Mixed signal double-path blind source separation system and method based on deep learning

Abstract

The invention belongs to the field of signal processing, and provides a mixed signal double-path blind source separation system and method based on deep learning in order to solve the problems that an existing blind source separation structure is not high in separation precision and not high in anti-noise capacity. The system at least comprises an encoder, a mask network and a decoder. And the encoder extracts mixed signal features and converts the mixed signal features into high-dimensional feature representation. And after entering the mask network, generating a mask representation through a mask encoder, a separator, a mask decoder and an output layer, multiplying the mask representation by the stacked high-dimensional feature representation, and performing conversion through the decoder to obtain a separated signal. And block processing in the mask encoder and the mask decoder is realized by adopting double paths. The system is also provided with a condition denoising module for removing signal noise. The method comprises the steps of configuring mixed signals, extracting features, carrying out block processing and mask generation, obtaining intermediate and original feature representations, realizing signal separation and the like. The blind source separation precision of signal processing can be improved, and the anti-noise capability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a mixed-signal dual-path blind source separation system and method based on deep learning. Background Technology

[0002] With the rapid advancement of signal processing technology and the widespread application of observation terminals, multi-source signal mixing scenarios are becoming increasingly complex, and the degree of signal component superposition is continuously deepening. The interaction of multi-source signals leads to a significant intensification of signal mixing effects within limited channels, and the superposition of different types of signals is becoming increasingly common. This complex signal environment not only strengthens the signal superposition effect but also easily generates composite signals, making signal detection, separation, and demodulation more challenging. Against this backdrop, the importance of Blind Source Separation (BSS) technology is becoming increasingly prominent. As a signal processing technique that requires no prior information, BSS can separate independent signal components from composite signals even in the absence of prior knowledge about the signal sources. This is of crucial value in multi-source signal mixing scenarios because traditional signal separation methods usually rely on specific signal models or training data, while BSS technology can extract different sources from the mixed signal through adaptive mechanisms, helping signal processing systems to more efficiently identify target signals and reduce the impact of interference on signal quality.

[0003] Blind source separation (BSS) technology has demonstrated wide applicability in signal demixing, interference suppression, and multi-source signal analysis. For example, in complex signal processing systems, BSS can extract effective signal components from background noise and interference signals, improving signal processing efficiency. In multi-channel signal acquisition scenarios, this technology can separate mixed signals generated by multi-path propagation, enhancing signal detection accuracy. In the field of signal security, BSS can also be used to detect and analyze spoofed signals, providing technical support for signal system security. With the continuous increase in the complexity of signal processing environments, BSS technology is gradually becoming a core tool for intelligent signal processing and system optimization.

[0004] In blind source separation technology, single-channel blind source separation is an important research direction. In this scenario, effective signal components need to be extracted from the observation signals acquired by a single sensor. Because the number of observation channels is less than the number of signal sources, the system is in an underconstrained state, significantly increasing the difficulty of signal separation. Although related research has made some progress, existing blind source separation architectures still suffer from insufficient separation accuracy and weak noise resistance, requiring further breakthroughs. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient separation accuracy and weak noise resistance in existing blind source separation structures during signal processing. This invention proposes a deep learning-based dual-path blind source separation system and method for mixed signals. The separation system includes at least an encoder, a masking network, and a decoder. The encoder extracts mixed signal features and transforms them into a high-dimensional feature representation. This high-dimensional feature representation enters the masking network and passes through a masking encoder, a separator, a masking decoder, and an output layer to generate a masking representation. This masking representation is multiplied by the stacked high-dimensional feature representation, and then converted back to the original feature representation by the decoder to obtain the separated signal. The masking encoder performs multiple block processing operations on the high-dimensional feature representation, with the block processing implemented using a dual-path approach. Furthermore, the system includes a conditional denoising module to remove signal noise and improve separation accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a mixed-signal dual-path blind source separation system based on deep learning, comprising at least: an encoder, a mask network, and a decoder; The encoder is used to extract the features of the received mixed signal and convert them into a high-dimensional feature representation. A mask network includes: a mask encoder, a splitter, a mask decoder, and an output layer; The mask encoder performs preprocessing and continuous processing on the high-dimensional feature representation sequentially. Each block processing and downsampling operation outputs the feature representation obtained after each block processing step. The second block processing outputs a total of One characteristic representation; separator receiving Each feature represents a different source. Each mask feature is represented and then input into the mask decoder; the mask decoder processes the received data... Each mask feature represents the execution The next block processing operation obtains the output of the mask decoder; the output layer performs post-processing on the output of the mask decoder to output the mask representation; The high-dimensional feature representations are stacked, and the mask representation is multiplied with the stacked high-dimensional feature representations to obtain the intermediate feature representation of each source signal in the mixed signal. The decoder receives the intermediate feature representation of each source signal and converts it into the original feature representation to obtain the separated signal.

