A method and system for echo cancellation and noise reduction in steel plant workshop environments.

Abstract

This invention belongs to the field of industrial noise control and relates to a method and system for echo cancellation and noise reduction in steel plant workshop environments. The method includes the following steps: (1) real-time acquisition of sound signals from on-site monitoring points in the steel plant workshop environment during operation; the on-site monitoring points include various noise sources and various target sources; (2) time alignment of the original sound signal and the workshop echo signal from the same on-site monitoring point to obtain an aligned signal; (3) filtering the aligned signal obtained in step (2) to suppress superimposed noise and echo signals in the channel; (4) training a neural network mask, using the mask frequency domain to process the sound signal to suppress residual noise signals, and outputting a clean sound signal. The method provided by this invention effectively solves the limitations of traditional sound noise reduction technology in complex industrial environments.

Description

Technical Field

[0001] This invention belongs to the field of industrial noise control, and specifically relates to an echo cancellation and noise reduction method and system for environmental sound in steel plant workshops. Background Technology

[0002] In heavy industries such as steel manufacturing, environmental noise levels are typically very high. This noise not only affects the health and safety of workers but also severely interferes with the early detection of equipment malfunctions and the effective operation of sound monitoring systems. Traditional noise reduction techniques, such as active noise control and passive acoustic isolation, focus primarily on reducing the impact of noise on the human ear, and are insufficient for extracting and enhancing signals in complex industrial noise environments.

[0003] In heavy industrial environments such as steel mills, multiple noise sources exist, including mechanical operating noise, impact noise, and electrical noise. These noise sources are characterized by a wide frequency range, high intensity, and rapid changes. Furthermore, the sound signals in these environments often contain important operational information, such as equipment operating status and fault warnings. Therefore, effective technologies are needed to improve the recognizability and clarity of these signals. Simultaneously, workshop environments are prone to echoes and reverberation, making acoustic signal quality enhancement a pressing challenge.

[0004] Currently, most sound noise reduction techniques focus on simple linear processing methods, such as filtering and echo cancellation. These methods are effective when dealing with single or predictive noise sources, but their effectiveness is limited in multi-source and high-dynamic environments. For example, linear filters cannot adapt to rapid changes in ambient noise, and while traditional neural network methods can improve the ability to process complex signals, they still face challenges in computational efficiency and resource consumption in real-time applications.

[0005] Furthermore, existing technologies often fail to fully utilize the frequency and temporal correlations of sound signals, resulting in an inability to effectively extract useful signals from high-noise backgrounds. Therefore, developing a technology that can effectively handle noise in the complex acoustic environment of steel plants is of great significance for improving production safety and efficiency. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses an echo cancellation and noise reduction method and system for environmental sound in steel plant workshops; by integrating a delay compensation module, a linear filter module, and a residual noise suppression module based on a deep neural network, it effectively solves the limitations of traditional sound noise reduction technology in complex industrial environments.

[0007] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0008] A method for echo cancellation and noise reduction in a steel plant workshop environment, the method comprising the following steps:

[0009] (1) Real-time acquisition of sound signals from on-site monitoring points in the steel plant workshop environment during the working process; the on-site monitoring points include various noise sources and various target sources;

[0010] (2) Align the original acoustic signal and the workshop echo signal from the same field monitoring point in time to obtain the aligned signal;

[0011] (3) Filter the aligned signal obtained in step (2) to suppress noise and echo signals superimposed in the channel;

[0012] (4) Train the neural network mask and use the mask frequency domain processing step (3) to filter the sound signal to suppress residual noise signals and output a clean sound signal.

[0013] Furthermore, in step (2), a generalized cross-correlation and phase transformation algorithm is used to calculate the reverberation time delay of the original acoustic signal and the workshop echo signal generated by the workshop on site.

[0014] Furthermore, step (2) specifically includes:

[0015] (2.1) Define generalized cross-correlation, calculate the statistical correlation between the original sound signal and the echo signal of the steel plant workshop, and establish the correlation between the signals; at each time step, update the cross-correlation value between the original sound signal and the echo signal of the steel plant workshop.

