A robust multimodal beamforming design method based on liquid neural network

Abstract

The present invention discloses a robust multi-modal beamforming design method based on liquid neural network, which aims to optimize the spectrum efficiency in communication systems and cope with dynamically changing channel conditions. The method includes steps such as data acquisition, variable initialization, forward propagation, back propagation, multiple iterations and continuous operation. By integrating manifold learning and liquid neural network technology, an adaptive beamforming strategy in a multi-user communication environment is realized, which is particularly suitable for communication scenarios with complex channel conditions and high dynamics. This method not only reduces the complexity of algorithm deployment, but also does not require pre-training, and can learn and adjust in real time when receiving changing channel samples, significantly improving the algorithm's ability to respond quickly to various channel states and user needs, and providing an efficient, flexible and easy-to-deploy solution for beamforming problems in communication systems, which is particularly suitable for modern communication networks that need to quickly adapt to complex and dynamic channel conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent wireless communication, and in particular to a robust multi-modal beamforming design method based on liquid neural network. Background Art

[0002] Faced with the growing demand for large-capacity, high-quality future wireless communication services, researchers have proposed the concept of 6G communication. In the next decade, AI and digital twins will form a dual-wheel drive to further promote technological breakthroughs. Among them, 6G will transcend human connection and Internet of Things, and move towards the ambitious goal of intelligent connection of all things. 6G communication technology is expected to support multiple communication modes, covering a wide range of application scenarios from ground to air and even space. These modes will not only include traditional mobile communications and Internet of Things connections, but will also expand to high-frequency communications, satellite network interconnection, and advanced virtual reality and augmented reality technologies, making 6G a key technology to support the future smart society.

[0003] However, in high-frequency bands such as millimeter-wave communications, beamforming faces many challenges in practical applications. The non-convexity of the problem, the deep coupling of variables, and the dependence of the optimization algorithm on accurate channel information have greatly increased the difficulty of the design. In addition, other technologies such as massive MIMO and network slicing also face their own challenges, such as the design complexity of the algorithm, the high dependence on channel state information, and the robustness issues in dynamic environments. Therefore, it is necessary to explore and solve how to design a robust beamforming strategy in multimodal 6G scenarios, especially considering the uncertainty and dynamic changes in practical applications, to optimize communication performance, reduce resource consumption, and improve the overall efficiency of the system. Summary of the invention

[0004] The present invention provides a robust multimodal beamforming design method based on liquid neural network, which can reduce the complexity of beamforming problems in multimodal communication systems including but not limited to multi-user, multiple-input and multiple-output, improve the overall spectrum efficiency of the system, enhance the robustness of the system to deal with imperfect channel information, and enhance the overall performance of the system in dynamic scenarios.

[0005] The system and method disclosed in the present invention embody a series of innovative features and aspects, which work together to reduce the complexity of beamforming problems in multi-modal communication systems including but not limited to multi-user, multi-input and multi-output, and improve system performance and efficiency. It should be made clear that the effectiveness and value of the present invention are not determined or fully responsible by any single aspect, angle or part.

[0006] Furthermore, the scope of the present invention should not be limited to the aspects or parts that directly and obviously achieve the expected effect. On the contrary, the scope of protection of the present invention should also be extended to all aspects and parts that indirectly contribute to the realization of the objectives and effects of the present invention through innovative combinations, configurations or synergies. In addition, even if some aspects or parts do not seem to have significant novelty or directly contribute to the expected effects of the present invention when considered alone, these aspects or parts, when combined with other aspects or parts of the present invention, may significantly enhance the overall performance and effects of the present invention in an unexpected manner. Therefore, these aspects and parts should be regarded as an indispensable part when evaluating the novelty and scope of the present invention.

[0007] The present invention provides a technical solution, and the purpose of the present invention can be achieved by the following technical solution: a robust multi-modal beamforming design method based on liquid neural network. The method comprises the following steps:

[0008] Step 1: Data acquisition: Assume that in the actual communication scenario there is a base station equipped with M antennas and K nodes each with N r The user terminal has a receiving antenna; channel data is collected from actual deployment scenarios, where the scenario is as follows Figure 1 As shown; Get the estimate of the channel from the base station to K users The channel of the kth user is (·) H Represents the conjugate transpose of a matrix or vector;

[0009] Step 2, variable initialization: first randomly initialize the precoding basis matrix X0 according to manifold learning and calculate the complete precoding matrix W0 through the formula, and finally initialize the internal parameters θ of the liquid neural network;

