Signal separation method and device, storage medium and computer program product

Abstract

The embodiment of the invention discloses a signal separation method and device, a storage medium and a computer program product. The method comprises the following steps: inputting a to-be-separated multi-source mixed signal into a trained signal separation model; the trained signal separation model comprises a coding module, a separation module and a decoding module; converting the multi-source mixed signal into intermediate feature representation by using an encoding module; using a separation module to estimate a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation for each source signal in the multi-source mixed signal based on the corresponding mask; and performing signal reconstruction on each source signal in the multi-source mixed signal based on the corresponding feature information by using a decoding module.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular, to a signal separation method, device, storage medium, and computer program product. Background Art

[0002] 100BASE-T1 is a new Ethernet protocol. It is widely concerned by the industry, especially the automotive field, because it supports the devices on both sides of the link to receive and send signals at the same time and supports higher signal transmission bandwidth. In the full-duplex communication stage of 100BASE-T1, the gateway and electronic sensor connected by a pair of unshielded twisted pair cables act as the master and slave devices respectively, and send and receive three-level electrical signals at the same time. The possible signals are [-1V, 0V, +1V]. The master-slave working mode realizes the timing synchronization between the physical layer of the master and slave devices. In each device, the single-ended signal sent by the opposite device can be clearly identified by echo cancellation. The master and slave devices use the master scrambling code mode and the slave scrambling code mode respectively, which also ensures the difference between the received and sent signals, so that the two will not cancel each other. However, only the five-level mixed signal sent by the devices on both sides can be observed on the link. The possible levels are [-2V, -1V, 0V, +1V, +2V] (the master and slave devices can send -1, 0, 1 simultaneously or in time division). It is not conducive to analyzing the specific information sent by each electronic sensor device, and it is also impossible to authenticate the legitimacy of the identity of different electronic sensor devices based on the signals sent by them. Therefore, it is necessary to study the signal separation method of 100BASE-T1.

[0003] At present, the signal separation of 100BASE-T1 mainly includes traditional blind source signal separation method and hardware separation method. The traditional blind source signal separation method is used to restore the mixed signal sources of N different positions into non-Gaussian signal components of N independent signal sources. The fast independent component analysis algorithm is one of the most classic algorithms in the field of blind source separation. This method modifies the cumulative distribution function and uses Newton's method to update, which accelerates the convergence speed. The independence criterion function is constructed by the statistics of the source signal so that the function can reach the optimal iteration. The separation effect is affected by the choice of cost function and optimization method. The hardware separation method mainly uses directional couplers, which are passive measurement devices used to sample signals on twisted pair cables. The directional coupler has three ports, namely the input end, the output end and the coupling end. The directional coupler can couple the input end signal to the coupling end without affecting the normal data transmission of the line, and isolate the influence of the output end signal. The signal at the coupling end is the pure signal of the device connected to the input end.

[0004] However, in the traditional blind signal separation problem, multiple signal sources will interfere with each other, and the separation effect will be poor when the distance between the sampling positions is set too close. The training phase requires independent signal training models of two or three signal sources. In the 100BASE-T1 full-duplex communication state, it is impossible to directly obtain the single-ended signals sent by both ends. Although the hardware separation method based on directional couplers has high separation accuracy, when this method is used in the actual use environment, a pair of directional couplers needs to be connected between the gateway and each electronic component. The entire circuit needs to be modified on a large scale, which greatly increases the cost. The signal loss introduced by the directional coupler itself will also increase, which will have an adverse effect on the normal transmission of the signal. Summary of the invention

[0005] The embodiments of the present application provide a signal separation method, device, storage medium and computer program product, which reconstruct different source signals in a multi-source mixed signal through a signal separation model to achieve signal separation. The method is applicable to various signal transmission scenarios and does not require the use of hardware such as directional couplers. It not only does not affect the normal transmission of the signal, but also improves the flexibility and accuracy of signal separation.

[0006] The technical solution of the embodiment of the present application is implemented as follows:

[0007] The present application provides a signal separation method, including:

[0008] Inputting the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model comprises an encoding module, a separation module and a decoding module;

[0009] Using the encoding module, converting the multi-source mixed signal into an intermediate feature representation;

[0010] Using the separation module, estimating a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal;

[0011] The decoding module is used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

[0012] In the above method, the intermediate feature representation includes multiple feature representations, each feature representation corresponds to a time domain length, and the converting of the multi-source mixed signal into the intermediate feature representation includes:

[0013] Dividing the multi-source mixed signal into multiple signal segments with the same length in the time domain;

[0014] Each signal segment in the multiple signal segments is converted into an N-dimensional feature representation to obtain the multiple feature representations; N is a natural number greater than 1.

