Mixed signal double-path blind source separation system, method and terminal

Abstract

The invention belongs to the technical field of signal processing, and relates to a mixed signal double-path blind source separation system and method, and a terminal. The blind source separation system comprises an encoder, a mask network and a decoder, the encoder receives a mixed signal, extracts features and then converts the features into high-dimensional feature representation, the mask network comprises a decomposition module and DPGCU blocks which are sequentially connected, each DPGCU block is provided with a dual-path LSTM structure and comprises a GCU block, and each GCU block comprises a GCU block; the decomposition module segments the high-dimensional features and inputs the segmented high-dimensional features to the first DPGCU block, a first path LSTM structure extracts local features of a plurality of segmented data and performs residual connection, and outputs local feature representation; the second path LSTM structure extracts global features represented by the local features and performs residual connection, and outputs global feature representation; the GCU block extracts global features of the global feature representation and outputs enhanced feature representation. And the output of each DPGCU block is used as the input of the next DPGCU block. The method has the advantages of strong signal separation capability, high separation precision and the like.

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, method and terminal. Background Technology

[0002] With the rapid increase in devices and traffic in wireless networks, the demand for spectrum resources is also constantly increasing. However, under the current static spectrum allocation strategy, spectrum resources are increasingly unable to meet the demand. Therefore, cognitive radio technology has been proposed to improve the utilization rate of spectrum resources. In cognitive radio, spectrum sensing is a crucial step. Cognitive users need to continuously sense their surrounding environment to opportunistically use the spectrum without interfering with the primary user, thereby improving spectrum resource utilization. Meanwhile, with the continuous development of blind source separation technology, researchers have begun to focus on its application potential in the field of spectrum sensing, such as using BSS to improve detection accuracy and solve problems such as the inability to communicate during sensing.

[0003] Single-channel blind source separation is an important part of blind source separation tasks. It refers to the process of separating multiple source signals from a single observed signal, where the number of observed signals is less than the number of source signals, making it an underdetermined blind source separation problem. Research methods for single-channel blind source separation can be broadly divided into traditional methods and deep learning-based methods. Traditional methods solve the problem by pre-establishing mathematical models, but they suffer from low separation accuracy and uncertain separation sequences. In recent years, with the continuous development of deep learning, some deep learning-based methods have been proposed to solve the single-channel blind source separation problem. However, current research on deep learning-based blind source separation is mostly concentrated in the field of speech separation, with relatively little research on wireless communication signal separation.

[0004] Therefore, there is an urgent need to propose a wireless communication signal separation method with high blind source separation accuracy and high separation sequence stability. Summary of the Invention

[0005] The purpose of this invention is to address the problems of low separation accuracy, uncertain separation sequences, and limited research on wireless communication signal separation in existing blind source separation technologies. This invention proposes a hybrid signal dual-path blind source separation system, method, and terminal. A DPGCU block with a dual-path structure is used to process intra-segment and inter-segment data. The GCU block is used to enhance the feature extraction capability of the system, which has the advantages of strong signal separation capability and high 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, comprising: An encoder receives a mixed signal and extracts its features, then transforms the extracted features into a high-dimensional feature representation. A masked network consists of decomposition modules and sequentially connected components. Each DPGCU block has a dual-path LSTM structure and includes a GCU block. The decomposition module is used to segment the received high-dimensional features and input the resulting segmented data into the first DPGCU block; The DPGCU block includes a first-path LSTM structure that extracts local features from multiple segmented data and performs residual connections to output local feature representations; a second-path LSTM structure that extracts global features from the local feature representations and performs residual connections to output global feature representations; and a GCU block that extracts global features from the global feature representations to output enhanced feature representations. No. The output of the DPGCU block is used as the first Input of each DPGCU block ,go through The separated signal characteristics are obtained after processing by each DPGCU block; And a decoder, used to convert the separated signal features processed by the masking network into the original feature representation.

[0007] As one possible implementation, the output of the first path LSTM structure is processed by the second path LSTM structure and then added to the output of the first path LSTM structure; The output of the second-path LSTM structure is processed by the GCU block and then added to the output of the second-path LSTM structure.

[0008] As one possible implementation, the GCU block includes a first path and a second convolutional layer connected in parallel. The first path includes a first convolutional layer and an inflated feedback shift memory network connected in sequence. The output of the first path performs an XOR operation with the output of the second convolutional layer.