[0010] Step 3: Perform forward propagation: Calculate the sum rate based on the initialized precoding matrix and the estimated channel information, and find the gradient of the sum rate. Then input the gradient into the liquid neural network and record the output of the neural network as ΔX i , and update the precoding basis matrix through specific rules, where i represents the current i-th forward propagation;

[0011] Step 4: Back propagation: According to the X obtained by forward propagation i Calculate the complete precoding matrix W i , and calculate a specific loss function, and update the gradient of the liquid neural network according to the value of the loss function;

[0012] Step 5, multiple iterations: Repeat steps 3 and 4 for a total of N e The last neural network output is processed according to specific rules. Restore to precoding matrix Used for base station precoding on channel H in real scenarios;

[0013] Step 6, continuous operation: reacquire the next set of channels through step 1, do not reinitialize the parameters of the liquid neural network, and directly proceed to the process described in steps 3 to 5.

[0014] Furthermore, in step 2, the liquid neural network is composed of three layers of liquid neurons, namely, an input layer, a command layer, and an action layer. Each liquid neuron is Figure 2 As shown, it consists of a central skeleton, g, f, h network fulcrums and nerve tails. For each liquid neuron layer, the calculation formula satisfies:

[0015] p(t)=σ(-f(p,I;θ f )t)⊙g(p,I;θ g )+[1-σ(-f(p,I;θ f )t)]⊙h(p,I;θ h ),

[0016] Where p(t), g, f, h, and I represent the hidden state of the neuron layer, the internal function of each neuron g, f, and h neural network pivot, and the vector formed by the stacking operation of the input within the layer; θ g ,θ f ,θ h They represent the parameters of the neural network pivots g, f, and h of all liquid neurons in the layer respectively, and ⊙ represents the element-by-element product operation.

[0017] Furthermore, in step 2 and step 4, the precoding base matrix and the base station precoding matrix satisfy:

[0018] W=H H X,

[0019] The dimension of the precoding matrix is ​​M × K, and the dimension of the precoding base matrix is ​​N × K. Since the number of base station antennas M is generally much larger than the number of receiving antennas N at the user end, this method reduces the optimization space.

[0020] Furthermore, in step 3 and step 5, the updating rule of the precoding base matrix described is:

[0021]

[0022] The subscript i represents the variable corresponding to the i-th cycle described in step 5, P represents the sum of the maximum powers that can be emitted by all transmitting antennas of the base station, and Tr(·) represents the result of summing the elements in the diagonal direction of the matrix.

[0023] Furthermore, in step 4, the calculation formula of the spectrum efficiency and loss function is:

[0024]

[0025]

[0026] where σ 2 The variance of the additive Gaussian circular complex noise, I is a square matrix (i.e., a diagonal matrix) with all values ​​on the diagonal being 1 and all other values ​​being 0, det(·) represents the determinant of the matrix, and Var(R) represents the variance of the spectral efficiency of all users, (·) -1 Represents the inverse of the matrix, R k represents the spectral efficiency of the kth user, α k represents the weight of the kth user, w k represents the kth column of W, max k R k and min k R k They represent the spectral efficiency of the users with the largest and smallest spectral efficiency respectively. ReLU(·)=max(·,0) is the activation function used to punish the neural network for not falling into low performance. β and γ are hyperparameters whose purpose is to promote the neural network to balance the overall effect and allocate resources to more users.

[0027] Furthermore, the liquid neurons have hidden state memory capability, and the state of each neuron depends not only on the current input, but also on the state of the neuron at the previous moment. This time-varying characteristic gives the liquid neural network a unique advantage in processing time series data.

[0028] Furthermore, in the input layer, command layer and action layer, the input received by each neuron in the input layer includes the input from the outside and the hidden states of all neurons in the input layer, the input received by each neuron in the command layer includes the output of the input layer and the hidden states of all neurons in the action layer, and the input received by each neuron in the action layer includes the output in the command layer and the hidden states of all neurons in the action layer.

[0029] Furthermore, in the liquid neural network, the number of liquid neurons in the input layer is K×2, the number of liquid neurons in the action layer is K×2, and the number of liquid neurons in the command layer is C, which is a hyperparameter.