[0015] In the above method, estimating a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal, includes:

[0016] Based on different feature representations in the multiple feature representations, estimating a mask corresponding to each source signal in the multi-source mixed signal at different time domain lengths;

[0017] For each source signal in the multi-source mixed signal, the mask corresponding to each time domain length is multiplied by the feature representation under the same time domain length, and the products obtained under different time domain lengths are combined into corresponding feature information.

[0018] In the above method, the separation module includes a gated convolution module for estimating the mask;

[0019] The gated convolution module is composed of one-dimensional dilated gated convolution blocks stacked one-way, and the output end of each dilated gated convolution block includes a residual path for outputting a convolution operation residual and a skip connection path for outputting a convolution operation result;

[0020] The residual path of the previous dilated gated convolution block is the input end of the next dilated gated convolution block;

[0021] The sum of the outputs of the skip connection paths of each of the dilated gated convolutional blocks is an estimated value of the mask.

[0022] In the above method, the dilated gated convolution block includes a depthwise convolution and two pointwise convolutions respectively connected before and after the depthwise convolution;

[0023] The point-by-point convolution includes two operation paths, one operation path includes a first convolution layer, and the other operation path includes a second convolution layer, a batch normalization layer and an activation function connected in sequence, and the operation results of the two operation paths are combined and output.

[0024] In the above method, it also includes:

[0025] Collecting a first source signal and a first mixed signal formed by mixing the first source signal and a second source signal;

[0026] Calculating a difference between the first mixed signal and the first source signal to obtain the second source signal;

[0027] The signal separation model to be trained is trained using the first source signal, the second source signal and the first mixed signal to obtain the trained signal separation model.

[0028] In the above method, it also includes:

[0029] collecting a third source signal and a second mixed signal formed by mixing the third source signal and a fourth source signal;

[0030] Inputting the second mixed signal into the trained signal separation model, and reconstructing the third source signal through the signal separation model to obtain a reconstructed signal;

[0031] The model performance of the trained signal separation model is tested by comparing the difference between the third source signal and the reconstructed signal.

[0032] The present application provides a signal separation device, including:

[0033] A signal processing unit, used for inputting the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model comprises an encoding module, a separation module and a decoding module;

[0034] The encoding module is used to convert the multi-source mixed signal into an intermediate feature representation;

[0035] The separation module is used to estimate the mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separate the corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal;

[0036] The decoding module is used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

[0037] The embodiment of the present application provides a signal separation device, including: a processor, a memory and a communication bus;

[0038] The communication bus is used to realize the communication connection between the processor and the memory;

[0039] The processor is used to execute one or more computer programs stored in the memory to implement the signal separation method.

[0040] An embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a signal separation method is implemented.

[0041] An embodiment of the present application provides a computer program product, including a computer program, which implements a signal separation method when executed.

[0042] The embodiment of the present application provides a signal separation method, device, storage medium and computer program product, the method includes: inputting the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model includes an encoding module, a separation module and a decoding module; using the encoding module, converting the multi-source mixed signal into an intermediate feature representation; using the separation module, estimating the mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and for each source signal in the multi-source mixed signal, separating the corresponding feature information from the intermediate feature representation based on the corresponding mask; using the decoding module, reconstructing the signal for each source signal in the multi-source mixed signal based on the corresponding feature information. The technical solution provided by the embodiment of the present application reconstructs different source signals in the multi-source mixed signal through a signal separation model to achieve signal separation, which can be applied to various signal transmission scenarios, and does not require the use of hardware such as directional couplers, which not only does not affect the normal transmission of the signal, but also improves the flexibility and accuracy of signal separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram of a flow chart of a signal separation method provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of an exemplary signal separation process provided in an embodiment of the present application;

[0045] Figure 3 A schematic diagram of the structure of an exemplary signal separation module provided in an embodiment of the present application;

[0046] Figure 4 A schematic diagram of the structure of an exemplary dilated gated convolution block provided in an embodiment of the present application;