[0007] As one possible implementation, block processing employs a dual-path approach, comprising an intra-block processing path and an inter-block processing path. Both the intra-block and inter-block processing paths include a bidirectional LSTM, a linear layer, and a normalization layer connected sequentially. Each block processing operation performs the following steps: The data to be processed is segmented to obtain multiple segmented data; Multiple segmented data are input into the block processing path to extract information within the segmented data and output the block features. The intra-block features are added to multiple segmented data and input into the inter-block processing path to extract information between segmented data, and the inter-block features are output. The inputs of inter-block features and inter-block processing paths are added together, and the sum is then superimposed before output.

[0008] As one possible implementation, the mask decoder processes the received... Each mask feature represents the execution The next block processing operation obtains the output of the mask decoder, specifically: for the first block... After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; let The above operations are performed iteratively until the output of the first block processing is obtained, which is the output of the mask decoder.

[0009] As one possible implementation, the output layer includes a first path and a second path connected in parallel, an output convolutional layer, and an output activation function layer, with the first path or the second path selected based on a gating mechanism; the first path includes a first path convolutional layer and a first path activation function, and the second path includes a second path convolutional layer and a second path activation function; Post-processing includes: processing the output of the mask decoder using the first path and the second path respectively to obtain the output of the first path and the output of the second path; performing an XOR operation on the output of the first path and the output of the second path, and inputting the operation result into the output convolutional layer and the output activation function layer to process and output the mask representation.

[0010] As one possible implementation, the deep learning-based mixed signal dual-path blind source separation system also includes a conditional denoising module, which is used to remove noise from the original mixed signal and then input it to the encoder. The conditional denoising module includes a denoising encoder and a denoising decoder. The conditional denoising module performs the following operations to remove noise from the original mixed signal: The denoising encoder receives the original mixed signal, performs multiple downsampling and expands the feature dimension before inputting it to the denoising decoder; The denoising decoder performs multiple upsampling operations and compresses the feature dimensions before outputting a noise estimate. The noise estimate is residually concatenated with the original mixed signal to output the denoised mixed signal.

[0011] As one possible implementation, starting from the input end, the denoising encoder includes three cross-connected one-dimensional convolutional layers and two activation functions, and the denoising decoder includes three cross-connected one-dimensional deconvolutional layers and two activation functions, both of which are ReLU activation functions.

[0012] Secondly, the present invention provides a deep learning-based method for separating dual-path blind sources in mixed signals, comprising: S1. Configure the mixed signal; S2. Extract the features of the mixed signal and transform them into a high-dimensional feature representation; S3. Perform preprocessing and continuous processing on the high-dimensional feature representations sequentially. Each block processing and downsampling operation outputs the feature representation obtained after each block processing step. The second block processing yielded a total of Each feature is represented; S4. Will These features indicate that the data were separated into different sources. Each mask feature is represented; S5. To Each mask feature is represented Secondary block processing operations, for The output obtained after the next block processing operation is post-processed to obtain a mask representation; S6. Stack the high-dimensional feature representations, multiply the mask representation with the stacked high-dimensional feature representations, and obtain the intermediate feature representation of each source signal in the mixed signal; S7. Convert the intermediate feature representation of each source signal in the mixed signal into the original feature representation to obtain the separated signal.

[0013] As one possible implementation, block processing includes: S30. Segment the data to be processed to obtain multiple segmented data; S31. Extract information from multiple segmented data using the bidirectional LSTM method to obtain intra-block features; S32. Based on the bidirectional LSTM method, extract intra-block features and inter-segment data information by summing multiple segmented data to obtain inter-block features; S33. Perform an overlay operation on the summation of inter-block features and intra-block features with multiple segmented data.

[0014] As one possible implementation, S5 includes: S50. Regarding After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; S51. Order Repeat S50 until the output of the first block processing is obtained; S52. Post-process the output of the first block to obtain a mask representation.

[0015] As one possible implementation, S52 includes: S520. The output of the first block processing is downsampled and activated by the Tanh function to obtain the first output feature representation; the output of the first block processing is downsampled and activated by the Sigmoid function to obtain the second output feature representation; S521. Perform an XOR operation on the first output feature representation and the second output feature representation to obtain the operation result; S522. The operation result is downsampled and activated by the ReLU function to obtain a mask representation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The deep learning-based dual-path blind source separation system for mixed signals proposed in this invention employs a mask network with deep learning capabilities, which can better learn the mask representations of different source signals and greatly improve the separation accuracy of mixed signals.

[0017] 2. The deep learning-based dual-path blind source separation system for mixed signals proposed in this invention employs dual-path multi-block processing operations in both the mask encoder and mask decoder, which can more fully extract the local and global features of the segmented data, thus helping to obtain more accurate separation results.