[0009] As one possible implementation, the first and second convolutional layers have the same structure, both including a normalized layer, a first linear layer, a gated linear unit layer and a depthwise convolutional layer connected in sequence. The normalization layer is used to normalize the received data. The first linear layer is used to perform feature transformation on the normalized data; Gated linear unit layers are used to control the feature extraction process of deep convolutional layers; Deep convolutional layers are used to extract fine-grained local features.

[0010] As one possible implementation, the dilated feedback shift memory network includes a second linear layer and a memory layer connected sequentially; the second linear layer receives the data output from the first convolutional layer and performs linear processing on it; the linearly processed data is then processed by the memory layer and residually connected with the linearly processed data.

[0011] As one possible implementation, the memory layer includes A stacked, interconnected two-dimensional convolutional layer, the first... The number of expansion factors in a two-dimensional convolutional layer is , .

[0012] As one possible implementation method, The two-dimensional convolutional layers are densely connected, and each two-dimensional convolutional layer uses the same output channel. , No. The output of each 2D convolutional layer is ,in, Indicates the original input. This indicates mapping processing.

[0013] Secondly, the present invention provides a method for separating blind sources in a mixed signal dual-path system, comprising: S1. Configure the mixed signal, extract the features of the mixed signal, and transform it into a high-dimensional feature representation; S2. Segment the high-dimensional feature representation to obtain multiple segmented data; S3. Extract local features from multiple segmented data and perform residual connections to obtain local feature representations; S4. Extract the global features from the local feature representations and perform residual connections to obtain the global feature representations; S5. Extract global features from the global feature representation based on the gated convolution method to obtain the enhanced feature representation; S6. Determine whether the iteration termination condition has been met. If yes, stop the iteration and obtain the separated signal characteristics; otherwise, proceed to step S3. S7. Convert the separated signal features into the original feature representation.

[0014] As one possible implementation, the iteration termination condition is that the result obtained after 10 consecutive iterations remains unchanged.

[0015] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the mixed-signal dual-path blind source separation method provided in the second aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The mixed-signal dual-path blind source separation system proposed in this invention uses... A DPGCU block with a dual-path structure is used to process intra-segment and inter-segment data. Then, a GCU block is used to enhance the feature extraction capability of the system. This can effectively solve the problems of low separation accuracy, uncertain separation sequence, and limited research on wireless communication signal separation in existing blind source separation technologies.

[0017] 2. The hybrid signal dual-path blind source separation system proposed in this invention segments the hybrid data and uses a dual-path structure to act on the intra-block and inter-block data of the segmented data respectively, which can more fully extract the local and global features of the segmented data and help to obtain more accurate separation results.

[0018] 3. This invention proposes a hybrid signal dual-path blind source separation system. A GCU block is introduced after the DPGCU block with a dual-path structure to further extract features from the integrated local and global feature information. This system can capture finer-grained local features. Experiments show that this significantly reduces the separation error rate and improves the quality of signal separation. 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 hybrid signal dual-path blind source separation system proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the DPGCU block structure in the mixed-signal dual-path blind source separation system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the GCU block included in the DPGCU block in an embodiment of the present invention; Figure 4 This is a schematic diagram of the convolutional layer structure included in the GCU block in an embodiment of the present invention; Figure 5 This is a schematic diagram of the expansion feedback shift memory network structure included in the GCU block in an embodiment of the present invention; Figure 6 This is a flowchart of the mixed-signal dual-path blind source separation method proposed in an embodiment of the present invention; Figure 7 This is a comparison chart of the separation performance of the system of the present invention and the StackLSTM model for 16QAM-64QAM mixed types in an embodiment of the present invention; Figure 8 and Figure 9 This is a comparison chart showing the effects of signal-to-noise ratio on different hybrid types in embodiments of the present invention. Detailed Implementation

[0020] 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.

[0021] 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.

[0022] 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.

[0023] The embodiments of the present invention aim to provide a mixed-signal dual-path blind source separation system, method, and terminal, using... A DPGCU block with a dual-path structure is used to process intra-segment and inter-segment data. The GCU block is used to enhance the feature extraction capability of the system, so as to solve the problems of low separation accuracy, uncertain separation sequence and limited research on wireless communication signal separation in existing blind source separation technology.

