Exhaust valve self-adaptive noise reduction method based on environmental noise recognition

Abstract

The invention discloses an exhaust valve self-adaptive noise reduction method based on environmental noise recognition, and belongs to the technical field of exhaust valve data noise reduction. The problem that noise reduction parameters of an exhaust valve cannot be automatically generated according to the actual environment noise condition at present, and then signals cannot be effectively processed is solved. The method comprises the following steps: acquiring a noisy exhaust valve audio signal, preprocessing and standardizing the signal, then sending the signal into a deep learning model for identification to obtain an environment category of a valve audio, and acquiring a binary mask corresponding to a prior environment category according to the environment category; performing short-time Fourier transform on the preprocessed audio to obtain a frequency amplitude spectrum Yn (f) and a phase phi n (f) of a signal; and performing point-by-point multiplication operation on the Yn (f) and a binary mask Mv (f) corresponding to an environment category to obtain a signal, and performing inverse STFT based on the phase phi n (f) to recover a time domain signal # imgabs0 # to synthesize a continuous waveform through an overlap-add method.

Description

Technical Field

[0001] The present invention belongs to the technical field of exhaust valve data noise reduction, and particularly relates to an adaptive noise reduction method for exhaust valves. Background Art

[0002] In the fields of industrial automation and equipment monitoring, the exhaust valve is an important part of the pipe network system, and accurate monitoring of its working state is crucial for ensuring the safe operation of the system. However, traditional methods for monitoring the state of exhaust valves mainly rely on directly collecting the action sounds or vibration signals of the valves for analysis, and have the following deficiencies:

[0003] During the actual operation of the exhaust valve, due to the influence of the mechanical structure and working environment, the amplitude of the action sound generated is usually low, and at the same time, it is often accompanied by various environmental noise interferences from inside and outside the factory. Existing methods are difficult to effectively distinguish the sound of the valve itself from the background noise during the acquisition and processing process, resulting in the monitoring signal being mixed with noise, thereby affecting the accuracy of fault diagnosis and status assessment.

[0004] Most existing noise suppression technologies use fixed thresholds or preset models for noise reduction processing. When facing the changing industrial environmental noise, these methods are often difficult to make adaptive adjustments. Especially in the case where the exhaust valve noise itself has a low amplitude and complex spectral characteristics, the fixed noise reduction strategy is likely to cause the loss of voice or action signal components, affecting the subsequent signal processing and recognition effects.

[0005] With the development of artificial intelligence and deep learning technologies, some studies have begun to attempt to combine environmental noise classification with signal processing. However, most applications mainly focus on the field of speech noise reduction, and there is still a lack of systematic research and application examples for the noise processing of industrial equipment such as exhaust valves. In the existing technology, there is no complete solution that can automatically generate noise reduction parameters according to the actual environmental noise situation and then adaptively adjust the signal processing flow. Summary of the Invention

[0006] The present invention aims to solve the problem that currently, it is impossible to automatically generate noise reduction parameters for exhaust valves according to the actual environmental noise situation, and thus it is impossible to effectively process the signals.

[0007] An adaptive noise reduction method for exhaust valves based on environmental noise recognition includes:

[0008] Obtain the noisy exhaust valve audio signal, perform preprocessing and normalization, and then send it to a deep learning model for recognition to obtain the environmental category where the valve audio is located, and obtain the binary mask M v (f);

[0009] Perform STFT short-time Fourier transform on the preprocessed audio to obtain the frequency amplitude spectrum of the signal as |Yn (f); sum phase φ n (f); then multiply |Y n (f) with the binary mask M corresponding to the environmental category v (f) point - by - point to obtain a signal Based on the phase φ n (f), perform inverse STFT to recover the time - domain signal Synthesize a continuous waveform through the overlap - add method

[0010] Furthermore, the pre - processing process includes pre - emphasis processing, and the pre - emphasis processing is as follows:

[0011] x1′(n) = x1(n)-α·x1(n - 1)

[0012] Where n is the sampling point index in the discrete time series, representing the nth sampling value obtained during the sampling of the audio signal, and n is a non - negative integer; x1(n) is the sound signal before pre - emphasis at the nth sampling point; x1′(n) is the sound signal after pre - emphasis; α is the pre - emphasis coefficient.