[0018] 3. The deep learning-based dual-path blind source separation system for mixed signals proposed in this invention incorporates a conditional denoising module to remove noise from the original mixed data. Experiments show that, under all mixing types, the separation error probability with the conditional denoising module is reduced compared to the separation error probability without it, indicating that this invention has high noise resistance. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the structure of a deep learning-based dual-path blind source separation system for mixed signals in an embodiment of the present invention. Figure 2 This is a schematic diagram of the mask network structure in an embodiment of the present invention; Figure 3This is a schematic diagram of the dual-path structure included in the block processing in an embodiment of the present invention; Figure 4 This is a schematic diagram of the separator structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the conditional denoising module structure in an embodiment of the present invention; Figure 6 This is a flowchart of a deep learning-based dual-path blind source separation method for mixed signals in an embodiment of the present invention. Figure 7 This is a comparison of the source signal waveform and the separated signal waveform under different mixing types in the embodiments of the present invention; Figure 8 The separation error probabilities of different mixing types under different signal-to-noise ratios in the embodiments of the present invention; Figure 9 This is the confusion matrix for separating and identifying signals in the absence of noise in this embodiment of the invention; Figure 10 This is the confusion matrix for signal separation identification under five different signal-to-noise ratios in the embodiments of the present invention.

[0020] Figure Labels 1-Encoder, 2-Mask Network, 20-Mask Encoder, 21-Separator, 210-First Convolutional Layer, 211-Linear Gating Unit, 212-Second Convolutional Layer, 213-View Layer, 214-Second Normalization Layer, 22-Mask Decoder, 23-Output Layer, 230-Output Convolutional Layer, 231-Output Activation Function Layer, 232-First Path Convolutional Layer, 233-First Path Activation Function Layer, 234-Second Path Convolutional Layer, 235-Second Path Activation Function Layer, 24-Bidirectional LSTM, 25-Linear Layer, 26-Normalization Layer, 3-Decoder, 4-Conditional Denoising Module, 40-Denoising Encoder, 41-Denoising Decoder. Detailed Implementation

[0021] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0022] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0024] This invention aims to address the problems of low separation accuracy and weak noise resistance in existing blind source separation structures during signal processing by providing a deep learning-based dual-path blind source separation system and method for mixed signals. The separation system includes at least an encoder, a masking network, and a decoder. The encoder extracts features from the mixed signal and transforms them into a high-dimensional feature representation. This high-dimensional feature representation enters the masking network and passes through a mask encoder, a separator, a mask decoder, and an output layer to generate a mask representation. This mask representation is multiplied by the stacked high-dimensional feature representation, and then converted back to the original feature representation by the decoder to obtain the separated signal. The mask encoder performs multiple block processing operations on the high-dimensional feature representation, with the block processing implemented using a dual-path approach. Furthermore, the system includes a conditional denoising module to remove signal noise and improve separation accuracy. In a first aspect, embodiments of the present invention provide a mixed-signal dual-path blind source separation system based on deep learning, see [link to relevant documentation]. Figure 1 The separation system includes at least: encoder 1, mask network 2, and decoder 3; Encoder 1 is used to extract features from the received mixed signal and convert them into a high-dimensional feature representation. For example, encoder 1 is implemented using one-dimensional convolution and the ReLU activation function. Assume the mixed signal received by encoder 1 is... After passing through encoder 1, the following is obtained: (1) in, , This indicates the length of the sequence after encoder 1. Indicates the output dimension of encoder 1. This represents one-dimensional convolution.

[0025] See Figure 2 The masking network 2 includes a mask encoder 20, a separator 21, a mask decoder 22, and an output layer 23; wherein, the mask encoder 20 performs preprocessing and continuous processing on the high-dimensional feature representation sequentially. Each block processing and downsampling operation outputs the feature representation obtained after each block processing step. The second block processing outputs a total of Each feature is represented; As an example, the high-dimensional feature representation output by encoder 1 Entering the mask encoder 20, the mask encoder 20 first performs bucket normalization and convolution operations to preprocess the high-dimensional feature representation, obtaining... : (2) See Figure 2 Next, the mask encoder has 20 pairs. Perform continuously The process involves several block processing and downsampling operations, followed by a final block processing operation. Each block processing operation yields a feature representation. The second block processing yielded a total of Each feature is represented.

[0026] See Figure 3 As one possible implementation, block processing employs a dual-path approach, comprising an intra-block processing path and an inter-block processing path. Both the intra-block and inter-block processing paths include a bidirectional LSTM24, a linear layer 25, and a normalized layer 26 connected sequentially. Each block processing operation performs the following steps: The data to be processed is segmented to obtain multiple segmented data; Multiple segmented data are input into the block processing path to extract information within the segmented data and output the block features. The intra-block features are added to multiple segmented data and input into the inter-block processing path to extract information between segmented data, and the inter-block features are output. The inputs of inter-block features and inter-block processing paths are added together, and the sum is then superimposed before being output.

[0027] As an example, the result after preprocessing Assuming the input data for the first block processing, Indicates the first Output after secondary block processing Then for the first The input to the next block processing operation is: First, the data is segmented using a sliding window method to obtain... ,in, Indicates the segment length. Indicates the number of segments. Segmented data. Enter the intra-block processing path and inter-block processing path respectively, extract the information in the segmented data in the intra-block processing path, and obtain the local dependency relationship. The operation process is shown in the following formula (3): (3) Then, after passing through linear layer 25 and normalization layer 26, the output is combined with the segmented data. Adding them together gives , Enter the inter-block processing path and extract global information in the inter-block processing path. The operation process is shown in the following formula (4): (4) Then, after passing through linear layer 25 and normalization layer 26, the output is combined with the input of the inter-block processing path. Adding them together yields the feature representation. Finally, the feature representation is achieved through an overlay operation. Restored to the original feature dimensions That is, the first The output of the second block processing is The output is retained as part of the input data in the mask decoder 22. Using a bidirectional LSTM24 to extract intra-block and inter-block features from the segmented data allows for better modeling of long sequences.