[0016] (2.2) Convert the frequency domain signal back to the time domain signal by inverse fast Fourier transform; find the position of the maximum cross-correlation value in the time domain signal, calculate the delay time between the signals, and adjust the workshop echo signal so that the workshop echo signal is aligned with the corresponding original sound signal in time.

[0017] Furthermore, step (2.2) specifically includes:

[0018] Fast Fourier Transform is performed on the original acoustic signal x1(t) and the workshop echo signal x2(t) from the same on-site monitoring point to obtain X1(f) and X2(f);

[0019] Cross-power spectral density of two signals The calculation formula is:

[0020]

[0021] in, It is the conjugate of X2(f);

[0022] By applying phase transformation, a regularized cross-power spectral density is obtained.

[0023]

[0024] For regularized cross-power spectral density Perform an inverse fast Fourier transform to convert from the frequency domain to the time domain, and obtain the cross-correlation function.

[0025]

[0026] Find the position τ of the peak in the cross-correlation function. max Position τ max This refers to the estimated time delay between the original sound signal x1(t) and the workshop echo signal x2(t);

[0027] Based on the calculated delay τ max Adjust the time axis of signal x2(t) so that x1(t) and x2(t+τ) are equal. max Align the signals in time to obtain the aligned signal x(t).

[0028] Furthermore, in step (3), a weighted recursive least squares filter is used for filtering.

[0029] Furthermore, step (3) specifically includes:

[0030] The weight vector w(0) and the inverse P(0) of the correlation matrix are initialized with the aligned signal x(t) as input. P(0) is usually set to a large scalar multiplied by the identity matrix, such as δI, where δ is a large positive number. Specifically, δ can be set to 10^6 to enhance the adaptability of the algorithm, which is suitable for situations where the input signal changes dynamically and faster convergence is required.

[0031] Calculate the gain vector k(n):

[0032]

[0033] Here, the input vector for each time step n is x(n), which means the alignment signal of all field monitoring points of time step n and the previous L time steps; P(n-1) represents the covariance matrix of the previous time step; λ is the forgetting factor, which is between 0.9 and 1;

[0034] Update the weight vector w(n):

[0035] w(n)=w(n-1)+k(n)[d(n)-x T (n)w(n-1)]

[0036] Where w(n-1) is the iterative update value of W(0); w(n) is the weight vector of the adaptive filter, representing the parameters of the filter at the nth time step; d(n) is the desired output, i.e. the target signal under noise-free conditions.

[0037] Update the inverse P(n) of the correlation matrix:

[0038]

[0039] Applying filter w(n) to input x(n) yields the filtered output y(n):

[0040] y(n)=w T (n)x(n).

[0041] Further, in step (4), the method for training the neural network mask is as follows: convert the short-time Fourier transform (STFT) of the training set data into a time-spectrum image; and input the time-spectrum image into the FRCRN model for cIRM target prediction in order to train and obtain the mask.

[0042] Further, in step (4), the method of using a mask frequency domain to process the sound signal is as follows:

[0043] The filtered output from step (3) is converted into a time-domain spectrum by STFT. The time-domain spectrum is multiplied with the trained mask, and the result of the multiplication is subjected to inverse short-time Fourier transform. The enhanced spectrum is then converted back to the time-domain signal output to eliminate residual noise signals.

[0044] An echo cancellation and noise reduction system for environmental sound in a steel plant workshop, the echo cancellation and noise reduction system comprising:

[0045] Sound signal sampling module: acquires sound signals from on-site monitoring points in the steel plant workshop environment during operation in real time; the on-site monitoring points include various noise sources and various target sources;

[0046] Delay compensation module: Aligns the original sound signal and the workshop echo signal from the same on-site monitoring point in time to obtain the aligned signal;

[0047] Linear filter module: The linear filter filters the aligned signal output by the delay compensation module to suppress noise and echo signals superimposed in the channel;

[0048] The residual noise suppression module based on deep neural networks is used to train the neural network mask. The mask is used to process the audio signal output by the linear filter module in the frequency domain to suppress residual noise signals and output a clean audio signal.

