5g fr1 form filter system and design method

Abstract

The application provides a 5G FR1 forming filter system and a design method, relates to the technical field of forming filters, and the system comprises a BBU end intelligent prediction module, an RRU end configuration management and dynamic coefficient switching module and an RRU end unified structure channel filter module; the BBU intelligent prediction module predicts an optimal FIR filter coefficient set in real time through a light DNN model and delivers the same to the RRU end; the RRU end configuration management and dynamic coefficient switching module realizes flexible switching of the coefficients based on a pre-stored initial filter coefficient set of the FR1 full frequency band and in combination with a dynamic configuration instruction; and the RRU end unified structure channel filter module performs spectral shaping filter processing on a downlink CP-OFDM signal after adjusting the target coefficient set into a FIR structure. Through the AI-driven dynamic optimization mechanism and the unified hardware architecture, the application solves the problems of dispersed filter structures and insufficient static configuration flexibility in the prior art.

Description

Technical Field

[0001] This document relates to the field of shaped filter technology, and in particular to a general 5G FR1 shaped filter system and its design method. Background Technology

[0002] 5G NR systems, with their multi-subcarrier spacing and flexible bandwidth configuration, achieve single-carrier bandwidth adjustment from 3MHz to 100MHz in the FR1 band to adapt to complex spectrum deployment requirements. In the base station transmit link, the digital channel filter located between the IFFT and digital upconversion modules undertakes the important task of downlink signal spectrum shaping, meeting spectrum specification requirements by limiting bandwidth and suppressing out-of-band power.

[0003] Current mainstream filter designs mostly employ fixed parameters and static configurations, adapting only to a single frequency band or bandwidth. This customized approach has significant drawbacks: new bandwidth configurations require redesigning coefficients and hardware logic, leading to increased system development and verification costs and difficulties in upgrades and maintenance. While some FPGA-based filters offer a degree of flexibility, they are not compatible with the full bandwidth configuration of the FR1 band, making it difficult to meet the highly flexible reuse requirements of 5G public / private networks and industry-customized deployments.

[0004] In digital communication systems, the design of digital filters for multi-bandwidth signal processing has been extensively studied, especially in scenarios involving high-speed link processing and diverse frequency resource configurations. Some solutions improve processing capabilities and configuration flexibility through interpolation structures or reconfigurable coefficient switching mechanisms. Although these designs have certain practical applications in engineering, they mainly focus on computational efficiency or static structure reuse, and have not yet addressed the key requirements of 5G NR base station transmit links, such as spectrum shaping accuracy, real-time adaptive capabilities, and structural uniformity, for channel filters.

[0005] For example, patent CN118041302A proposes a digital filtering structure combining 4x interpolation and a FIFO scheduling mechanism, aiming to improve the system's parallel processing capability and throughput efficiency. This solution exhibits high processing performance under fixed bandwidth conditions, but the interpolation factor is fixed, lacking adaptability to multi-bandwidth scenarios. Its filter structure requires independent design and verification for each bandwidth, resulting in high repetitive development costs, non-reusable architecture, and difficulty in supporting the dynamic requirements of frequent bandwidth switching in 5G NR systems.

[0006] Patent CN116545412A proposes a reconfigurable FIR filter scheme that supports multiple carrier bandwidth configurations by loading different coefficient tables, providing a certain degree of parameter flexibility. However, this scheme typically requires configuring multiple filter modules to switch in parallel, lacks structural uniformity, has complex resource scheduling and control logic, is prone to FPGA resource redundancy, and its coefficient configuration is still statically designed, lacking the ability to adaptively optimize based on runtime channel conditions.

[0007] In summary, existing solutions either prioritize throughput performance or focus on parameter reuse, failing to achieve an effective balance between structural versatility, multi-bandwidth support, and intelligent optimization performance. This makes it difficult to meet the comprehensive requirements of efficient spectrum shaping and dynamic operational adaptation in the 5G FR1 band. Therefore, there is an urgent need for a channel filter design scheme that integrates AI adaptive mechanisms under a unified architecture, supports real-time channel awareness, and possesses dynamic parameter optimization capabilities to address the flexible deployment needs of 5G base stations in various scenarios and configurations. Summary of the Invention

[0008] This invention provides a 5G FR1 shaping filter system and design method, achieving a unified architecture to cover the entire FR1 frequency band and bandwidth configuration. By establishing a flexible coefficient configuration mechanism, it can quickly adapt to different bandwidths. By introducing AI dynamic optimization, it can intelligently adjust filter parameters based on channel conditions, thereby improving spectrum shaping accuracy and resource utilization.

[0009] According to an embodiment of the present invention, a 5G FR1 shaped filter system is provided, comprising:

[0010] The BBU-side intelligent prediction module, the RRU-side configuration management and dynamic coefficient switching module, and the RRU-side unified structure channel filter module;

[0011] The BBU-end intelligent prediction module is used to collect channel context information during the operation of the wireless link, input the channel context information into a pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU-end configuration management and dynamic coefficient switching module.