[0047] Figure 5 A schematic diagram of an exemplary model training and testing framework provided for an embodiment of the present application;

[0048] Figure 6 A schematic diagram of a signal separation device provided in an embodiment of the present application Figure 1 ;

[0049] Figure 7 A schematic diagram of a signal separation device provided in an embodiment of the present application Figure 2 . DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0051] The following will specifically describe the technical solution of the present application and how the technical solution of the present application solves the above technical problems through embodiments and in conjunction with the accompanying drawings. The following embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0052] In addition, the technical solutions described in the embodiments of the present application can be combined arbitrarily without conflict.

[0053] The present application embodiment provides a signal separation method, which is implemented by a signal separation device. The present application embodiment does not limit the specific type of the signal separation device, which can be any electronic device.

[0054] Figure 1 A flow chart of a signal separation method provided in an embodiment of the present application. Figure 1 As shown, in the embodiment of the present application, the signal separation method mainly includes the following steps:

[0055] S101, inputting a multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model includes a coding module, a separation module and a decoding module.

[0056] In an embodiment of the present application, the signal separation device may input the multi-source mixed signal to be separated into a trained signal separation model.

[0057] It should be noted that, in an embodiment of the present application, the multi-source mixed signal to be separated is formed by mixing multiple source signals, wherein each source signal originates from one device. Specifically, the multi-source mixed signal can be formed by mixing two source signals, and the two source signals can be source signals transmitted from one device to another device in two devices communicating in two directions.

[0058] It should be noted that, in the embodiment of the present application, the signal separation device is configured with a trained signal separation model, see Figure 2 ,The model includes a coding module, a separation module and a decoding module, and the model can ,be used to separate source signals in multi-source mixed signals.

[0059] S102: Using an encoding module, convert the multi-source mixed signal into an intermediate feature representation.

[0060] In an embodiment of the present application, a signal separation device inputs a multi-source mixed signal into a trained signal separation model, which includes an encoding module. The encoding module can be used to convert the multi-source mixed signal into an intermediate feature representation.

[0061] In an embodiment of the present application, the intermediate feature representation includes multiple feature representations, each feature representation corresponds to a time domain length, and the encoding module can perform the following steps to convert the multi-source mixed signal into an intermediate feature representation: divide the multi-source mixed signal into multiple signal segments with the same time domain length; convert each signal segment in the multiple signal segments into an N-dimensional feature representation to obtain multiple feature representations; N is a natural number greater than 1.

[0062] It can be understood that in an embodiment of the present application, the encoding module can divide the multi-source mixed signal according to the time domain length, and perform N-dimensional feature representation conversion on each divided signal segment, wherein the smaller the time domain length, the shorter the length of each signal segment, and the more feature representations obtained, which is conducive to improving the accuracy of signal separation.

[0063] In the examples of this application, see Figure 2 , multi-source mixed signal mix(t)∈R 1×T It can be represented by two source signals x1(t), x2(t)∈R 1×T Mixed formation, the multi-source mixed signal can be expressed according to the following formula (1):

[0064] mix(t)=x1(t)+x2(t) (1)

[0065] The encoding module can divide the multi-source mixed signal mix(t) into multiple signal segments with a time domain length of L. Each signal segment is converted into an N-dimensional feature representation through a one-dimensional convolution. For any segment of the divided signal mix x It can be expressed as the following formula (2):

[0066] w x =H(mix x U) (2)

[0067] Among them, w x ∈R 1×N is the characteristic representation of a signal, and the encoding module basis function U∈R N×L It includes N vectors, and H(·) is a nonlinear rectification function.

[0068] S103, using a separation module, estimating a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation for each source signal in the multi-source mixed signal based on the corresponding mask.

[0069] In an embodiment of the present application, the trained signal separation model includes a separation module, and the separation module can be used to separate feature information corresponding to each source signal in the multi-source mixed signal from the intermediate feature representation.

[0070] In an embodiment of the present application, a mask corresponding to each source signal in a multi-source mixed signal is estimated based on an intermediate feature representation, and for each source signal in the multi-source mixed signal, corresponding feature information is separated from the intermediate feature representation based on the corresponding mask, including: estimating the mask corresponding to each source signal in the multi-source mixed signal at different time domain lengths based on different feature representations in a plurality of feature representations; for each source signal in the multi-source mixed signal, multiplying the mask corresponding to each time domain length by the feature representation at the same time domain length, and forming the products obtained at different time domain lengths into corresponding feature information.

