Lightweight design method of ICA-R algorithm and related assembly

Abstract

The invention discloses a lightweight design method of an ICA-R algorithm and related components, and relates to the field of signal processing. According to the method, prior information of a target signal is introduced to construct a reference signal, a separation vector is initialized based on a minimum mean square error criterion, iterative optimization of an ICA-R algorithm is realized on an FPGA, dependence on complex linear algebraic operations such as pseudo-inverse, matrix product and Cholesky decomposition is not needed, and occupation of computing resources and logic resources of the FPGA is reduced. Meanwhile, the initialization mode under the guidance of the prior information avoids the problem of local optimum caused by random initialization, and improves the precision and convergence speed of target signal extraction. Visibly, occupation of computing resources and logic resources of the FPGA can be reduced, the FPGA can extract the target signal from the observation signal more quickly, and the parallel processing advantage of the FPGA is combined, so that the signal processing capacity of the wireless communication equipment and the signal transmission reliability in a wireless communication scene are improved.

Description

Technical Field

[0001] The present invention relates to the field of signal processing, and in particular to a lightweight design method for an ICA-R algorithm and related components. Background Art

[0002] With the widespread deployment of wireless communication systems in radio stations, mobile terminals, and satellite communications, co-channel interference in complex electromagnetic environments is becoming increasingly serious, severely impacting communication quality. To effectively extract signals, blind source separation (BSS) algorithms have become a key anti-interference technology. Independent component analysis (ICA), a typical BSS algorithm, can separate mixed signals without requiring prior information. However, this separation process requires simultaneous extraction of all signal sources, resulting in significant computational redundancy and inaccurate target signal separation, making it difficult to meet the demands of real-time communication under resource-constrained conditions.

[0003] To improve separation performance, the reference independent component analysis (ICA-R) algorithm constructs a reference signal by introducing prior information of the target signal (such as frequency, symbol rate, phase, etc.). This is used as an optimization constraint to directly extract the target signal of interest and reduce the separation operations of unnecessary signals, thereby significantly reducing the algorithm execution time and system complexity, becoming the preferred BSS solution in complex communication systems.

[0004] However, in practical engineering applications, the performance of the ICA-R algorithm is highly dependent on the initialization strategy for the separation vector. Traditional random initialization methods suffer from problems such as being prone to falling into local optimality, poor separation accuracy, and slow convergence. While non-random initialization methods based on prior information can improve accuracy, they often involve complex matrix operations such as pseudo-inverse calculations of the observation signal matrix and Cholesky decomposition. This results in high computational complexity, lengthy data paths, high resource consumption, and complex logic control when implemented on hardware platforms such as field-programmable gate arrays (FPGAs). These algorithms struggle to meet the signal processing module requirements of low-power, high-real-time wireless communication devices, resulting in slow signal processing and transmission rates.

[0005] Therefore, how to reduce resource usage and computational delay on the FPGA platform and improve the signal processing and transmission rate while ensuring the target signal separation accuracy has become a key technical problem that needs to be solved urgently in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a lightweight design method and related components of the ICA-R algorithm, which can reduce the occupation of FPGA computing resources and logic resources, enable the FPGA to extract the target signal from the observed signal more quickly, and combine the parallel processing advantages of the FPGA to improve the signal processing capability of wireless communication equipment and the signal transmission reliability in wireless communication scenarios.

[0007] In a first aspect, the present application provides a lightweight design method for an ICA-R algorithm, which is applied to an FPGA in a wireless communication device, the method comprising:

[0008] Acquire an observation signal and a reference signal containing prior information of a target signal; the observation signal is a signal received by the wireless communication device and sent and mixed by multiple other wireless communication devices, and the observation signal includes a target signal to be extracted and at least one interference signal;

[0009] Initializing a separation vector according to a minimized mean square error between an estimated signal and the reference signal; the estimated signal is calculated based on the reference signal and the observation signal;

[0010] The initialized separation vector is used to perform iterative optimization of the ICA-R algorithm to extract the target signal from the observed signal.

[0011] In a second aspect, the present application provides a lightweight design system for an ICA-R algorithm, which is applied to an FPGA in a wireless communication device. The method includes:

[0012] an acquisition module, configured to acquire an observation signal and a reference signal containing prior information of a target signal; the observation signal is a signal received by the wireless communication device and transmitted and mixed by multiple other wireless communication devices, and the observation signal includes a target signal to be extracted and at least one interference signal;

[0013] an initialization module, configured to initialize a separation vector based on a minimized mean square error between an estimated signal and the reference signal; the estimated signal is calculated based on the reference signal and the observation signal;

[0014] The extraction module is used to perform iterative optimization of the ICA-R algorithm using the initialized separation vector to extract the target signal from the observation signal.