[0012] The RRU end configuration management and dynamic coefficient switching module is used to receive the optimal FIR filter coefficient group, convert the format of the optimal FIR filter coefficient group based on the preset FR1 full carrier bandwidth initial filter coefficient group and store it, and obtain the corresponding target coefficient group from the optimal FIR filter coefficient group based on the obtained configuration command and the preset mapping table.

[0013] The unified structure channel filter module at the RRU end loads the target coefficient group into the FIR structure and then performs spectrum shaping and filtering processing on the downlink CP-OFDM signal.

[0014] According to an embodiment of the present invention, a design method for a 5G FR1 shaped filter is provided, comprising:

[0015] S1. Collect channel context information during the operation of the wireless link, input the channel context information into the pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU end configuration management and dynamic coefficient switching module.

[0016] S2. Receive the optimal FIR filter coefficient set, perform format conversion on the optimal FIR filter coefficient set based on the preset FR1 full carrier bandwidth initial filter coefficient set, and store it, based on the acquired configuration command and the preset mapping table.

[0017] Obtain the corresponding target coefficient set from the optimal FIR filter coefficient set;

[0018] S3. After loading the target coefficient group into the FIR structure, perform spectrum shaping and filtering processing on the downlink CP-OFDM signal.

[0019] By employing embodiments of the present invention, a channel filter hardware platform covering the entire FR1 frequency band and all bandwidth configurations is realized through unified order design and module instantiation, improving system consistency and logic reuse efficiency. It supports preloading multiple sets of static filter coefficients into the FPGA module and switching them on demand in conjunction with configuration control logic, reducing redundant development and enhancing bandwidth adaptability. By deploying a lightweight deep neural network model on the baseband processing unit (BBU) side, the channel context (CQI, load, PRB distribution, etc.) is perceived in real time, and the optimal filter parameter coefficient set is predicted, improving spectrum shaping accuracy and resource utilization. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a framework diagram of a 5G FR1 shaped filter system according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart of a 5G FR1 molding filter design method according to an embodiment of the present invention;

[0023] Figure 3This is a flowchart illustrating the working principle of a 5G FR1 molding filter according to an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0025] Existing filter designs lack a unified architecture, making them difficult to reuse and maintain: Most current filter designs adopt customized solutions for specific bandwidths, requiring the development and verification of filter structures and coefficients separately for each carrier bandwidth. This results in fragmented modules and poor resource reusability. The system structure is complex, and the costs of field deployment, upgrades, expansions, and long-term operation and maintenance are high, making it difficult to support the rapid deployment and flexible expansion requirements of 5G networks.

[0026] Lack of dynamic adaptation capability: Traditional filters adopt a static configuration method, which cannot dynamically adjust the filtering performance according to different service loads, modulation methods, spectrum utilization and other operating parameters. This results in unstable spectrum shaping effect, low spectrum utilization efficiency and inability to adapt to the ever-changing wireless environment.

[0027] Lack of intelligent optimization mechanism: Existing solutions do not introduce artificial intelligence or context-aware technology, and rely on offline static modeling and manual experience configuration. They cannot dynamically optimize passband ripple, stopband attenuation or transition band characteristics based on real-time channel quality, which limits the further improvement of spectrum shaping accuracy and system performance.

[0028] This invention proposes a design and implementation method for 5G FR1 digital channel filters that combines a general architecture with artificial intelligence optimization mechanisms, thereby constructing a solution that can simultaneously address the aforementioned problems.

[0029] System Implementation Examples

[0030] According to an embodiment of the present invention, a 5G FR1 shaped filter system is provided. Figure 1 This is a system framework diagram of the 5GFR1 molding filter according to an embodiment of the present invention. Figure 1 As shown, the 5G FR1 shaping filter system of this embodiment of the invention specifically includes: a BBU-end intelligent prediction module, an RRU-end configuration management and dynamic coefficient switching module, and an RRU-end unified structure channel filter module. The baseband unit is the BBU end, and the radio frequency unit is the RRU end.

[0031] The BBU-end intelligent prediction module is used to collect channel context information during the operation of the wireless link, input the channel context information into a pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU-end configuration management and dynamic coefficient switching module.

[0032] The BBU-side intelligent prediction module is deployed on the processor resources of the BBU. It obtains multi-dimensional context parameters such as link quality, spectrum usage, and transmission configuration from the physical and MAC layers, normalizes them, and then inputs them into a pre-trained lightweight DNN model. The model adopts a three-layer network structure of input layer, hidden layer, and output layer, outputting a set of coefficients for a fixed-order FIR filter, which is finally sent to the RRU through the fronthaul interface.

[0033] The BBU-side intelligent prediction module is specifically used for:

[0034] Obtain context parameters and standardize them. The context parameters include: link quality parameters, spectrum usage parameters, and transmit configuration parameters.

