Single-channel blind source separation method based on improved Conv-TasNet network

Abstract

The invention relates to a single-channel blind source separation method based on an improved Conv-TasNet network, and belongs to the technical field of power grid monitoring. According to the single-channel blind source separation method based on the improved Conv-TasNet network, one-dimensional convolution of four different convolution kernels is connected in parallel, and features extracted by different convolution are fused by using an AFF feature fusion module to replace an encoder part of the Conv-TasNet network, so that the network can utilize features of input signals more effectively. In addition, the mean value of the time-frequency entropy of the signal is combined with the Mel-frequency spectrogram to serve as an evaluation index, the uniformity of energy distribution of the signal on a time-frequency plane can be quantified, and an intuitive mode is provided to visually evaluate the signal separation effect.

Description

Technical Field

[0001] The invention relates to a single-channel blind source separation method based on an improved Conv-TasNet network, and belongs to the technical field of power grid monitoring. Background Art

[0002] Smart grids improve the efficiency and safety of grid management by integrating electronics, information, sensors, and system engineering technologies. However, existing video sensors have limited performance in bad weather and light conditions, which affects the effectiveness of grid monitoring. To solve this problem, the ultra-weak grating optical fiber distributed acoustic sensing (uwDAS) system has emerged as a low-cost, easy-to-maintain alternative. It senses vibration signals and can continuously provide data even in poor visibility conditions. Its high temporal and spatial resolution provides a new solution for all-weather, comprehensive grid monitoring.

[0003] In practical applications, due to the high sensitivity of DAS and the strong noise background, multi-source interference and unknown underground conditions in urban areas, a large number of unknown interference sources and vibration sources may act on the sensor at the same time, causing the signal to mix in space and time, making the signal received by DAS complex, and facing the problem of difficulty in accurately identifying and separating useful information from the mixed signal. Some researchers have combined signal separation technology with DAS signal processing and made a series of progress. However, the existing traditional separation methods have problems such as poor separation effect in underdetermined conditions and unknown separation effect when the real mixed signal exceeds two sources. At present, deep learning methods have achieved good results in fields such as speech separation, and in laboratory environments, Conv-TasNet networks have achieved excellent results in the separation of multi-source mixed vibration signals of uwDAS systems, but only in time-aligned mixed DAS vibration signals. Mixed signals in real environments are usually mixed in unknown ways and the signals are not aligned in time, which shows that deep learning-based models have great potential in dealing with complex signal separation problems, and there is also room for further optimization to improve the performance and robustness of separation.

[0004] In terms of blind separation of real mixed vibration signals of uwDAS, the defects and shortcomings of existing methods are as follows: (1) Signal complexity: The vibration signals collected by uwDAS usually have complex frequency distribution and time-varying characteristics, and the mixing mode is unknown. Other methods are usually more suitable for mixed signals of linear instantaneous mixing, which makes it difficult to accurately capture these complexities and leads to inaccurate separation.

[0005] (2) Noise interference: Vibration signals in real environments are often accompanied by various noises. Other methods are prone to misjudgment in high-noise conditions, thus affecting the accuracy of separation.

[0006] (3) Non-Gaussian signal: The uwDAS vibration signal may have non-Gaussian characteristics, which may not be consistent with the assumptions of the method and lead to deviations in the separation results.

[0007] (4) Evaluation indicators: The mixing mode of mixed signals in real environments is usually unknown and the time of each signal is not aligned, which makes it difficult to provide accurate labels for these signals and it is impossible to objectively and accurately evaluate the separation effect. Summary of the invention

[0008] The purpose of the present invention is to propose a single-channel blind source separation method based on an improved Conv-TasNet network to solve some of the above problems.

[0009] The technical solution of the present invention is: A single-channel blind source separation method based on an improved Conv-TasNet network, characterized in that the method comprises the following steps: Step 1: Collect single-source and multi-source vibration signals using uwDAS equipment.

[0010] Step 2: Make the collected data into three data sets: training set, validation set, and test set.

[0011] Step 3: Input the training set into the improved Conv-TasNet network model for training, adjust the parameters to optimize the model, verify the model effect on the validation set, and output the separation model after training.

[0012] Step 4: Compare the trained separation model with the original Conv-TasNet network model on the test set, and compare and evaluate the separation model with the original Conv-TasNet network model through evaluation indicators.

[0013] In step 1, single-source and multi-source vibration signals are collected using uwDAS equipment.

[0014] 1.1. Use uwDAS equipment to collect vibration signals of 5 different audios separately. In each experiment, the audio was placed about 10 cm away from the fiber optic disk of the uwDAS equipment and played to obtain 5 independent single-source vibration signal data.