[0015] In a third aspect, the present application provides a lightweight design device for an ICA-R algorithm, comprising:

[0016] Memory for storing computer programs;

[0017] The processor is configured to implement the steps of the lightweight design method of the ICA-R algorithm as described above when executing the computer program.

[0018] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the lightweight design method of the ICA-R algorithm as described above.

[0019] The present invention provides a lightweight design method and related components of an ICA-R algorithm, which relate to the field of signal processing. By introducing prior information of a target signal to construct a reference signal, initializing a separation vector based on the minimum mean square error criterion, and implementing iterative optimization of the ICA-R algorithm on an FPGA, there is no need to rely on complex linear algebra operations such as pseudo-inverse, matrix product, and Cholesky decomposition, thereby reducing the occupation of computing resources and logic resources of the FPGA. At the same time, the initialization method guided by prior information avoids the local optimal problem caused by random initialization, and improves the accuracy and convergence speed of target signal extraction. It can be seen that the present application can reduce the occupation of computing resources and logic resources of the FPGA, so that the FPGA can extract the target signal from the observed signal faster, and combined with the parallel processing advantage of the FPGA, it improves the signal processing capability of wireless communication equipment and the signal transmission reliability in wireless communication scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0021] Figure 1 A schematic diagram of an ICA-R algorithm processing module provided by the present invention;

[0022] Figure 2 A flowchart of an ICA-R algorithm provided by the present invention;

[0023] Figure 3 This is a flow chart of a lightweight design method for an ICA-R algorithm provided by the present invention;

[0024] Figure 4 This is a schematic diagram of the theoretical computational complexity changing with n when m=20 provided by the present invention;

[0025] Figure 5 This is a schematic diagram of the theoretical computational complexity changing with m when n is fixed at 1000 provided by the present invention;

[0026] Figure 6 A schematic diagram of a lightweight design system for an ICA-R algorithm provided by the present invention;

[0027] Figure 7 A schematic diagram of a lightweight design device for an ICA-R algorithm provided by the present invention;

[0028] Figure 8A schematic diagram of a computer-readable storage medium provided by the present invention. DETAILED DESCRIPTION

[0029] The core of this invention is to provide a lightweight design method for the ICA-R algorithm and related components, which can reduce the occupation of FPGA computing resources and logic resources, enabling the FPGA to extract the target signal from the observed signal more quickly. Combined with the parallel processing advantages of the FPGA, it improves the signal processing capability of wireless communication equipment and the signal transmission reliability in wireless communication scenarios.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0031] Before describing this application, the principles of BSS (Blind Source Separation) and ICA-R are introduced.

[0032] BSS, a popular technology in signal processing, aims to recover unknown independent source signals from multi-channel observations. Its core challenge lies in the dual unknowns of the mixing matrix and the source signals. This technology has demonstrated significant application value in a variety of fields. In wireless communications, massive MIMO systems face the problem of multi-user signal aliasing. BSS improves spectral efficiency by separating multi-user signals, enabling efficient channel equalization and interference cancellation. In biomedical engineering, electroencephalogram (EEG) signals contain a large number of mixed signals of neuronal activity. BSS can isolate specific EEG components (such as alpha and beta waves), providing key technical support for epileptic lesion localization and brain-computer interfaces. In the industrial Internet of Things (IIoT), multi-source vibration signals collected by sensor networks are often contaminated by environmental noise. BSS separates equipment vibration signals for fault diagnosis and equipment health monitoring. In acoustic processing, for speech enhancement in noisy environments, BSS can effectively separate target speech from background noise, improving speech recognition accuracy and communication quality. These application scenarios place stringent demands on the robustness, computational efficiency, and real-time performance of BSS algorithms.

[0033] Mathematically, the BSS model can be described as: , (1).

[0034] in, is the matrix of the observation signal (m is the number of observation channels, n is the number of sampling points), is the unknown mixing matrix, is a statistically independent source signal matrix, is the additive noise matrix. Its essence is to estimate the separation matrix W through matrix inversion operation so that the estimated signal Approximate the source signal S.

[0035] Independent Component Analysis (ICA) is the most commonly used BSS algorithm. ICA's theoretical basis is the assumption of statistical independence of signals. It achieves signal separation by optimizing objective functions (such as maximizing non-Gaussianity, minimizing mutual information, or maximizing negative entropy). BSS algorithms based on ICA are typically complex and have low accuracy in noisy environments. Therefore, they struggle to meet the high real-time and high reliability requirements of communication scenarios, limiting their practical application in real-world communication systems.