[0030] Furthermore, in step 4, the parameters of the liquid neural network are updated by the following formula:

[0031]

[0032] Among them, Adam represents the Adaptive Moment Estimation optimizer, κ represents the learning rate, which is a hyperparameter. represents the gradient of the loss function, θ * Represents the updated neural network parameters, ρ represents the regularization rate, which is a hyperparameter that penalizes the tendency of neural network parameters to become larger in long sequences through L2 regularization.

[0033] Furthermore, this algorithm does not require pre-training, but instead uses the characteristics of liquid neural networks to learn and adjust in real time as the received channel samples change, thereby achieving rapid response to various channel states and user needs, allowing the algorithm to learn directly from a dynamically changing environment without relying on a large amount of pre-collected data sets for training, thereby significantly reducing the threshold and complexity of algorithm deployment.

[0034] The present invention provides an analysis of the beneficial effects of the present invention and the technical methods provided by the present invention. It should be noted that the beneficial aspects involved in the technical methods provided by the present invention should not be regarded as isolated characteristics or single advantages, but should be understood and evaluated as the result of the interaction and synergy of multiple factors and characteristics. Each component, each feature of the present invention and the interaction between these parts and features together constitute the unique innovative nature and technological progress of the present invention. Therefore, when analyzing the beneficial aspects of the present invention and the technical methods provided by the present invention, these interrelated factors should be fully considered to ensure a complete and in-depth understanding of the true value and potential of the present invention. Further, the design and implementation of the present invention take into account a wide range of application scenarios and possible technical challenges, aiming to provide a solution that is not only effective in the current technical context, but also has the flexibility and scalability to adapt to future technological development. Therefore, the beneficial aspects of the present invention also include its compatibility and adaptability to future technological advances, and can maintain its effectiveness and efficiency in different technical environments and application requirements. It should be noted that the beneficial aspects provided by the present invention include but are not limited to improving performance, reducing costs, enhancing user experience, providing safer and more reliable technical solutions, simplifying operating procedures, reducing energy consumption, and promoting technological innovation and development. The realization of these beneficial aspects is the result of the combined effects of a series of innovative designs, methodologies and technical strategies introduced in the present invention.

[0035] Analysis of the beneficial effects of the present invention:

[0036] The robust multi-modal beamforming design method based on liquid neural network provided by the present invention embodies several significant beneficial effects.

[0037] The present invention utilizes the time-varying characteristics of liquid neural networks, and the method can adapt to the dynamic changes of the channel environment in real time. This adaptive capability is particularly suitable for mobile communications and rapidly changing communication environments, improving the communication quality and stability of the system. By introducing the precoding base matrix, the method effectively reduces the dimension of the optimization space. This is especially important when the number of base station antennas is much larger than the number of user-side receiving antennas, because it directly reduces the computational burden, making the algorithm more efficient and practical in actual deployment.

[0038] The present invention adopts a specific loss function and spectrum efficiency calculation formula, and the method can optimize the spectrum efficiency of the entire system. In particular, by considering the variance of spectrum efficiency, the method promotes fairer resource allocation and ensures the consistency and satisfaction of user experience; the design of hyperparameters considers the fairness of resource allocation and promotes resource allocation to more users by adjusting the neural network. This not only improves user satisfaction, but also enhances the system's adaptability to different user needs.

[0039] The design of liquid neurons gives the network the ability to remember previous states, which enables the network to use historical information to optimize the current decision-making process. This memory ability has significant advantages for processing data with strong temporal characteristics, improving the accuracy of prediction and decision-making. By adopting the Adam optimizer and a specific regularization method, this method can effectively update network parameters, prevent overfitting, and maintain the stability of the network in long sequences. This is crucial to maintaining the performance of the network during long-term operation. The design of the input layer, command layer, and action layer, as well as the interconnection of neurons between layers, enhances the network's ability to process complex data. This structural design enables the network to extract features from different levels and perform beamforming design more accurately.

[0040] In summary, the present invention realizes efficient, fair, and robust beamforming design in a dynamically changing communication environment through a series of innovative designs. These beneficial effects jointly improve the performance of the communication system and provide users with more stable and high-quality communication services. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Details of one or more implementations of the subject matter described in the present invention are set forth in the accompanying drawings and the following description. Other aspects, features, and advantages will be obtained or understood from the description, drawings, and claims. It should be noted that the following drawings are schematic diagrams and their relative dimensions may not be drawn to scale.

[0042] Figure 1 A scene graph is provided for implementing the method provided by the embodiment of the present invention.

[0043] Figure 2 Schematic diagram of liquid neurons used in an embodiment of the present invention.