[0015] 1.2. When collecting multi-source mixed vibration signals, these five signals are combined in different ways, including combinations of 2, 3, 4 and 5 audio sources. At the same time, the playback position of all audios is the same as that of single-source vibration signals, which is kept about 10 cm away from the fiber optic disk. The uwDAS device is used for data collection to obtain data of all vibration signals mixed with 2, 3, 4 and 5 sources.

[0016] In step 2, the collected data is made into three data sets: training set, validation set, and test set.

[0017] 2.1. Convert the vibration signal data in step 1 into a WAV audio file for subsequent processing.

[0018] 2.2. Make training set and validation set: 2.2.1. Cut the five single-source vibration signal audios collected in 1.1 into 1-second segments and save them in five files respectively.

[0019] 2.2.2. Shuffle the audio clips in these files to ensure the randomness of the order.

[0020] 2.2.3. Randomly select audio clips from 5 files for mixing. The mixing method is linear instantaneous mixing. Label the mixed audio according to the number of mixed sources. The label is the source signal corresponding to the mixed signal. Suppose there are two audio signals and , its linear instantaneous mixing can be expressed by formula (1), where is the mixed signal, a and b are the weight coefficients during mixing, and their value range is [0,1).

[0021]

[0022] 2.2.4. Divide the randomly mixed data into training set and validation set in a ratio of 8:2.

[0023] 2.3. Cut the mixed vibration signal audio from 2, 3, 4, and 5 sources collected in 1.2 into 1-second segments to make a test set.

[0024] In step 3, the specific improvement scheme and network training settings of the Conv-TasNet network are as follows: 3.1. Network Improvement: 3.1.1. In the encoder part, multiple one-dimensional convolutional layers are used to process the same input signal in parallel to extract multi-scale features. The input signal is a three-dimensional tensor (B, C, T), where B represents the batch size, C represents the number of channels, and T represents the time length. By applying convolution kernels of different sizes, four multi-scale features are obtained respectively. , the dimensions of each feature are (B, C, H, W), where H and W are the height and width of the feature map respectively. These features are converted into activation features after being processed by the ReLU activation function. .

[0025] 3.1.2. The one-dimensional convolutional features of different scales are fused through the AFF feature fusion module.

[0026] The Attention Feature Fusion (AFF) module uses a two-stage attention mechanism to weight and fuse the activated features. and As an example, the fusion process of AFF includes two main steps: the first stage is to and Add element by element, then pass the local ( ) and global attention ( ) network, use sigmoid Function to generate weights , which is used to adjust the ratio of feature map fusion. The mathematical expression of this step is as follows:

[0027] In the second stage, the above attention mechanism is repeated, but this time it is for the output of the first stage Iterate the process to further refine the weighting and fusion of feature maps to obtain the final output (correspond Figure 3 The output of 1) is expressed mathematically as follows:

[0028] This two-stage attention mechanism gives the network a deeper understanding, enabling it to identify important information in the feature map, thereby improving the overall feature representation quality.

[0029] 3.2 Network training: The loss function of network training is a combination of SISDR and MSE. By adjusting the weight between the two, the model can maintain the quality of source signal separation while also focusing on the details and accuracy of the signal. The mathematical expression of SISDR is as follows:

[0030] in the formula represents the original signal, represents the signal estimated by the model, and for Decomposed into two mutually perpendicular vectors, where and Same direction, if and The more parallel, will be smaller, The larger the value, the larger the calculated SISDR value will be. During training, the negative SISDR value is taken as the loss.

[0031] The mathematical expression of MSE is as follows:

[0032] in, is the true value, is the predicted value.

[0033] The loss function of network training is expressed as:

[0034] in q Represents weight.

[0035] In step 4, the time-frequency entropy is calculated by Hilbert transform, and the time domain signal is converted into Mel spectrum representation using the melSpectrogram function in MATLAB.

[0036] The advantages of the present invention are: (1) The present invention replaces the encoder part of the ConvTasNet network by performing one-dimensional convolution with four different convolution kernels in parallel and then using the AFF feature fusion module to fuse the features extracted by different convolutions, so that the network can more effectively utilize the features of the input signal.

[0037] (2) The present invention combines SISDR and MSE as the training loss function to minimize the negative SI-SDR, thereby improving the separation quality. At the same time, MSE, as a time domain loss function, can make the waveform of the separated signal closer to the waveform of the original signal, which can improve the auditory naturalness and clarity.