[0071] It can be understood that in an embodiment of the present application, as described in the above step S102, the intermediate feature representation may include multiple feature representations, each feature representation corresponds to a time domain length. Based on this, the separation module can estimate the mask corresponding to each source signal at different time domain lengths based on different feature representations. For each source signal, the product of the mask corresponding to the same time domain length and the feature representation can be calculated, and then the products at different time domain lengths can be combined into corresponding feature information.

[0072] In the embodiment of the present application, the intermediate feature representation may include multiple feature representations. For any feature representation w x , the separation module estimates C masks m ix ∈R 1×N , i = 1, ..., C, where C is the number of source signals in the multi-source mixed signal, where C = 2, m ix Indicates that the i-th source signal is in the feature representation w x The mask corresponding to the time domain length, that is, m ix And the feature representation w x Corresponding to the same time domain length, then m ix Applied to the feature representation w x superior:

[0073] d ix =w x ⊙m i (3)

[0074] Among them, ⊙ represents the multiplication of corresponding points, d ix ∈R 1×N , the above method can be used to process each feature representation, so as to estimate the mask of the i-th source signal at different time domain lengths, calculate the product of the mask corresponding to each time domain length and the feature representation at the same time domain length, and calculate the d ixThe corresponding characteristic information d i .

[0075] S104: Using a decoding module, reconstruct a signal for each source signal in the multi-source mixed signal based on corresponding feature information.

[0076] In an embodiment of the present application, the trained signal separation model includes a decoding module, and the decoding module can be used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

[0077] In an embodiment of the present application, the waveform of each source signal is reconstructed from the representation using a one-dimensional transposed convolution operation by a decoding module:

[0078]

[0079] in, is the ith source signal x i The estimated value of (t), that is, the signal reconstruction result of the i-th source signal, V∈R N×L The rows are decoding module basis functions, and the length of each basis function is L. It should be noted that, as described in the above formula (3), d i It can be composed of multiple products. Based on this, when it is substituted into formula (4), It can include multiple results, each of which is the estimated value of the i-th source signal in a time domain length.

[0080] As described in the above steps S101 to S104, in the embodiment of the present application, the trained signal separation model can be used to reconstruct the source signal in the multi-source mixed signal to achieve signal separation. Figure 3 Describe the model structure.

[0081] In the examples of this application, see Figure 3 The encoding module and the decoding module each include a one-dimensional convolution block. The one-dimensional convolution block of the encoding module can realize the conversion of feature representation, and the one-dimensional convolution block of the decoding module can process the feature information to reconstruct the signal.

[0082] In the examples of this application, see Figure 3 The separation module includes a gated convolution module for estimating the mask; the gated convolution module is composed of one-dimensional dilated gated convolution blocks stacked front and back, and the output end of each dilated gated convolution block includes a residual path for outputting the convolution operation residual and a skip connection path for outputting the convolution operation result; the residual path of the previous dilated gated convolution block is the input end of the next dilated gated convolution block; the sum of the outputs of the skip connection paths of each dilated gated convolution block is the estimated value of the mask.

[0083] It should be noted that in the embodiments of the present application, see Figure 3 , a stack of one-dimensional dilated gated convolution blocks is used in the separation module to estimate the source signal mask. This design can process data in parallel and improve data processing efficiency. In addition, no cyclic state information is required, so no memory unit is required, which reduces the number of parameters. The dilation coefficient of each layer gradually increases exponentially according to the integer power of 2, such as Figure 3 As shown, the dilated gated convolution blocks in the left column are the first dilated gated convolution blocks of each layer, with a dilation coefficient of 1, the dilated gated convolution blocks in the middle column are the second dilated gated convolution blocks of each layer, with a dilation coefficient of 2, and the dilated gated convolution blocks in the left column are the Rth dilated gated convolution of each layer, with a dilation coefficient of 2 to the Rth power.