[0049] Furthermore, the sound signal sampling module employs a network microphone, which converts the acquired sound signal into a digital signal.

[0050] Beneficial technical effects of the present invention:

[0051] This invention provides an echo noise reduction method for environmental sound in steel plant workshops, which eliminates workshop echoes while suppressing noise interference, thereby enhancing the acoustic quality of the steel plant workshop environment. It can automatically sample and process on-site audio online in real time using a microphone, and can achieve rapid detection through multi-threaded parallel computing to ensure real-time detection.

[0052] This invention provides an echo noise reduction system for environmental sounds in steel plant workshops. It overcomes audio distortion in existing solutions, allowing operators to hear pure, target-interest sounds while providing a clean dataset creation solution that can be used for subsequent machine learning to further improve automation levels.

[0053] In summary, the method provided by this invention effectively reduces noise levels within the work area while suppressing echoes and maintaining the clarity of important environmental sounds. Through large-scale training with field sound data, the network can adapt to various harsh acoustic conditions in the industrial environment of steel mill workshops. The implementation of this system facilitates efficient noise treatment in steel mills and improves the acoustic quality of the steel mill environment. Attached Figure Description

[0054] Figure 1 This is a structural diagram of the delay compensation module based on generalized cross-correlation in an embodiment of the present invention.

[0055] Figure 2 This is the correlation detection result in an embodiment of the present invention.

[0056] Figure 3 This is a schematic diagram of the frame division operation in an embodiment of the present invention.

[0057] Figure 4 This is a structural diagram of the residual noise suppression module based on a deep neural network in an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0060] Example 1: A method for echo cancellation and noise reduction in a steel plant workshop, the method comprising the following steps:

[0061] (1) Real-time acquisition of sound signals from on-site monitoring points in the steel plant workshop environment during the working process; the on-site monitoring points include various noise sources and various target sources; specifically, by arranging network microphones at various noise sources and various target sources in the steel plant workshop environment, the sound signals from the on-site monitoring points during the working process are transmitted to the network recording host in real time via the network, so as to realize storage, playback and analysis on the computer.

[0062] (2) Align the original acoustic signal and the workshop echo signal from the same field monitoring point in time to obtain the aligned signal;

[0063] (3) Filter the aligned signal obtained in step (2) to suppress noise and echo signals superimposed in the channel;

[0064] (4) Train the neural network mask and use the mask frequency domain processing step (3) to filter the sound signal to suppress residual noise signals and output a clean sound signal.

[0065] In step (2) of this embodiment, a generalized cross-correlation and phase transformation algorithm is used to calculate the reverberation time delay of the original acoustic signal and the workshop echo signal generated by the workshop.

[0066] In this embodiment, step (2) specifically includes:

[0067] Step (2) specifically includes:

[0068] (2.1) Using the GCC-PHAT algorithm, a generalized cross-correlation is defined, and the statistical correlation between the original acoustic signal and the echo signal in the steel plant workshop is calculated to establish the association between the signals; such as Figure 2The results of the correlation detection are shown, with each peak corresponding to a compensation point for the echo. At each time step, the cross-correlation value between the original sound signal from the steel mill workshop and the echo signal is updated. To achieve real-time calculation when establishing the correlation between signals, a smoothing parameter is used to balance the signal data from the current and past times, except during the alignment process of the steel mill workshop echo and the target sound signal. A weighted averaging method is used to smooth the cross-correlation value, thus adapting to dynamic changes in the signal. The GCC-PHAT algorithm used in this invention is a conventional time delay compensation algorithm; this invention simply uses a weighted averaging method to meet real-time requirements.

[0069] (2.2) Convert the frequency domain signal back to the time domain signal by inverse fast Fourier transform; find the position of the maximum cross-correlation value in the time domain signal, calculate the delay time between the signals, and adjust the workshop echo signal according to the delay time so that the workshop echo signal is aligned with the corresponding original sound signal in time.