[0038] (3) The present invention combines the mean time-frequency entropy of the signal with the Mel-spectrogram as an evaluation indicator, which can not only quantify the uniformity of the energy distribution of the signal on the time-frequency plane, but also provide an intuitive way to visually evaluate the effect of signal separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the present invention.

[0040] Figure 2 This is a diagram of the uwDAS system for collecting vibration signals in the present invention.

[0041] Figure 3This is a diagram of the encoder part of the improved Conv-TasNet network of the present invention.

[0042] Figure 4 Comparison chart of training set and validation set loss before and after improvement.

[0043] Figure 5 Comparison chart of the mean time-frequency entropy of the estimated signal in the test set before and after improvement.

[0044] Figure 6 Comparison of the mel spectra of a random 1-second signal in the test set before and after improvement. DETAILED DESCRIPTION

[0045] like Figure 1 As shown in the figure, the process of single-channel blind source separation method based on the improved Conv-TasNet network is as follows: In step 1, single-source and multi-source vibration signals are collected using uwDAS equipment.

[0046] 1.1. Use uwDAS equipment to collect vibration signals of 5 different audios separately. In each experiment, the audio was placed about 10 cm away from the fiber optic disk of the uwDAS equipment and played to obtain 5 independent single-source vibration signal data.

[0047] 1.2. When collecting multi-source mixed vibration signals, these five signals are combined in different ways, including combinations of 2, 3, 4 and 5 audio sources. At the same time, the playback position of all audios is the same as that of single-source vibration signals, which is kept about 10 cm away from the fiber optic disk. The uwDAS device is used for data collection to obtain data of all vibration signals mixed with 2, 3, 4 and 5 sources.

[0048] In step 2, the collected data is made into three data sets: training set, validation set, and test set.

[0049] 2.1. Convert the vibration signal data in step 1 into a WAV audio file for subsequent processing.

[0050] 2.2. Make training set and validation set: 2.2.1. Cut the five single-source vibration signal audios collected in 1.1 into 1-second segments and save them in five files respectively.

[0051] 2.2.2. Shuffle the audio clips in these files to ensure the randomness of the order.

[0052] 2.2.3. Randomly select audio clips from 5 files for mixing. The mixing method is linear instantaneous mixing. Label the mixed audio according to the number of mixed sources. The label is the source signal corresponding to the mixed signal. Suppose there are two audio signals and , its linear instantaneous mixing can be expressed by formula (1), where is the mixed signal, a and b are the weight coefficients during mixing, and their value range is [0,1).

[0053]

[0054] 2.2.4. Divide the randomly mixed data into training set and validation set in a ratio of 8:2.

[0055] 2.3. Cut the mixed vibration signal audio from 2, 3, 4, and 5 sources collected in 1.2 into 1-second segments to make a test set.

[0056] In step 3, the specific improvement scheme and network training settings of the Conv-TasNet network are as follows: 3.1. Network Improvement: 3.1.1. In the encoder part, multiple one-dimensional convolutional layers are used to process the same input signal in parallel to extract multi-scale features. The input signal is a three-dimensional tensor (B, C, T), where B represents the batch size, C represents the number of channels, and T represents the time length. By applying convolution kernels of different sizes, four multi-scale features are obtained respectively. , the dimensions of each feature are (B, C, H, W), where H and W are the height and width of the feature map respectively. These features are converted into activation features after being processed by the ReLU activation function. .

[0057] 3.1.2. The one-dimensional convolutional features of different scales are fused through the AFF feature fusion module.

[0058] The Attention Feature Fusion (AFF) module uses a two-stage attention mechanism to weight and fuse the activated features. and As an example, the fusion process of AFF includes two main steps: the first stage is to and Add element by element, then pass the local ( ) and global attention ( ) network, use sigmoid Function to generate weights , which is used to adjust the ratio of feature map fusion. The mathematical expression of this step is as follows:

[0059] In the second stage, the above attention mechanism is repeated, but this time it is for the output of the first stage Iterate the process to further refine the weighting and fusion of feature maps to obtain the final output (correspond Figure 3 The output of 1) is expressed mathematically as follows:

[0060] This two-stage attention mechanism gives the network a deeper understanding, enabling it to identify important information in the feature map, thereby improving the overall feature representation quality.

[0061] 3.2 Network training: The loss function of network training is a combination of SISDR and MSE. By adjusting the weight between the two, the model can maintain the quality of source signal separation while also focusing on the details and accuracy of the signal. The mathematical expression of SISDR is as follows:

[0062] in the formula represents the original signal, represents the signal estimated by the model, and for Decomposed into two mutually perpendicular vectors, where and Same direction, if and The more parallel, will be smaller, The larger the value, the larger the calculated SISDR value will be. During training, the negative SISDR value is taken as the loss.