[0036] It's worth noting that wireless communication is an interactive process that follows a predetermined protocol. This means that the receiver typically has some prior information about the communication signal and often only needs to focus on that specific signal. Independent Component Analysis with Reference (ICA-R) incorporates prior information about the target signal as a reference signal into the ICA constraints. With the help of the reference signal, ICA-R can search for and extract the target signal more quickly and accurately. Furthermore, the ICA-R algorithm directly extracts the target signal, eliminating the need to separate all source signals and then identify them, as ICA does. This significantly reduces the algorithm's execution time and simplifies the complexity of the separation system, making ICA-R a preferred BSS algorithm for communication systems in engineering applications.

[0037] However, in the ICA-R algorithm, the initial value setting of the separation vector has a crucial impact on the separation accuracy, convergence speed, and computational efficiency of the algorithm. Because the objective function of the ICA-R algorithm is usually non-convex, traditional gradient-based algorithms often fall into certain local optimal solutions. Even the same algorithm may give different separation results under different initial conditions. In addition, the quality of the initial value will directly affect the distance between the initial iteration point and the optimal point, and thus affect the convergence speed. A good initial value means faster convergence. Therefore, separation vector initialization is a key preprocessing step of the ICA-R algorithm. Its core role is to provide a high-quality initial iteration starting point for the subsequent optimization process, which directly affects the convergence speed, separation accuracy, and robustness of the algorithm.

[0038] The ICA-R-based blind signal extraction method can effectively address the shortcomings of the ICA-based blind source separation method. By introducing a reference signal containing prior information about the target signal, the ICA-R-based blind source separation method can accurately separate and extract a unique target signal. The key to the success of this method is the selection of an appropriate reference signal and the appropriate initialization of the separation vector. The reference signal must be a non-Gaussian signal containing prior information about the target signal. A review of existing public literature shows that the main methods for generating reference signals include the following: based on the signal's sign function, based on the statistical characteristics of the target signal, based on the signal frequency, etc.

[0039] The initialization methods of separation vectors mainly include random initialization methods and non-random initialization methods based on prior information. Conventional ICA-R algorithm uses random initialization methods, such as , m is the number of mixed source signals. Randomly selecting the initial value of the mixing matrix will cause the algorithm to fall into a local minimum during convergence, resulting in low separation accuracy, slow convergence speed, and poor real-time performance of the algorithm.

[0040] like Figure 1 and Figure 2 , Figure 1 This is a schematic diagram of the ICA-R algorithm processing module provided by the present invention. The main idea is to introduce a proximity measure between the estimated signal y(t) and the reference signal r(t) into the difference function of traditional ICA. The separation vector is then learned in a loop-iterative manner. After the separation vector converges, it is multiplied by the observed signal x(t) to achieve separation and extraction of the target signal s(t). The specific steps are as follows:

[0041] (1) Initialize Lagrange parameters ( ), learning rate ( ), randomly generate separation vectors ( ), and set the convergence criteria;

[0042] (2) Select the proximity metric function ( ), compares the proximity of the estimated signal y(t) to the reference signal r(t). When the output signal y(t) is the desired target signal s(t), the proximity metric reaches its minimum. The proximity metric can be selected from the following three forms:

[0043] , (2);

[0044] in, represents the mathematical expectation.

[0045] (3) Construct inequality constraints ( ). The output estimation signal and the reference signal are controlled to be close to each other within a certain threshold ( ), which is used to distinguish the target signal from other communication signals and is expressed as:

[0046] , (3);

[0047] (4) Constructing equality constraints ( ). This constraint can ensure that the separated and extracted signal has unit variance, and its expression is:

[0048] , (4);

[0049] (5) Construct the optimization objective function ( ). It can be a difference function based on negative entropy, and its expression is:

[0050] , (5);

[0051] in, is a positive constant, is any non-quadratic function, is a Gaussian variable with zero mean and unit variance.

[0052] (6) Using the augmentation method, the optimization problem under the above constraints is transformed into a Lagrangian parameter ( ). The extreme value is:

[0053] , (6);

[0054] In the formula and is the Lagrange multiplier corresponding to the two constraints, is the penalty parameter, is the Euclidean norm.

