A Ф-OTDR signal denoising method, system and storage medium

Through adaptive Gabor transformation and fundamental pitch analysis combined with the Self-Attention and CRF module of the neural network model, the loss function is optimized, and the problem of insufficient denoising performance of Ф-OTDR signal is solved, achieving higher signal-to-noise ratio and denoising effect. Suitable for long-term deployment and complex noise environments in the field of fiber logging.

CN118643279BActive Publication Date: 2024-11-08JARI AUTOMATION CO LTD CHINA
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411118680.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-08
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The existing Ф-OTDR signal denoising algorithms lack performance in adapting to the characteristics of the fiber logging field, making it difficult to achieve high signal-to-noise ratio and real-time denoising, especially in long-term deployment and complex noise environments.

Method used

Adaptive Gabor transformation is used to process Ф-OTDR signals, combined with fundamental pitch analysis and multi-dimensional feature extraction, input the pre-trained neural network model, and denoising it through the Self-Attention module and the CRF module, using band gain interpolation and inverse adaptive Gabor transformation , optimize the loss function to improve the signal-to-noise ratio.

Benefits of technology

It significantly improves the signal-to-noise ratio of the signal, enhances the denoising performance and robustness, maintains the naturalness and integrity of the signal, avoids excessive smoothing or distortion, and improves the overall performance of the communication system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118643279B_ABST
    Figure CN118643279B_ABST
Patent Text Reader

Abstract

The present application discloses a Ф‑OTDR signal denoising method, system and storage medium, which relate to the field of signal processing. The method comprises the following steps: processing the Ф‑OTDR signal by using an adaptive Gabor transform to obtain a first processing feature; performing a pitch analysis on the Ф‑OTDR signal to obtain a second processing feature; performing feature extraction on the first processing feature and the second processing feature to obtain a multidimensional feature, inputting the multidimensional feature into a pre-trained neural network model to obtain a multidimensional gain, and performing band gain interpolation to process the multidimensional gain to obtain a third processing feature; performing pitch filtering on the first processing feature and the second processing feature to obtain a fourth processing feature; applying the third processing feature to the fourth processing feature and then performing an inverse adaptive Gabor transform to output a denoised signal; the present application improves the signal-to-noise ratio of the Ф‑OTDR signal, which is beneficial to improving the overall performance of the communication system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a Φ-OTDR signal denoising method, system and storage medium. Background Art

[0002] In the field of fiber optic logging, in addition to the preliminary analysis of downhole signals, the task of continuously monitoring the parameters of production wells is also very important, which requires the Ф-OTDR system to provide higher-precision signals to achieve a better grasp of the production status. Therefore, the Ф-OTDR system that can run in a fixed position for a long time requires a denoising algorithm with better denoising effect and stronger real-time performance, and because it needs to be deployed for a long time, it has the conditions to obtain noise distribution and noise data sets. Therefore, in this case where the data set can be obtained, the neural network denoising algorithm provides a higher denoising potential and faster denoising speed for the Ф-OTDR signal.

[0003] The huge size of neural networks often makes real-time applications difficult without GPU support. The RNNoise denoising algorithm aims to solve real-time application problems in low-complexity situations. The RNNoise algorithm improves the quality of noise suppression through deep learning technology, and the RNNoise algorithm also has good robustness. The RNNoise algorithm has significant advantages in noise suppression, including improved signal quality, reduced computational complexity, good robustness and real-time performance. These advantages make the algorithm have broad application prospects in speech processing, audio enhancement and related fields.

[0004] When solving the problem of Ф-OTDR signal denoising, we found that the RNNoise network model has potential but still needs to be improved. We need to further improve the algorithm's performance in denoising so that it can better adapt to the characteristics of Ф-OTDR signals and thus achieve higher performance in denoising tasks. Summary of the invention

[0005] The purpose of the present application is to provide a Φ-OTDR signal denoising method, system and storage medium to improve the signal-to-noise ratio of the Φ-OTDR signal.

[0006] In order to achieve the above objectives, this application is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present application discloses a Φ-OTDR signal denoising method, comprising:

[0008] Adaptive Gabor transform is used to process the Ф-OTDR signal to obtain a first processing feature; fundamental tone analysis is performed on the Ф-OTDR signal to obtain a second processing feature;

[0009] Performing feature extraction on the first processing feature and the second processing feature to obtain a multidimensional feature, inputting the multidimensional feature into a pre-trained neural network model to obtain a multidimensional gain, and processing the multidimensional gain by using a frequency band gain interpolation to obtain a third processing feature;

[0010] Performing pitch filtering on the first processing feature and the second processing feature to obtain a fourth processing feature;

[0011] Applying the third processing feature to the fourth processing feature and then performing an inverse adaptive Gabor transform to output a denoised signal;

[0012] The architecture of the neural network model includes a Dense input layer, a first GRU layer, a second GRU layer, a third GRU layer, and a Self-Attention module, a Dense output layer, and a CRF module connected in sequence; the Self-Attention module receives the outputs of the Dense input layer, the first GRU layer, the second GRU layer, and the third GRU layer;

[0013] The input end of the first GRU layer is connected to the Dense input layer, the output end of the first GRU layer is connected to the second GRU layer and the third GRU layer, and the output end of the second GRU layer is connected to the third GRU layer; the Dense input layer, the second GRU layer and the third GRU layer can all receive external input.

