Aluminum liquid leakage detection method and system based on deep learning

Abstract

The present invention proposes a method and system for detecting aluminum liquid leakage based on deep learning, the method comprising: collecting multimodal data and preprocessing the multimodal data; using sound data to build an adaptive noise suppression model; fusing multimodal data to extract key features of aluminum liquid leakage; using transfer learning technology to pretrain the adaptive noise suppression model and fine-tune it in combination with labeled data; monitoring the data in real time, and triggering an alarm if there is an abnormal aluminum liquid leakage. The present invention proposes a method and system for detecting aluminum liquid leakage based on deep learning, the system integrates multiple innovative points, aims to overcome the limitations of traditional monitoring methods, and achieve high sensitivity, high accuracy and timely detection of aluminum liquid leakage.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent early warning technology, and in particular relates to a method and system for detecting aluminum liquid leakage based on deep learning. Background Art

[0002] In modern industrial production, the handling and transportation of molten aluminum is a key link, especially in high-temperature smelting and casting processes, where the safe management of molten aluminum is crucial. Leakage of molten aluminum not only leads to significant waste of resources, but also may cause fires and worker injuries, which has a significant impact on production safety and corporate economic benefits. Traditional methods of monitoring molten aluminum leakage mainly rely on visual inspection and regular maintenance, which often cannot respond to leakage incidents in real time. In addition, in complex industrial environments, workers may find it difficult to find leakage points in time, leading to leakage accidents.

[0003] With the advancement of technology, some automated monitoring systems have begun to be used in the detection of aluminum liquid leakage. These systems mostly use physical sensors such as temperature sensors and pressure sensors for monitoring. However, these methods based on physical sensors have certain limitations. First, they have high requirements for installation location and environmental conditions, and the accuracy of the sensors will be affected by changes in the surrounding environment, such as temperature fluctuations or mechanical vibrations, which may cause false alarms. Secondly, physical sensors often cannot provide a fast enough response speed to detect the early stage of leakage in time, especially when the amount of aluminum liquid leakage is small, the detection sensitivity of traditional sensors may not be enough to trigger an alarm.

[0004] In recent years, the development of sound recognition technology has provided a new solution for aluminum liquid leakage detection. By analyzing the specific sound characteristics generated when aluminum liquid comes into contact with the cooling medium, real-time monitoring of leakage events can be achieved. However, the existing monitoring systems based on sound recognition still face many challenges in practical applications. On the one hand, the background noise interference in the industrial environment is large, which can easily affect the accuracy of sound recognition; on the other hand, the generalization ability of the existing system is limited. For different production environments and equipment configurations, the system's adaptability and flexibility are insufficient, and a large number of customized adjustments are required to meet the actual application needs. Therefore, how to improve the accuracy and adaptability of the sound recognition system in a noisy environment has become a problem that needs to be solved in the current technological development. Summary of the invention

[0005] The purpose of the present invention is to design a method and system for detecting aluminum liquid leakage based on deep learning, which integrates the innovations of deep learning and multimodal data fusion, aiming to overcome the limitations of traditional monitoring methods and achieve high-sensitivity, high-accuracy and timely detection of aluminum liquid leakage.

[0006] In order to achieve the above object, in a first aspect of the present invention, a method for detecting aluminum liquid leakage based on deep learning is provided, and the method comprises:

[0007] S1. Install a microphone under the coolant in the aluminum well for sound collection, and arrange temperature and pressure sensors to collect temperature and pressure, combine the sound, temperature and pressure into multimodal data, and pre-process the multimodal data;

[0008] S2, constructing an adaptive noise suppression model using sound data;

[0009] S3, perform data fusion on multimodal data to extract key features of aluminum liquid leakage;

[0010] S4. Pre-train the adaptive noise suppression model using transfer learning technology and fine-tune it with the labeled data; the loss function of the training is

[0011]

[0012] Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i ,n i ) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss;

[0013] S5. Monitor the data in real time and trigger an alarm if there is any abnormal leakage of aluminum liquid.