[0063] The mathematical expression of MSE is as follows:

[0064] in, is the true value, is the predicted value.

[0065] The loss function of network training is expressed as:

[0066] in q Represents weight.

[0067] In step 4, the time-frequency entropy is calculated by Hilbert transform, and the time domain signal is converted into Mel spectrum representation using the melSpectrogram function in MATLAB.

[0068] Embodiment 1: The present invention is further described with reference to the accompanying drawings and embodiments.

[0069] Step 1 Figure 2 The uwDAS device used in the present invention has the following principle: the continuous light emitted by the narrowband light source is modulated into an optical pulse signal by a semiconductor amplifier, the pulse power is increased by an erbium-doped fiber amplifier, and injected into the UW-FBGs array in the optical fiber. The reflected pulses of two adjacent gratings pass through an unbalanced Michelson interferometer and interfere in a 3×3 coupler to form three interference optical signals with a phase difference of 120°. After photoelectric conversion, they are synchronously collected by three high-speed A / Ds. At the same time, the embedded circuit completes the inverse tangent, digital filtering, and phase reverse calculation in real time, and then outputs and displays through the network port. The uwDAS device was used to separately collect vibration signals of 5 different audios. These 5 audios are pure audios without noise, namely, fans, mechanical operation, sirens, footsteps, and wind. Each audio was placed separately at a distance of about 10 cm from the optical fiber disk of the uwDAS device for playback. Finally, the duration of each vibration signal was 180 seconds; the duration of the vibration signal mixed from 2 sources was 2261 seconds in total; the duration of the vibration signal mixed from 3 sources was 1867 seconds in total; the duration of the vibration signal mixed from 4 sources was 1139 seconds in total; the duration of the vibration signal mixed from 5 sources was 425 seconds in total; the vibration signals collected in all experiments were recorded at a sampling frequency of 10 kHz.

[0070] Step 2: Taking 4-source mixed signal separation as an example, the duration of each data in the data set is 1 second, the number of training set data is 6048, the number of validation set data is 1512, and the number of test set data is 1000.

[0071] Step 3: The input signal is a three-dimensional tensor (8, 1, 1), and the convolution kernels of the four parallel one-dimensional convolutions are 4, 8, 16, and 32. Training parameter settings: learning rate 0.001, optimizer Adam, training set batch 8, validation set batch 1, training round number 200, loss function weight 0.6.

[0072] Comparison of training set and validation set losses of the improved network and the original network Figure 4 As shown in the figure, it can be seen that after the improvement, the loss convergence speed becomes faster and the minimum value of the convergence loss is also lower.

[0073] Step 4: Compare the trained separation model with the original Conv-TasNet network model on the test set, and compare and evaluate the separation model with the original Conv-TasNet network model through evaluation indicators.

[0074] Use the network model obtained in step 4 to separate the mixed signal of the test set and calculate its time-frequency entropy mean. Use the melSpectrogram function in MATLAB to convert the time domain signal into Mel spectrum representation. The window function uses a Hamming window, and the windowSize is set to 256. The number of overlapping samples between each frame is set to half of the window size, that is, the overlap length is 128, and the FFT length is set to 2048. According to the principle that the closer the time-frequency entropy mean is to the source signal, the better, Figure 5 In the figure, the mean time-frequency entropy of signals 1, 3, and 4 in the improved separated signals are closer to the source signal, indicating that the signals separated by the model obtained by the improved network training are closer to the source signal. Figure 6 The signal-free part of the Mel-spectrogram of the last separated signal is cleaner, which shows that the improved network performs better in separating real mixed signals. Figure 5 and Figure 6 The results show that the separation effect of the improved network is significantly better than that before the improvement.

[0075] This single-channel blind source separation method based on the improved Conv-TasNet network replaces the encoder part of the ConvTasNet network by connecting four one-dimensional convolutions with different convolution kernels in parallel, and then uses the AFF feature fusion module to fuse the features extracted by different convolutions, so that the network can more effectively utilize the features of the input signal. In addition, it combines the mean time-frequency entropy of the signal with the Mel spectrum as an evaluation indicator, which can not only quantify the uniformity of the energy distribution of the signal on the time-frequency plane, but also provide an intuitive way to visually evaluate the effect of signal separation.