[0014] Furthermore, the training process of the neural network model includes:

[0015] Set the loss function and activation function of the neural network model;

[0016] In each iteration, the preprocessed data is input into the architecture of the neural network model to update the parameters of the neural network model, and after the iteration, the neural network model is obtained.

[0017] Furthermore, the loss function of the neural network model is:

[0018] ;

[0019] in, and To fine-tune the loss weight parameters; is the loss function considering the gain, is the loss function considering the signal-to-noise ratio;

[0020] ;

[0021] ;

[0022] in, The value of is 0.5. represents the output gain, represents the estimated gain, and is a balance parameter.

[0023] Furthermore, the loss function is a loss function that adds a signal-to-noise ratio parameter; wherein t Frame signal-to-noise ratio [0,1] is defined as:

[0024] ;

[0025] ;

[0026] in, and is the signal-to-noise ratio of all-noise speech The mean and standard deviation of and are the amplitude spectra of pure speech and noise, respectively.

[0027] Furthermore, the activation functions of the neural network model include tanh function, sig function and relu function;

[0028] The Dense input layer adopts the tanh function, the Dense output layer adopts the sig function, and the first GRU layer, the second GRU layer and the third GRU layer all adopt the relu function.

[0029] Furthermore, the process of processing the Φ-OTDR signal by using adaptive Gabor transform to obtain the first processing feature includes:

[0030] Split the Ф-OTDR signal into multiple short-time signals;

[0031] Adaptively adjust the Gabor window of the first short-term signal, perform point multiplication on the adjusted Gabor window and the short-term signal, and perform short-time Fourier transform to obtain a spectrum representation of the short-term signal;

[0032] The frequency spectrum representations of the remaining short-term signals are repeatedly obtained, and the frequency spectrum representations of all the short-term signals are combined to obtain the first processing feature represented in the frequency domain.

[0033] Furthermore, the multi-dimensional features are 88-dimensional features, and the 88-dimensional features include:

[0034] 34 ERB cepstral coefficients, zero-crossing rates of 34 ERB frequency bands, first and second time derivatives of the first 6 of the 34 ERB cepstral coefficients, first 6 pitch correlations, 1 pitch period, and 1 special non-stationary value.

[0035] Furthermore, before the Ф-OTDR signal is processed by the adaptive Gabor transform, the method further includes: performing overlapping window processing on the Ф-OTDR signal;

[0036] The process between the inverse adaptive Gabor transform and the output of the denoised signal also includes: performing window overlapping and adding processing on the signal.

[0037] In a second aspect, the present application discloses a Φ-OTDR signal denoising system, characterized in that it comprises:

[0038] A memory for storing instructions;

[0039] A processor is used to operate according to the instructions to execute the steps of the method as described in any one of the first aspects.

[0040] In a third aspect, the present application discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the denoising method as described in any one of the first aspects is implemented.

[0041] This application uses adaptive Gabor transform to process Ф-OTDR signals and uses inverse adaptive Gabor transform to process the signals before outputting the results. The PESQ of the output denoised signal is improved by 0.07 and the STOI is improved by 0.26%. This design shows that the adaptive Gabor transform is more effective in retaining local time-frequency information, improving the signal-to-noise ratio of the Ф-OTDR signal, and is conducive to improving the overall performance of the communication system.

[0042] The present application adopts adaptive Gabor transform to process Ф-OTDR signal. Adaptive Gabor transform is efficient and accurate in extracting downhole Ф-OTDR signal features and cooperating with neural network model for denoising. The architecture of the neural network model includes Self-Attention module, Dense output layer and CRF module connected in sequence. The Self-Attention module receives the output of Dense input layer, first GRU layer, second GRU layer and third GRU layer. Compared with the traditional RNNoise network, it can effectively remove the noise and interference of Ф-OTDR signal and improve the signal-to-noise ratio of Ф-OTDR signal. Compared with the RNN network of the traditional RNNoise algorithm, the addition of Self-Attention module and CRF module can make it more intelligent when processing sequence data, not only considering the information of the current time point, but also integrating the context of the entire sequence, so as to generate a more reasonable and coherent output sequence. In the denoising task, this helps to remove noise while maintaining the naturalness and integrity of the signal, avoiding excessive smoothing or unnecessary distortion. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the overall flow chart of the algorithm of this application;

[0044] Figure 2 This is the architecture diagram of the neural network model for this application;

[0045] Figure 3 It is the ERB frequency band division rule diagram;

[0046] Figure 4 This is a comparison chart of the denoising performance of this application and the original RNNoise algorithm;

[0047] Figure 5 This is a schematic diagram of the Self-Attention layer structure of this application. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, objectives and effects achieved by this application easy to understand, this application is further explained below in conjunction with specific implementation methods.