[0013] Furthermore, the pre - processing process also includes down - sampling processing, and the sampling rate during down - sampling is 44.1KHz.

[0014] Furthermore, the binary mask M corresponding to the prior environmental category v (f) is obtained through the following steps:

[0015] Obtain a pure exhaust valve audio signal, perform pre - processing and normalization to get the normalized valve sound signal y1(n); normalize different environmental noise data to get the normalized different environmental noise data y2(n);

[0016] Perform weighted superposition of the valve sound signal with different environmental noise signals respectively:

[0017] y(n) = β·y1(n)+(1 - β)·y2(n)

[0018] Where SNR is the signal - to - noise ratio, and k is the adjustment coefficient;

[0019] The signal - to - noise ratio SNR is as follows:

[0020]

[0021] Where N represents the total number of sampling points of the signal;

[0022] Frame the noise signal y2(n) and perform windowing with a Hamming window:

[0023] y k y(n) = y2(n + kR)w(n), 0 ≤ n < L

[0024] where y2(·) is the standardized noise signal, k is the frame index; R is the frame shift, L is the frame length, and y k (n) is the k-th frame signal after windowing, and w(n) is the window function;

[0025] Based on Y k (f), calculate the magnitude spectrum |N k (f)| of the noise for each frame through short-time Fourier transform, and take the mean of the spectral magnitudes of multiple frames of noise to obtain the average magnitude spectrum where k represents the number of frames; thus, obtain the typical energy distribution of the noise

[0026] Frame the noisy signal y(n) and perform windowing processing:

[0027] y m y(n) = y(n + mR)w(n), 0 ≤ n < L

[0028] where m is the frame index, and y m (n) is the m-th frame signal after windowing;

[0029] Based on y m (n), calculate the magnitude spectrum |Y m (f)| of the noisy signal for each frame through short-time Fourier transform;

[0030] Generate ideal binary masks for different environmental noises respectively:

[0031]

[0032] where δ is a coefficient used to adjust the strictness of noise suppression;

[0033] Through the above process, obtain the binary mask M(f) for different environmental noises, and denote the binary masks for different environmental noises as M v (f).

[0034] Furthermore, the coefficient used to adjust the strictness of noise suppression is as follows:

[0035]

[0036] where |N k (f)| is the magnitude spectrum of the k-th frame; σ N (f) is the standard deviation.

[0037] Furthermore, when framing the noise signal y2(n), the frame length L is 20 - 40 ms.

[0038] Further, the window function

[0039] Further, the process of calculating the magnitude spectrum |N k (f)| of each frame of noise by short-time Fourier transform includes:

[0040] First, calculate the frequency spectrum

[0041]

[0042] where Y k (f) is the frequency spectrum of the k-th frame, f is the frequency component, and e -j2πfn is the exponential term of the Fourier transform;

[0043] Take the magnitude value of each frame of noise, that is, |N k (f)| = |Y k (f)|.

[0044] Further, the process of calculating the magnitude spectrum |Y m (f)| of each frame of noisy signal by short-time Fourier transform includes:

[0045] First, calculate the frequency spectrum:

[0046]

[0047] where Y m (f) is the complex frequency spectrum of the m-th frame, f is the frequency component, and e -j2πfn is the exponential term of the Fourier transform; expressed in the form of magnitude spectrum and phase spectrum:

[0048]

[0049] where |Y m (f)| is the magnitude spectrum of the m-th frame, and φ m (f) is the phase spectrum of the m-th frame;

[0050] Furthermore, obtain the magnitude spectrum |Y m (f)|.