[0070] Step (2.2) specifically includes: performing a fast Fourier transform on the original acoustic signal x1(t) and the workshop echo signal x2(t) from the same on-site monitoring point to obtain X1(f) and X2(f);

[0071] Cross-power spectral density of two signals The calculation formula is:

[0072]

[0073] in, It is the conjugate of X2(f);

[0074] By applying phase transformation, a regularized cross-power spectral density is obtained.

[0075]

[0076] For regularized cross-power spectral density Perform an inverse fast Fourier transform to convert from the frequency domain to the time domain, and obtain the cross-correlation function.

[0077]

[0078] Find the position τ of the peak in the cross-correlation function. max Position τ max This refers to the estimated time delay between the original sound signal x1(t) and the workshop echo signal x2(t);

[0079] Based on the calculated delay τ max Adjust the time axis of signal x2(t) so that x1(t) and x2(t+τ) are equal.max Align the signals in time to obtain the aligned signal x(t).

[0080] The original acoustic signals and workshop echo signals from all field monitoring points are aligned in time to obtain all aligned signals corresponding to all time steps.

[0081] The aligned signal obtained after this step can be more accurately processed for subsequent noise suppression, ensuring that all signals reach optimal synchronization before entering the linear filter processing step.

[0082] Step (3) in this embodiment specifically includes:

[0083] Taking the aligned signal x(t) as input, the input signal is first divided into frames and windowed; for example... Figure 3 As shown, the purpose of framing and windowing of the signal is to divide the continuous audio signal into several 1-second frames for local signal analysis and processing. Frames overlap partially by setting frame shifts, ensuring continuity between frames, reducing boundary effects caused by framing, and enhancing the capture of signal details. Applying a window function (in this example, a Hamming window) to each frame reduces spectral leakage and optimizes frequency domain analysis.

[0084] Initialize the filter's weight vector w(0) and the inverse P(0) of the correlation matrix. Typically, P(0) is set to a large scalar multiplied by the identity matrix, such as δI, where δ is a large positive number. Specifically, δ can be set to 10^6 to enhance the algorithm's adaptability, making it suitable for situations where the input signal has large dynamic changes and faster convergence is required.

[0085] Calculate the gain vector k(n):

[0086]

[0087] Here, the input vector for each time step n is x(n), where x(n) represents the aligned signal of all field monitoring points at time step n and the previous L time steps, and x(n) is the sampled discrete signal; P(n-1) represents the covariance matrix of the previous time step; λ is the forgetting factor, which is between 0.9 and 1, and is used to adjust the algorithm's memory of old data; x T (n) is the transpose of the input vector.

[0088] Update the weight vector w(n):

[0089] w(n)=w(n-1)+k(n)[d(n)-x T (n)w(n-1)]

[0090] Where w(n-1) is the iterative update value of W(0); w(n) is the weight vector of the adaptive filter, representing the parameters of the filter at the nth time step; d(n) is the desired output, i.e. the target signal under noise-free conditions.

[0091] Update the inverse P(n) of the correlation matrix:

[0092]

[0093] Among them, by recursively updating P(n), the weight vector can be adjusted according to the new input signal, so that the filter can better adapt to the data; P(n-1) is the inverse of the correlation matrix of the previous time step, which contains the statistical information of the input signal of the previous time step; P(n) is an update of P(n-1), which uses information of discarding the historical data part to update.

[0094] Applying filter w(n) to input x(n) yields the filtered output y(n):

[0095] y(n)=w T (n)x(n).

[0096] This step involves calculating the correlation matrix and gain vector of the input echo reverberation signal at each time step; updating the filter coefficients by minimizing the error using a recursive least squares algorithm; and iteratively updating the filter coefficients at each time step based on the current correlation matrix and correlation vector. The updated filter coefficients are then applied to filter the time-aligned near-end signal, extracting the noise-suppressed target signal. The delayed far-end signal is weighted using the filter coefficients to compensate for the influence of noise paths, and the filtered near-end target signal is normalized and enhanced. Through these steps, the wRLS linear filter effectively suppresses noise in the steel mill environment, providing a clear sound signal and clean input for subsequent deep learning processing.