[0049] This application is improved based on the RNNoise algorithm, and the concept is as follows: the Ф-OTDR signal first provides support for subsequent fundamental tone filtering through fundamental tone analysis, and extracts the characteristics of the fundamental tone; the Ф-OTDR signal captures the characteristics within the time-frequency range through adaptive Gabor changes; a total of 88-dimensional features are input into the neural network model and a 34-dimensional gain is output; the spectral gain calculated by the neural network model is used to suppress noise, and the suppression process is concentrated on the low-resolution spectral envelope; a comb filter is used to suppress noise interference between harmonics and improve the output quality.

[0050] Specifically, the present application discloses a Φ-OTDR signal denoising method, comprising the following steps: using adaptive Gabor transform to process the Φ-OTDR signal to obtain a first processing feature; performing pitch analysis on the Φ-OTDR signal to obtain a second processing feature; performing feature extraction on the first processing feature and the second processing feature to obtain a multidimensional feature, inputting the multidimensional feature into a pre-trained neural network model to obtain a multidimensional gain, and using band gain interpolation to process the multidimensional gain to obtain a third processing feature; performing pitch filtering on the first processing feature and the second processing feature to obtain a fourth processing feature; applying the third processing feature to the fourth processing feature and then performing inverse adaptive According to Gabor transformation, a denoised signal is output; wherein, the architecture of the neural network model includes a Self-Attention module, a Dense output layer and a CRF module connected in sequence; the Self-Attention module receives the outputs of the Dense input layer, the first GRU layer, the second GRU layer and the third GRU layer; the input end of the first GRU layer is connected to the Dense input layer, the output end of the first GRU layer is connected to the second GRU layer and the third GRU layer, and the output end of the second GRU layer is connected to the third GRU layer; the Dense input layer, the second GRU layer and the third GRU layer can all receive external input.

[0051] This application uses adaptive Gabor transform to process Ф-OTDR signals and uses inverse adaptive Gabor transform to process the signals before outputting the results. The PESQ of the output denoised signal is improved by 0.07 and the STOI is improved by 0.26%. This design shows that the adaptive Gabor transform is more effective in retaining local time-frequency information, improving the signal-to-noise ratio of the Ф-OTDR signal, and is beneficial to improving the overall performance of the communication system. The architecture of the neural network model includes a Self-Attention module, a Dense output layer and a CRF module connected in sequence; and the Self-Attention module receives the outputs of the Dense input layer, the first GRU layer, the second GRU layer and the third GRU layer; compared with the traditional RNNoise network, it can effectively remove the noise and interference of the Ф-OTDR signal and improve the signal-to-noise ratio of the Ф-OTDR signal.

[0052] Example 1

[0053] The present application is described below by a specific embodiment. Figure 1 shown.

[0054] Step 1: Adaptive Gabor transform is used to process the Ф-OTDR signal to obtain a first processing feature; and pitch analysis is performed on the Ф-OTDR signal to obtain a second processing feature, specifically:

[0055] The input Ф-OTDR signal is subjected to adaptive Gabor transform to obtain the first processing feature represented in the frequency domain. Adaptive Gabor transform allows the width of the window to change with the frequency. A wider window is used at low frequencies to obtain better frequency resolution, while a narrower window is used at high frequencies to obtain better time resolution, thereby solving the time-frequency resolution trade-off problem of STFT.

[0056] The adaptive Gabor transform process for Ф-OTDR signals specifically includes the following operations: When processing Ф-OTDR signals, the adaptive Gabor transform process cleverly combines dynamic adjustment with signal characteristic analysis. First, the parameters are initialized, including the center frequency, bandwidth, and time position, and the control parameters of the adaptive algorithm are planned, laying the foundation for subsequent dynamic adjustment.

[0057] Next, the Ф-OTDR signal is divided into a series of short time periods, and a separate Gabor window is prepared for each period. Then the key step of adaptively adjusting the Gabor window is entered. According to the local characteristics of the signal, the Gabor window parameters, such as window width and center frequency, are dynamically adjusted. This process involves the calculation of instantaneous frequency or power spectrum density, and the formulation of criterion functions to optimize the window function parameters.