[0051] Further, the deep learning model is pre-trained, and the training process of the deep learning model includes:

[0052] Based on the valve sound signal y(n) obtained by weighted superposition of the valve sound signal and different environmental noise signals respectively, a training set is constructed and then converted into a Mel spectrogram; the Mel spectrogram is used as the input of the deep learning model, and the classification result of the environmental noise is used as the output of the deep learning model. The classification result of the environmental noise is different environmental noise categories. Based on the training set converted into a Mel spectrogram, the deep learning model will be trained to obtain a trained deep learning model.

[0053] Beneficial effects:

[0054] The present invention identifies the exhaust valve signal to determine the classification model of environmental noise, then uses the ideal binary mask for different environmental noises respectively, and realizes the adaptive generation of noise reduction parameters and noise reduction through short-time Fourier transform, effectively improving the limitations of traditional methods under low signal-to-noise ratio conditions. Furthermore, it can effectively process the signal, not only improving the recognition effect, but also expanding the application scope of the present invention, thus providing a more accurate and intelligent solution for the state monitoring of industrial exhaust valves. Description of the drawings

[0055] Figure 1 It is a hardware system structure diagram.

[0056] Figure 2 It is a software operation interface diagram.

[0057] Figure 3 It is a software operation logic diagram.

[0058] Figure 4 It is a sound signal preprocessing flow chart.

[0059] Figure 5 It is a model framework diagram. Specific embodiments

[0060] Aiming at the problems in the background technology, the present invention proposes an adaptive noise reduction method for exhaust valves based on environmental noise recognition. By combining a high-performance hardware platform and an environmental noise classification model based on deep learning, it adaptively generates noise reduction parameters for different environmental noises, effectively improving the limitations of traditional methods under low signal-to-noise ratio conditions, thus providing a more accurate and intelligent solution for the state monitoring of industrial exhaust valves. The following is a detailed description of the present invention in combination with specific embodiments. Specific embodiment one:

[0062] The specific embodiment is an adaptive noise reduction method for exhaust valves based on environmental noise recognition, including the following steps:

[0063] S1. Acquisition of valve sound signal:

[0064] To collect a high-quality and pure exhaust valve data set, a pipeline valve sound signal recognition system based on NVIDIA Jetson NANOB01 is used to collect sound data.

[0065] As Figure 1 shown, the hardware of the pipeline valve sound signal recognition system mainly includes five parts: a power supply module, a sound collection microphone, a Jetson NANO core module, a liquid crystal display touch screen, and a TF card data storage module. In this embodiment, the software part of the system is written in python code.

[0066] The audio sampling rate is 48KHz. The interface is designed based on the PyQt5 library function to control the acquisition of audio signals, making the operation more convenient. Then, the interface is packaged into an application program using pyinstall, enabling the collection of audio without the need for a person with code knowledge. The software running interface is as Figure 2 shown, and the process of collecting audio signals is as Figure 3 shown.

[0067] Jetson NANO is equivalent to a small computer. After connecting to a monitor, it can collect and analyze the sound of valves running in the factory through an omnidirectional microphone. The acquisition system can be placed in different environments to obtain the sound signals of valves in various states in real time. It should be noted here that the pure valve sound signals collected in the valve factory are free from external environmental interference, while the acquisition of audio signals in different outdoor scenarios includes different environmental noises.

[0068] S2. Obtain environmental noise data:

[0069] To improve the generalization ability of the method of the present invention and meet the noise reduction requirements of valve noise signals in various different scenarios, this embodiment realizes data augmentation by simulating diverse background noises in the environment where the valve is located. For this purpose, the public data set UrbanSound8K, which is widely used in automatic urban environmental sound classification research, is selected. The data set contains 8,732 precisely labeled sound clips with a duration of no more than 4 seconds, covering 10 common urban sound categories, including air conditioner, car horn, children's games, dog barking, drilling, engine noise, gunshots, hammering, siren, and street music.