Claims

1. A single-channel blind source separation method based on an improved Conv-TasNet network, characterized in that: The method comprises the following steps: Step 1: Collect single-source and multi-source vibration signals using uwDAS equipment; Step 2: Make the collected data into three data sets: training set, validation set, and test set; Step 3: Input the training set into the improved Conv-TasNet network model for training, adjust the parameters to optimize the model, verify the model effect on the validation set, and output the separation model after training; Step 4: Compare the trained separation model with the original Conv-TasNet network model on the test set, and compare and evaluate the separation model with the original Conv-TasNet network model through evaluation indicators.

2. According to claim 1, a single-channel blind source separation method based on an improved Conv-TasNet network is characterized in that: In step 1, single-source and multi-source vibration signals are collected using uwDAS equipment, which includes the following steps: 1.

1. Use uwDAS equipment to separately collect vibration signals of 5 different audios. In each experiment, the audio is placed about 10 cm away from the optical fiber disk of the uwDAS equipment to play, and 5 independent single-source vibration signal data are obtained. 1.

2. When collecting multi-source mixed vibration signals, these five signals are combined in different ways, including combinations of 2, 3, 4 and 5 audio sources. At the same time, the playback position of all audios is the same as that of single-source vibration signals, which is kept about 10 cm away from the fiber optic disk. The uwDAS device is used for data collection to obtain data of all vibration signals mixed with 2, 3, 4 and 5 sources.

3. According to claim 1, a single-channel blind source separation method based on an improved Conv-TasNet network is characterized in that: In step 2, the collected data is made into three data sets: a training set, a validation set, and a test set, which includes the following steps: 2.

1. Convert the vibration signal data in step 1 into a WAV audio file for subsequent processing; 2.

2. Make training set and validation set: 2.2.

1. Cut the five single-source vibration signal audios collected in 1.1 into 1-second segments and save them in five files respectively; 2.2.

2. Shuffle the audio clips in these files to ensure the randomness of the order; 2.2.

3. Randomly select audio clips from 5 files for mixing. The mixing method is linear instantaneous mixing. Label the mixed audio according to the number of mixed sources. The label is the source signal corresponding to the mixed signal. Suppose there are two audio signals and , its linear instantaneous mixing can be expressed by formula (1), where is the mixed signal, a and b are the weight coefficients during mixing, and their value range is [0,1); ； 2.2.

4. Divide the randomly mixed data into training set and validation set in a ratio of 8:2; 2.

3. Cut the mixed vibration signal audio from 2, 3, 4, and 5 sources collected in 1.2 into 1-second segments to make a test set.

4. According to claim 1, a single-channel blind source separation method based on an improved Conv-TasNet network is characterized in that: In step 3, the specific improvement scheme and network training settings of the Conv-TasNet network are as follows: 3.

1. Network Improvement: 3.1.

1. In the encoder part, multiple one-dimensional convolutional layers are used to process the same input signal in parallel to extract multi-scale features; The input signal is a three-dimensional tensor (B, C, T), where B represents the batch size, C represents the number of channels, and T represents the time length; By applying convolution kernels of different sizes, four multi-scale features are obtained respectively: , the dimension of each feature is (B, C, H, W), where H and W are the height and width of the feature map respectively; These features are transformed into activation features after being processed by the ReLU activation function. ; 3.1.

2. The one-dimensional convolutional features of different scales are fused through the AFF feature fusion module; Feature fusion is Attention Feature Fusion. The AFF module uses a two-stage attention mechanism to weight and fuse the activated features. and As an example, the fusion process of AFF includes two main steps: the first stage is to and Add element by element, and then pass the local and global attention Network, use sigmoid Function to generate weights , which is used to adjust the ratio of feature map fusion. The mathematical expression of this step is as follows: ； In the second stage, the above attention mechanism is repeated, but this time it is for the output of the first stage Iterate the process to further refine the weighting and fusion of feature maps to obtain the final output , the mathematical expression of this step is as follows: ； This two-stage attention mechanism gives the network a deeper understanding ability, enabling it to identify important information in the feature map, thereby improving the overall feature representation quality; 3.2 Network training: The loss function of network training is a combination of SISDR and MSE. By adjusting the weight between the two, the model can maintain the quality of source signal separation while also focusing on the details and accuracy of the signal. The mathematical expression of SISDR is as follows: ； in the formula represents the original signal, represents the signal estimated by the model, and for Decomposed into two mutually perpendicular vectors, where and Same direction, if and The more parallel, will be smaller, The larger the value, the larger the calculated SISDR value. During training, the negative SISDR value is taken as the loss. The mathematical expression of MSE is as follows: ； in, is the true value, is the predicted value; The loss function of network training is expressed as: ； in q represents weight; In step 4, the time-frequency entropy is calculated by Hilbert transform, and the time domain signal is converted into Mel spectrum representation using the melSpectrogram function in MATLAB.

Images