Blind Source Separation Method Based on Autoencoder Joint Fast Independent Component Analysis

Abstract

This invention provides a blind source separation method based on autoencoder joint fast independent component analysis. The method includes: constructing and training an autoencoder-based network; performing blind source separation on the aliased signal to be separated to obtain the recovered original independent signal; the network includes a network and a network; training includes: generating the aliased signal, which, after blind source separation, yields the original independent signal to be recovered; inputting the aliased signal into the network to obtain a latent space representation, fusing it with the original independent signal to be recovered, and then inputting it into the network to obtain the recovered original independent signal; calculating the loss between the recovered original independent signal and the original independent signal to be aliased, performing backpropagation, and iteratively optimizing the network parameters. This invention improves the separation accuracy of aliased signals.

Description

Technical Field

[0001] This invention relates to the field of blind source separation technology, and more specifically, to a blind source separation method based on autoencoder combined with fast independent component analysis. Background Technology

[0002] Currently, blind source separation (BSS) algorithms are commonly used to separate unobserved original signals from a mixture of multiple observed signals.

[0003] Blind source separation algorithms are commonly used to solve the well-known "cocktail party problem" in computer speech recognition. This problem describes a situation where multiple people speak simultaneously at a cocktail party, with multiple recording devices recording the conversation, and the algorithm is used to identify the speech content afterward. However, while current speech recognition technology can accurately identify the speech of a single person, the accuracy drops significantly when multiple people are speaking simultaneously. Therefore, in scenarios where multiple people are speaking at the same time, it is necessary to reconstruct the individual original speech of each person from the mixed audio files to achieve more accurate recognition of the individual original speech content.

[0004] Secondly, with the development of communication technology, a large number of military and civilian electronic devices have been put into use, leading to a diversity of signal modulation patterns and thus creating a complex electromagnetic environment. Especially for non-cooperative broadband reception, the observed signals are often unknown and aliased in the time and frequency domains. Therefore, blind source separation algorithms are frequently used to recover the original independent signals from the aliased signals received by the communication system.

[0005] Independent Component Analysis (ICA) is currently the mainstream method for separating blind source signals. In 1996, Finnish researchers first proposed the Fast-ICA algorithm based on degree, and in 1999, they proposed an improved Fast-ICA algorithm based on negative entropy. This algorithm decomposes multi-channel observed signals into several independent components according to the principle of statistical independence and through optimization. Although the Fast-ICA algorithm can decompose aliased signals into multiple uncorrelated original independent signals, it suffers from two uncertainties: poor accuracy of the separated original independent signals' amplitude and phase inversion, meaning the accuracy of the separated original independent signals is relatively poor.

[0006] In order to solve the above problems, people have been seeking an ideal technological solution. Summary of the Invention

[0007] Therefore, it is necessary to provide a blind source separation method based on autoencoder-based fast independent component analysis to address the aforementioned technical problems. This invention constructs an AE-ICA network based on an autoencoder (AE) to repair the amplitude and phase inversion defects of the original independent signals obtained through the Fast-ICA algorithm, resulting in a repaired signal with higher accuracy.

[0008] To achieve the above objectives, this invention provides a blind source separation method based on autoencoder joint fast independent component analysis, comprising:

[0009] Construct and train an autoencoder-based AE-ICA network;

[0010] The input aliased signal to be separated is blindly separated using an AE-ICA network to obtain the recovered original independent signal.

[0011] The autoencoder-based AE-ICA network includes an Encoder network, a Fast-ICA network, and a Decoder network;

[0012] The method for training the autoencoder-based AE-ICA network includes:

[0013] (1) Generate an aliasing signal based on the original independent signals to be aliased;

[0014] (2) Input the aliased signal into the Fast-ICA network for blind source separation to obtain the original independent signal to be recovered;

[0015] (3) Input the aliased signal into the Encoder network to obtain the latent spatial representation of the aliased signal;

[0016] (4) The latent space representation is fused with the original independent signal to be recovered, and the fused signal is then input into the Decoder network to obtain the recovered original independent signal;

[0017] (5) Calculate the loss between the recovered original independent signal and the original independent signal to be aliased and perform backpropagation to optimize the parameters of the autoencoder-based AE-ICA network.