[0084] It should be noted that in the embodiments of the present application, see Figure 3 , the N-dimensional feature representation generated by the encoding module enters the separation module and first passes through a normalization layer to normalize the data in this layer. It is not affected by batches and is more stable. The normalization formula is:

[0085]

[0086] Among them, x i is the i-th feature of the sample, μ L and σ L are the mean and standard deviation of the layer, γ and β are learnable scaling and translation parameters, and ∈ is a small constant to avoid instability caused by division by zero. This is followed by a 1×1 convolutional layer, which is a bottleneck layer with fewer neurons. The function of this layer is to compress the input high-dimensional data and reduce the data dimension, thereby reducing the computational burden of the network, speeding up training, and improving the generalization ability of the model.

[0087] It should be noted that in the embodiments of the present application, see Figure 3 , the 1×1 convolutional layer is connected to the stacked one-dimensional dilated gated convolutional blocks. By stacking multiple dilated gated convolutional blocks in the network, dependencies of different ranges can be learned, thereby improving the expressiveness of sequence data.

[0088] In the examples of this application, see Figure 4 The dilated gated convolution block shown includes a depthwise convolution and two pointwise convolutions respectively connected before and after the depthwise convolution; the pointwise convolution includes two operation paths, one operation path includes the first convolution layer, and the other operation path includes the second convolution layer, the batch normalization layer and the activation function connected sequentially, and the operation results of the two operation paths are merged and output.

[0089] It should be noted that in the embodiments of the present application, see Figure 4 , point-by-point convolution is a gated structure, which is specifically divided into two operation paths. One operation path directly uses 1×1 convolution (the first convolution layer) without other processing, and the other operation path adds a batch normalization layer and activation function after the 1×1 convolution layer (the second convolution layer). Finally, the operation results of the two paths are merged and passed to the next layer. The gated structure can effectively improve the modeling ability because it can select the input features and retain the most important information, which helps to improve the computational efficiency of the model and reduce the risk of overfitting. Gated convolution can also improve the robustness of the model because it can suppress unimportant feature channels and emphasize features related to the target task, which can reduce the interference of noise on the model. The input and output positions of each dilated gated convolution block adopt a gated structure to control the input data and the output after data processing respectively, which can effectively reduce the number of parameters and computation of the model and improve the efficiency and accuracy of the model. Thereby saving computing resources and memory consumption.

[0090] It should be noted that in the embodiments of the present application, see Figure 4 ,The dilated gated convolution block usually also includes a normalization layer and an activation function, which helps improve the stability of the model.

[0091] In an embodiment of the present application, the above-mentioned trained signal separation model is obtained by training the signal separation model to be trained. The model structure of the signal separation model to be trained is the model structure introduced above. Model training is to continuously optimize the parameters therein. The model training process is detailed below.

[0092] In an embodiment of the present application, the signal separation device can execute the steps of: collecting a first source signal and a first mixed signal formed by mixing the first source signal and the second source signal; calculating the difference between the first mixed signal and the first source signal to obtain the second source signal; using the first source signal, the second source signal and the first mixed signal to train the signal separation model to be trained to obtain a trained signal separation model.

[0093] It should be noted that in the embodiments of the present application, see Figure 5, in the training phase, the mixed signal and one of the source signals can be collected at the same time. Specifically, taking the bidirectional communication between the master device and the slave device as an example, a pair of directional couplers are connected between the devices. A directional coupler is a passive measurement device used for signal sampling on a transmission line. It has three ports, the input end, the output end and the coupling end. The directional coupler can couple the input end signal to the coupling end without affecting the normal data transmission of the line, and only transmits the signal entering the coupler at the output end without coupling energy to the coupling end. That is, the signal sent by the device corresponding to the coupler input end is the real single-ended signal, that is, the source signal, which will be output on the data acquisition device such as an oscilloscope through the coupling end. This source signal is also the target true value that the signal separation expects to approximate. In the training phase, the first mixed signal of the communication between the two devices and the source signal of one device at the coupling end of the coupler (the first source signal) are collected, and then the source signal of another device (the second source signal) is obtained by difference. It should be noted that the other source signal obtained by difference here is actually an estimated value, which also includes the noise interference and echo generated by the entire line. These three signals are input into the signal separation model to be trained for model training.