[0028] Then, a convolutional downsampling operation is performed to reduce the sequence length. For the first... The input data for the next downsampling process is: After downsampling, a new feature representation is obtained. The operation process is shown in equation (5) below: (5) in, , .

[0029] See Figure 2 In order to obtain a better mask representation, a separator 21 is used in the mask network 2 to separate the mask representations of different source signals in advance.

[0030] As an example, see Figure 4The separator 21 includes a first convolutional layer 210, a linear gating unit 211, a second convolutional layer 212, a view layer 213, and a second normalization layer 214 connected sequentially, wherein the first convolutional layer 210 and the second convolutional layer 212 are both one-dimensional convolutional layers; The result of the secondary block processing The first feature is input to separator 21, with the second feature being the first feature. Feature representation of the output of the secondary block processing , To illustrate the changes in data within separator 21, the data entering the first convolutional layer 210 is... After passing through the first convolutional layer 210, the output is The output after passing through the linear gating unit 211 is After passing through the second convolutional layer 212, the output is After passing through the view layer 213 and the second normalization layer 214, the output is a mask feature representation. , which serves as the input to the mask decoder 22.

[0031] See Figure 2 The mask decoder 22 receives the... Each mask feature represents the execution The next block processing operation yields the output of mask decoder 22; As one possible implementation, the above process specifically involves: for the first... After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; let The above operations are performed iteratively until the output of the first block processing is obtained, which is the output of the mask decoder 22.

[0032] As an example, different sources Each mask feature is gradually restored to its original tensor size in the mask decoder 22 through continuous upsampling, tensor concatenation, convolution, and the same block processing operations as in the mask encoder 20. The mask decoder 22 corresponds to the mask encoder 20 and contains... The upsampling, concatenation, and block processing operations are performed; assuming the original input of the mask decoder 22 is... For the first This process Its input is Upsampling yields: After that and After concatenating along the first dimension, we obtain: (6) Then, it is processed in blocks. The block processing operation is the same as that in the mask encoder 20, and will not be described in detail here. The result is the first... Output of the second processing .go through After repeating the same process several times, the output of the first block processing is obtained, which is the final output of the mask decoder 22. .

[0033] See Figure 2 The output of the mask decoder 22 enters the output layer 23, which performs post-processing on it and outputs the mask representation. As one possible implementation, the output layer 23 includes a first path and a second path connected in parallel, an output convolutional layer 230, and an output activation function layer 231. Exemplarily, the first path or the second path is selected based on a gating mechanism. The first path includes a first path convolutional layer 232 and a first path activation function layer 233, and the second path includes a second path convolutional layer 234 and a second path activation function layer 235. Post-processing includes: processing the output of the mask decoder 22 using the first path and the second path respectively to obtain the output of the first path and the output of the second path; performing a multiplicative OR operation on the output of the first path and the output of the second path, and inputting the result into the output convolutional layer 230 and the output activation function layer 231 to process and output the mask representation.

[0034] As an example, the first path activation function layer 233 is The activation function, the second path activation function layer 235 is... Activation function, output activation function layer 231 is Activation function. Final output of mask decoder 22. Upon entering the output layer 23, the output layer 23 first filters information based on a gating mechanism to further improve the feature representation capability of this system. The operation of filtering information based on the gating mechanism is shown in equation (7): (7) Finally, it undergoes one more convolution and... The activation function yields a mask representation, as shown in equation (8): (8) Next, the high-dimensional feature representations are stacked, and the mask representation is multiplied with the stacked high-dimensional feature representations to obtain the intermediate feature representation of each source signal in the mixed signal. As an example, for high-dimensional feature representation Stacking process to obtain ,in Assume the mask representation obtained after passing through mask network 2 is as follows: ,in, Multiplying them together, we get: (9) in, , Indicates the first The intermediate feature representation of a source signal.

[0035] See Figure 1 Decoder 3 receives the intermediate feature representation of each source signal and converts it into the original feature representation to obtain the separated signal.

[0036] As an example, decoder 3 is internally implemented using a one-dimensional transposed convolution, therefore, ,in, Finally, all the separated signals are spliced ​​together to obtain... , .

[0037] See Figure 1 and Figure 5 As one possible implementation, the deep learning-based mixed-signal dual-path blind source separation system also includes a conditional denoising module 4, used to remove noise from the original mixed signal before inputting it to the encoder 1. The conditional denoising module 4 includes a denoising encoder 40 and a denoising decoder 41. Starting from the input, the denoising encoder 40 includes three cross-connected one-dimensional convolutional layers and two activation functions, and the denoising decoder 41 includes three cross-connected one-dimensional deconvolutional layers and two activation functions, both of which are... Activation function.