[0055] (7) By solving equation (5) for The extreme value of The iterative learning formula is:

[0056] , (7);

[0057] in, and They are right The first and second order partial derivatives of is the covariance matrix of the whitened observation matrix z, is the learning rate. At the same time, the Lagrange parameter and The iterative update formula is:

[0058] , (8);

[0059] (8) Iterate formula (6) and formula (7) and judge Whether the convergence criteria are met, if so, stop the loop calculation and save the final separation vector , otherwise continue iterative calculation.

[0060] (9) Finally, through the formula Calculation can directly obtain the reference signal The closest expected independent component , that is, the target signal The recovered samples are then separated and extracted, and the target signal is extracted.

[0061] When the existing technology uses the method of non-random initialization of separation vector to initialize the separation vector, it is assumed that the reference signal is the convergence target of the ICA-R algorithm. According to the signal extraction formula , the initialization formula of the separation vector can be obtained through the Moore-Penrose generalized inverse operation: , (9).

[0062] in called The Moore-Penrose generalized inverse matrix of is the observed signal after whitening. Substituting the expression into formula (9) yields: , (10).

[0063] Compared to the original ICA-R algorithm with random initialization, this method uses reference prior information not only in the contrast function but also in the initialization of the separation vector. It is worth noting that this method requires calculating the pseudo-inverse of the signal matrix, which is computationally complex and very complex for hardware implementation such as FPGAs, requiring significant computational and logic resources.

[0064] Specifically, completing the calculation of formula (10) requires completing the following four steps in sequence.

[0065] Step 1. Calculation :X is a The matrix, is a Matrix, calculate Get one Each element of the matrix The calculation of requires n multiplications and n-1 additions. Therefore, The computational complexity is .

[0066] Step 2. Calculation (Using Cholesky decomposition): First, Cholesky decomposition converts a The symmetric positive definite matrix Decompose into ,in is a lower triangular matrix. The computational complexity of this process is approximately So the solution is Equivalent to solving , requires two inversion processes of the triangular matrix, so the calculation The total computational complexity is approximately .

[0067] Step 3. Calculation : is a The matrix, is a Column vector of Get one A column vector of . Each element of this column vector is A row vector and The inner product of the column vectors needs to be times multiplication and Additions. Since the result vector has elements, a total of times multiplication and Therefore, the computational complexity of this step is approximately .

[0068] Step 4. Calculation : is a The matrix, is a Column vectors, calculate their product to get a Column vector of This matrix-vector multiplication requires times multiplication and Therefore, the computational complexity of this step is approximately .

[0069] Adding up the multiplication calculations of each of the above steps, the total multiplication calculation amount is approximately: .when When , the total computational complexity of the algorithm is , that is, calculation of Complexity will become the main computational bottleneck.

[0070] In addition, the matrix inversion in step 2 is a relatively computationally expensive operation, especially when the matrix dimension m is large, even with the relatively efficient Cholesky decomposition. The complexity will become a bottleneck.

[0071] Mapping the above computational steps to FPGA hardware presents the following challenges:

[0072] (1) Calculation A large number of multipliers and adders are required to execute in parallel to increase the computing speed. and The multiplication operation at this level consumes a lot of FPGA logic cells, especially when m and n are large.

[0073] (2) Calculation using the Cholesky decomposition method Cholesky decomposition and triangular matrix inversion involve division and square root operations. These operations are typically more complex to implement on an FPGA than multiplication and addition, requiring more resources and increasing latency. Implementing the matrix inversion unit may require complex iterative algorithms or lookup tables, consuming significant resources. Furthermore, Cholesky decomposition is a serial process, with subsequent calculations dependent on the results of previous ones. This makes it difficult to fully parallelize the process, limiting computational speed improvements.

[0074] (3) To improve throughput, the computation process needs to be pipelined. However, the dependency between matrix multiplication and matrix inversion complicates pipeline design and may introduce long delays.

[0075] (4) Implementing high-precision floating-point operations on FPGAs consumes a lot of resources. If fixed-point operations are used, careful consideration must be given to numerical quantization and overflow issues to ensure computational accuracy. Matrix inversion is sensitive to numerical precision, and small numerical errors may be magnified.

[0076] In other words, initializing the separation vector using the above algorithm consumes a significant amount of hardware resources. Implementing a large number of parallel multipliers and adders consumes significant FPGA logic resources (LUTs, FFs), especially for large m and n values. This results in high computational latency. The serial nature of matrix inversion, complex data dependencies, and complex operations (such as division and square root) can also lead to long computational latency, making it difficult to meet the requirements of real-time applications. Complex logic control design is also required to implement complex matrix operations and data flow management, increasing the design complexity.