[0014] Furthermore, the collected sound is first pre-emphasized and windowed, and then the spectrum information is obtained by fast Fourier transform and Mel frequency cepstrum coefficients are extracted to capture the characteristics of the aluminum liquid leakage sound. The formula is as follows:

[0015] Assume s(t) is the sound signal, and the feature MFCC extraction process of the sound is expressed as:

[0016]

[0017] Among them, DCT represents discrete cosine transform, FilterBank represents Mel filter bank, t represents time threshold, and MFCC represents the characteristics of sound.

[0018] Furthermore, the preprocessing includes normalization processing, which is expressed as follows:

[0019]

[0020] Where X represents any sensor data, μ and σ represent the mean and standard deviation of X, respectively.

[0021] Furthermore, the S2 specifically includes:

[0022] S21, calculate the power spectrum density PSD(f) of the current environmental background noise in real time, expressed as follows:

[0023]

[0024] Where x(t) represents the sound signal in the time domain t, FFT represents fast Fourier transform, N represents the number of samples, and f represents the frequency;

[0025] S22. Calculate the dynamic filter coefficient according to the power spectrum density PSD(f) to adapt to the current noise environment, as shown below:

[0026]

[0027] Wherein, α(f) represents the filter coefficient at frequency f, and θ represents the threshold value preset according to the ambient noise level;

[0028] S23, using dynamic filter coefficients to perform denoising on the sound signal, as shown below:

[0029] S clean (f) = S raw (f)·α(f)

[0030] Among them, S raw (f) represents the frequency domain representation of the original sound signal, S clean (f) represents the signal after denoising;

[0031] S24, convert the denoised signal back to the time domain through inverse Fourier transform, which is expressed as follows:

[0032] s clean (t) = IFFT(S clean (f)).

[0033] Furthermore, in S3, the fusion feature vector F of the multimodal data fusion is c It is expressed as follows:

[0034] F c =DNN(X′1,X′2,...,X′ n )

[0035] Among them, X′ i is the standardized data of the i-th sensor, DNN represents a deep neural network; wherein the deep neural network includes a convolutional neural network and a recurrent neural network.

[0036] Furthermore, the autoencoder AE is used to process the deep fusion feature vector F c , extract the key feature F that is most conducive to aluminum liquid leakage detection opt , which is expressed as follows:

[0037] F opt =AE(F c )

[0038] Further, through the key feature F opt Input into the classifier to judge the leakage of aluminum liquid, as shown below:

[0039] Leak_Detection=C(F opt )

[0040] Among them, Leak_Detection represents the leak detection result, and C represents the classifier.

[0041] Furthermore, in S4, the loss function is optimized using a stochastic gradient descent method with an adaptive learning rate, which is expressed as:

[0042]

[0043] Among them, θ t represents the model parameters of the tth iteration; α t Represents the adaptive learning rate, which is dynamically adjusted according to the training progress; represents the gradient of the loss function L with respect to the parameter θ.

[0044] Furthermore, the S5 specifically includes: inputting the fused features into the trained model prediction score, judging whether the score is higher than a set threshold, and thereby judging whether there is aluminum liquid leakage.

[0045] In a second aspect of the present invention, a deep learning-based aluminum liquid leakage detection system is provided, the system comprising:

[0046] The data acquisition unit is used to install a pickup under the coolant of the aluminum well for sound collection, and arrange temperature and pressure sensors to collect temperature and pressure, combine the sound, temperature and pressure into multi-modal data, and pre-process the multi-modal data;

[0047] A model building unit, used for building an adaptive noise suppression model using sound data;

[0048] Data fusion unit, used to fuse multi-modal data and extract key features of aluminum liquid leakage;

[0049] The model training unit is used to train the adaptive noise suppression model using transfer learning technology and fine-tune it in combination with the labeled data; the training loss function is

[0050]

[0051] Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i , n i ) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss;

[0052] The real-time early warning unit is used to monitor the data in real time and trigger an alarm if there is an abnormal leakage of molten aluminum.