[0070] These categories comprehensively reflect various types of background noise in real life. For example, the air conditioner sound simulates the continuous, low-frequency background noise generated by central air conditioning or ventilation equipment in an indoor environment; the car horn sound simulates the environmental noise near busy urban roads, such as in traffic congestion areas, intersections, or the city center; the children's game sound simulates the environment around playgrounds, parks, or schools, reflecting the noise background brought by the gathering of people and laughter in children's activity areas; the dog barking sound simulates the background noise in residential areas or open areas of communities, reflecting the natural sound interference in the environment where pets are active and neighbors live; the drilling sound simulates the impact noise generated by construction machinery operations in construction sites or urban renovation areas, which is suitable for reflecting the possible construction environment interference around valves; the engine sound simulates the continuous background noise generated by equipment operation, vehicle startup, or mechanical equipment operation near industrial areas, parking lots, or factories, representing the sound environment of heavy industrial sites; the gunshot sound simulates occasional high-intensity pulse noise scenarios, which are relatively rare in normal environments but can be used to test the system's ability to suppress sudden noise and ensure stability in extreme situations; the hammer sound simulates the knocking sound generated by manual tools during operation in construction sites or repair workshops, which is suitable for reflecting the possible noise background during mechanical maintenance or repair in a small area; the siren sound simulates the long-ring signal background emitted by warning devices or large transportation tools in industrial parks, along railway lines, or port areas, reflecting the noise characteristics in emergency warning or transportation systems; the street music simulates the sound generated by playing background music or street artist performances in public areas such as commercial pedestrian streets, squares, or markets, reflecting the noise environment in areas with intensive cultural and entertainment activities. By preprocessing and synthesizing the audio in UrbanSound8K with the collected valve action signals, representative noisy valve audio data can be generated, which can then provide sufficient training and test samples for the subsequent adaptive noise reduction model.

[0071] S3. Sound preprocessing and normalization:

[0072] S31. This part mainly synthesizes the pure valve sound signal and the environmental noise signal to simulate the valve's state in different environments, providing data support for the subsequent adaptive environmental noise training model. For example, Figure 4 As shown, the preprocessing includes:

[0073] In real life, the key high-frequency components of the valve are above 2000Hz, which are high-frequency signals. Therefore, in order to compensate for the high-frequency attenuation of the valve sound signal during transmission and enhance its key high-frequency characteristics (such as the transient sound of valve operation), pre-emphasis processing is first performed:

[0074] x1′(n) = x1(n) - α·x1(n - 1)

[0075] Among them, n is the sampling point index in the discrete time series, representing the nth sampling value obtained during the sampling process of the audio signal. n is a non - negative integer (n = 0, 1, 2, …); x1(n) is the pre - emphasis sound signal at the nth sampling point; x1′(n) is the sound signal after pre - emphasis; α is the pre - emphasis coefficient, usually taking 0.9 < α < 1.0, and here α = 0.97 is selected.

[0076] S312. Ensure that the time lengths of the valve sound signal and the noise signal are the same, and clip the noise signal and the valve sound signal to the same length.

[0077] S313. Unify the sampling rates of the valve sound signal and the noise signal. The sampling rate of the noise signal is 44.1KHz. Therefore, we need to down - sample the valve audio to 44.1KHz and down - sample the collected valve sound signal as follows:

[0078] x1″(n)=resample(x1′(n),f s )

[0079] Among them, x1″(n) is the sampled signal, x1′(n) represents the sound signal before sampling; resample(·) represents down - sampling; f s is the down - sampling frequency. In this embodiment, the down - sampling frequency is 44.1KHz.