[0044] Figure 3 The figure is a flow chart of an algorithm according to an embodiment of the present invention.

[0045] Figure 4 A spectrum efficiency diagram of an embodiment of the present invention provided under different degrees of channel estimation error.

[0046] Figure 5 A spectrum efficiency diagram of online operation in a dynamic scenario provided by an embodiment of the present invention.

[0047] Figure 6 A spectrum efficiency diagram of an embodiment of the present invention provided under different signal-to-noise ratio conditions with perfect channel estimation. DETAILED DESCRIPTION

[0048] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0049] The present invention provides an embodiment of a robust multimodal beamforming design method based on a liquid neural network. The present invention elaborates on its core concepts and key technical aspects in the following sections, aiming to provide a comprehensive understanding of the present invention by listing specific technical details. The description of the present invention includes a detailed description of various components, operating steps, and how these steps work together to achieve the desired goals. In addition, this document may also provide data, parameter settings, and expected effects of specific embodiments to demonstrate the potential and advantages of the present invention in practical applications. However, it should be clear that although the present invention provides these specific details in the embodiments to enhance the understanding of the present invention, the implementation of the present invention is not limited to the specific embodiments described or the specific technical forms adopted. On the contrary, according to the basic principles of the present invention, those skilled in the art can recognize that a variety of different ways, configurations, and technical details can be used to implement and optimize the present invention without departing from the spirit and scope of the present invention. For example, although a specific algorithm, hardware configuration, or software platform may be used in some embodiments, those skilled in the art should understand that the core concepts of the present invention can be applied in other algorithms, hardware, or software environments, and such changes or adjustments still fall within the scope of protection defined by the present invention. In addition, some steps or operations involved in the present invention may be performed in a different order, or some steps may be added or omitted without affecting the overall effect and purpose of the present invention. Therefore, the scope of the present invention should be understood to include all such changes, alternatives and modifications, which are based on the basic principles and introduction of the present invention. Specifically, the scope of protection of the present invention should not be limited to the specific embodiments described herein, but should be based on the concept and scope of the present invention defined in the attached claims. For the algorithm flow chart, see Figure 3, which provides a graphical overview of the implementation of the present invention, showing the key steps included in the present invention and their logical relationship with each other. Figure 3 As part of the description of the present invention, it is intended to assist in understanding the operating logic and implementation process of the present invention, but it is not intended to limit the possible implementation methods of the present invention. Based on the teachings of the present invention, those skilled in the art can flexibly adjust or redesign these processes to adapt to specific application requirements or technical environments, and these adjustments or designs also belong to the scope of the present invention.

[0050] Specifically, the embodiments provided by the present invention are as follows:

[0051] Step 1: Data acquisition: Assume that in the actual communication scenario there is a base station equipped with M antennas and K nodes each with N r The user terminal has a receiving antenna; channel data is collected from actual deployment scenarios, where the scenario is as follows Figure 1 As shown; Get the estimate of the channel from the base station to K users The channel of the kth user is (·) H Represents the conjugate transpose of a matrix or vector.

[0052] Step 2, variable initialization: first randomly initialize the precoding basis matrix X0 according to manifold learning and calculate the complete precoding matrix W0 through the formula, and finally initialize the internal parameters θ of the liquid neural network. The liquid neural network is composed of three layers of liquid neurons, namely the input layer, the command layer and the action layer. Each liquid neuron is as follows Figure 2 As shown, it consists of a central skeleton, g, f, h network fulcrums and nerve tails. For each liquid neuron layer, the calculation formula satisfies:

[0053] p(t)=σ(-f(p,I;θ f )t)⊙g(p,I;θ g )+[1-σ(-f(p,I;θ f )t)]⊙h(p,I;θ h ),

[0054] Where p(t), g, f, h, and I represent the hidden state of the neuron layer, the internal function of each neuron g, f, and h neural network pivot, and the vector formed by the stacking operation of the input within the layer; θ g ,θ f ,θ h represents the parameters of the g, f, and h neural network pivots of all liquid neurons in the layer, and ⊙ represents the element-by-element product operation. The precoding base matrix and the base station precoding matrix satisfy: W = H HX, where the dimension of the precoding matrix is ​​M × K, and the dimension of the precoding base matrix is ​​N × K. Since the number of base station antennas M is generally much larger than the number of receiving antennas N at the user end, this method reduces the optimization space.