[0053] The beneficial technical effects of the present invention are at least as follows:

[0054] (1) The present invention significantly improves the recognition accuracy of aluminum liquid leakage events by fusing sound data with multiple sensor data such as temperature and pressure, and using deep learning technology for data fusion and feature extraction. This multimodal fusion method utilizes the complementarity of various sensor data, enhances the robustness of the system in complex environments, and effectively avoids false alarms caused by environmental changes in a single sensor.

[0055] (2) Taking into account the interference of background noise in industrial environments, the present invention adopts an adaptive noise suppression model based on deep learning, which can dynamically adjust the noise suppression strategy according to the real-time environmental noise level. This technology significantly improves the accuracy of the sound recognition module in noisy environments and ensures the effective detection of aluminum liquid leakage sound.

[0056] (3) In order to improve the adaptability and flexibility of the system in different production environments, the present invention adopts transfer learning and rapid adaptation technology, so that the system can use pre-trained models and quickly adjust to new monitoring environments through a small amount of localized data. This method greatly reduces the time and cost of system deployment and improves the practicality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0058] Figure 1 This is a flow chart of an aluminum liquid leakage detection method based on deep learning in the present invention.

[0059] Figure 2 This is a framework diagram of an aluminum liquid leakage detection system based on deep learning in the present invention. DETAILED DESCRIPTION

[0060] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0061] In one or more embodiments, Figure 1 As shown, a method for detecting aluminum liquid leakage based on deep learning is disclosed in the present invention, and the method comprises:

[0062] S1. Install a microphone under the coolant in the aluminum well for sound collection, and arrange temperature and pressure sensors to collect temperature and pressure. Combine the sound, temperature and pressure into multimodal data, and pre-process the multimodal data.

[0063] Traditional monitoring systems often rely on only a single type of sensor data, which is easily disturbed by environmental factors, resulting in false alarms or missed alarms. Therefore, in this step, multiple sensors (sound, temperature, pressure, etc.) are deployed for data collection, and advanced signal processing technology is used to pre-process the collected data, including filtering, denoising, normalization and other steps, to provide an accurate and clean data source for subsequent data analysis.

[0064] Specifically, the microphone is fixed below the coolant in the aluminum well, ensuring a distance of 5 to 20 cm from the liquid surface to optimize the sound capture effect. The selection and installation position of the microphone is designed to maximize the capture of the specific "hissing sound" produced by aluminum liquid dripping into the coolant, while reducing equipment loss and noise interference caused by direct contact.

[0065] In addition to sound collection, the system also integrates temperature and pressure sensors and deploys them at key locations to collect environmental data. These data will be used to assist in sound data analysis and improve the accuracy and reliability of leak detection.

[0066] The collected sound is first pre-emphasized and windowed, and then the spectrum information is obtained by fast Fourier transform and Mel frequency cepstrum coefficients are extracted to capture the characteristics of the aluminum liquid leakage sound. The formula is as follows:

[0067] Assume s(t) is the sound signal, and the feature MFCC extraction process of the sound is expressed as:

[0068]

[0069] Among them, DCT represents discrete cosine transform, FilterBank represents Mel filter bank, t represents time threshold, and MFCC represents the characteristics of sound.

[0070] Among them, for the data collected from different sensors, timestamps are used for synchronization to ensure data symmetry. At the same time, all sensor data are normalized to unify the data scale for subsequent analysis. Therefore, the preprocessing includes normalization processing, which is expressed as follows:

[0071]

[0072] Where X represents any sensor data, μ and σ represent the mean and standard deviation of X, respectively.