[0097] Furthermore, during the filter update process for the linear filter targeting the reverberation signal in the workshop, a semi-blind source separation method is employed to improve the filter's adaptability to different noise environments. Different components in the signal are separated through independent component analysis.

[0098] In step (4) of this embodiment, the method for training the neural network mask is as follows: convert the short-time Fourier transform (STFT) of the training set data into a time-spectrum image; and input the time-spectrum image into the FRCRN model for cIRM target prediction to train and obtain the mask.

[0099] In step (4) of this embodiment, the method of using a mask frequency domain to process the sound signal is as follows:

[0100] The filtered output from step (3) is converted into a time-domain spectrum by STFT. The time-domain spectrum is multiplied with the trained mask, and the result of the multiplication is subjected to inverse short-time Fourier transform. The enhanced spectrum is then converted back to the time-domain signal output to eliminate residual noise signals.

[0101] Example 2: An echo cancellation and noise reduction system for environmental sound in a steel plant workshop, the echo cancellation and noise reduction system comprising:

[0102] Sound signal sampling module: used to acquire sound signals from on-site monitoring points in the steel plant workshop environment during operation in real time;

[0103] Delay compensation module: Used in multi-source sound environments in workshops to align echo signals from different sound sources in time; such as... Figure 2 As shown, the delay compensation module utilizes the Generalized Cross-Correlation and Phase Transform (GCC-PHAT) algorithm to process multi-source signals, estimating and compensating for the time delay between sound sources. This module ensures that signals from different sound sources are time-aligned, laying the foundation for efficient noise suppression. The module performs a Fast Fourier Transform on each signal and calculates the cross-power spectral density of the two signals. A phase transform is applied to regularize the cross-power spectral density. An Inverse Fast Fourier Transform is performed to transform from the frequency domain to the time domain, obtaining the cross-correlation function. The peak position τ is found in the cross-correlation function. max This position represents the estimated delay between the two signals. Based on the calculated delay τ... max The time axes of the two signals are adjusted to align them in time. Aligned signals can then be processed more accurately for subsequent noise suppression. This ensures that all signals reach optimal synchronization before entering the linear filter processing step.

[0104] The linear filter module processes the aligned signal output from the delay compensation module using a linear filter to suppress superimposed noise and echo signals in the channel. The linear filter module employs weighted recursive least squares filtering technology to adjust the filter coefficients in real time to suppress superimposed noise and echo signals. This module is particularly suitable for dynamically changing industrial noise environments, automatically optimizing filter performance based on real-time changes in environmental noise. The linear filter module windows the signal at each step, using the signal aligned by the delay compensation module as input. The process involves initializing the filter weights and the inverse of the correlation matrix for both the target signal and noise signal. This is typically set as a large scalar multiplied by the identity matrix, such as δI, where δ is a large positive number. For each time step n, considering the input vector x(n), which includes samples from the current and previous L time steps, the gain vector is calculated, the weight vector is updated, the target signal under noise-free conditions is obtained, and the inverse of the correlation matrix is ​​updated. The filter is then applied to the input to obtain the filtered output. Through these steps, the linear filter module effectively suppresses noise in the steel plant environment, providing a clear sound signal and a clean input for subsequent deep learning processing.