[0094] It should be noted that in the embodiment of the present application, the loss function of the model training selects the scale-invariant signal-to-noise ratio. It is similar to the signal-to-noise ratio, but it has scale invariance. The scale-invariant signal-to-noise ratio can obtain different source signals at different scales, thereby determining the scale of the signal-to-noise ratio, so as to better understand the quality of signal separation. The formula of the scale-invariant signal-to-noise ratio can be expressed as:

[0095]

[0096] Among them, x is the real source signal collected, is the source signal reconstructed by the model, It is an element-wise product and then sum operation. The higher the scale-invariant signal-to-noise ratio value, the higher the similarity between the estimated signal and the true value signal, and the better the separation effect.

[0097] It should be noted that in the traditional single-channel two-source signal mixed signal separation problem, two source signals are reconstructed each time for training. These two need to be arranged and combined with two real collected source signals, and the sum of the loss values ​​in different situations is calculated. The smallest group is selected, and the average is taken as the loss for this time to solve the problem of signal arrangement order. Considering that in actual applications, the signal separation problem of a certain device may be focused on, such as the signal separation problem at one end of the slave device, the training objectives are different. Only the two source signals estimated each time and the real source signal at one end of the slave device can be used to calculate the loss separately, and the smaller value can be taken. There is no need to pay attention to the single-ended signal emitted by the main device, the echo interference and noise in the entire line, etc., which is conducive to speeding up the calculation speed, improving the efficiency of the model, and more in line with the actual application scenario. Of course, whether the double-end calculation loss or the single-end calculation loss can be based on the actual demand equipment, and the embodiment of the present application is not limited.

[0098] It should be noted that in the embodiments of the present application, considering that the cloud database ClickHouse can store massive amounts of data, it is possible to choose to use a large amount of data used in the storage model training phase. And ClickHouse will compress the data when storing it. Data compression in columnar databases is easier to implement, and a large amount of training data can be stored in a smaller storage space, saving costs. ClickHouse uses a large number of read optimization algorithms, which makes the data reading speed faster than other cloud databases. The computing resources of the ClickHouse cluster can process data in parallel, further improving the data query speed.

[0099] In the examples of this application, see Figure 5 , not only can the model be trained, but also tested after model training. In the testing stage, it is only necessary to input the collected mixed signal into the model for signal separation, and then compare the collected source signal with the reconstructed source signal output by the model. The steps are as follows: collect the third source signal, and the second mixed signal formed by the third source signal and the fourth source signal; input the second mixed signal into the trained signal separation model, and reconstruct the third source signal through the signal separation model to obtain the reconstructed signal; by comparing the difference between the third source signal and the reconstructed signal, test the model performance of the trained signal separation model.

[0100] Exemplarily, in an embodiment of the present application, device 1 is used as a master device at both ends of the twisted pair, and device 2 is used as a slave device. A signal differential board is used to collect the mixed signal (first mixed signal) of the master-slave device communication and the source signal (first source signal) on the master device 1 side, and then the obtained signal is regarded as the source signal (second source signal) on the slave device 2 side by difference. In the training phase, the above three signals are input into the signal separation model to be trained for model training. In the test phase, device k is used to replace device 1 and device 2 respectively, and the signal separation effect after the master device and the slave device are replaced is observed. In a test environment, device k is used to replace device 1 as the master device in the test phase, and the mixed signal (second mixed signal) of device k and device 2 is input, and the estimated value (reconstructed signal) of the signal sent by device k (third source signal) is output, and the reconstructed signal and the collected third source signal are compared. In another test environment, device 2 is replaced by device k as the slave device in the test phase. The mixed signal of device 1 and device k (the second mixed signal) is input, and the estimated value (reconstructed signal) of the signal sent by device 1 (the third source signal) is output. The reconstructed signal and the collected third source signal are compared to verify the effectiveness of the proposed scheme.

[0101] The technical solution provided by the embodiment of the present application, first, realizes signal separation through a model, solves the problem that the restriction on the source in the blind signal separation method does not conform to the actual network environment, and significantly improves the accuracy of signal separation. Second, a separation method based on a gated convolution signal separation network is proposed. In the training stage, a mixed signal and a source signal are collected, and another source signal is obtained by difference, and the signal separation model is trained together, and the mixed signal is directly separated during testing. Reduce the signal loss due to the large number of directional couplers connected, and reduce the adverse effects on the normal transmission of the signal. Third, only the loss value of the separation result and the real signal from one end of the device can be calculated, which speeds up the calculation speed, improves the model efficiency, and is more in line with the actual application scenarios of vehicle-mounted Ethernet; Fourth, the convolution gate structure is introduced in the separation module, which effectively improves the modeling ability and the characterization ability of the features, making the signal separation more accurate. It solves the problem that when there is a difference between the input data and the training data, the separation performance will be affected and it is not suitable for the network environment.