[0073] S2. Using sound data to build an adaptive noise suppression model, specifically including:

[0074] S21, calculate the power spectrum density PSD(f) of the current environmental background noise in real time, expressed as follows:

[0075]

[0076] Where x(t) represents the sound signal in the time domain t, FFT represents fast Fourier transform, N represents the number of samples, and f represents the frequency;

[0077] S22. Calculate the dynamic filter coefficient according to the power spectrum density PSD(f) to adapt to the current noise environment, as shown below:

[0078]

[0079] Wherein, α(f) represents the filter coefficient at frequency f, and θ represents the threshold value preset according to the ambient noise level;

[0080] S23, using dynamic filter coefficients to perform denoising on the sound signal, as shown below:

[0081] S clean (f) = S raw (f)·α(f)

[0082] Among them, S raw (f) represents the frequency domain representation of the original sound signal, S clean (f) represents the signal after denoising;

[0083] S24, convert the denoised signal back to the time domain through inverse Fourier transform, which is expressed as follows:

[0084] s clean (t) = IFFT(S clean (f)).

[0085] S3. Perform data fusion on multimodal data to extract key features of aluminum liquid leakage.

[0086] Specifically, a deep neural network (DNN) is constructed to achieve deep fusion of different sensor data. This network takes the standardized form of each sensor data as input, learns the intrinsic correlation and complementarity between different data sources through a series of hidden layers, and thus generates a comprehensive feature representation F c , which can fully reflect the characteristics of aluminum liquid leakage. The fusion feature vector F of the multi-modal data fusion c It is expressed as follows:

[0087] F c =DNN(X′1,X′2,...,X′ n )

[0088] Among them, X′ i is the standardized data of the i-th sensor, DNN represents a deep neural network; wherein the deep neural network includes a convolutional neural network and a recurrent neural network.

[0089] Furthermore, the autoencoder (AE) is used to further process the deep fused feature vector F c , in order to extract the key features F that are most conducive to aluminum liquid leakage detection opt This step aims to reduce the feature dimension, remove noise, and optimize the input of the detection model, as shown below:

[0090] F opt =AE(F c )

[0091] Among them, AE structure: According to F c According to the dimensions and characteristics of the image, the number of layers and nodes of the encoder and decoder are designed to achieve efficient feature compression and reconstruction.

[0092] Furthermore, the optimized feature F opt The data is input into the classifier to make the final judgment on aluminum liquid leakage. The classifier here can be a traditional machine learning model, such as a support vector machine (SVM), or a specially designed deep learning classification network, and then the judgment on aluminum liquid leakage is made, as shown below:

[0093] Leak_Detection=C(F opt )

[0094] Among them, Leak_Detection represents the leak detection result, and C represents the classifier.

[0095] The design of the classifier is selected or designed according to the actual application requirements. For complex aluminum liquid leakage sound patterns, deep learning methods such as convolutional neural network (CNN) or recurrent neural network (RNN) can be used to improve classification performance.

[0096] S4. Use transfer learning technology to pre-train the adaptive noise suppression model and fine-tune it based on the labeled data.

[0097] Among them, the loss function of training is

[0098]

[0099] Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i , n i ) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss.

[0100] Furthermore, the loss function is optimized using a stochastic gradient descent method with an adaptive learning rate, which is expressed as:

[0101]

[0102] Among them, θ t represents the model parameters of the tth iteration; α t Represents the adaptive learning rate, which is dynamically adjusted according to the training progress; represents the gradient of the loss function L with respect to the parameter θ.

[0103] Furthermore, for a small amount of labeled data for a specific environment, the last few layers of the pre-trained model are fine-tuned to adapt to the characteristics of the new environment.

[0104] S5. Monitor the data in real time. If there is an abnormal leakage of molten aluminum, trigger an alarm. Specifically, input the fused features into the trained model prediction score, determine whether the score is higher than the set threshold, and thus determine whether there is a leakage of molten aluminum.

[0105] Specifically, the system receives data streams from microphones and other sensors in real time, including sound signals, temperature, pressure, and other data. Feature extraction is performed on the received data in real time, using the methods in the adaptive multi-dimensional signal fusion framework (ALLF) defined earlier, such as MFCC extraction, temperature change monitoring, etc. Combining the features of all sensors, feature fusion is performed using a deep learning model to extract more meaningful information for subsequent leak identification.