[0024] In a first aspect, embodiments of the present invention provide a mixed-signal dual-path blind source separation system, see [link to previous document]. Figure 1 It includes: encoder 1, mask network 2 and decoder 3; In this process, encoder 1 receives the mixed signal and extracts its features, transforming the extracted features into a high-dimensional feature representation that is easy to process. Masking network 2 is used to learn the masking representation and separate the signal features from different sources. Then, the learned masking representation is multiplied by the high-dimensional feature representation obtained by the encoder to retain the feature information of the specific signal source. Finally, the masked features are converted into the original feature representation by decoder 3.

[0025] As an example, the mixed signal is a complex signal, i.e., IQ data. Single-channel blind source separation tasks are typically modeled as follows: (1) in, Indicates the input mixed signal, , Indicates the length of the mixed signal. Represents the mixing coefficient. Indicates the source signal, Indicates the number of source signals. This represents Gaussian white noise. Represents the set of complex numbers. This represents each sampling point.

[0026] For ease of understanding, this embodiment will be described using two scenarios, namely... and fix Assume the input mixed signal is Mixed signals First, the signal is processed by encoder 1, which uses a one-dimensional convolutional structure to extract signal features and transform them into high-dimensional feature representations. ,in, ,in This indicates the sequence length of the data after convolution. This represents the dimension of the encoded data features, and is set accordingly. The encoded high-dimensional feature representation is then fed into the mask network 2.

[0027] See Figure 1 As one possible implementation, the masking network 2 includes a decomposition module 20 and sequentially connected components. Each DPGCU block 21 has a dual-path LSTM structure and includes a GCU block 210; the decomposition module 20 is used to segment the received high-dimensional features and input the resulting segmented data into the first DPGCU block 21.

[0028] As an example, the decomposition module 20 uses a sliding window approach to segment the received high-dimensional features, assuming the segment length is... The sliding step size is The final segmented data ,in, Indicates the number of segments. The data after segmentation. Enter the first DPGCU block 21.

[0029] See Figure 1 Each DPGCU block 21 uses a dual-path structure to process intra-segment and inter-segment data respectively, in order to extract local and global dependencies of the data. The first path LSTM structure 211 included in each DPGCU block 21 extracts local features of multiple segment data and performs residual connections to output local feature representations; the second path LSTM structure 212 extracts global features from the local feature representations and performs residual connections to output global feature representations; the GCU block 210 extracts global features from the global feature representations and outputs enhanced feature representations.

[0030] As one possible implementation, the output of the first path LSTM structure 211 is processed by the second path LSTM structure 212 and then added to the output of the first path LSTM structure 211; the output of the second path LSTM structure 212 is processed by the GCU block 210 and then added to the output of the second path LSTM structure 212.

[0031] See Figures 1 to 2 As an example, for the first 21 DPGCU blocks, whose inputs are First, process the intra-block features of each data segment to extract local information. , The first path LSTM structure 211 is used to extract local features and perform residual connections to obtain the local feature representation. As shown in the following formula: (2) Then, extract the features between each data segment, i.e., the global features of the data, and input them. , The global feature representation is obtained by using the second-path LSTM structure 212 to extract global features and perform residual connections. As shown in the following formula: (3) Finally, to further enhance the feature extraction capability of this system, the global features of the global feature representation are extracted using GCU block 210, thereby further extracting overall feature information, strengthening the feature representation capability, and obtaining an enhanced feature representation. As shown in the following formula: (4) See Figure 3 As one possible implementation, the GCU block 210 includes a first path and a second convolutional layer 2100 connected in parallel. The first path includes a first convolutional layer 2101 and an inflated feedback shift memory network 2102 connected in sequence. The output of the first path performs an XOR operation with the output of the second convolutional layer 2100.