[0077] Based on this, Figure 3 This application provides a lightweight design method for the ICA-R algorithm, simplifying the separation vector initialization formula of the ICA-R algorithm to the product of the observation signal and the reference signal. Based on the principle of statistical independence of signal sources, this method aims to solve the interference problem caused by multiple signal sources operating at the same frequency in wireless communication systems. It is particularly suitable for resource-constrained FPGA platforms.

[0078] Specifically, the method includes:

[0079] S11: Acquire an observation signal and a reference signal containing prior information of the target signal;

[0080] Specifically, two key inputs for subsequent signal separation are obtained: the observation signal and the reference signal. The observation signal is sent simultaneously by multiple other wireless communication devices and mixed in the channel, and contains the target signal and one or more interference signals. The reference signal is an auxiliary signal constructed based on the prior information of the target signal, which is used to guide the separation process to converge to the target signal. The prior information can be its carrier frequency, phase, modulation type or symbol rate, etc. The observation signal is usually represented as a multidimensional time series matrix, denoted as , where m is the number of observation channels and n is the number of sampling points.

[0081] The purpose of this step is to provide dual information input containing both mixed information and the directionality of the target signal at the beginning of the signal processing process, providing a basis for subsequent minimum mean square error calculation and vector initialization, thereby enhancing the ICA-R algorithm's ability to extract the directionality of the target signal and laying the foundation for the entire lightweight design.

[0082] S12: Initialize the separation vector based on the minimized mean square error between the estimated signal and the reference signal; the estimated signal is calculated based on the reference signal and the observed signal;

[0083] Specifically, based on the minimum mean square error criterion, the correlation between the observed signal and the reference signal is used to initialize the separating vector in the ICA-R algorithm. Specifically, by constructing an error signal between the estimated signal (i.e., the output of the current separating vector acting on the observed signal) and the reference signal, and using the mean square value of this error as the optimization target, the initial value of the separating vector that minimizes the error is determined. This allows the algorithm to initially focus on the target signal, effectively improving the convergence speed and separation accuracy of subsequent iterations while avoiding falling into local optima. This process avoids complex matrix inversion and decomposition operations, achieving efficient and low-complexity initialization of the separating vector.

[0084] S13: Using the initialized separation vector, the ICA-R algorithm is iteratively optimized to extract the target signal from the observed signal.

[0085] Specifically, based on the initial separation vector that has been obtained, the separation effect is further improved through an iterative optimization process in the FPGA platform, and the target signal is accurately extracted from the observed signal. Specifically, the objective function of the ICA-R algorithm (such as maximum non-Gaussianity or minimum mean square error) is used to continuously adjust the direction and amplitude of the separation vector so that the estimated signal gradually approaches the reference signal or meets the independence requirements. In each iteration, the error or independence index is evaluated based on the current separation result, and the separation vector is updated accordingly until the preset convergence condition is met. This process effectively improves the separation accuracy and stability of the algorithm and achieves accurate extraction of the target signal. This method reduces complex operations such as matrix pseudo-inverse and Cholesky decomposition, improves hardware implementation efficiency, and is suitable for communication scenarios with real-time performance and low resource requirements.

[0086] In one embodiment, the wireless communication device is a satellite communication terminal deployed at a ground receiving station, configured to receive downlink data signals from multiple near-Earth satellites. Under frequency reuse conditions, multiple satellites may operate simultaneously in adjacent orbits, causing co-channel interference in downlink signals at the receiving end.

[0087] The satellite communication terminal receives a mixed observation signal through an FPGA chip. This mixed observation signal is a linear superposition of modulated signals emitted by multiple satellite terminals, including the communication signal of the target satellite and the signal of at least one interfering satellite. Because the downlink frequency, symbol rate, and modulation parameters of the target satellite are known, a reference signal can be constructed based on this prior information.

[0088] By using the mean squared error between the estimated and reference signals as an optimization criterion, the FPGA module initializes the separation vector, eliminating the need for large-scale matrix inversion operations and significantly reducing computational latency. Subsequently, the FPGA performs iterative ICA-R operations based on this initial vector, enabling rapid and accurate separation of the target satellite signal. This improves communication quality and stability at the receiving end in complex interference environments, making it particularly suitable for dense deployments and bandwidth-constrained scenarios.

[0089] The above is only a specific application scenario provided by the embodiment of the present application. The wireless communication device may also be a radio station or a mobile communication terminal.