[0018] (6) Iteratively optimize the parameters of the autoencoder-based AE-ICA network until the number of iterations reaches a preset number and stop iterating, or determine whether to continue iteratively optimizing the parameters of the autoencoder-based AE-ICA network based on the difference between the recovered original independent signal corresponding to the current iteration and the original independent signal to be aliased.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention fuses the latent spatial representation and the corresponding original independent signals to be recovered to obtain a fused signal. This fused signal is then input into the decoder network of an autoencoder to obtain the recovered original independent signals. The autoencoder parameters are iteratively optimized by backpropagating the loss between the recovered original independent signals and the corresponding original independent signals to be aliased, resulting in an autoencoder-based AE-ICA network with higher accuracy in separating aliased signals. Finally, the trained autoencoder-based AE-ICA network can more accurately obtain the original independent signals to be separated from the aliased signals, with more accurate amplitude and phase separation. Attached Figure Description

[0021] Figure 1 This is a simplified schematic diagram of the AE-ICA network structure based on an autoencoder according to the present invention;

[0022] Figure 2 This is a simplified schematic diagram of the encoder network structure of the self-encoder of the present invention;

[0023] Figure 3 This is a simplified schematic diagram of the Decoder network structure of the autoencoder of the present invention;

[0024] Figure 4 These are schematic diagrams of the two original modulation signals of this invention;

[0025] Figure 5 This is a schematic diagram showing the result of the first aliasing of the two original modulation signals of the present invention;

[0026] Figure 6 This is a schematic diagram showing the result of the second aliasing of the two original modulation signals of the present invention;

[0027] Figure 7 This is a comparative schematic diagram of the original modulation signal 1 of the present invention, the original independent signal corresponding to the original modulation signal 1 separated by the Fast-ICA algorithm, and the original independent signal corresponding to the original modulation signal 1 separated by the AE-ICA algorithm of the present invention;

[0028] Figure 8 This is a comparative diagram of the original modulation signal 2 of the present invention, the original independent signal corresponding to the original modulation signal 2 separated by the Fast-ICA algorithm, and the original independent signal corresponding to the original modulation signal 2 separated by the AE-ICA algorithm of the present invention;

[0029] Figure 9This is a comparative diagram of the original modulation signal 3 of the present invention, the original independent signal corresponding to the original modulation signal 3 separated by the Fast-ICA algorithm, and the original independent signal corresponding to the original modulation signal 3 separated by the AE-ICA algorithm of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0031] The terms "first," "second," "third," "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application.

[0032] To facilitate understanding, the interactive parties and / or terms and / or custom terms involved in this invention will first be explained in conjunction with the technical solution of this invention:

[0033] An autoencoder (AE) is a deep learning model that learns feature representations of data through unsupervised learning. An autoencoder consists of an encoder and a decoder. The encoder maps the input data to a low-dimensional latent space, while the decoder reconstructs the representation from the latent space back to the original input data.

[0034] It should be noted that this invention uses an autoencoder (AE), widely used in deep learning, to compensate for defects in the Fast-ICA algorithm. After defect compensation, the amplitude and phase accuracy of the separated original independent signals are higher.

[0035] This embodiment provides a specific implementation of a blind source separation method based on autoencoder joint fast independent component analysis. The network structure designed in this invention is as follows: Figure 1 As shown, the network consists of two main parts: an autoencoder and a Fast-ICA algorithm, where x(t) represents the aliased signal. This represents the separated original independent signals. The method includes:

[0036] Construct and train an autoencoder-based AE-ICA network;

[0037] The input aliased signal to be separated is blindly separated using an AE-ICA network to obtain the recovered original independent signal.

[0038] The autoencoder-based AE-ICA network includes an Encoder network, a Fast-ICA network, and a Decoder network;

[0039] The method for training the autoencoder-based AE-ICA network includes:

[0040] (1) Generate an aliasing signal based on the original independent signals to be aliased;

[0041] (2) Input the aliased signal into the Fast-ICA network for blind source separation to obtain the original independent signal to be recovered;

[0042] (3) Input the aliased signal into the Encoder network to obtain the latent spatial representation of the aliased signal;

[0043] (4) The latent space representation is fused with the original independent signal to be recovered, and the fused signal is then input into the Decoder network to obtain the recovered original independent signal;

[0044] (5) Calculate the loss between the recovered original independent signal and the original independent signal to be aliased and perform backpropagation to optimize the parameters of the autoencoder-based AE-ICA network.