[0080] S32. Standardize the down - sampled valve sound signal and noise signal:

[0081]

[0082] Among them, μ is the signal mean, μ1 is the mean of the valve signal after down - sampling, μ2 is the mean of the environmental noise signal; σ is the signal standard deviation, σ1 is the standard deviation of the valve sound signal after down - sampling, σ2 is the standard deviation of the environmental noise signal; y1(n) is the standardized valve signal, y2(n) is the standardized environmental noise signal, and x2(n) is the environmental noise signal before standardization.

[0083] S305. Finally, perform weighted superposition of the valve sound signal with different environmental noise signals respectively:

[0084] y(n)=β·y1(n)+(1 - β)·y2(n)

[0085] Among them, SNR is the signal - to - noise ratio, k is the adjustment coefficient, and here k = 0.1;

[0086] The signal - to - noise ratio SNR is as follows:

[0087]

[0088] Among them, y1(n) is the standardized valve signal, y2(n) is the standardized environmental noise signal, and N represents the total number of signal sampling points. Since the sampling rates of the two downsampled signals are the same and they have been cropped to the same duration, the total number of sampling points of the two signals is equal.

[0089] S4. Feature extraction and construction of environmental classification noise model:

[0090] The noisy valve sound signal y(n) is divided into a training set and a validation set at a ratio of 8:2, and then it is converted into a Mel spectrogram, which captures the basic features of the audio. Based on the training set, the Mel spectrogram is input into the CNN model we constructed for training. The model is as Figure 5 shown. Finally, the classification result of the environmental noise is output, and the classification results include: air conditioner, car horn, children's game, dog barking, drilling, engine sound, gunshot, hammer sound, steam whistle sound, and street music.

[0091] S5. Generation of dynamic frequency domain mask:

[0092] The ideal binary mask (IBM) is a noise suppression method based on time-frequency analysis. Its core idea is to distinguish the frequency domain energy distribution of speech and noise by setting a threshold, so as to directly retain the speech dominant component and suppress the noise dominant component at the frequency points; it includes the following steps:

[0093] S501. Noise spectrum modeling: For the noise signal y2(n), it is framed, the frame length is usually 20 - 40 ms, and windowing is performed using a Hamming window, as shown in the formula:

[0094] y k (n) = y2(n + kR)w(n), 0 ≤ n < L

[0095] Among them, y2(·) is the standardized noise signal (discrete time series), k is the frame index (the k-th frame); R is the frame shift (the overlapping length between frames), L is the frame length (usually 20 - 40 ms), and y k (n) is the k-th frame signal after windowing, and w(n) is the windowing function. The windowing formula for the Hamming window is:

[0096]

[0097] After that, the magnitude spectrum |N k (f)| of each frame of noise is calculated through the short-time Fourier transform (STFT):

[0098] First, calculate the spectrum:

[0099]

[0100] Among them, Y k (f) is the spectrum of the k-th frame (including amplitude and phase information), f is the frequency component, and e -j2πfn is the exponential term of the Fourier transform.

[0101] Take the amplitude value of the noise for each frame, that is, |N k (f)| = |Y k (f)|;

[0102] After that, take the average of the spectral amplitudes of multiple frames of noise to obtain the typical energy distribution of the noise As shown in the formula:

[0103]

[0104] Among them, k is the number of frames used to calculate the average value, and N(f) is the average amplitude spectrum of the noise.

[0105] After that, calculate its standard deviation σ N (f) to quantify the noise fluctuation range, as shown in the formula:

[0106]

[0107] Among them, |N k (f)| is the amplitude spectrum of the k-th frame, is the average amplitude spectrum.