[0090] In an exemplary embodiment, initializing the separation vector according to minimizing the mean square error between the estimated signal and the reference signal includes:

[0091] According to minimizing the mean square error between the estimated signal and the reference signal, an optimization model is constructed, and the optimization model includes an objective function and constraint conditions;

[0092] According to the optimization model calculation, the initialized separation vector is obtained;

[0093] The expression of the optimization model is: ,in, represents the mathematical expectation, is the initial separation vector, represents the transposed matrix of the initialized separation vector, X is the observed signal, r is the reference signal, st represents the constraint condition, and min MSE is the minimized mean square error.

[0094] Specifically, in order to make the estimated signal obtained after the separation vector is initialized as close to the reference signal as possible, the mean square error between the estimated signal and the reference signal is used as the objective function, and the degree of closeness between the estimated signal and the reference signal is ensured by minimizing the error, because the mean square error can effectively measure the difference between the two and guide the optimization direction; at the same time, the additional unit variance constraint condition is to ensure the numerical stability of the separated signal and avoid the subsequent separation effect being affected by the abnormal signal amplitude, so as to derive the initial value of the separation vector that can both meet the error minimization and ensure the signal specification, and realize the optimal design of the initialization process.

[0095] That is, the process of determining the initialization separation vector in this application is the problem of solving the above-mentioned optimization model.

[0096] In an exemplary embodiment, obtaining an initialized separation vector by calculation according to the optimization model includes:

[0097] According to the optimization model, the augmented Lagrangian function is constructed by the Lagrangian multiplier method; the expression of the augmented Lagrangian function is: ;in, For constraints The Lagrange multiplier of is the penalty factor, is the Euclidean form, is the quadratic penalty term, MSE is the mean square error;

[0098] The expression of the augmented Lagrangian function is about Find the partial derivative and set it equal to zero;

[0099] Solve the initialized separation vector based on the partial derivative being equal to zero .

[0100] Specifically, based on the core idea of ​​the Lagrange multiplier method, this embodiment transforms the constrained optimization problem into an unconstrained extreme value problem for solution. First, in order to minimize the objective function of the mean square error between the estimated signal and the reference signal, and the constraint that the separated signal must satisfy the unit variance, the Lagrange multiplier λ and the penalty factor σ are introduced to construct an augmented Lagrangian function. The function strengthens the satisfaction of the constraint through the penalty term to ensure the standardization of the signal during the optimization process; then, the partial derivative of the augmented Lagrangian function with respect to the separation vector is calculated and set to zero. Through mathematical derivation, the influence of the Lagrange multiplier is eliminated, and the linear relationship expression between the separation vector and the whitened observation signal matrix and the reference signal is obtained, thereby deriving the separation vector initialization formula that only requires the multiplication of the whitened observation signal matrix and the transpose of the reference signal, thereby significantly reducing the computational complexity and improving the engineering feasibility.

[0101] In an exemplary embodiment, the initial separation vector is solved based on the partial derivative being equal to zero. , including: solving the intermediate form of the separating vector according to the partial derivative being equal to zero, the expression of the intermediate form of the separating vector is: ; Normalize the intermediate separation vector to obtain the initialized separation vector.

[0102] Specifically, when deriving the separation vector initialization formula, the principle of normalization is to rescale the initially obtained separation vector so that it satisfies normalization constraints such as unit variance or unit norm, eliminating the influence of parameters such as Lagrange multipliers on the vector amplitude, and ensuring the numerical stability and consistency of the initialization vector. This step avoids numerical anomalies caused by parameter differences by adjusting the vector to a standard scale, while simplifying the formula expression so that the final initialization formula only retains the core linear operation relationship between the whitened observation signal matrix and the reference signal, thereby improving computational efficiency and engineering practicality while ensuring algorithm accuracy.

[0103] In an exemplary embodiment, the expression of the initialized separation vector is: ;in, is the initial separation vector, X is the observation signal, and r is the reference signal.

[0104] Compared with the initialization formula of separation vector in the prior art (10) ( ) It can be seen that the initialization formula (16) of the separation vector of this application ( ) omits the calculation of matrix multiplication and then inversion ( ) This complex operation is not only simpler and clearer, but also has lower computational complexity and is easier to implement in engineering.

[0105] Specifically, X is a The matrix, is a Column vector of Get one Column vector of Each element of this matrix-vector multiplication requires n multiplications and n-1 additions. Since the result vector has m elements, a total of m×n multiplications and Therefore, the computational complexity of formula (16) is approximately .

[0106] Compared with formula (10), formula (16) has the following advantages in terms of computational complexity:

[0107] (1) Order of magnitude reduction: The computational complexity of formula (16) Much lower than the main complexity term of formula (10) and (when (2) Avoids matrix inversion: Formula (16) directly calculates the matrix-vector multiplication, avoiding the calculation The inverse matrix of , thus avoiding the Cholesky decomposition complexity.