[0038] As an example, starting from the input, the parameters of the three one-dimensional convolutional layers are as follows: the first one-dimensional convolutional layer has a kernel size of 5, a stride of 2, a padding size of 2, and 32 output channels; the second one-dimensional convolutional layer has a kernel size of 5, a stride of 2, a padding size of 2, and 64 output channels; the third one-dimensional convolutional layer has a kernel size of 3, a stride of 1, a padding size of 1, and 128 output channels. Similarly, the parameters of the three one-dimensional deconvolutional layers are as follows: the first one-dimensional deconvolutional layer has a kernel size of 3, a stride of 2, a padding size of 1, an output padding size of 1, and 64 output channels; the second one-dimensional deconvolutional layer has a kernel size of 5, a stride of 2, a padding size of 2, an output padding size of 1, and 32 output channels; the third one-dimensional deconvolutional layer has a kernel size of 5, a stride of 2, a padding size of 2, an output padding size of 1, and 2 output channels.

[0039] Conditional denoising module 4 performs the following operations to remove noise from the original mixed signal: The denoising encoder 40 receives the original mixed signal, performs multiple downsampling and expands the feature dimension, and then inputs it to the denoising decoder 41. The denoising decoder 41 performs multiple upsampling operations and compresses the feature dimension before outputting a noise estimate. The noise estimate is residually concatenated with the original mixed data to output the denoised mixed signal.

[0040] As an example, suppose the original mixed signal is The denoising encoder 40 extracts the original mixed signal by continuously downsampling and expanding the feature dimension. The high-dimensional feature representation is then used by the denoising decoder 41 to recover the original mixed signal features using one-dimensional deconvolution, thus restoring the original mixed signal. The noise estimate is obtained after passing through the denoising encoder 40 and the denoising decoder 41: (10) in, Finally, after residual connection, the denoised mixed signal is obtained: .

[0041] Secondly, embodiments of the present invention provide a method for separating dual-path blind sources in mixed signals based on deep learning, see [link to relevant documentation]. Figure 6 ,include: S1. Configure the mixed signal; As an example, the mixed signal is a noise-free mixed signal. If the mixed signal contains noise, noise reduction processing is performed first to obtain a noise-free mixed signal before step S2 is executed.

[0042] S2. Extract the features of the mixed signal and transform them into a high-dimensional feature representation; S3. Perform preprocessing and continuous processing on the high-dimensional feature representations sequentially. Each block processing and downsampling operation outputs the feature representation obtained after each block processing step. The second block processing yielded a total of Each feature is represented; As one possible implementation, S3 includes: S30. Segment the data to be processed to obtain multiple segmented data; S31. Extract information from multiple segmented data using the bidirectional LSTM method to obtain intra-block features; S32. Based on the bidirectional LSTM method, extract intra-block features and inter-segment data information by summing multiple segmented data to obtain inter-block features; S33. Perform an overlay operation on the summation of inter-block features and intra-block features with multiple segmented data; S4. Will These features indicate that the data were separated into different sources. Each mask feature is represented; S5. To Each mask feature is represented Secondary block processing operations, for The output obtained after the next block processing operation is post-processed to obtain a mask representation; As one possible implementation, S5 includes: S50. Regarding After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; S51. Order Repeat S50 until the output of the first block processing is obtained; S52. Post-process the output of the first block to obtain a mask representation.

[0043] As one possible implementation, S52 includes: S520. The output of the first block processing is downsampled and activated by the Tanh function to obtain the first output feature representation; the output of the first block processing is downsampled and activated by the Sigmoid function to obtain the second output feature representation; S521. Perform an XOR operation on the first output feature representation and the second output feature representation to obtain the operation result; S522. The operation result is downsampled and activated by the ReLU function to obtain a mask representation.

[0044] S6. Stack the high-dimensional feature representations, multiply the mask representation with the stacked high-dimensional feature representations, and obtain the intermediate feature representation of each source signal in the mixed signal; S7. Convert the intermediate feature representation of each source signal in the mixed signal into the original feature representation to obtain the separated signal.

[0045] The following experiments further illustrate the effectiveness of the deep learning-based dual-path blind source separation system and method for mixed signals proposed in this embodiment. The experimental procedure is as follows: In single-channel blind source separation, the received mixed signal can be represented as: (11) To facilitate the evaluation of the separation effect, five modulation types—BPSK, QPSK, 8PSK, 16-QAM, and 64-QAM—were generated as source signals using MATLAB. 10,000 data points were generated for each signal type, with a sampling rate of 1MHz. The signal length was standardized to 1024, resulting in a final signal size of 1024×2. The final experimental data were obtained by mixing these source signals.

