mapped to the frequency domain of the preset target sound to construct a spectrum energy density comparison matrix; The frequency overlap intervals in the spectral energy density comparison matrix are analyzed, the amplitude masking effect value is calculated, and the amplitude masking effect value is subjected to frequency domain weighted integration to obtain the overlap ratio. If the overlap ratio is higher than a preset deviation threshold, it is determined that the overlap ratio causes an extraction deviation.

5. The industrial noise identification method according to claim 1, characterized in that, The step of calculating the energy intensity value based on the precise noise spectrum and determining the intensity assessment result includes: The precise noise spectrum is analyzed to obtain discrete frequency point data and the spectral density function is calculated. Integrating the spectral density function yields the interval frequency band energy value, and a full-band energy accumulation sequence is constructed based on the interval frequency band energy value; The full-band energy accumulation sequence is smoothed to generate an instantaneous intensity fluctuation curve, and dynamic intensity features are extracted based on the instantaneous intensity fluctuation curve. The dynamic strength characteristic quantity is subjected to strength evaluation to obtain the strength evaluation result.

6. The industrial noise identification method according to claim 1, characterized in that, The step of comparing the intensity assessment result with a preset pollution threshold, calculating the noise pollution level, and outputting the noise identification result includes: Based on the intensity assessment results, the standard pollution threshold is indexed from the preset pollution threshold matrix; Calculate the Euclidean distance between the intensity assessment result and the standard pollution threshold, generate a difference distribution vector, and input the difference vector into a preset risk quantification function to obtain the impact index; The noise pollution level is determined based on the impact index, and the intensity assessment results are weighted using a nonlinear correction factor to obtain a corrected intensity value. The modified intensity value is associated with the noise pollution level and encapsulated to output the noise identification result.

7. An industrial noise identification system, characterized in that, include: The audio acquisition module is used to acquire audio signals from industrial scenarios and perform wavelet basis function matching to generate frequency coefficients. Calculate the modulus maxima of the frequency coefficients, determine the singularity locations based on the propagation law of the modulus maxima, and separate the frequency coefficients according to the singularity locations to obtain preliminary decomposition data; The coupling classification module is used to construct coupling feature vectors based on the preliminary decomposition data, and to classify the coupling feature vectors using a preset support vector machine to obtain the boundary values ​​of the coupling region. The filtering module is used to extract the coupling frequency band data sequence corresponding to the coupling region boundary value and perform adaptive iterative decomposition if the boundary value of the coupling region exceeds a preset boundary threshold, so as to obtain the intrinsic mode component set; and to filter out the noise-dominant component based on the instantaneous frequency fluctuation variance of the intrinsic mode component set. The dominant noise component is subjected to Hilbert transform and stripping to obtain a background noise sequence, which is then converted into a frequency domain energy distribution to obtain separated noise spectrum data. The overlap determination module is used to calculate the overlap ratio between the separated noise spectrum data and the preset target sound, and to determine whether the overlap ratio causes an extraction deviation. The dynamic correction module is used to extract the time-varying envelope and frequency modulation features of the separated noise spectrum data and map them to the frequency domain when the overlap ratio causes extraction deviation; calculate the amplitude fluctuation trajectory and instantaneous frequency components; calculate the correction gain factor and inverse filter coefficient based on the amplitude fluctuation trajectory and instantaneous frequency components; construct a spectrum reconstruction matrix based on the correction gain factor and inverse filter coefficient; and parse the spectrum reconstruction matrix to obtain the accurate noise spectrum. When there is no overlap ratio causing extraction deviation, the separated noise spectrum data is directly output as the accurate noise spectrum. The intensity assessment module is used to calculate the intensity value of the energy based on the precise noise spectrum and determine the intensity assessment result; The result output module is used to compare the intensity assessment result with the preset pollution threshold, calculate the noise pollution level, and output the noise identification result.

Images