[0106] The fused features are input into the trained model (e.g., a CNN-based model) in real time for leak identification. The model will output a prediction value indicating whether there is a possibility of aluminum liquid leakage at the current monitoring point. The prediction value of the model is evaluated by setting a predefined threshold, such as 0.9. If the prediction value exceeds this threshold, it is considered that aluminum liquid leakage is detected.

[0107] Once the system determines that aluminum liquid has leaked at a certain monitoring point, the alarm mechanism will be activated immediately. This includes but is not limited to sound alarms, sending alarm information to the control center or the mobile devices of relevant personnel, etc. All alarm events and related data will be recorded in the system log, including alarm time, location, predicted leakage probability and impact range, etc., for subsequent analysis and improvement.

[0108] User feedback and model updates are also set up to allow operators to verify alarm events and provide feedback. This feedback data is used to further train and optimize the model to improve the accuracy and reliability of the system. At the same time, based on the collected feedback, the monitoring model is updated regularly or as needed. This may include retraining the model or using transfer learning technology to quickly adapt to new monitoring environments.

[0109] In one or more embodiments, Figure 2 As shown, the present invention discloses a deep learning-based aluminum liquid leakage detection system, the system comprising:

[0110] The data acquisition unit 101 is used to install a pickup below the aluminum well coolant for sound collection, and to arrange temperature and pressure sensors to collect temperature and pressure, to form multimodal data from the sound, temperature and pressure, and to pre-process the multimodal data;

[0111] A model building unit 102, used to build an adaptive noise suppression model using sound data;

[0112] A data fusion unit 103 is used to fuse the multi-modal data and extract key features of aluminum liquid leakage;

[0113] The model training unit 104 is used to train the adaptive noise suppression model using transfer learning technology and fine-tune it in combination with the labeled data; wherein the loss function of the training is

[0114]

[0115] Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i σn i) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss;

[0116] The real-time warning unit 105 is used to monitor the data in real time and trigger an alarm if there is an abnormal leakage of aluminum liquid.

[0117] The above disclosures are only some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. A method for detecting aluminum liquid leakage based on deep learning, characterized in that: The method comprises: S1. Install a microphone under the coolant in the aluminum well for sound collection, and arrange temperature and pressure sensors to collect temperature and pressure, combine the sound, temperature and pressure into multimodal data, and pre-process the multimodal data; S2, constructing an adaptive noise suppression model using sound data; S3, perform data fusion on multimodal data to extract key features of aluminum liquid leakage; S4. Use transfer learning technology to pre-train the adaptive noise suppression model and fine-tune it with the labeled data; the loss function of the training is Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i ,n i ) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss; S5. Monitor the data in real time and trigger an alarm if there is any abnormal leakage of aluminum liquid; The S2 specifically includes: S21, calculate the power spectrum density PSD(f) of the current environmental background noise in real time, expressed as follows: Where x(t) represents the sound signal in the time domain t, FFT represents fast Fourier transform, N represents the number of samples, and f represents the frequency; S22. Calculate the dynamic filter coefficient according to the power spectrum density PSD(f) to adapt to the current noise environment, as shown below: Wherein, α(f) represents the filter coefficient at frequency f, and θ represents the threshold value preset according to the ambient noise level; S23, using dynamic filter coefficients to perform denoising on the sound signal, as shown below: S clean (f)=S raw (f)·α(f) Among them, S raw (f) represents the frequency domain representation of the original sound signal, S clean (f) represents the signal after denoising; S24, convert the denoised signal back to the time domain through inverse Fourier transform, which is expressed as follows: s clean (t)=IFFT(S clean (f))。 2. The aluminum liquid leakage detection method based on deep learning according to claim 1 is characterized in that: The collected sound is first pre-emphasized and windowed, and then the spectrum information is obtained by fast Fourier transform and Mel frequency cepstrum coefficients are extracted to capture the characteristics of the sound of aluminum liquid leakage. The formula is as follows: Assume s(t) is the sound signal, and the feature MFCC extraction process of the sound is expressed as: Among them, DCT represents discrete cosine transform, FilterBank represents Mel filter bank, t represents time threshold, and MFCC represents the characteristics of sound.