[0108] It can be seen that the initialization formula (16) of the separation vector of the present application significantly reduces the computational complexity and avoids the inverse calculation of the matrix, which can simplify the calculation process and improve the computational efficiency.

[0109] Accordingly, FPGA hardware implementation has the following advantages.

[0110] (1) Fewer computing units are required. multipliers and Adders (which can be parallelized) avoid the implementation of complex Cholesky decomposition units, matrix inversion units, and additional matrix multiplication units compared to formula (10). (2) Lower computational latency: Since there are fewer computational steps, the complex matrix inversion process is avoided, and there is no data dependency, lower computational latency can be achieved. (3) Simpler control logic: The control logic of matrix-vector multiplication is relatively simple, making it easier to implement data flow scheduling and synchronization on FPGAs. (4) Lower resource consumption: Since fewer computational units and storage resources are required, formula (16) typically consumes less logic resources (LUTs, FFs,) and on-chip storage resources when implemented on FPGAs.

[0111] Compared with formula (10), formula (16) has overwhelming advantages in both mathematical computation complexity and FPGA hardware implementation, because it avoids the highly complex matrix inversion operation, reduces the amount of computation, computational delay, hardware resource consumption, and control logic design, and can achieve higher throughput and real-time performance.

[0112] At the same time, like formula (10), formula (16) also includes the reference signal Therefore, the separation vector initialization method derived based on the embodiment of the present application can also improve the accuracy of the ICA-R algorithm to about 95% like formula (10). Compared with the random initialization method, the robustness of the ICA-R algorithm is greatly improved.

[0113] The computational complexity of the two separation vector initialization formulas (Equation (10) and Equation (16)) is shown in Table 1.

[0114] Table 1 Comparison of computational complexity of two separation vector initialization formulas

[0115]

[0116] The computational complexity of the two separation vector initialization formulas (10) and (16) varies with the dimension m of the matrix X and the number of samples n as shown in the following curves: Figure 4 and Figure 5 As shown in the figure, when the dimension of the fixed matrix X is m=20, both formulas change linearly with the number of samples, but the slope of formula 10 is steeper, which is higher than that of formula 16. times. When the number of samples n is fixed at 1000, Formula 16 varies linearly with the matrix dimension m, while Formula 10 varies quadratically with the matrix dimension m. It can be seen that in either case, the computational complexity of Formula 16 is lower than that of Formula 10, and the difference becomes more significant as the values ​​of m and n increase.

[0117] This application makes a lightweight design of the robust ICA-R algorithm, and changes the formula of separating vector initialization from Simplified to In terms of computational complexity, it avoids the need for large amounts of computational resources. and its complex inversion process, the computational complexity is reduced from down to , which is reduced by m times, significantly reducing the computational complexity. Accordingly, when implemented in hardware on an FPGA, it can bring benefits such as reducing the demand for computing units, obtaining lower computational latency, simplifying logic control design, and reducing logic resource consumption. It is especially suitable for the real-time processing requirements of communication systems and resource-constrained platforms of communication terminals. In addition, the separation vector initialization formula 16 of the present application has better numerical stability, avoids computational errors caused by matrix pathology, and naturally supports parallel computing (such as SIMD instructions or GPU acceleration), which makes it more adaptable in edge devices and distributed systems.

[0118] Therefore, the separation vector initialization formula proposed in this application offers greater engineering feasibility and efficiency advantages in resource-constrained scenarios, large-scale data, or real-time systems, and is particularly suitable for interference signal separation in communication systems. In other words, the separation vector initialized in this application can significantly reduce computational complexity and resource consumption of the hardware devices (such as FPGAs) that carry out these computations.

[0119] Suitable, such as Figure 6 , this application also provides a lightweight design system of the ICA-R algorithm, including:

[0120] An acquisition module 61 is configured to acquire an observation signal and a reference signal containing prior information of a target signal;

[0121] Initialization module 62, configured to initialize a separation vector based on a minimized mean square error between an estimated signal and a reference signal; the estimated signal is calculated based on the reference signal and the observed signal;

[0122] The extraction module 63 is used to perform iterative optimization of the ICA-R algorithm using the initialized separation vector to extract the target signal from the observation signal.

[0123] For an introduction to the lightweight design system of the ICA-R algorithm, please refer to the above embodiment, and this application will not go into details here.