[0102] The technical solution provided by the embodiment of the present application, on the one hand, by directly modeling the mixed signal in the time domain, calculating the mask of the single-side signal, and multiplying it with the intermediate feature representation of the mixed signal, and finally reconstructing the source signal through the decoding model, significantly improves the accuracy of the separation result, and on the other hand, introduces a directional coupler in the model training stage, which can be performed offline, and uses the trained model in the actual production environment to separate the mixed signal by software. Since most of the features are retained, the accuracy of signal separation is achieved, and the impact of introducing a directional coupler on the signal in the network of the actual production environment is reduced.

[0103] An embodiment of the present application provides a signal separation device. Figure 6 A schematic diagram of a signal separation device provided in an embodiment of the present application Figure 1 .like Figure 6 As shown, the signal separation device 1 comprises:

[0104] A signal processing unit 11 is used to input the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model includes an encoding module, a separation module and a decoding module;

[0105] The encoding module is used to convert the multi-source mixed signal into an intermediate feature representation;

[0106] The separation module is used to estimate the mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separate the corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal;

[0107] The decoding module is used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

[0108] In one embodiment of the present application, the intermediate feature representation includes multiple feature representations, each feature representation corresponds to a time domain length, and the encoding module is used to divide the multi-source mixed signal into multiple signal segments with the same time domain length; convert each signal segment in the multiple signal segments into an N-dimensional feature representation to obtain the multiple feature representations; N is a natural number greater than 1.

[0109] In one embodiment of the present application, the separation module is used to estimate the mask corresponding to each source signal in the multi-source mixed signal at different time domain lengths based on different feature representations in the multiple feature representations; for each source signal in the multi-source mixed signal, the mask corresponding to each time domain length is multiplied by the feature representation at the same time domain length, and the products obtained at different time domain lengths are used to form corresponding feature information.

[0110] In an embodiment of the present application, the separation module includes a gated convolution module for estimating a mask;

[0111] The gated convolution module is composed of one-dimensional dilated gated convolution blocks stacked one-way, and the output end of each dilated gated convolution block includes a residual path for outputting a convolution operation residual and a skip connection path for outputting a convolution operation result;

[0112] The residual path of the previous dilated gated convolution block is the input end of the next dilated gated convolution block;

[0113] The sum of the outputs of the skip connection paths of each of the dilated gated convolutional blocks is an estimated value of the mask.

[0114] In one embodiment of the present application, the dilated gated convolution block includes a depthwise convolution, and two pointwise convolutions respectively connected before and after the depthwise convolution;

[0115] The point-by-point convolution includes two operation paths, one operation path includes a first convolution layer, and the other operation path includes a second convolution layer, a batch normalization layer and an activation function connected in sequence, and the operation results of the two operation paths are combined and output.

[0116] In one embodiment of the present application, the signal processing unit is also used to collect a first source signal and a first mixed signal formed by mixing the first source signal and the second source signal; calculate the difference between the first mixed signal and the first source signal to obtain the second source signal; use the first source signal, the second source signal and the first mixed signal to train the signal separation model to be trained to obtain the trained signal separation model.

[0117] In one embodiment of the present application, the signal processing unit is also used to collect a third source signal and a second mixed signal formed by mixing the third source signal and a fourth source signal; input the second mixed signal into the trained signal separation model, and reconstruct the third source signal through the signal separation model to obtain a reconstructed signal; and test the model performance of the trained signal separation model by comparing the difference between the third source signal and the reconstructed signal.

[0118] Figure 7 A schematic diagram of a signal separation device provided in an embodiment of the present application Figure 2 .like Figure 7 As shown, the signal separation device 1 includes: a processor 12, a memory 13 and a communication bus 14;

[0119] The communication bus 14 is used to realize the communication connection between the processor 12 and the memory 13;

[0120] The processor 12 is used to execute one or more computer programs stored in the memory 13 to implement the signal separation method.

[0121] An embodiment of the present application provides a computer program product, including a computer program, which implements a signal separation method when executed.