[0108] S502. Time-frequency decomposition of the noisy signal: Frame the noisy signal y(n) and window it (the window function needs to be consistent with the noise analysis) to reduce spectral leakage, and then it becomes y m (n). The formula is:

[0109] y m (n) = y(n + mR)w(n), 0 ≤ n < L

[0110] Among them, y(n) is the original noise signal (discrete time series), m is the frame index, R is the frame shift (the overlapping length between frames), L is the frame length (usually 20 - 40 ms), and y m (n) is the m-th frame signal after windowing, and w(n) is the window function. The windowing formula for the Hamming window is:

[0111]

[0112] Perform STFT on each frame of the signal to obtain the complex spectrum, and its formula is:

[0113]

[0114] Among them, Y m(f) is the complex frequency spectrum of the m-th frame, and e -j2πfn is the exponential term of the Fourier transform. The above formula can also be expressed in the form of the amplitude spectrum and the phase spectrum as follows:

[0115]

[0116] where, |Y m (f)| is the amplitude spectrum of the m-th frame, and φ m (f) is the phase spectrum of the m-th frame.

[0117] S503. Generate ideal binary masks for different environmental noises respectively:

[0118]

[0119] Threshold setting: δ is a coefficient used to adjust the strictness of noise suppression, specifically as follows:

[0120]

[0121] Because the noise variances σ N (f) of different environmental noises are different, the values of δ for different types are also different. The larger the noise variance, the larger the δ coefficient, and the smaller the noise variance, the smaller the δ coefficient.

[0122] Physical meaning: If the energy of the noisy signal at a certain frequency point is significantly higher than the noise baseline (exceeding δ times), it is determined that this frequency point is dominated by speech and is retained, that is, M(f) = 1, otherwise it is set to zero, M(f) = 0.

[0123] Through the above process, we can obtain binary masks M(f) for ten different environmental noises, and denote the binary masks of different environmental noises as M v (f).

[0124] S6. Testing stage:

[0125] In the testing stage, we use the valve audio acquisition system to collect audio signals in different environments where the exhaust valve is located, which are noisy sound signals. Then, the noisy valve sound signals are preprocessed, including pre-emphasis, clipping, and downsampling; then unified standardization processing is performed;

[0126] After that, we send its audio into the CNN model for recognition to obtain the environmental category where the valve audio is located.

[0127] Before this, we have obtained corresponding different binary masks for different environmental categories, assumed to be M v (f); in addition, the preprocessed audio is subjected to STFT short-time Fourier transform to obtain the frequency amplitude spectrum of the signal as |Y n (f)| and the phase φn (f), and then perform a point-by-point multiplication operation on the collected signal and the corresponding binary mask:

[0128]

[0129] Based on the original phase φ n (f), perform the inverse STFT (ISTFT) to restore the time-domain signal Synthesize a continuous waveform by the overlap-add method (the frame overlap rate is usually 50%) Eliminate the boundary effect caused by frame segmentation, so as to obtain a noise-reduced signal, and then realize the adaptive noise reduction of the valve under different environmental noises.

[0130] The above examples of the present invention are only used to illustrate in detail the calculation model and calculation process of the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation manners here. Any obvious changes or modifications derived from the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. An adaptive noise reduction method for an exhaust valve based on environmental noise recognition, characterized in that, including: Obtain the noisy exhaust valve audio signal, perform preprocessing and normalization, and then send it into the deep learning model for recognition to obtain the environmental category where the valve audio is located. According to the environmental category, obtain the binary mask M corresponding to the prior environmental category v (f); Performing STFT (Short-Time Fourier Transform) on the preprocessed audio yields the frequency magnitude spectrum of the signal as |Y n (f)| and the phase φ n (f); subsequently, multiply |Y n (f)| point-by-point with the binary mask M v (f) corresponding to the environmental category to obtain the signal Based on the phase φ n (f), perform inverse STFT to recover the time-domain signal Synthesize a continuous waveform through the overlap-and-add method 2. An adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 1, characterized in that, including: The preprocessing process includes pre-emphasis processing, and the pre-emphasis processing is as follows: x1′(n) = x1(n) - α·x1(n - 1) where n is the sampling point index in the discrete time series, representing the nth sampling value obtained during the sampling of the audio signal, and n is a non-negative integer; x1(n) is the sound signal before pre-emphasis at the nth sampling point; x1′(n) is the sound signal after pre-emphasis; and α is the pre-emphasis coefficient.