[0124] Suitable, such as Figure 7 , this application also provides a lightweight design device for the ICA-R algorithm, including:

[0125] Memory 101, used for storing computer programs;

[0126] The processor 102 is configured to implement the steps of the aforementioned lightweight design method for the ICA-R algorithm when executing the computer program.

[0127] For example, the specific implementation of the lightweight design device of the ICA-R algorithm can be a hardware device with computing capabilities such as FPGA, or other hardware, and this application does not limit it here.

[0128] For an introduction to the lightweight design device of the ICA-R algorithm, please refer to the above embodiment, and this application will not go into details here.

[0129] Suitable, such as Figure 8 The present application also provides a computer-readable storage medium 201, on which a computer program 202 is stored. When the computer program 202 is executed by a processor, the steps of the lightweight design method of the above-mentioned ICA-R algorithm are implemented.

[0130] For an introduction to computer-readable storage media, please refer to the above embodiments, and this application will not go into details here.

[0131] It should also be noted that, in this specification, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0132] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightweight design method for the ICA-R algorithm, characterized in that: The method for an FPGA used in a wireless communication device includes: Acquire an observation signal and a reference signal containing prior information of a target signal; the observation signal is a signal received by the wireless communication device and sent and mixed by multiple other wireless communication devices, and the observation signal includes a target signal to be extracted and at least one interference signal; Initializing a separation vector according to a minimized mean square error between an estimated signal and the reference signal; the estimated signal is calculated based on the reference signal and the observation signal; The initialized separation vector is used to perform iterative optimization of the ICA-R algorithm to extract the target signal from the observed signal.

2. The lightweight design method of the ICA-R algorithm according to claim 1, characterized in that: The wireless communication device is a radio station, a mobile communication terminal, or a satellite communication terminal.

3. The lightweight design method of the ICA-R algorithm according to claim 1, characterized in that: The expression of the initialized separation vector is: ;in, is the initialized separation vector, X is the observation signal, and r is the reference signal.

4. The lightweight design method of the ICA-R algorithm according to claim 1, characterized in that: Initializing a separation vector according to a minimized mean square error between the estimated signal and the reference signal, comprising: Constructing an optimization model based on minimizing the mean square error between the estimated signal and the reference signal, wherein the optimization model includes an objective function and constraints; Calculating according to the optimization model to obtain an initialized separation vector; The expression of the optimization model is: ,in, represents the mathematical expectation, is the initial separation vector, represents the transposed matrix of the initialized separation vector, X is the observation signal, r is the reference signal, st represents the constraint condition, and min MSE is the minimized mean square error.

5. The lightweight design method of the ICA-R algorithm according to claim 3, characterized in that: The initial separation vector is obtained by calculation according to the optimization model, including: According to the optimization model, an augmented Lagrangian function is constructed by the Lagrangian multiplier method; the expression of the augmented Lagrangian function is: ;in, For constraints The Lagrange multiplier of is the penalty factor, is the Euclidean form, is the quadratic penalty term, MSE is the mean square error; An initialized separation vector is calculated according to the augmented Lagrangian function.

6. The lightweight design method of the ICA-R algorithm according to claim 5, characterized in that: Calculating an initialized separation vector according to the augmented Lagrangian function includes: The expression of the augmented Lagrangian function is about Find the partial derivative and set it equal to zero; Solve the initialized separation vector based on the partial derivative being equal to zero .

7. The lightweight design method of the ICA-R algorithm according to claim 5, characterized in that: Solve the initialized separation vector based on the partial derivative being equal to zero ,include: The intermediate form of the separating vector is solved according to the partial derivative being equal to zero. The expression of the intermediate form of the separating vector is: ; The intermediate separation vector is normalized to obtain an initialized separation vector.

8. A lightweight design system based on ICA-R algorithm, characterized in that: The method for an FPGA used in a wireless communication device includes: an acquisition module, configured to acquire an observation signal and a reference signal containing prior information of a target signal; the observation signal is a signal received by the wireless communication device and transmitted and mixed by multiple other wireless communication devices, and the observation signal includes a target signal to be extracted and at least one interference signal; an initialization module, configured to initialize a separation vector based on a minimized mean square error between an estimated signal and the reference signal; the estimated signal is calculated based on the reference signal and the observation signal; The extraction module is used to perform iterative optimization of the ICA-R algorithm using the initialized separation vector to extract the target signal from the observation signal.

9. A lightweight design device for ICA-R algorithm, characterized in that: include: Memory for storing computer programs; A processor is configured to implement the steps of the lightweight design method for the ICA-R algorithm according to any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the lightweight design method of the ICA-R algorithm according to any one of claims 1 to 7.

Images