[0055] Step 3: Perform forward propagation: Calculate the sum rate based on the initialized precoding matrix and the estimated channel information, and find the gradient of the sum rate. Then input the gradient into the liquid neural network and record the output of the neural network as ΔX i , and update the precoding base matrix through specific rules, where i represents the current i-th forward propagation. Specifically, the update rule of the precoding base matrix described is:

[0056]

[0057] The subscript i represents the variable corresponding to the i-th cycle described in step 5, P represents the sum of the maximum powers that can be emitted by all transmitting antennas of the base station, and Tr(·) represents the result of summing the elements in the diagonal direction of the matrix.

[0058] Step 4: Back propagation: According to the X obtained by forward propagation i Calculate the complete precoding matrix W i , and calculate a specific loss function, and update the gradient of the liquid neural network according to the value of the loss function. Specifically, the calculation formula of the spectral efficiency and loss function is

[0059]

[0060]

[0061] where σ 2 The variance of the additive Gaussian circular complex noise, I is a square matrix (i.e., a diagonal matrix) with all values ​​on the diagonal being 1 and all other values ​​being 0, det(·) represents the determinant of the matrix, and Var(R) represents the variance of the spectral efficiency of all users, (·) -1 Represents the inverse of the matrix, R k represents the spectral efficiency of the kth user, α k represents the weight of the kth user, w k represents the kth column of W, max k R k and min k R k ReLU(·)=max(·,0) is the activation function used to punish the neural network for not falling into low performance. β and γ are hyperparameters, the purpose of which is to promote the neural network to balance the overall effect and allocate resources to more users. The parameters of the liquid neural network are updated by the following formula:

[0062]

[0063] Among them, Adam represents the Adaptive Moment Estimation optimizer, κ represents the learning rate, which is a hyperparameter. represents the gradient of the loss function, θ * Represents the updated neural network parameters, ρ represents the regularization rate, which is a hyperparameter that penalizes the tendency of neural network parameters to become larger in long sequences through L2 regularization.

[0064] Step 5, multiple iterations: Repeat steps 3 and 4 for a total of N e The last neural network output is processed according to specific rules. Restore to precoding matrix Used to perform base station precoding on channel H in real scenarios.

[0065] Step 6, continuous operation: reacquire the next set of channels through step 1, do not reinitialize the parameters of the liquid neural network, and directly proceed to the process described in steps 3 to 5.

[0066] Furthermore, the liquid neurons have hidden state memory capability, and the state of each neuron depends not only on the current input, but also on the state of the neuron at the previous moment. This time-varying characteristic gives the liquid neural network a unique advantage in processing time series data.

[0067] Furthermore, in the input layer, command layer and action layer, the input received by each neuron in the input layer includes the input from the outside and the hidden states of all neurons in the input layer, the input received by each neuron in the command layer includes the output of the input layer and the hidden states of all neurons in the action layer, and the input received by each neuron in the action layer includes the output in the command layer and the hidden states of all neurons in the action layer.

[0068] Furthermore, in the liquid neural network, the number of liquid neurons in the input layer is K×2, the number of liquid neurons in the action layer is K×2, and the number of liquid neurons in the command layer is C, which is a hyperparameter.

[0069] Furthermore, this algorithm does not require pre-training, but instead uses the characteristics of liquid neural networks to learn and adjust in real time as the received channel samples change, thereby achieving rapid response to various channel states and user needs, allowing the algorithm to learn directly from a dynamically changing environment without relying on a large amount of pre-collected data sets for training, thereby significantly reducing the threshold and complexity of algorithm deployment.

[0070] This embodiment provides a test for evaluating the spectrum efficiency under different setting conditions. It should be understood that the test provided by the present disclosure is only used as a means of facilitating the understanding of the present disclosure. The test method provided in this embodiment is intended to evaluate the spectrum efficiency of the present invention under different setting conditions. This test is to facilitate the understanding and evaluation of the technical benefits brought by the present invention. By evaluating the performance of the present invention under a variety of different settings and environmental conditions, these test results help to demonstrate the practicality and advantages of the present invention, and also provide an empirical basis to verify the innovation and improvement of the present invention. It should be emphasized that the design and implementation of these tests do not mean to limit the scope of application or effect of the present invention. On the contrary, these tests are only used as a tool and method to assist in understanding the working principle and potential application value of the present invention. The test results are intended to provide an intuitive way to show the performance of the present invention under specific conditions, rather than defining or limiting the full potential and effect that the present invention can achieve. In addition, it should be understood that the test settings and conditions provided by the present disclosure only represent one possible application scenario. In view of the wide applicability and flexibility of the present invention, those skilled in the art can adjust or modify these test conditions according to specific application requirements and technical backgrounds. These adjustments may include but are not limited to changes in the test environment, adjustments to parameters, and changes in evaluation indicators. This flexibility further demonstrates the applicability and effectiveness of the present invention in different technical fields and application scenarios. Therefore, the tests provided in this disclosure should be regarded as an exemplary description to assist in understanding and evaluating the present invention, rather than limiting or defining the technical scope of the present invention. The scope of protection of the present invention should be determined by the concept and scope of the present invention defined in the attached claims, rather than based solely on these test results or specific test conditions.