[0045] (6) Iteratively optimize the parameters of the autoencoder-based AE-ICA network until the number of iterations reaches a preset number and stop iterating, or determine whether to continue iteratively optimizing the parameters of the autoencoder-based AE-ICA network based on the difference between the recovered original independent signal corresponding to the current iteration and the original independent signal to be aliased.

[0046] It should be noted that the present invention first optimizes the parameters of the autoencoder, and then uses the trained autoencoder-based AE-ICA network to separate the aliased signal to be separated, so as to obtain the corresponding recovered original independent signal, thereby improving the accuracy of the amplitude and phase of the separated original independent signal.

[0047] In some embodiments, (1) generating an aliased signal based on the original independent signal to be aliased includes: generating an aliased signal x(t) using the following formula:

[0048]

[0049] Where n is the number of signal sources (i.e., the total number of original independent signals to be aliased), m is the number of receiving channels (i.e., the total number of aliased signals received), v(t) is additive white Gaussian noise (AWGN), and T is the signal length. To simplify the above formula, it is rewritten as x(t) = As(t) + v(t). Where A ∈ R m×n Let A be a linear aliasing matrix, s1(t), s2(t)...s n (t) represents n independent modulation signals (in other words, n is the total number of independent modulation signal sources).

[0050] In some embodiments, n = 2, m = 2.

[0051] In some embodiments, before inputting the aliased signal into the Fast-ICA network, the method further includes: decentralizing the aliased signal to achieve a mean of 0; and performing a "whitening" preprocessing on the aliased signal. The purpose of decentralization is to simplify and facilitate subsequent computation. The purpose of whitening preprocessing is to remove the correlation between the aliased signals, simplify the subsequent extraction process of the original independent signals, and improve the convergence of the algorithm.

[0052] It should be noted that the aliased signal input to the encoder network of the autoencoder is unprocessed (i.e., it is not decentralized and whitened preprocessed).

[0053] In some embodiments, the Fast-ICA algorithm comes from the popular modular machine learning library Scikit-learn in the Python environment, specifically from the sklearn.decomposition.FastICA module. It should be noted that this module already includes whitening preprocessing.

[0054] (2) The aliased signal is input into the Fast-ICA network for blind source separation to obtain the original independent signal to be recovered, including:

[0055] S ICA =FastICA[x(t)],{t=1,2,…,T}

[0056] Among them, S ICA The original independent signal to be recovered (in the form of a matrix) is represented by FastICA[], which represents the Fast-ICA algorithm (Fast-ICA network).

[0057] (3) Inputting the aliased signal into the Encoder network to obtain the latent spatial representation of the aliased signal includes:

[0058] F enc =Enc[x(t)],{t=1,2,…,T}

[0059] F enc Enc[] represents the latent spatial representation of the aliased signal (in the form of a matrix), and Enc[] represents the encoder (Encoder network).

[0060] like Figure 2 and Figure 3 As shown, in some embodiments, the encoder network includes a first LSTM network (Long Short-Term Memory Network), a linear layer, an activation function ReLU, and another linear layer connected in sequence; the first LSTM network is LSTM1;

[0061] The first LSTM network consists of 3 LSTM layers, with an input dimension of 2 (number of receive channels) and 256 hidden layers.

[0062] The Decoder network comprises a linear layer (Linear), an activation function (ReLU), another linear layer (Linear), a second LSTM network (LSTM2), and another linear layer (Linear) connected in sequence.

[0063] The second LSTM network consists of 3 LSTM layers, with an input dimension of 2 (number of receive channels) and 256 hidden layers.

[0064] In some embodiments, to correct the defects of the Fast-ICA algorithm, (4) the latent space representation is fused with the original independent signal to be recovered to obtain a fused signal. The fusion of the latent space representation and the original independent signal to be recovered is a dot product fusion. For example, the calculation formula of the fused signal is:

[0065] F = F enc ⊙S ICA

[0066] Where F represents the fused signal, ⊙ represents matrix dot product, F enc S represents the latent space representation. ICA This represents the original independent signal to be recovered.