3. The aluminum liquid leakage detection method based on deep learning according to claim 1 is characterized in that: The preprocessing includes normalization processing, which is expressed as follows: Where X represents any sensor data, μ and σ represent the mean and standard deviation of X, respectively.

4. The method for detecting aluminum liquid leakage based on deep learning according to claim 1, characterized in that: In S3, the fusion feature vector F of the multimodal data is obtained by performing data fusion on the multimodal data. c It is expressed as follows: F c =DNN(X′1,X′2,…,X′ n ) Among them, X' i is the standardized data of the i-th sensor, DNN represents a deep neural network; wherein the deep neural network includes a convolutional neural network and a recurrent neural network.

5. The method for detecting aluminum liquid leakage based on deep learning according to claim 4, characterized in that: Use autoencoder AE to process deep fusion feature vector F c , extract the key feature F that is most conducive to aluminum liquid leakage detection opt , which is expressed as follows: F opt =AE(F c )。 6. The method for detecting aluminum liquid leakage based on deep learning according to claim 5, characterized in that: Passing through the key feature F opt Input into the classifier to judge the leakage of aluminum liquid, as shown below: Leak_Detection=C(F opt ) Among them, Leak_Detection represents the leak detection result, and C represents the classifier.

7. The method for detecting aluminum liquid leakage based on deep learning according to claim 1, characterized in that: In S4, the loss function is optimized using a stochastic gradient descent method with an adaptive learning rate, which is expressed as: Among them, θ t represents the model parameters of the tth iteration; α t Represents the adaptive learning rate, which is dynamically adjusted according to the training progress; represents the gradient of the loss function L with respect to the parameter θ.

8. The method for detecting aluminum liquid leakage based on deep learning according to claim 1, characterized in that: The S5 specifically includes: inputting the fused features into the trained model prediction score, judging whether the score is higher than a set threshold, and thus judging whether there is aluminum liquid leakage.

9. A deep learning-based aluminum liquid leakage detection system, characterized in that: The system comprises: The data acquisition unit is used to install a pickup under the coolant of the aluminum well for sound collection, and arrange temperature and pressure sensors to collect temperature and pressure, combine the sound, temperature and pressure into multi-modal data, and pre-process the multi-modal data; A model building unit, used for building an adaptive noise suppression model using sound data; Data fusion unit, used to fuse multi-modal data and extract key features of aluminum liquid leakage; The model training unit is used to train the adaptive noise suppression model using transfer learning technology and fine-tune it in combination with the labeled data; the training loss function is Among them, y i represents the true label of the i-th sample; p i represents the probability predicted by the model; n i represents the estimated ambient noise level; H(p i ,n i ) represents the noise adaptation term, which is used to adjust the model's sensitivity to noise; λ represents the adjustment coefficient, which is used to balance the classification loss and the noise adaptation loss; Real-time warning unit, used to monitor data in real time and trigger an alarm if there is an abnormal leakage of aluminum liquid; The model building unit executes the following steps: S21, calculate the power spectrum density PSD(f) of the current environmental background noise in real time, expressed as follows: Where x(t) represents the sound signal in the time domain t, FFT represents fast Fourier transform, N represents the number of samples, and f represents the frequency; S22. Calculate the dynamic filter coefficient according to the power spectrum density PSD(f) to adapt to the current noise environment, as shown below: Wherein, α(f) represents the filter coefficient at frequency f, and θ represents the threshold value preset according to the ambient noise level; S23, using dynamic filter coefficients to perform denoising on the sound signal, as shown below: S clean (f)=S raw (f)·α(f) Among them, S raw (f) represents the frequency domain representation of the original sound signal, S clean (f) represents the signal after denoising; S24, convert the denoised signal back to the time domain through inverse Fourier transform, which is expressed as follows: s clean (t)=IFFT(S clean (f))。

Images