[0046] In the experiment, it is assumed that the mixed signal is formed by mixing two source signals. In the specific mixing process, the mixing factor is randomly generated. Since the generated signal is a complex signal, including real and imaginary data, the specific mixing process is as follows (12): (12) Different types of signals were mixed during the mixing process, resulting in ten types of mixed signals: BPSK-QPSK, BPSK-8PSK, BPSK-16QAM, BPSK-64QAM, QPSK-8PSK, QPSK-16QAM, QPSK-64QAM, 8PSK-16QAM, 8PSK-64QAM, and 16QAM-64QAM. A total of 100,000 mixed data points were generated, with 80% used for training, 10% for validation, and 10% for testing. To compare the model performance under different signal-to-noise ratios (SNRs), five SNR scenarios were set, ranging from SNR = 0dB to SNR = 20dB, with a step size of 5dB.

[0047] All experiments were implemented using PyTorch and run on an NVIDIA GeForce RTX 4090 GPU. Early stopping was used during training, and the patience was set to 10. Mean squared error (MSE) was used as the loss function during training to accurately recover the source signal. In the model part, the encoder output dimension C = 256, and the number of layers in the mask network N = 6. To evaluate the separation effect, the separated signal was decoded, and the probability of the number of separation error bits was calculated as the evaluation metric, as shown in Equation (13): (13) in, This represents the number of bits contained in a signal. This indicates the number of erroneous bits. Fewer erroneous bits indicate higher quality of the separated signal and better separation performance of the model.

[0048] The experiment consisted of two parts. The first part involved signal separation to separate the mixed signal, and the second part involved identifying the type of the separated signal. Three sets of experiments were conducted in the signal separation experiment. The first set compared the performance of different models under noise-free conditions to verify the effectiveness of the proposed model, comparing two different models: one for speech separation and one for wireless signal separation. The second set verified the results of different models under different signal-to-noise ratios to evaluate the noise resistance of each model. The third set of experiments included ablation experiments to verify the effectiveness of the conditional denoising module.

[0049] The first set of experiments compared the performance of different models under noise-free conditions to verify the effectiveness of the model proposed in this invention. The experimental results are shown in Table 1: Table 1 Comparison of Separation Error Probabilities of Different Models As can be seen from Table 1, the model proposed in this invention has a significant improvement in separation performance compared to other models, with the separation error probability decreasing from... The order of magnitude decreased to The order of magnitude.

[0050] To further analyze the experimental results, Table 2 shows the separation error probability results for different models under different mixing types: Table 2 Comparison of error probabilities for different mixture types in different models As shown in Table 2, under noise-free conditions, the separation performance is relatively good for most mixed signal types, with the best separation performance for PSK-PSK mixed signals. The separation error probability for all three mixed signals is less than [value missing]. These technologies effectively separate data. Among PSK-QAM hybrid types, BPSK-16QAM, QPSK-16QAM, and QPSK-64QAM hybrid types also show relatively good separation performance. The magnitude is higher than that of the others, which are relatively worse. The order of magnitude is around 16QAM-64QAM. The worst performing type is the mixed type, with a separation error probability of around 100%. The order of magnitude is around 1 / 2.

[0051] To more intuitively demonstrate the separation effect of the model, Figure 7The figure shows the first 100 points of the source signal and the separated signal for ten different mixing types. As can be seen from the figure, the source signal and the separated signal largely overlap, indicating relatively good separation. Among all types, the separated signal at the beginning differs significantly from the source signal. Furthermore, comparing the ten mixing types, the 16QAM-64QAM mixing type shows a significantly larger deviation from the source signal.

[0052] Next, we compare the experimental results of different models under different signal-to-noise ratios to verify the impact of noise on the models. The experimental results are shown in Table 3. Table 3 Separation error probabilities of different models under different signal-to-noise ratios As shown in Table 3, the separation error probability decreases continuously with the increase of the signal-to-noise ratio. Compared with other models, the separation error probability of this invention is lower, indicating that the model has stronger noise resistance.

[0053] See Figure 8 The results show the impact of signal-to-noise ratio on different hybrid types. Figure 8 (a) Statistics on the separation of PSK-QAM and QAM-QAM hybrid types were presented. Figure 8 (b) presents statistics on the separation of PSK-PSK hybrid types. Figure 8 As can be seen, the PSK-PSK hybrid type performs better in separation than the PSK-QAM hybrid and QAM-QAM hybrid types. Specifically, the PSK-QAM and QAM-QAM hybrid types have relatively high overall separation error probabilities; at SNR = 20 dB, the separation error probability of most hybrid types is still greater than 0.005. The separation error probability of the PSK-PSK hybrid type changes more significantly compared to the PSK-QAM and QAM-QAM hybrid types. BPSK-QPSK performs best in separation, with a separation error probability of less than 0.005 at SNR = 20 dB. .

[0054] To verify the effectiveness of the conditional denoising module in the system, ablation experiments were conducted under different signal-to-noise ratios. The experimental results are shown in Table 4. Table 4 Ablation experimental results under different signal-to-noise ratios As shown in the table, the average separation error probability decreased from SNR = 0 dB to SNR = 20 dB, with a decrease of 0.0042 at SNR = 20 dB. This result indicates that the denoising module improves the model's robustness to noise and enhances its separation performance in noisy environments.