[0058] The Gabor transform is followed by the point multiplication of the adjusted Gabor window and the short-term signal, followed by the short-time Fourier transform, to obtain the spectrum representation under a specific window; on this basis, the window is continuously moved to the next position of the signal, and the above adaptive adjustment and transformation process is repeated until the complete signal is covered. The spectrum information of each window is collected, and a time-frequency diagram is constructed to form a two-dimensional representation with time as the horizontal axis, frequency as the vertical axis, and color or grayscale to display amplitude or energy. This time-frequency diagram intuitively shows the time-frequency distribution of the signal; finally, through post-processing and analysis, the time-frequency diagram is deeply analyzed, which may include threshold processing, filtering, feature extraction, etc., to strengthen or refine key information and provide a solid foundation for subsequent decision support or application analysis.

[0059] This application modifies the original Fast Fourier Transform (FFT) into Adaptive Gabor Transform (AGT). Adaptive Gabor Transform dynamically adjusts the window width, optimizes the frequency resolution in the low-frequency area, and improves the time resolution in the high-frequency area; the transient characteristics of downhole signals are significant, and the adaptive Gabor transform is better than STFT in capturing transient events to avoid feature blur; for signal characteristics such as frequency modulation, the adaptive Gabor transform flexibly selects window functions to improve the accuracy of time-frequency representation; the fixed window introduces artifacts, and the adaptive Gabor transform reduces artifacts and provides clear time-frequency representation; the nonlinear adaptive Gabor transform of downhole signals deeply mines signal details to support the denoising task of the neural network model; the adaptive Gabor transform is efficient and accurate in extracting the characteristics of downhole sound signals (Ф-OTDR signals) and cooperating with the neural network model for denoising, and is the preferred solution for complex environment sound signal processing.

[0060] Perform a fundamental tone analysis on the Ф-OTDR signal to obtain the second processing feature. The fundamental tone analysis is used to detect the basic frequency components of the signal, which is very helpful for identifying the key modes and features of the signal. The fundamental tone analysis process is the process of the conventional RNNoise algorithm, and this application will not be further described.

[0061] Before step 1, the Ф-OTDR signal is also subjected to overlapping window processing, which is: dividing the Ф-OTDR signal into multiple signal segment windows of the same length, each window covering a part of the original signal, and adjacent windows having a certain degree of overlap, with an overlap degree of 60-70%. Through overlapping window processing, more dense frequency analysis samples can be obtained on the time axis, thereby improving the time resolution, which is particularly important for capturing rapidly changing Ф-OTDR signals.

[0062] Step 2: extract the first processing feature and the second processing feature to obtain a multidimensional feature, input the multidimensional feature into a neural network model to obtain a multidimensional gain, and use frequency band gain interpolation to process the multidimensional gain to obtain a third processing feature.

[0063] Feature extraction usually involves operations such as dimensionality reduction, normalization, and selection of feature subsets that best reflect signal characteristics. This process is a routine step of the RNNoise algorithm and will not be further described in this application. It is different from the 42-dimensional features extracted by the conventional RNNoise algorithm. The feature extraction of this application obtains 88-dimensional features, which include: 34 ERB cepstral coefficients obtained by logarithmic transformation of 34 frequency band energies and then discrete cosine transformation, the zero crossing rate of 34 ERB frequency bands, the first-order time derivatives and second-order time derivatives of the first 6 ERB cepstral coefficients among the 34 ERB cepstral coefficients, the first 6 fundamental frequency correlations, 1 fundamental frequency period, and 1 special non-stationary value.

[0064] The conventional RNNoise algorithm extracts 22-dimensional features and divides them into Bark bands. The difference of this application is that the Bark band is expanded to the ERB band that is more in line with the equivalent bandwidth of human hearing, which increases the dimension of the features and enhances the learning ability of the RNN network.

[0065] Specifically, this application selects 34 frequency bands to approximate the spectral envelope of speech and noise signals. These frequency bands cover the range from 0 to 20 kHz, and the intervals are set based on the concept of equivalent rectangular bandwidth of human hearing. ERB is a measurement method used to describe a person's ability to perceive sound frequencies, which takes into account the differences in the sensitivity of the human ear to different frequencies. By using ERB-based frequency band division, the human ear's perception of sound can be more accurately simulated.

[0066] In order to avoid the possibility of overly detailed spectral analysis introducing noise and increasing computational complexity, this application sets the minimum bandwidth to 100Hz. This setting can reduce the complexity of the calculation while ensuring the accuracy of the analysis. In order to make the enhanced signal perceptually closer to pure speech, this application further proposes a constraint: the signal in each frequency band should be consistent in terms of total energy and periodicity. Such processing helps to ensure that the frequency bands of the enhanced signal are more harmonious in hearing and closer to the real speech signal. In this way, the dimension of the feature can be increased, so that the model can learn more information and have a higher denoising ability. The ERB band division rules are as follows: Figure 3 shown.