[0067] The fused signal is input into the Decoder network to obtain the recovered original independent signals, including: S ′ =Dec[F]=Dec[F] enc ⊙S ICA ], where S ′ S represents the recovered original independent signal (i.e., the original independent signal separated by the decoder). ′The form is a matrix, where Dec[] represents the decoder (Decoder network) and F represents the fused signal (i.e., the fused latent space representation). This step uses the decoder to remap the fused signal back to the original independent signals.

[0068] In some implementations, the loss of the recovered original independent signal and the original independent signal to be aliased The calculation formula is:

[0069]

[0070] Here, s(t) is denoted as S, where S represents the original independent signal to be aliased, ∥∥2 represents the vector 2-norm (L2 norm), n represents the total number of original independent signals to be aliased, and i represents the index of the original independent signal.

[0071] This represents the mean of the squares of the 2-norm values ​​of the vectors of each recovered original independent signal and the corresponding original independent signal to be aliased.

[0072] It should be noted that the smaller the loss (i.e., difference) between the recovered original independent signal and the original independent signal to be aliased, the greater the accuracy of the parameters of the autoencoder-based AE-ICA network. Therefore, backpropagation is performed based on the loss (i.e., difference) between the recovered original independent signal and the original independent signal to be aliased to optimize the parameters of the autoencoder-based AE-ICA network.

[0073] It should be noted that the technique of backpropagating based on the loss function to optimize the parameters of the autoencoder is an existing technique and will not be elaborated further.

[0074] In some embodiments, the parameters of the autoencoder-based AE-ICA network are iteratively optimized until the number of iterations reaches a preset number (which can be 50), at which point the iteration stops. Those skilled in the art can set this preset number of iterations. Through multiple iterations, a trained autoencoder-based AE-ICA network can be obtained, and the defects of the Fast-ICA algorithm can then be corrected based on the trained autoencoder-based AE-ICA network.

[0075] In some embodiments, the difference between the recovered original independent signal corresponding to the current iteration (i.e., the latest) and the original independent signal to be aliased is used to determine whether to continue iteratively optimizing the parameters of the autoencoder-based AE-ICA network.

[0076] Specifically, in some embodiments, if the difference between the recovered original independent signal and the original independent signal to be aliased in the current iteration is less than a preset threshold, then the iterative optimization of the parameters of the autoencoder-based AE-ICA network is stopped. Those skilled in the art can set this preset threshold themselves. The smaller the value of this preset threshold, the better the optimization effect of the parameters of the autoencoder-based AE-ICA network.

[0077] For example, in the first iteration: if the difference between the recovered original independent signal and the original independent signal to be aliased is greater than or equal to a preset threshold, then the parameters of the autoencoder-based AE-ICA network are optimized; in the second iteration: based on the first optimized autoencoder-based AE-ICA network, the recovered original independent signal is obtained (i.e., the recovered original independent signal corresponding to the current iteration). If the difference between the recovered original independent signal and the original independent signal to be aliased is less than a preset threshold, then the optimization of the parameters of the autoencoder-based AE-ICA network is stopped.

[0078] Specifically, for ease of understanding, the process of separating the aliased signal is explained. The input aliased signal is subjected to blind source separation using an AE-ICA network to obtain the recovered original independent signal, including:

[0079] The process involves: acquiring the aliased signal to be separated; using a Fast-ICA network based on an autoencoder AE-ICA network to separate the aliased signal, obtaining the original independent signal to be recovered; inputting the aliased signal to be separated into the Encoder network of the autoencoder AE-ICA network to obtain the latent spatial representation of the aliased signal; fusing the latent spatial representation with the original independent signal to be recovered to obtain the fused signal; and inputting the fused signal into the Decoder network of the autoencoder AE-ICA network to obtain the recovered original independent signal. This completes the separation of the aliased signal. It should be noted that the aliased signal input to the Fast-ICA network of the autoencoder AE-ICA network may have undergone past centering and whitening preprocessing, while the aliased signal input to the Encoder network of the autoencoder AE-ICA network is unprocessed (i.e., without decentralization and whitening preprocessing).