[0105] A residual noise suppression module based on a deep neural network is used to train a neural network mask. This mask processes the sound signal in the frequency domain to suppress residual noise and output a clean sound signal. The input signal to this module is the sound signal output from a linear filter module. Deep learning training employs a Frequency Recurrent Convolutional Recurrent Network (FRCRN) model, which performs complex ideal ratio mask prediction in the time-frequency domain. The mask is multiplied by the spectrogram of the workshop environmental noise, and the enhanced spectrogram is converted back to a time-domain signal using an inverse short-time Fourier transform. This module further suppresses residual noise using deep learning techniques. The FRCRN model training uses a time-frequency domain complex ideal ratio mask (cIRM) to effectively extract the target sound signal and convert it back to a time-domain signal using an inverse short-time Fourier transform. Figure 4 As shown, this module employs a residual noise suppression method based on deep neural networks. First, it converts the historical audio signal into a time-frequency graph and decomposes it into the real and imaginary parts of the signal. Then, these signals are processed by a FRCRN neural network model to generate an ideal ratio mask. Finally, by combining the time-frequency graph of the real-time audio signal, a time-frequency graph of the clean audio signal is output, thereby achieving effective noise suppression and clear extraction of the target sound.

[0106] In this embodiment, the sound signal sampling module employs a network microphone. By placing a network microphone near the sound source, it captures the sound signal of the cutting process and transmits it in real time to a network recording host. The sound signal sampling module converts the collected sound signal into a digital signal, enabling storage, playback, and analysis on a computer. The sound signal sampling module collects sound signals from various noise sources in real time, including mechanical operating sounds, impact sounds, and sound signals from target sources such as fertilizer falling during the cutting process.

[0107] This invention provides an echo noise reduction system for steel plant workshop environments, designed to effectively reduce noise levels in the work area while suppressing echoes to maintain the clarity of important environmental sounds. The system integrates four modules: a sound signal sampling module, a time delay compensation module, a linear filtering module, and a residual noise suppression module based on a deep neural network. The sound signal sampling module acquires on-site sound signals in real time using a microphone. The time delay compensation module performs time alignment on sounds and their echoes from multiple sound sources in the steel plant, ensuring that all sound signals are synchronized before entering subsequent processing modules. The linear filter module adjusts its filtering parameters to adapt to the unique sound characteristics of steel plants, reducing noise interference, enhancing the sounds of concern in the workshop environment, and improving the overall quality of the sound signals. The residual noise suppression module is specifically designed to further suppress residual noise remaining after filtering by the linear filter module. Through large-scale training with on-site sound data, the network can adapt to various harsh acoustic conditions in the industrial environment of steel plant workshops. The system provided by this invention achieves efficient processing of steel plant noise and improves the acoustic quality of the steel plant environment.

[0108] In one specific embodiment, a window length of 20ms and a frame shift of 10ms were used for wideband voice input. The total delay of the wideband FRCRN model was limited to within 30ms; to improve spectral resolution, the STFT length was set to 640, and zero-padding was employed. The model had one input channel. The output obtained from the linear filter was converted into a time-frequency graph by the STFT and multiplied with a mask; the inverse short-time Fourier transform of the result converted the enhanced spectrum back into a time-domain signal output to eliminate residual noise signals.

[0109] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, modifications or variations can still be made to the technical solutions described above, and these modifications and variations all fall within the protection scope of the present invention.

Claims

1. A method for echo cancellation and noise reduction in steel plant workshop environments, characterized in that, The method includes the following steps: (1) Real-time acquisition of sound signals from on-site monitoring points in the steel plant workshop environment during the working process; the on-site monitoring points include various noise sources and various target sources; (2) Align the original acoustic signal and the workshop echo signal from the same field monitoring point in time to obtain the aligned signal; (3) Filter the aligned signal obtained in step (2) to suppress noise and echo signals superimposed in the channel; (4) Train the neural network mask, use the mask frequency domain processing step (3) to filter the sound signal to suppress residual noise signal and output a clean sound signal; Step (2) specifically includes: (2.1) Define generalized cross-correlation, calculate the statistical correlation between the original sound signal and the echo signal of the steel plant workshop, and establish the correlation between the signals; at each time step, update the cross-correlation value between the original sound signal and the echo signal of the steel plant workshop. (2.2) Convert the frequency domain signal back to the time domain signal by inverse fast Fourier transform; find the position of the maximum cross-correlation value in the time domain signal, calculate the delay time between the signals, and adjust the workshop echo signal so that the workshop echo signal is aligned with the corresponding original sound signal in time. Step (3) specifically includes: The weight vector w(0) of the filter and the inverse P(0) of the correlation matrix are initialized with the aligned signal x(t) as input. Calculate the gain vector k(n): Here, the input vector for each time step n is x(n), which means the alignment signal of all field monitoring points of time step n and the previous L time steps; P(n-1) represents the covariance matrix of the previous time step; λ is the forgetting factor, which is between 0.9 and 1; Update the weight vector w(n): w(n)=w(n-1)+k(n)[d(n)-x T (n)w(n-1)] Where w(n-1) is the iterative update value of W(0); w(n) is the weight vector of the adaptive filter, representing the parameters of the filter at the nth time step; d(n) is the desired output, i.e. the target signal under noise-free conditions. Update the inverse P(n) of the correlation matrix: Applying filter w(n) to input x(n) yields the filtered output y(n): y(n)=w T (n)x(n)。 2. The method for echo cancellation and noise reduction in steel plant workshop environment according to claim 1, characterized in that, In step (2), a generalized cross-correlation and phase transformation algorithm is used to calculate the reverberation time delay of the original acoustic signal and the workshop echo signal generated by the workshop.