[0067] On the basis of the original ERB feature, the zero-crossing rate is calculated on the 34 frequency bands of the signal. This strategy greatly enriches the characteristic expression of the signal, improves the ability to capture the transient characteristics of nonlinear signals, and also refines the description of the spectrum characteristics, making the signal analysis more comprehensive. It not only focuses on the spectrum energy distribution, but also takes into account the fluctuation and transient behavior in the time domain, improves the feature discrimination, and is conducive to the model to more accurately distinguish the signal type, especially in complex noise environments. In addition, the zero-crossing rate feature on multiple frequency bands increases the pertinence of the algorithm during denoising, and can be more effectively processed according to the transient characteristics of noise components of different frequencies, improving the denoising effect. At the same time, this feature combination also enhances the adaptability to complex signal structures, and can provide rich descriptions even when facing signals with multiple frequency components intertwined. In addition, the calculation of the zero-crossing rate is relatively stable and is not affected by changes in the signal amplitude, thereby enhancing the robustness of the feature and ensuring that the algorithm can maintain stable performance under large fluctuations in signal strength. In summary, the ERB feature set that incorporates the zero-crossing rate not only enriches the feature dimension, but also significantly improves the accuracy and robustness of signal processing and analysis.

[0068] The 88-dimensional features are input into the neural network model to obtain a 34-dimensional gain. The neural network model is further explained below.

[0069] The RNNoise algorithm improves the quality of noise suppression through deep learning technology. Due to the powerful feature learning and pattern recognition capabilities of deep learning, the algorithm can more accurately estimate and eliminate noise, improving the clarity and intelligibility of speech or audio signals.

[0070] In the RNNoise algorithm, in order to simulate the time dependency of the signal, the RNN neural network introduces a GRU layer. In order to better learn the characteristics of the signal to distinguish between the signal and noise, another GRU layer is used to capture the frequency evolution of the spectral pattern.

[0071] When dealing with the problem of Ф-OTDR signal denoising, it was found that the RNNoise network model has potential but still needs to be improved. In order to further improve the performance of the algorithm in denoising, it was decided to make necessary supplements and improvements to the RNNoise network model. Through such changes, the RNNoise network can better adapt to the characteristics of the Ф-OTDR signal, thereby exerting higher efficiency in the denoising task.

[0072] Specific improvements include: This application improves the original GRU layer of the neural network in the RNNoise algorithm, and innovatively adds the Self-Attention module and the CRF module. Specifically, the architecture of the neural network model includes a Dense input layer, a first GRU layer, a second GRU layer, a third GRU layer, and a Self-Attention module, a Dense output layer, and a CRF module connected in sequence; the Self-Attention module receives the output of the Dense input layer, the first GRU layer, the second GRU layer, and the third GRU layer; the input end of the first GRU layer is connected to the Dense input layer, the output end of the first GRU layer is connected to the second GRU layer and the third GRU layer, and the output end of the second GRU layer is connected to the third GRU layer; the Dense input layer, the second GRU layer, and the third GRU layer can all receive external input.

[0073] In the neural network architecture for signal denoising, the Self-Attention module is as follows Figure 5As shown in the figure, the core function of the Self-Attention module is to calculate the attention weights, apply these weights, and update the model state, which is similar to the task of processing natural language. First, the attention weights are calculated based on the similarity between the query vector, the key vector, and the value vector. The query vector can be derived from the intermediate layer output of the signal denoising network, representing the signal characteristics at the current moment; the key vector is usually derived from the feature representation of the original signal, which reflects the characteristics of the signal at different time points or in the frequency domain; the value vector is also rooted in the feature representation of the signal and is used for subsequent weighted averaging. The attention weights are calculated by dot product, additive or multiplicative methods and normalized into probability distributions by the softmax function. Subsequently, the attention weights are applied to the value vector, and a comprehensive signal feature is formed by weighted summation. This feature vector focuses on the key information in the signal and filters out the noise components. Finally, the comprehensive signal feature is combined with the current state of the network to update the internal representation of the network or directly as the input for subsequent processing. This combination can be achieved by concatenation or dot product.

[0074] In the preprocessing stage, the Self-Attention module makes the features of each signal segment not only limited to itself, but also takes into account other segments of the entire signal sequence, thereby generating richer and more closely related key and value vectors. In the post-processing stage, the current state of the network is used as a query to match the previously generated keys, and the value vectors are weighted and summed using the weights calculated by the Self-Attention module to generate a denoised signal feature, which is then fused with the network state to guide the next denoising operation. This design not only enhances the network's expressive power, but also optimizes the signal processing process. In particular, when processing long signals, the self-attention mechanism (Self-Attention module) can effectively reduce the computational complexity while maintaining or improving the denoising effect.

[0075] Adding a CRF module after the Dense output layer can capture the sequence dependency between outputs. For sequence prediction tasks such as denoising, it can optimize the coherence of the final output sequence. The conditional random field (CRF module) can optimize the output of the entire sequence, ensuring the coherence and rationality of the output sequence, rather than just element-by-element prediction.