[0080] Experiment and results:

[0081] This invention uses four common modulation schemes (FM, 4ASK, OOK, and 64QAM) from the publicly available dataset RadioML 2018.01A for experimental verification. A dual-channel IQ quadrature modulation signal with a signal length of 1024 in the dataset is used as the basic signal format for the network, with a basic signal size of shape(x) = [2, 1024]. To facilitate network processing, the dual-channel signal is separated and concatenated to obtain a new signal size shape(x') = [1, 2048].

[0082] Two original modulation signals (original modulation signal 1 and original modulation signal 2) are as follows Figure 4 As shown, 401 represents the original modulation signal 1, and 402 represents the original modulation signal 2. Two modulation schemes are randomly selected from FM, 4ASK, OOK, and 64QAM to alias these two original modulation signals. The result of the first aliasing is shown below. Figure 5 As shown, 501 is the signal after the original modulated signal 1 is modulated for the first time, and 502 is the signal after the original modulated signal 2 is modulated for the first time; the result of the second aliasing is as follows. Figure 6 As shown, 601 is the signal after the original modulated signal 1 is modulated a second time, and 602 is the signal after the original modulated signal 2 is modulated a second time. The two modulation methods corresponding to the first aliasing are different from the two modulation methods corresponding to the second aliasing.

[0083] The original independent signals separated by the Fast-ICA algorithm are compared with the original independent signals separated by the AE-ICA algorithm of this invention (i.e., the trained autoencoder-based AE-ICA network).

[0084] like Figure 7 As shown, Figure 7 The diagram shows the original modulated signal 1, the original independent signal 701 corresponding to the original modulated signal 1 separated by the Fast-ICA algorithm, and the original independent signal 702 corresponding to the original modulated signal 1 separated by the AE-ICA algorithm of this invention.

[0085] like Figure 8 As shown, Figure 8 The diagram shows the original modulated signal 2, the original independent signal 801 corresponding to the original modulated signal 2 separated by the Fast-ICA algorithm, and the original independent signal 802 corresponding to the original modulated signal 2 separated by the AE-ICA algorithm of this invention.

[0086] This further demonstrates that the amplitude of the original independent signal separated by the Fast-ICA algorithm differs significantly from the amplitude of the original modulated signal, while the amplitude of the original independent signal separated by the AE-ICA algorithm of this invention is closer to the amplitude of the original modulated signal.

[0087] For the phase inversion problem of the original independent signals separated by the Fast-ICA algorithm: such as Figure 9 As shown, Figure 9 The diagram shows the original modulated signal 3, the original independent signal 902 corresponding to the original modulated signal 3 separated by the Fast-ICA algorithm, and the original independent signal 903 corresponding to the original modulated signal 3 separated by the AE-ICA algorithm of this invention. Here, 901 represents the original modulated signal 3. It should be noted that, for example, the original modulated signal 3 can be aliased with another original modulated signal to obtain two aliased signals. Separating these two aliased signals using the corresponding algorithm yields the original independent signal corresponding to the original modulated signal 3.

[0088] like Figure 9 As shown, the phase of the original independent signal corresponding to the original modulation signal 3 separated by the Fast-ICA algorithm is reversed compared to the phase of the original modulation signal 3; however, the difference between the phase of the original independent signal corresponding to the original modulation signal 3 separated by the AE-ICA algorithm of this invention and the phase of the original modulation signal 3 is small.

[0089] Meanwhile, a comprehensive comparison was conducted between the Fast-ICA algorithm and the AE-ICA algorithm of this invention on a test set (a total of 2358 signal pairs). Specifically, the mean L2 distance between the separation results of the 2358 signal pairs obtained using the Fast-ICA algorithm (i.e., the original independent signals after separation) and the corresponding original independent signals to be aliased (i.e., the signals before aliasing) was calculated, as was the mean L2 distance between the separation results of the 2358 signal pairs obtained using the AE-ICA algorithm of this invention and the corresponding original independent signals to be aliased, as shown in Table 1. Here, the signal pairs include signal 1 and signal 2, and the L2 distance is the Euclidean distance.

[0090] Table 1

[0091]

[0092] As shown in Table 1, the quality of the signal separated by the AE-ICA algorithm of this invention is better than that of the Fast-ICA algorithm (with a significant improvement).