[0055] Table 5 below shows the separation error probability and average error probability for different mixing types at SNR = 0 dB: Table 5 Ablation experimental results under different hybrid types when SNR = 0 dB As can be seen from Table 5, the separation effect of most hybrid types was improved after adding the noise reduction module. Only the separation error probability of QPSK-16QAM, QPSK-64QAM and 16QAM-64QAM was improved. However, overall, the separation error probability decreased by 0.0008.

[0056] Table 6 below shows the different mixing types and the average separation error probability at SNR = 20 dB: Table 6 Ablation experimental results for different hybrid types at SNR = 20 dB As shown in Table 6, under all hybrid types, the separation error probability with the denoising module added was lower than that without the denoising module. Overall, the average separation error probability decreased by 0.0042. This indicates that the denoising module improved the model's noise resistance.

[0057] Comparing the two scenarios, the denoising module performs more stably under high signal-to-noise ratio (SNR), with a reduced separation error probability in all cases. Under low SNR, its performance is less stable, and the error probability increases in some situations. However, overall, the average separation error probability decreases in both scenarios, validating the denoising module's noise reduction effect.

[0058] In this embodiment, the first set of experiments compared the model separation performance under noise-free conditions. The experimental results show that the separation error probability is significantly reduced under noise-free conditions, while the separation error probability of other models is approximately [missing information]. The order of magnitude of left and right, this invention has decreased to The magnitude of the error demonstrates the effectiveness of this invention. Further analysis reveals that the separation effect is best for PSK-PSK hybrid models, relatively poor for PSK-QAM, and worst for QAM-QAM hybrid models, with a separation error probability of 0.0053. Overall, it can be seen that the separation error probability increases with the modulation order. The second set of experiments compared the model separation performance under noisy conditions, considering five signal-to-noise ratio (SNR) levels. Compared to the other four models, this invention achieved the best performance at all five SNR levels. At SNR = 0dB, the separation error probability decreased by 0.0003 compared to the best-performing model among the other four models; at SNR = 20dB, the separation error probability decreased by 0.0014. Overall, this invention exhibits better noise immunity compared to other models. Furthermore, the performance of different hybrid types was analyzed. The PSK-PSK hybrid signal performed better than other hybrid types, with the BPSK-QPSK hybrid type showing the best performance. At an SNR of 20 dB, its separation error probability was less than [value missing]. The worst performing type was the 16QAM-64QAM hybrid, with a separation error probability of 0.3183 at SNR = 20 dB. The third set of experiments conducted ablation experiments on the denoising module under five signal-to-noise ratios. The results showed that the model with the denoising module achieved better separation performance at all five SNRs, with the separation error probability decreasing by 0.0042 at SNR = 20 dB. Furthermore, the classification error probabilities of different hybrid types were compared at SNR = 0 dB and SNR = 20 dB. The results indicated that the denoising module's performance was more stable at higher SNRs, and the separation error probability decreased for all hybrid types.

[0059] Two sets of experiments were then conducted to identify the types of separated signals and further verify the generalization ability of the invention. The first set of experiments identified separated signals under noise-free conditions; the second set of experiments compared the recognition performance of separated signals under different signal-to-noise ratios.

[0060] See Figure 9 The confusion matrix for signal identification under noise-free conditions is shown. The final accuracy is 99.965%, and the macro-precision is also 99.965%. It can be seen that in the absence of Gaussian white noise, this invention performs very well in identifying and separating signals, with only some confusion in the separation of 16-QAM and 64-QAM signals. The classification accuracy for other signal types reaches 100%.

[0061] The identification results of the separated signals under different signal-to-noise ratios are shown in Table 7: Table 7. Identification results of separated signals under different signal-to-noise ratios As shown in Table 7, the classification effect improves with the increase of signal-to-noise ratio. When SNR = 0 dB, the effect is relatively poor, only 66.85%. When SNR = 5 dB, the classification accuracy is significantly improved, reaching 81.68%. When SNR = 15 dB, the classification accuracy exceeds 99%.

[0062] Figure 10 The confusion matrix represents the classification results under five signal-to-noise ratio (SNR) conditions. Specifically, the BPSK signal exhibits the best recognition performance, maintaining excellent separation even at SNR = 0 dB, with a recall rate of 99%. Other signal types perform relatively poorly. When SNR = 15 dB, the overall classification performance begins to improve significantly, with only a small number of signal recognition errors.

[0063] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0064] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A deep learning based hybrid signal dual-path blind source separation system, characterized in that, At least comprising: an encoder, a mask network, and a decoder; wherein the encoder is configured to extract features of a received mixed signal and convert the features into a high-dimensional feature representation; the mask network comprises a mask encoder, a separator, a mask decoder, and an output layer; The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, The mask encoder sequentially performs pre-processing, successive block processing and down-sampling, block processing operations on the high-dimensional feature representation, and outputs a feature representation obtained after each block processing, stacking the high-dimensional feature representation, multiplying the mask representation by the stacked high-dimensional feature representation to obtain an intermediate feature representation of each source signal in the mixed signal; the decoder receives the intermediate feature representation of each source signal and converts it into an original feature representation to obtain a separated signal.