[0076] In the neural network model, the core advantage of the CRF module is that it can build a probabilistic model unit that defines the interdependence between elements in the output sequence, which means that in the denoising process, it not only depends on the information of the current frame, but also takes into account the mutual influence of the previous and next frames, thereby generating a smoother and more coherent signal. Compared with the sequence directly output by the RNN, the CRF layer is committed to finding the most likely output sequence under the given input sequence conditions through a global optimization strategy, namely the maximum a posteriori probability (MAP) inference, rather than simply pursuing the maximization of the output probability at a single time point. This feature ensures the overall quality and consistency of the output signal. In addition, the CRF layer allows explicit modeling of sequence constraints, such as the physical properties of the signal or specific noise patterns, and is coordinated with the rest of the network through parameter training and decoding using the Viterbi algorithm to ensure that the denoised signal reaches the optimal state as a whole, avoiding the limitations of local optimal solutions.

[0077] In general, the addition of the Self-Attention module and the CRF module can make the RNN network more intelligent when processing sequence data. It not only considers the information at the current time point, but also integrates the context of the entire sequence, thereby generating a more reasonable and coherent output sequence. In the denoising task, this helps to remove noise while maintaining the naturalness and integrity of the signal, avoiding over-smoothing or unnecessary distortion. The modified neural network model framework is as follows Figure 2 shown.

[0078] The training process of the neural network model includes: setting the loss function and activation function of the neural network model; in each iteration, inputting the preprocessed data into the architecture of the neural network model to update the parameters of the neural network model, and obtaining the neural network model after the iteration. The training process of the neural network is a conventional process, and this application will not be further explained. What is important is that this application has also made improvements to the neural network model architecture function.

[0079] In addition, the original optimization process of RNNoise focuses too much on whether the network can correctly estimate the correct gain, while the choice of gain is largely affected by feature selection. Therefore, the loss function of the neural network model architecture of this application is to add signal-to-noise ratio related parameters to the loss function of the traditional RNN neural network, so as to predict the frame-level signal-to-noise ratio under the learning framework to maintain good voice quality. Frame signal-to-noise ratio [0,1] is defined as:

[0080] ;

[0081] ;

[0082] in, and is the signal-to-noise ratio of all-noise speech The mean and standard deviation of and are the amplitude spectra of pure speech and noise, respectively.

[0083] This application further improves the original loss function of the RNN model in RNNoise. For the gain, it can be considered that the perceived loudness of the signal is proportional to the energy, that is, the energy is The power of which 0.5 is used. Therefore, before calculating the indicator, the gain is increased to In addition to the square error, the loss function also uses the fourth power to target losses that produce larger errors, such as completely attenuated speech. The loss function for gain is as follows:

[0084] ;

[0085] in, represents the output gain, represents the estimated gain, and To balance the parameters, we use and =10 to balance the two losses. The loss function implicitly incorporates many features, including scale invariance, power-law compression, and nonlinear frequency resolution.

[0086] The predicted signal-to-noise ratio of each frame can be used for the signal frame level signal-to-noise ratio and the actual signal-to-noise ratio The MSE loss is used to train the network for signal-to-noise ratio guidance. The loss function for signal-to-noise ratio is as follows:

[0087] ;

[0088] Loss function that takes gain into account And the loss function of signal-to-noise ratio , and finally the improved RNN model is trained using the following overall loss function:

[0089] ;

[0090] in, and To fine-tune the loss weight parameter.

[0091] In summary, the improved RNN network in this application has changed the architecture of the neural network on the one hand, and improved the loss function on the other hand.

[0092] The 34-dimensional gain is processed by band gain interpolation to obtain the third processing feature. Band gain interpolation is a technique for smoothing and adjusting the band gain to make it more suitable for application to the original signal. Band gain interpolation is a step of the conventional RNNoise algorithm and will not be further described in this application.

[0093] Step 3: Perform pitch filtering on the first processing feature and the second processing feature to obtain a fourth processing feature.

[0094] Pitch filtering is usually used to eliminate low-frequency noise in a signal or enhance the high-frequency part. This step is also a step of the conventional RNNoise algorithm and will not be further described in this application.

[0095] Step 4: Apply the third processing feature to the fourth processing feature and then perform inverse adaptive Gabor transform to output a denoised signal.

[0096] In this step, compared with the conventional RNNoise algorithm, IFFT is replaced by IAGT (Inverse Adaptive Gabor Transform), which is used to convert the frequency domain representation back to the time domain signal for subsequent processing and visualization.

[0097] In constructing the training data set, this application uses the public data set used by the RNNoise model to ensure alignment with the original research. These clean speech samples are mainly derived from the McGill TSP speech database and NTT's multilingual speech database, aiming to capture rich and diverse speech characteristics. In terms of simulating real-world noise scenarios, a variety of common noise sources are brought together, including the hum of computer fans, office noise, crowd noise, airplane roars, car horns, train whistles, and construction site noise. In addition, this application also pays special attention to the signal-free noise collected by the Ф-OTDR system, and includes it in the data set as a noise source with a large weight. In order to simulate the noise conditions in actual applications, these noises are mixed with pure speech signals to create training and test data sets with a wide range of signal-to-noise ratios. By combining a variety of gain adjustment and filtering techniques, the authenticity and validity of the data are ensured, and the sampling frequency is uniformly adjusted to 48kHz.