[0032] As an example, in order to extract the overall sequence features of the signal, GCU block 210 first transforms the data into one-dimensional data through reshaping, obtaining... ,in, The overall processing of GCU block 210 is shown in the following formula: (5) See Figure 4 As one possible implementation, the first convolutional layer 2101 and the second convolutional layer 2100 have the same structure, both including a normalization layer 21000, a first linear layer 21001, a gated linear unit layer 21002, and a deep convolutional layer 21003 connected in sequence. The normalization layer 21000 is used to normalize the received data; the first linear layer 21001 is used to perform feature transformation on the normalized data; the gated linear unit layer 21002 is used to control the deep convolutional layer to extract features; and the deep convolutional layer 21003 is used to extract fine-grained local features. This embodiment utilizes a deep convolutional structure to capture fine-grained local features. After data is processed by deep convolution, it obtains... , This embodiment uses a normalization layer 21000 to normalize the input data, making the data distribution more stable and reducing the risk of gradient explosion or vanishing. The first linear layer 21001 mainly performs feature transformation to improve the expressive power of the system. The gated linear unit layer 21002 can use the SiLU activation function to increase the nonlinearity of the system. Finally, a deep convolutional layer 21003 is used to perform deep convolution to capture local temporal features. The deep convolutional layer can focus more on feature extraction and recognition of local segments, and more intuitively capture short-term patterns.

[0033] See Figure 5 As one possible implementation, the dilated feedback shift memory network 2102 includes a second linear layer 21020 and a memory layer 21021 connected sequentially. The second linear layer 21020 receives the data output by the first convolutional layer 2101 and performs linear processing on it. The linearly processed data is then processed by the memory layer 21021 and residually connected with the linearly processed data.

[0034] See Figure 5 As one possible implementation, the memory layer 21021 includes A stacked, interconnected two-dimensional convolutional layer 210210, the first The number of expansion factors in a two-dimensional convolutional layer is , . The two-dimensional convolutional layers 210 and 210 are densely connected, and each two-dimensional convolutional layer 210 and 210 uses the same output channel. , No. The output of each 2D convolutional layer is ,in, Indicates the original input. This indicates mapping processing.

[0035] As an example, data processed by depthwise convolution After processing by the second linear layer 21020, the following is obtained , Represents the original input, for the ... A two-dimensional convolutional layer, the output of which is Dense connections can enhance the memory capacity of the expanded feedback shift memory network 2102, ultimately achieving the... The output of the layer is passed through residual connections to obtain the output of the dilated feedback shift memory network 2102: (6) No. The output of the DPGCU block is used as the first Input of each DPGCU block ,go through The separated signal is obtained after processing by each DPGCU block. For example, The range of values ​​for is not specifically limited. That's all, usually The larger the value of , the stronger the system's modeling ability, but it will also increase the amount of computation accordingly. The specific application depends on the actual situation.

[0036] The decoder converts the separated signal features processed by the masking network into the original feature representation.

[0037] Secondly, embodiments of the present invention provide a method for separating blind sources in a mixed-signal dual-path system, see [link to relevant documentation]. Figure 6 ,include: S1. Configure the mixed signal, extract the features of the mixed signal, and transform it into a high-dimensional feature representation; S2. Segment the high-dimensional feature representation to obtain multiple segmented data; S3. Extract local features from multiple segmented data and perform residual connections to obtain local feature representations; S4. Extract the global features from the local feature representations and perform residual connections to obtain the global feature representations; S5. Extract global features from the global feature representation based on the gated convolution method to obtain enhanced feature representation; S6. Determine whether the iteration termination condition has been met. If yes, stop the iteration and obtain the separated signal characteristics; otherwise, proceed to step S3. S7. Convert the separated signal features into the original feature representation.

[0038] As one possible implementation, the iteration termination condition is that the result obtained after 10 consecutive iterations remains unchanged.

[0039] Next, a specific experiment was conducted using the hybrid signal dual-path blind source separation system and method provided in this embodiment. The source data used in the experiment included five types of modulation signals: BPSK, QPSK, 8PSK, 16QAM, and 64QAM. Matlab was used to generate the five types of signals, with 10,000 data points for each type and 1,024 sampling points for each data point. The source signals were mixed between each pair of types to generate a total of ten hybrid signals (BPSK-QPSK, BPSK-8PSK, BPSK-16QAM, BPSK-64QAM, QPSK-8PSK, QPSK-16QAM, QPSK-64QAM, 8PSK-16QAM, 8PSK-64QAM, 16QAM-64QAM), and the hybrid signals were generated according to formula (1).

[0040] To compare the separation performance of this system under different signal-to-noise ratios (SNRs), five SNR scenarios were set, with SNR ranging from 0dB to 20dB and a step size of 5dB. The final dataset contains six sets of data: one set of noise-free mixed data, five sets of mixed data under different SNRs, each set containing 100,000 data points, for a total of 600,000 data points.