3. The method for echo cancellation and noise reduction in a steel plant workshop environment according to claim 1, characterized in that, Step (2.2) specifically includes: Fast Fourier Transform is performed on the original acoustic signal x1(t) and the workshop echo signal x2(t) from the same on-site monitoring point to obtain X1(f) and X2(f); Cross-power spectral density of two signals The calculation formula is: in, It is the conjugate of X2(f); By applying phase transformation, a regularized cross-power spectral density is obtained. For regularized cross-power spectral density Perform an inverse fast Fourier transform to convert from the frequency domain to the time domain, and obtain the cross-correlation function. Find the position τ of the peak in the cross-correlation function. max Position τ max This refers to the estimated time delay between the original sound signal x1(t) and the workshop echo signal x2(t); Based on the calculated delay τ max Adjust the time axis of signal x2(t) so that x1(t) and x2(t+τ) are equal. max Align the signals in time to obtain the aligned signal x(t).

4. The method for echo cancellation and noise reduction in a steel plant workshop environment according to claim 3, characterized in that, In step (3), a weighted recursive least squares filter is used for filtering.

5. The method for echo cancellation and noise reduction in a steel plant workshop environment according to claim 1, characterized in that, In step (4), the method for training the neural network mask is as follows: convert the short-time Fourier transform of the training set data into a time-spectrum image; and input the time-spectrum image into the FRCRN model for cIRM target prediction in order to train and obtain the mask.

6. The method for echo cancellation and noise reduction in a steel plant workshop environment according to claim 1, characterized in that, In step (4), the method of using a mask frequency domain to process the sound signal is as follows: The filtered output from step (3) is converted into a time-domain spectrum by STFT. The time-domain spectrum is multiplied with the trained mask, and the result of the multiplication is subjected to inverse short-time Fourier transform. The enhanced spectrum is then converted back to the time-domain signal output to eliminate residual noise signals.

7. An echo cancellation and noise reduction system for environmental sound in a steel plant workshop, employing the method described in any one of claims 1-6, characterized in that, The echo cancellation and noise reduction system includes: Sound signal sampling module: acquires sound signals from on-site monitoring points in the steel plant workshop environment during operation in real time; the on-site monitoring points include various noise sources and various target sources; Delay compensation module: Aligns the original sound signal and the workshop echo signal from the same on-site monitoring point in time to obtain the aligned signal; Linear filter module: The linear filter filters the aligned signal output by the delay compensation module to suppress noise and echo signals superimposed in the channel; The residual noise suppression module based on deep neural networks is used to train the neural network mask. The mask is used to process the audio signal output by the linear filter module in the frequency domain to suppress residual noise signals and output a clean audio signal.

8. The echo cancellation and noise reduction system for steel plant workshop environment sound according to claim 7, characterized in that, The sound signal sampling module uses a network microphone and converts the collected sound signal into a digital signal.

Images