[0098] In terms of frame extraction, an adaptive Gabor transform that is independent of the time-frequency characteristics of the signal is used. The window function uses an adaptive window function, which can dynamically adjust its shape and width according to the frequency characteristics of the signal, thereby providing different resolutions at different frequencies, which is very suitable for processing non-stationary signals.

[0099] The training process uses the Adam optimizer, the initial yield is set to 0.001, and the loss function weight is , , , , They are set to 0.5, 1, 10, 5, and 1 respectively. To make the results comparable, all other configurations are the same as the original RNNoise. PESQ and STOI are used to measure speech quality and intelligibility, respectively.

[0100] Explanation of STOI: Short-Time Objective Intelligibility (STOI) is an important indicator for measuring speech intelligibility. For a word in a speech signal, there are only two situations: it can be understood or it cannot be understood. From this perspective, intelligibility can be considered to be binary, so the value range of STOI is quantified between 0 and 1, representing the percentage of words that are correctly understood. When the value is 1, it means that the speech can be fully understood. STOI itself is a number between 0 and 1. It is a convention to use % to more intuitively represent the specific value. The improvement is a specific value, not a "percentage" of the original level.

[0101] Detailed baseline results analysis such as Figure 4 As shown in Figure 2, the entire test set is divided into four levels of signal-to-noise ratio ranges to observe the denoising performance of the algorithm under different signal-to-noise ratio conditions. Figure 4 It can be found that at a lower signal-to-noise ratio level, the denoising performance of the algorithm of the present application is always stronger than the original RNNoise algorithm, but the denoising effect decreases at 10-15dB samples, which shows that the weight of the signal-to-noise ratio in the improved method needs to be adjusted at different levels. In the case of a high signal-to-noise ratio, an excessively high weight value will cause the algorithm to not run well.

[0102] Table 1. PESQ and STOI (%) of various models on the test set

[0103]

[0104] Table 1 shows the performance comparison of all the improved methods for the RNNoise model proposed in this paper. Compared with the traditional RNNoise algorithm, the improved algorithm proposed in this application significantly improves the PESQ score from 2.19 to 2.35, and the STOI from 92.68% to 93.31%. Specifically, replacing FFT with AGT improves PESQ by 0.06 and STOI by 0.21%. The introduced ERB features lead to an absolute increase of 0.05 and 0.21% in PESQ and STOI, respectively. With the help of the signal-to-noise ratio loss function, this application obtains a PESQ of 0.02 and an STOI improvement of 0.13%. When the Self-Attention and CRF structures are added, PESQ and STOI are further improved by 0.03 and 0.10% to 2.35 and 93.31%, respectively. In addition, on the test set, the improved methods proposed in this application all achieved positive benefits, with an overall PESQ improvement of 0.16 and an STOI improvement of 0.63%. Therefore, it can be concluded that the performance of the algorithm in this application has been improved without significantly increasing the parameters of the neural network.

[0105] In summary, the denoising performance of the model is improved by improving signal feature extraction, feature selection, network structure and threshold function. The experimental results show that the improvements proposed in this application have achieved positive benefits. The improved algorithm has good performance in acoustic signal denoising, can effectively remove noise and interference, and improve the signal-to-noise ratio of the signal.

[0106] Example 2

[0107] In one embodiment of the present application, the present application further discloses a Φ-OTDR signal denoising system, comprising: a memory for storing instructions; a processor for operating according to the instructions to execute the steps of the method described in Example 1.

[0108] The system may include, but is not limited to: at least one processor, at least one memory, and a bus connecting different system components (including the memory and the processor). The memory stores program code, which can be executed by the processor so that the processor executes the steps of various exemplary embodiments of the present invention described in the "Exemplary Method" section above. For example, the processor may execute the following steps: Figure 1 Steps shown.

[0109] The memory may include readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory, and may further include read only memory (ROM).

[0110] The memory may also include a program / utility having a set (at least one) of program modules, such program modules including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0111] The bus may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0112] The system may also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the system, and / or any device that enables the system to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via input / output (I / O) interfaces. Furthermore, the system may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the system via a bus.

[0113] It should be understood that although not shown in the figures, other hardware and / or software modules may be used in conjunction with the system, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0114] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0115] Example 3

[0116] In an exemplary embodiment of the present disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the above method of the present specification is stored. In some possible implementations, various aspects of the present invention can also be implemented in the form of a program product, which includes a program code, and when the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present invention described in the above "Exemplary Method" section of the present specification.