3. An adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 2, characterized in that, including: The preprocessing process further includes downsampling processing, and the sampling rate during downsampling is 44.1KHz.

4. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to any one of claims 1 to 3, characterized in that The binary mask M corresponding to the prior environment category v (f) is obtained by the following steps: Obtain a pure exhaust valve audio signal, perform preprocessing and normalization to obtain a normalized valve sound signal y1(n); perform normalization on different environmental noise data to obtain normalized different environmental noise data y2(n); Perform weighted superposition on the valve sound signal and different environmental noise signals respectively: y(n) = β·y1(n) + (1 - β)·y2(n) Among them, SNR is the signal-to-noise ratio, and k is the adjustment coefficient; The signal-to-noise ratio SNR is as follows: where N represents the total number of sampling points of the signal; Frame the noise signal y2(n) and perform windowing using a Hamming window: y k (n) = y2(n + kR)w(n), 0 ≤ n < L where y2(·) is the normalized noise signal, k is the frame index; R is the frame shift, L is the frame length, and y k (n) is the k-th frame signal after windowing, and w(n) is the window function; Based on Y k (f), calculate the magnitude spectrum |N k (f)| of the noise for each frame through short-time Fourier transform, and take the mean of the spectral magnitudes of multiple frames of noise to obtain the average magnitude spectrum where k represents the number of frames; and then obtain the typical energy distribution of the noise Frame the noisy signal y(n) and perform windowing: y m y(n) = y(n + mR)w(n), 0 ≤ n < L where m is the frame index, and y m (n) is the m-th frame signal after windowing; Based on y m (n), calculate the magnitude spectrum |Y m (f)| of the noisy signal for each frame through short-time Fourier transform; Generate an ideal binary mask for different environmental noises respectively: where δ is a coefficient used to adjust the strictness of noise suppression; Through the above process, the binary mask M(f) of different environmental noises is obtained, and the binary masks of different environmental noises are denoted as M v (f).

5. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, characterized in that The coefficient used to adjust the strictness of noise suppression is as follows: Among them, |N k (f) is the amplitude spectrum of the k-th frame; σ N (f) is the standard deviation.

6. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, wherein, When framing the noise signal y2(n), the frame length L is 20 - 40ms.

7. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, wherein The window function 8. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, characterized in that Calculating the amplitude spectrum |N(f)| of the noise for each frame through short-time Fourier transform includes: k (f)| First calculate the spectrum Among them, Y k (f) is the spectrum of the k-th frame, f is the frequency component, and e -j2πfn is the exponential term of the Fourier transform; Take the amplitude value of the noise for each frame, i.e., |N k (f)| = |Y k (f)|.

9. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, characterized in that, The process of calculating the magnitude spectrum |Y(f)| of the noisy signal for each frame through short-time Fourier transform includes: m (f)| First calculate the spectrum: Among them, Y m (f) is the complex frequency spectrum of the m-th frame, f is the frequency component, and e -j2πfn is the exponential term of the Fourier transform; Express it in the form of an amplitude spectrum and a phase spectrum: Among them, |Y m (f) is the amplitude spectrum of the m-th frame, and φ m (f) is the phase spectrum of the m-th frame; Furthermore, the amplitude spectrum |Y m (f)| is obtained.

10. The adaptive noise reduction method for an exhaust valve based on environmental noise recognition according to claim 4, wherein The deep learning model is pre-trained, and the training process of the deep learning model includes: Based on the valve sound signal y(n) obtained by weighted superposition of the valve sound signal and different environmental noise signals respectively, construct a training set, and then convert it into a Mel spectrogram; use the Mel spectrogram as the input of the deep learning model, and use the classification result of the environmental noise as the output of the deep learning model. The classification result of the environmental noise is different environmental noise categories. Based on the training set converted into a Mel spectrogram, train the deep learning model to obtain a trained deep learning model.

Images