[0122] The embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements a signal separation method when executed by a processor. The computer-readable storage medium may be a volatile memory (volatile memory), such as a random access memory (Random-Access Memory, RAM); or a non-volatile memory (non-volatile memory), such as a read-only memory (Read-Only Memory, ROM), a flash memory (flash memory), a hard disk (Hard Disk Drive, HDD) or a solid-state drive (Solid-State Drive, SSD); or may be a respective device including one or any combination of the above memories, such as a mobile phone, a computer, a tablet device, a personal digital assistant, etc.

[0123] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of hardware embodiments, software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0124] The present application is described with reference to implementation flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0125] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which is implemented in the implementation flow diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0126] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process described in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0127] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the utility application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A signal separation method, characterized in that: include: Inputting the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model comprises an encoding module, a separation module and a decoding module; Using the encoding module, converting the multi-source mixed signal into an intermediate feature representation; Using the separation module, estimating a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal; The decoding module is used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

2. The method according to claim 1, characterized in that The intermediate feature representation includes a plurality of feature representations, each feature representation corresponds to a time domain length, and the converting the multi-source mixed signal into the intermediate feature representation includes: Dividing the multi-source mixed signal into multiple signal segments with the same length in the time domain; Each signal segment in the multiple signal segments is converted into an N-dimensional feature representation to obtain the multiple feature representations; N is a natural number greater than 1.

3. The method according to claim 2, characterized in that The estimating a mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separating corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal, comprises: Based on different feature representations in the multiple feature representations, estimating the mask corresponding to each source signal in the multi-source mixed signal at different time domain lengths; For each source signal in the multi-source mixed signal, the mask corresponding to each time domain length is multiplied by the feature representation under the same time domain length, and the products obtained under different time domain lengths are combined into corresponding feature information.

4. The method according to claim 1, characterized in that: The separation module includes a gated convolution module for estimating a mask; The gated convolution module is composed of one-dimensional dilated gated convolution blocks stacked one-way, and the output end of each dilated gated convolution block includes a residual path for outputting a convolution operation residual and a skip connection path for outputting a convolution operation result; The residual path of the previous dilated gated convolution block is the input end of the next dilated gated convolution block; The sum of the outputs of the skip connection paths of each of the dilated gated convolutional blocks is an estimated value of the mask.

5. The method according to claim 4, characterized in that The dilated gated convolution block includes a depthwise convolution and two pointwise convolutions respectively connected before and after the depthwise convolution; The point-by-point convolution includes two operation paths, one operation path includes a first convolution layer, and the other operation path includes a second convolution layer, a batch normalization layer and an activation function connected in sequence, and the operation results of the two operation paths are combined and output.

6. The method according to claim 1, characterized in that The method further comprises: Collecting a first source signal and a first mixed signal formed by mixing the first source signal and a second source signal; Calculating a difference between the first mixed signal and the first source signal to obtain the second source signal; The signal separation model to be trained is trained using the first source signal, the second source signal and the first mixed signal to obtain the trained signal separation model.

7. The method according to claim 6, characterized in that The method further comprises: Collecting a third source signal and a second mixed signal formed by mixing the third source signal and a fourth source signal; Inputting the second mixed signal into the trained signal separation model, and reconstructing the third source signal through the signal separation model to obtain a reconstructed signal; The model performance of the trained signal separation model is tested by comparing the difference between the third source signal and the reconstructed signal.

8. A signal separation device, characterized in that: include: A signal processing unit, used for inputting the multi-source mixed signal to be separated into a trained signal separation model; the trained signal separation model comprises an encoding module, a separation module and a decoding module; The encoding module is used to convert the multi-source mixed signal into an intermediate feature representation; The separation module is used to estimate the mask corresponding to each source signal in the multi-source mixed signal based on the intermediate feature representation, and separate the corresponding feature information from the intermediate feature representation based on the corresponding mask for each source signal in the multi-source mixed signal; The decoding module is used to perform signal reconstruction for each source signal in the multi-source mixed signal based on corresponding feature information.

9. A signal separation device, characterized in that: include: processor, memory, and communication bus; The communication bus is used to realize the communication connection between the processor and the memory; The processor is used to execute one or more computer programs stored in the memory to implement the signal separation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the signal separation method according to any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed, the signal separation method according to any one of claims 1 to 7 is implemented.

Images