[0117] The program product for implementing the above method according to an embodiment of the present invention can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto, and in this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0118] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0119] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may output, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0120] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

[0121] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0122] It is known from common technical knowledge that the present application can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present application or within the scope equivalent to the present application are included in the present application.

Claims

1. A Ф-OTDR signal denoising method, characterized in that: include: Adopting adaptive Gabor transform to process Ф-OTDR signal, obtaining the first processing feature; Performing fundamental tone analysis on the Ф-OTDR signal to obtain a second processing feature; Performing feature extraction on the first processing feature and the second processing feature to obtain a multidimensional feature, inputting the multidimensional feature into a pre-trained neural network model to obtain a multidimensional gain, and processing the multidimensional gain by using a frequency band gain interpolation to obtain a third processing feature; Performing pitch filtering on the first processing feature and the second processing feature to obtain a fourth processing feature; Applying the third processing feature to the fourth processing feature and then performing an inverse adaptive Gabor transform to output a denoised signal; The architecture of the neural network model includes a Dense input layer, a first GRU layer, a second GRU layer, a third GRU layer, and a Self-Attention module, a Dense output layer, and a CRF module connected in sequence; the Self-Attention module receives the outputs of the Dense input layer, the first GRU layer, the second GRU layer, and the third GRU layer; The input end of the first GRU layer is connected to the Dense input layer, the output end of the first GRU layer is connected to the second GRU layer and the third GRU layer, and the output end of the second GRU layer is connected to the third GRU layer; the Dense input layer, the second GRU layer and the third GRU layer can all receive external input.

2. The Φ-OTDR signal denoising method according to claim 1, characterized in that: The training process of the neural network model includes: Set the loss function and activation function of the neural network model; In each iteration, the preprocessed data is input into the architecture of the neural network model to update the parameters of the neural network model, and after the iteration, the neural network model is obtained.

3. The Φ-OTDR signal denoising method according to claim 2, characterized in that: The loss function of the neural network model is: ; in, and To fine-tune the loss weight parameters; is the loss function considering the gain, is the loss function considering the signal-to-noise ratio; ; ; in, The value of is 0.

5. represents the output gain, represents the estimated gain, and is the balance parameter, represents the predicted signal-to-noise ratio of each frame, Indicates the actual signal-to-noise ratio of each frame.

4. The Φ-OTDR signal denoising method according to claim 3, characterized in that: The loss function is a loss function with the signal-to-noise ratio parameter added; t Frame signal-to-noise ratio [0,1] is defined as: ; ; in, and is the signal-to-noise ratio of all-noise speech The mean and standard deviation of and are the amplitude spectra of pure speech and noise, respectively.

5. The Φ-OTDR signal denoising method according to claim 2, characterized in that: The activation functions of the neural network model include tanh function, sig function and relu function; The Dense input layer adopts the tanh function, the Dense output layer adopts the sig function, and the first GRU layer, the second GRU layer and the third GRU layer all adopt the relu function.

6. The Φ-OTDR signal denoising method according to claim 1, characterized in that: The method of using adaptive Gabor transform to process the Φ-OTDR signal to obtain the first processing feature includes: Split the Ф-OTDR signal into multiple short-time signals; Adaptively adjust the Gabor window of the first short-term signal, perform point multiplication on the adjusted Gabor window and the short-term signal, and perform short-time Fourier transform to obtain a spectrum representation of the short-term signal; The frequency spectrum representations of the remaining short-term signals are repeatedly obtained, and the frequency spectrum representations of all the short-term signals are combined to obtain the first processing feature represented in the frequency domain.

7. The Φ-OTDR signal denoising method according to claim 1, characterized in that: The multi-dimensional features are 88-dimensional features, and the 88-dimensional features include: 34 ERB cepstral coefficients, zero-crossing rates of 34 ERB frequency bands, first and second time derivatives of the first 6 of the 34 ERB cepstral coefficients, first 6 pitch correlations, 1 pitch period, and 1 special non-stationary value.

8. The Φ-OTDR signal denoising method according to claim 1, characterized in that: Before the Ф-OTDR signal is processed by the adaptive Gabor transform, the method further includes: performing overlapping window processing on the Ф-OTDR signal; The process between the inverse adaptive Gabor transform and the output of the denoised signal also includes: performing window overlapping and adding processing on the signal.

9. A Ф-OTDR signal denoising system, characterized in that: include: A memory for storing instructions; A processor, configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the denoising method as described in any one of claims 1 to 8 is implemented.

Citation Information

Patent Citations

  • Audio noise reduction system and method based on structure of GRU network

    CN110120225A

  • Voice noise reducing method and device

    CN110136737A

  • Speech enhancement method and device, equipment and medium

    CN115223583A

  • Water turbine monitoring method and device

    CN116517746A

  • Weak signal processing systems and methods

    US10879946B1