[0041] In the experiment, training, validation, and test sets were set up, accounting for 80%, 10%, and 10% of the total data, respectively. All experiments were implemented using PyTorch and run on an NVIDIA GeForce RTX 4090 GPU. An early stopping mechanism 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.

[0042] To evaluate the effectiveness of the separation, the probability of the number of separation error bits is calculated as the evaluation metric for the decoded signal, as shown in the following formula: (7) 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 system.

[0043] Based on the above settings, three sets of experiments were conducted. The first set of experiments compared the performance of different models under noise-free conditions to verify the effectiveness of the system proposed in this application. The best-performing models were selected from both the speech signal separation model and the wireless signal separation model for comparison. The second set of experiments 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 conducted ablation experiments to verify the effectiveness of the GCU block.

[0044] The results of the first group of experiments are shown in Table 1. Table 1 Comparison results of the first group of experiments As shown in Table 1, the system proposed in this application has the lowest separation error probability, which is reduced by 0.0015 compared to other models, achieving the best separation performance. It can also be seen that its average separation error probability is relatively high. To further analyze the experimental results, Table 2 below shows the separation error probability results under different hybrid types: Table 2 Separation error probability results under different mixing types As can be seen from the table, 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 1×10. −6 It can separate data very well. The separation error probability for PSK-QAM mixed types is mostly around 10. −5 The separation performance is relatively good for samples of the same order of magnitude; however, it performs the worst for 16QAM-64QAM mixed types, with a separation error probability of 1.118 × 10⁻⁶. −2 .

[0045] Next, the separation performance of the system proposed in this application and the Stacked Long Short-Term Memory Network (StackLSTM) on mixed types of 16QAM-64QAM was compared. The experimental results are as follows: Figure 7 As shown in the figure, the separation results of sampling points 200 to 300 of a randomly selected mixed sample are illustrated. The figure shows that both methods achieve relatively good separation, largely overlapping with the source signal, with most deviations occurring at signal extrema. Furthermore, compared to the two methods, this system performs better with less deviation, as indicated by the arrows in the figure.

[0046] The second set of experiments: To verify the impact of noise on the model, this set of experiments compared the experimental results of different models under different signal-to-noise ratios, as shown in Table 3: Table 3 Experimental results of different models at different signal-to-noise ratios As can be seen, the separation error probability decreases continuously with the increase of signal-to-noise ratio. Compared with other models, this system achieves the best results starting from SNR=5 dB.

[0047] Figures 8 to 9 The results show the impact of signal-to-noise ratio on different hybrid types. Figure 8 The separation of PSK-QAM and QAM-QAM hybrid types is shown. Figure 9 This demonstrates the separation of PSK-PSK hybrid types. From... Figure 8 and Figure 9 As can be seen, the PSK-PSK hybrid type performs better in separation than the PSK-QAM hybrid type and the QAM-QAM hybrid type. Specifically, the PSK-QAM and QAM-QAM hybrid types have relatively high overall separation error probabilities; at SNR = 20dB, 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 1×10 at SNR = 20dB. −7 .

[0048] The third group of experiments: To verify the effectiveness of the GCU block in the system, ablation experiments were conducted under noise-free conditions. The specific results are shown in Table 4. Table 4 Ablation Experiment Results As can be seen from the table, the separation performance of most hybrid types improved after adding the GCU block. Only the QPSK-16QAM hybrid type showed a worse signal separation performance, but it did not affect the overall separation performance. The average separation error probability decreased by 0.0003, indicating that the GCU block significantly improved the model's ability to extract features and improved the quality of signal separation.

[0049] The first set of experiments compared the separation performance of different models under noise-free conditions. The experimental results show that the system proposed in this application achieved the best performance, with the average separation error probability decreasing by approximately 1.5 × 10⁻⁶. −2The second set of experiments compared the results of different models under different signal-to-noise ratios, and studied the impact of noise on the models. The experimental results show that the system proposed in this application performs better than other models under high signal-to-noise ratios. The third set of experiments conducted ablation experiments to study the effectiveness of the GCU block. The experimental results show that the average separation error probability decreased by 0.0003 after adding the GCU block, proving the effectiveness of the GCU block.