[0071] In order to comprehensively evaluate the performance of the proposed method (labeled as NCP in the figure), several baselines were set in the simulation experiment for comparison. These baselines include the implementation using the weighted minimum mean square error algorithm (RWMMSE), the scheme using the long short-term memory network (LSTM), and the method using the gated recurrent unit (GRU).

[0072] For the simulation provided by the present disclosure, the set parameters satisfy:

[0073] N e =2,κ=1×10 -2 ,ρ=0.01,β=0.7,γ=0.4,C=30,

[0074] At the same time, the spacing between the antennas is set to half the wavelength corresponding to the baseband frequency of the emitted electromagnetic wave in a vacuum. In the simulation data provided in this disclosure, the scenarios used meet the assumptions: there are 4 users in each scenario, there is 1 base station in each scenario, and the base station has 64 transmitting antennas, and each user has 2 receiving antennas. The channel estimation error (CEE) is calculated and measured using the following formula:

[0075]

[0076] where H is the real channel, and It is the estimated channel. CEE is measured in dB. The larger the value, the more accurate the channel estimation.

[0077] The present disclosure performs simulation based on this setting, and takes the arithmetic average method to process and record the results of all samples. Figure 4 The spectral efficiency of operating under different degrees of channel estimation error is demonstrated; Figure 5 The spectral efficiency of dynamic scenarios switched at different speeds (0km / h, 60km / h, 120km / h) is shown, where CEE is set to 10dB; Figure 6 The spectral efficiency of the system under different signal-to-noise ratios under perfect channel estimation is shown. The analysis of the above tests of the present invention indicates that: Figure 4 It is shown that under different degrees of channel estimation error, this embodiment can obtain higher spectrum efficiency than the reference baseline; Figure 5 It is demonstrated that in a dynamic scenario of switching at different speeds (0km / h, 60km / h, 120km / h), this embodiment can achieve higher spectrum efficiency than other solutions when CEE is set to 10dB, which is conducive to adapting to the changing channel status in actual scenarios; Figure 6 It is demonstrated that under different signal-to-noise ratios under perfect channel estimation, this embodiment has obvious advantages over the reference solution.

[0078] In the specification and drawings of the present disclosure, in addition to the detailed description of the present invention, the spectrum efficiency and rate data of some other algorithms are also provided. The provision of these data is intended to provide a reference framework so that it is easier to understand and evaluate the improvements and advantages of the present invention and its embodiments compared with the prior art. By comparing the performance of the present invention with the performance of other algorithms, the innovation and effect of the present invention in improving spectrum efficiency, optimizing signal processing or enhancing communication system performance can be intuitively demonstrated. It should be emphasized that the spectrum efficiency data of these other algorithms are not intended to limit or define the technical scope of the present invention. On the contrary, they are only provided as background information to help build a comprehensive understanding of the technical contribution of the present invention. The purpose of these reference data is to facilitate comparison and understanding, rather than as the only criterion for evaluating the merits of the present invention. Further, the present disclosure provides an analysis of these test results based on the aforementioned test results. This analysis section aims to explore in depth the performance of the present invention under different conditions and settings, and how the present invention optimizes spectrum efficiency in various challenging scenarios. The analysis content is intended to reveal the technical details and working principles of the present invention, as well as the advantages and improvements of the present invention relative to other algorithms. It should be understood that the analysis described in the present disclosure is only used to assist in understanding and evaluating the simulation effects of the present invention and its embodiments. This analysis provides a method for interpreting test results from a theoretical and practical perspective, but does not mean that the application or effect of the present invention is limited to specific test conditions or results. On the contrary, the technical scope and application potential of the present invention should be based on the comprehensive description of the present disclosure and the appended claims.