2. The deep learning based mixed signal dual-path blind source separation system of claim 1, wherein, The block processing is implemented by a double path, which includes an intra-block processing path and an inter-block processing path; both the intra-block processing path and the inter-block processing path comprise a bidirectional LSTM, a linear layer, and a normalization layer connected in sequence; each block processing performs the following operations: segmenting the data to be processed to obtain a plurality of segmented data; inputting the plurality of segmented data into the intra-block processing path to extract intra-segmented data information, and outputting intra-block features; inputting the intra-block features and the plurality of segmented data into the inter-block processing path to extract inter-segmented data information, and outputting inter-block features; adding the inter-block features and the input of the inter-block processing path, and outputting after superposition operation on the addition result.

3. The deep learning based mixed signal dual-path blind source separation system of claim 1, wherein, The mask decoder receives the data. Each mask feature represents the execution The next block processing operation obtains the output of the mask decoder, specifically: for the first block... After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; let The above operations are performed iteratively until the output of the first block processing is obtained, which is the output of the mask decoder.

4. The deep learning based mixed signal dual-path blind source separation system of claim 1, wherein, The output layer comprises a first path and a second path connected in parallel, an output convolution layer, and an output activation function layer, and the first path or the second path is selected based on a gating mechanism; the first path comprises a first path convolution layer and a first path activation function, and the second path comprises a second path convolution layer and a second path activation function; The post-processing comprises: the first path and the second path process the output of the mask decoder respectively to obtain the output of the first path and the output of the second path; performing an exclusive OR operation on the output of the first path and the output of the second path, and inputting the operation result into the output convolution layer and the output activation function layer for processing to output the mask representation.

5. The deep learning based mixed signal dual-path blind source separation system of claim 1, wherein, The deep learning-based mixed signal double-path blind source separation system further comprises a conditional denoising module for removing noise in the original mixed signal and inputting the original mixed signal into the encoder; the conditional denoising module comprises a denoising encoder and a denoising decoder; The conditional denoising module performs the following operations to remove noise in the original mixed signal: The denoising encoder receives the original mixed signal, performs multiple downsampling and expands the feature dimension, and then inputs the result into the denoising decoder; The denoising decoder performs multiple upsampling and compresses the feature dimension, and then outputs a noise estimate; The noise estimate and the original mixed signal are connected in residual to output a denoised mixed signal.

6. The deep learning based mixed signal dual-path blind source separation system of claim 5, wherein, From the input end, the denoising encoder comprises three one-dimensional convolution layers and two activation functions connected in cross, and the denoising decoder comprises three one-dimensional deconvolution layers and two activation functions connected in cross, and the activation functions are ReLU activation functions.

7. A deep learning based hybrid signal dual-path blind source separation method, characterized in that, Comprising: S1. configuring a mixed signal; S2. extracting features of the mixed signal and converting the features into a high-dimensional feature representation; S3. sequentially pre-process, successively block process and down-sample, block process operation on the high-dimensional feature representation, output the feature representation obtained after each block process; a total offeature representations are obtained;​ S4. separating the feature representation into different source mask feature representations;​ S5. Regarding the aforementioned Each mask feature is represented Secondary block processing operations, for The output obtained after the next block processing operation is post-processed to obtain a mask representation; S6. stacking the high-dimensional feature representation, multiplying the mask representation by the stacked high-dimensional feature representation to obtain an intermediate feature representation of each source signal in the mixed signal; S7. Convert the intermediate feature representation of each source signal in the mixed signal to the original feature representation to obtain a separated signal.

8. The method of claim 7, wherein the deep learning-based mixed signal dual-path blind source separation method is characterized by, The block processing includes: S30. Segment the data to be processed to obtain a plurality of segmented data; S31. Extract information within the plurality of segmented data based on a bidirectional LSTM method to obtain intra-block features; S32. Extract information between the intra-block features and the sum of the plurality of segmented data based on a bidirectional LSTM method to obtain inter-block features; S33. Perform a superposition operation on the sum of the inter-block features, the intra-block features, and the plurality of segmented data.

9. The method of claim 7, wherein the deep learning-based mixed signal dual-path blind source separation method is characterized by, The S5 includes: S50. Regarding After upsampling, the mask feature representation is compared with the first... The mask feature representations are concatenated along the first dimension using convolution. The concatenated feature representations are then processed into blocks to obtain the nth... The output of the next block processing; S51. Let S50 is repeatedly executed until the output of the 1st block processing is obtained. S52. Post-process the output of the first block processing to obtain a mask representation.

10. The method of claim 9, wherein the deep learning-based mixed signal dual-path blind source separation method is characterized by, The S52 includes: S520. Downsample and Tanh function activation process the output of the first block processing to obtain a first output feature representation; downsample and Sigmoid function activation process the output of the first block processing to obtain a second output feature representation; S521. Perform an exclusive OR operation on the first output feature representation and the second output feature representation to obtain an operation result; S522. Downsample and ReLU function activation process the operation result to obtain a mask representation.

Images