[0093] In summary, the experimental results show that the present invention can more accurately separate the original independent signal from the aliased signal, which is of great significance for speech or signal recognition.

[0094] This application also provides a blind source separation device, including: a memory and a processor, the memory and the processor being connected via a bus communication connection, the memory storing a computer program that can run on the processor, thereby implementing the steps of the blind source separation method based on autoencoder joint fast independent component analysis disclosed in this application.

[0095] This application also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the blind source separation method based on autoencoder joint fast independent component analysis disclosed in this application.

[0096] This application also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the blind source separation method based on autoencoder joint fast independent component analysis as disclosed in this application.

[0097] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0098] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, storage media, and program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A blind source separation method based on autoencoder combined with fast independent component analysis, characterized in that, include: Construct and train an autoencoder-based AE-ICA network; The input aliased signal to be separated is blindly separated using an AE-ICA network to obtain the recovered original independent signal. The autoencoder-based AE-ICA network includes an Encoder network, a Fast-ICA network, and a Decoder network; The method for training the autoencoder-based AE-ICA network includes: (1) Generate an aliased signal based on the original independent signals to be aliased; (2) The aliased signal is input into the Fast-ICA network for blind source separation to obtain the original independent signal to be recovered; (3) Input the aliased signal into the Encoder network to obtain the latent spatial representation of the aliased signal; (4) The latent space representation is fused with the original independent signal to be recovered, and the fused signal is then input into the Decoder network to obtain the recovered original independent signal; (5) Calculate the loss between the recovered original independent signal and the original independent signal to be aliased and perform backpropagation to optimize the parameters of the autoencoder-based AE-ICA network; (6) Iteratively optimize the parameters of the autoencoder-based AE-ICA network until the number of iterations reaches a preset number and stop iterating, or determine whether to continue iteratively optimizing the parameters of the autoencoder-based AE-ICA network based on the difference between the recovered original independent signal corresponding to the current iteration and the original independent signal to be aliased.

2. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1, characterized in that, The Encoder network comprises a first LSTM network, a linear layer, a ReLU activation function, and another linear layer connected in sequence. The first LSTM network consists of 3 LSTM layers, with an input dimension of 2 (number of receive channels) and 256 hidden layers. The Decoder network comprises a linear layer (Linear), an activation function (ReLU), a linear layer (Linear), a second LSTM network, and a linear layer (Linear) connected in sequence. The second LSTM network consists of 3 LSTM layers, with an input dimension of 2 (number of receive channels) and 256 hidden layers.

3. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1 or 2, characterized in that, Before inputting the aliased signal into the Fast-ICA network, the method further includes: The aliased signal is decentered so that the mean of the processed signal is 0.

4. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1 or 2, characterized in that, Before performing blind source separation on the aliased signal, the Fast-ICA network further includes: The aliased signal is preprocessed by "whitening".

5. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1 or 2, characterized in that, The fusion of the latent spatial representation with the original independent signal to be recovered is a dot product fusion.

6. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1 or 2, characterized in that, The formula for calculating the loss between the recovered original independent signal and the original independent signal to be aliased is as follows: in, This represents the loss between the recovered original independent signal and the original independent signal to be aliased. S represents the original independent signal after recovery, and S represents the original independent signal to be aliased. It is represented as the 2-norm of a vector, where n represents the total number of original independent signals to be aliased, and i represents the index of the original independent signal.

7. The blind source separation method based on autoencoder joint fast independent component analysis according to claim 1 or 2, characterized in that, The decision to continue iteratively optimizing the parameters of the autoencoder-based AE-ICA network is based on the difference between the recovered original independent signal and the original independent signal to be aliased in the current iteration, including: If the difference between the recovered original independent signal and the original independent signal to be aliased in the current iteration is less than a preset threshold, then the iterative optimization of the parameters of the autoencoder-based AE-ICA network is stopped.

8. A blind source separation device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the blind source separation method based on autoencoder joint fast independent component analysis as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the blind source separation method based on autoencoder joint fast independent component analysis as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the blind source separation method based on autoencoder joint fast independent component analysis as described in any one of claims 1 to 7.

Images