[0079] The embodiment of the robust multimodal beamforming design method based on liquid neural network based on manifold learning of the present invention can be applied to any device with data processing capability. The present invention demonstrates an innovative solution for improving spectral efficiency and signal processing performance in communication systems. The design concept of this embodiment is highly versatile, allowing it to be applied to various devices equipped with data processing capabilities. These devices may include a series of hardware and software components to support the implementation and operation of the present invention. For example, these devices may be mobile communication devices, base stations, satellite communication devices, nodes in wireless sensor networks, or any other system that requires efficient beamforming processing. Specifically, the hardware elements in these devices may include circuit boards, application specific integrated circuits (ASICs), digital signal processors (DSPs), and general purpose or special purpose processors. These hardware elements work together to provide the required computing power and resources to implement the beamforming design method proposed in the present invention. In addition, the device may also include software components, such as operating systems, drivers, and dedicated software applications that execute the algorithms of the present invention. These software components can control the operation of hardware elements and perform necessary data processing and analysis tasks to achieve beamforming optimization. The operational description shown in the accompanying drawings provides a detailed step-by-step guide to illustrate how the beamforming design method of the present invention is implemented. These operational steps are designed to be executed by corresponding hardware and software functional components. For example, data acquisition and preprocessing may be completed by sensors and data interface hardware, while manifold learning and liquid neural network training and inference may be implemented by processors and related algorithm software. It is worth noting that the implementation of the present invention is not limited by specific hardware or software platforms. Based on the technical principles and design concepts of the present invention, technicians in this field can select and configure the devices and components according to actual application requirements and available technical resources to adapt to specific application scenarios and performance requirements. Therefore, the present invention provides a flexible and efficient solution that is suitable for a wide range of data processing equipment and communication systems, and can optimize the performance of multi-modal beamforming while maintaining robustness.

[0080] The device embodiment is roughly corresponding to the method embodiment in structure. For similar content, see the relevant description of the method embodiment. The above-mentioned device embodiment is a schematic display, in which the units defined as separate components may or may not be physically independent; the components identified as units may or may not be physical units. These components can be concentrated in one location or dispersed in multiple network units. According to actual needs, some or all modules can be selected to achieve the goals of the scheme of the present invention. For ordinary technicians in this technical field, it can be understood and practiced without too much innovative effort. Various modifications to the implementation described in the present disclosure may be obvious to those skilled in the art, and the universal principles defined herein can be applied to other implementations without departing from the spirit or scope of the present disclosure. Therefore, the claims are not intended to be limited to the implementations shown herein, but should be granted a scope consistent with the present disclosure, the principles disclosed herein and the novel features.

[0081] An embodiment of the present invention further provides a computer-readable storage medium, which stores a computer program. When this program is executed on a processor, a RIS-assisted robust multi-modal beamforming design method based on manifold learning described above can be implemented. The computer-readable storage medium can be an internal storage element of any device with data processing capabilities in the aforementioned embodiments, such as a hard disk or memory. In addition, the computer-readable storage medium may also refer to external storage devices, such as plug-in hard disks, smart memory cards (SMCs), SD cards, flash memory cards, CDs, laser discs, optical discs, DVDs, floppy disks, and Blu-ray discs. This computer-readable storage medium may also cover the internal storage portion of the device and external storage devices. In addition to storing the above-mentioned computer programs, the storage medium can also save other required programs and data, or temporarily store data that has been output or is to be output.

[0082] For those skilled in the art, they should understand that the functions described in this application can be implemented by hardware, software, firmware or a combination of these components. When implemented by software, these functions can be stored on a computer-readable medium, or transmitted as one or more instructions or codes on the medium. The computer-readable medium includes both computer-readable storage media and communication media, which is a medium that facilitates the transfer of computer programs from one location to another. This storage medium can be any general or special medium that is accessible to a computer.

[0083] Obviously, the above description and embodiments are only some specific examples of the present invention and are not meant to limit it. For professionals in the field of the present technology, based on the above description, they can make various changes or adjustments. Obviously, it is impossible to list all possible implementations in detail, but all reasonable changes made within the scope of the description should be considered within the scope of the present invention. It is worth noting that although the drawings may describe certain operations in a specific order, this does not mean that these operations must be performed in this specific order, nor does it mean that all described operations must be performed to obtain the expected effect. The drawings may only describe some operation processes in a schematic manner, but may include other steps that are not explicitly described. For example, some steps may be performed before, after, or in the middle of the described operations. In some situations, it may be beneficial to use multitasking or parallel processing. At the same time, although the independence of certain components may be described in the text, this does not mean that such independence is required in all cases. In fact, the various parts or systems described can usually be integrated into one software or divided into multiple different software. In general, other implementations and changes should be considered within the scope of the present invention. In some instances, the operations described in the claims may not be performed in a fixed order, but the expected results can still be obtained.