[0050] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the mixed-signal dual-path blind source separation method provided in the second aspect.

[0051] This invention adds a GCU block to the dual-path model to enhance the overall feature extraction capability of the system. Based on the extraction of features within and between blocks, the feature representation of the overall sequence is extracted through the GCU block, so that features under multiple data representations are integrated in the module, thereby enhancing the signal separation capability.

[0052] 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.

[0053] 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 mixed-signal dual-path blind source separation system, characterized by, The method comprises the following steps: An encoder receives a mixed signal and extracts features of the mixed signal, and converts the extracted features into a high-dimensional feature representation; The mask network comprises a decomposition module and DPGCU blocks connected in sequence Each DPGCU block has a double-path LSTM structure and comprises a GCU block. The decomposition module is configured to segment the received high-dimensional features, and input the obtained segmented data into a first DPGCU block; The first-path LSTM structure included in the DPGCU block extracts local features of the segmented data and performs residual connection, outputs a local feature representation; the second-path LSTM structure extracts global features of the local feature representation and performs residual connection, outputs a global feature representation; the GCU block extracts global features of the global feature representation, and outputs an enhanced feature representation; No. The output of the DPGCU block is used as the first Input of each DPGCU block ,go through The separated signal characteristics are obtained after processing by each DPGCU block; A decoder is configured to convert the separated signal features processed by the masking network into original feature representations.

2. The mixed-signal dual-path blind source separation system of claim 1, wherein, The output of the first-path LSTM structure is processed by the second-path LSTM structure and added to the output of the first-path LSTM structure; The output of the second-path LSTM structure is processed by the GCU block and added to the output of the second-path LSTM structure.

3. The mixed-signal dual-path blind source separation system of claim 1, wherein, The GCU block includes a first path and a second convolutional layer in parallel, the first path includes a first convolutional layer and an expanded feedback shift memory network connected in sequence, and the output of the first path is subjected to an exclusive or operation with the output of the second convolutional layer.

4. The mixed-signal dual-path blind source separation system of claim 3, wherein, The first convolutional layer and the second convolutional layer have the same structure, and each includes a normalization layer, a first linear layer, a gated linear unit layer and a deep convolutional layer connected in sequence; The normalization layer is configured to perform normalization processing on the received data; The first linear layer is configured to perform feature transformation on the normalized data; The gated linear unit layer is configured to control the deep convolutional layer to extract features; The deep convolutional layer is configured to extract fine-grained local features.

5. The mixed-signal dual-path blind source separation system of claim 3, wherein, The expanded feedback shift memory network includes a second linear layer and a memory layer connected in sequence; the second linear layer receives data output by the first convolutional layer and performs linear processing thereon; the linearly processed data is subjected to residual connection with the linearly processed data after being processed by the memory layer.

6. The mixed-signal dual-path blind source separation system of claim 5, wherein, The memory layer includes stacked two-dimensional convolution layers, the first two-dimensional convolution layer has an expansion factor of , .

7. The mixed-signal dual-path blind source separation system of claim 6, wherein, dense connections are used between the two-dimensional convolution layers, and each two-dimensional convolution layer uses the same output channel , the output of the first two-dimensional convolution layer is , wherein, represents the original input, represents the mapping processing.

8. A mixed signal dual-path blind source separation method, characterized by, The method comprises the following steps: S1. Configure a mixed signal, extract features of the mixed signal and convert them into a high-dimensional feature representation; S2. Segment the high-dimensional feature representation to obtain a plurality of segmented data; S3. Extract local features of the plurality of segmented data and perform residual connection to obtain a local feature representation; S4. Extract global features of the local feature representation and perform residual connection to obtain a global feature representation; S5. Extract global features of the global feature representation based on a gated convolution method to obtain an enhanced feature representation; S6. Determine whether an iteration termination condition is met, if yes, stop iteration and obtain a separated signal feature; Otherwise, perform step S3; S7. Convert the separated signal feature into an original feature representation.

9. The mixed-signal dual-path blind source separation method of claim 8, wherein, The iteration termination condition is that the results obtained by 10 consecutive iterations are unchanged.

10. A terminal comprising a processor and a communication interface coupled to the processor, the processor being configured to run a computer program or instructions to implement the mixed signal double-path blind source separation method according to any one of claims 8 to 9.

Images