Claims

1. A robust multimodal beamforming method based on liquid neural network, characterized in that: The following steps are involved: S1. Obtain channel estimation from the base station to each user terminal; S2. Randomly initialize the precoding base matrix according to streaming learning and obtain an initialized precoding matrix; S3. According to the initialized precoding matrix and the gradient of the channel estimation and rate, the gradient is input into the liquid neural network, and the precoding base matrix is ​​updated according to the output of the liquid neural network and the maximum total power of the antenna output; the update rule of the precoding base matrix is ​​specifically as follows: The subscript i represents the corresponding variable in the i-th cycle, and the output of the neural network is denoted as ΔX i ; P represents the maximum power that can be emitted by all transmitting antennas of the base station, Tr(·) represents the result of summing the elements in the diagonal direction of the matrix, H represents the channel, (·) H Represents the conjugate transpose of a matrix or vector; S4, obtaining an updated precoding matrix through the updated precoding base matrix, and performing a gradient update of the liquid neural network according to the precoding matrix and the user spectrum efficiency calculation loss function; The parameter θ of the liquid neural network is updated by the following formula: i * =θ+κAdam(▽ θ L,θ,ρ) Among them, κ represents the learning rate, which is a hyperparameter, ▽ θ L represents the gradient of the loss function, θ * represents the updated neural network parameters, ρ represents the regularization rate, which is a hyperparameter that penalizes the tendency of neural network parameters to become larger in long sequences through L2 regularization, and Adam represents the Adaptive Moment Estimation optimizer; S5. Repeat S3-S4 until a preset number of iterations is reached, and perform base station precoding on the channel according to the precoding matrix obtained in the last iteration.

2. The method for robust multimodal beamforming based on liquid neural network according to claim 1, characterized in that: The liquid neural network is composed of three layers of liquid neurons, namely an input layer, a command layer and an action layer. Each liquid neuron is composed of a central skeleton, a network fulcrum and a nerve tail.

3. The robust multi-modal beamforming method based on liquid neural network according to claim 2, characterized in that: The liquid neuron has a hidden state memory capability, and the state of each neuron depends not only on the current input but also on the state of the neuron at the previous moment.

4. The method for robust multimodal beamforming based on liquid neural network according to claim 2, characterized in that: In the input layer, command layer and action layer, the input received by each neuron in the input layer includes the input from the outside and the hidden states of all neurons in the input layer, the input received by each neuron in the command layer includes the output of the input layer and the hidden states of all neurons in the action layer, and the input received by each neuron in the action layer includes the output in the command layer and the hidden states of all neurons in the action layer.

5. The robust multi-modal beamforming method based on liquid neural network according to claim 2, characterized in that: The number of liquid neurons in the input layer and the command layer is twice the number of user terminals acquired each time.

6. The method for robust multimodal beamforming based on liquid neural network according to claim 1, characterized in that: The precoding matrix W is obtained by the following formula: W=H H X, Where X is the precoding basis matrix, H represents the channel, (·) H Represents the conjugate transpose of a matrix or vector.

7. The robust multi-modal beamforming method based on liquid neural network according to claim 1, characterized in that: The loss function calculated according to the precoding matrix and the user spectrum efficiency is specifically: Where K is the number of user terminals, σ 2 The variance of the additive Gaussian circular complex noise, I is a diagonal matrix, det(·) represents the determinant of the matrix, Var(R) represents the variance of the spectral efficiency of all users, (·) -1 Represents the inverse of the matrix, R k represents the spectral efficiency of the kth user, α k represents the weight of the kth user, w k represents the kth column of the precoding basis matrix W, max k R k represents the spectral efficiency of the user with the highest spectral efficiency, ReLU(·)=max(·,0) is the activation function, and β and γ are hyperparameters.

8. The robust multi-modal beamforming method based on liquid neural network according to claim 1, characterized in that: When data is acquired for the first time, the internal parameters of the liquid neural network need to be initialized. After the base station precoding is completed, the internal parameters of the liquid neural network are not reinitialized when the next set of channel data is acquired again.

Images