[0035] The preprocessed context parameters are input into a preset lightweight DNN model, and the optimal FIR filter coefficient set is output. The preset lightweight DNN model construction process includes:

[0036] Network structure: An MLP structure consisting of an input layer, hidden layers, and an output layer is adopted. The dimension of the input layer is determined by the context parameters. The hidden layer contains a number of nodes and the activation function is ReLU. The number of nodes in the output layer is consistent with the order of the FIR filter. The lightweight DNN model structure is as follows:

[0037] ;

[0038] in, ∈R^d is the input feature vector, where d is the input dimension and contains channel context information; ∈R^N represents the output FIR filter coefficient set, with length N consistent with the filter order. For a feedforward neural network (MLP) function with parameter set θ;

[0039] Model training: Based on the target passband ripple, stopband attenuation, and transition band indices, optimal filter coefficients are generated as the supervision target. The mean square error between the predicted coefficients and the target coefficients is calculated as the loss function. The model parameters are adjusted using the backpropagation algorithm. The loss function used is as follows:

[0040] ;

[0041] This represents the coefficients of the i-th target FIR filter, which are continuously adjusted during the training phase using the backpropagation algorithm, so that... Minimize the mean square error between the target coefficient and the target coefficient.

[0042] The RRU end configuration management and dynamic coefficient switching module is used to receive the optimal FIR filter coefficient group, convert the format of the optimal FIR filter coefficient group based on the preset FR1 full carrier bandwidth initial filter coefficient group and store it, and obtain the corresponding target coefficient group from the optimal FIR filter coefficient group based on the obtained configuration command and the preset mapping table.

[0043] The RRU-side configuration management and dynamic coefficient switching module is deployed within the RRU's internal FPGA logic resources. It is responsible for monitoring configuration commands such as "bandwidth index" or "AI prediction ID" received from the fronthaul interface (e.g., eCPRI link). Based on the command content and a preset mapping table, it controls the loading of FIR coefficient sets from the internal memory into the IP core filtering module.

[0044] The process of constructing the initial filter coefficient set of the preset FR1 full carrier bandwidth includes:

[0045] Calculate the transmission bandwidth, sampling rate, and normalized transition band ratio for each carrier bandwidth and subcarrier spacing combination within the FR1 band, and generate a normalized transition band ratio table.

[0046] Select the carrier bandwidth and subcarrier spacing combination corresponding to the minimum normalized transition band ratio from the normalized transition band ratio table as the design constraint.

[0047] Based on the design constraints and combined with the preset passband ripple and stopband attenuation indices, the unified filter order that meets the requirements of the entire FR1 frequency band is calculated using filter design tools.

[0048] For each carrier bandwidth, the combination with the smallest normalized transition band ratio among all its subcarrier spacings is selected as the design condition. The corresponding FIR filter coefficients are generated using digital filter design tools, and all coefficients adopt the unified order mentioned above.

[0049] The generated bandwidth coefficient groups are stored in a preset format to form an initial filter coefficient group covering the entire FR1 frequency band.

[0050] The preset mapping table construction process includes:

[0051] The carrier bandwidth and subcarrier spacing within the FR1 band are grouped and assigned unique coefficient numbers;

[0052] In the FPGA, a memory block is allocated for each coefficient group to record the starting address and length;

[0053] Establish a mapping relationship between associated carrier bandwidth, subcarrier spacing, coefficient sequence number, and storage block number.

[0054] The unified structure channel filter module at the RRU end loads the target coefficient group into the FIR structure and then performs spectrum shaping and filtering processing on the downlink CP-OFDM signal.

[0055] The unified structure channel filter module at the RRU end is implemented primarily on an FPGA, using an FIR IP core to complete the filtering calculation logic. The FPGA design initializes and loads multiple sets of FIR filter coefficients supporting the full bandwidth of 5G FR1, and supports dynamic switching of the currently active coefficient based on parameter configuration or AI-predicted input. The filter structure remains unified, the processing clock settings are compatible with all FR1 sampling rate requirements, and both input and output data use two's complement fixed-point format.

[0056] The unified structure channel filter module at the RRU end is specifically used for:

[0057] The target coefficient set is loaded into the FIR IP core of the FPGA;

[0058] Based on the bandwidth index or AI prediction ID in the configuration instructions, the storage address of the target coefficient group is determined through the mapping table, and the active filter coefficients are switched in real time.

[0059] Using FPGA as the hardware platform and employing two's complement fixed-point data format, the downlink CP-OFDM signal is subjected to spectrum shaping and filtering, with the processing clock frequency set to an integer multiple covering the highest sampling rate of FR1.

[0060] The embodiments of the present invention have the following beneficial effects:

[0061] By unifying the order design and module instantiation, a channel filter hardware platform covering the entire FR1 frequency band and all bandwidth configurations is realized, improving system consistency and logic reuse efficiency. It supports preloading multiple sets of static filter coefficients into the FPGA module and switching them on demand in combination with configuration control logic, reducing redundant development and enhancing bandwidth adaptability. By deploying a lightweight deep neural network model on the baseband processing unit (BBU) side, it can perceive channel context such as CQI, load, PRB distribution, etc. in real time, predict the optimal filter parameter coefficient set, and improve spectrum shaping accuracy and resource utilization.

[0062] Method Implementation Examples

[0063] According to an embodiment of the present invention, a design method for a 5G FR1 shaped filter is provided. Figure 2 This is a flowchart of a 5G FR1 shaped filter design method according to an embodiment of the present invention. Figure 2 As shown, the 5G FR1 shaped filter design method of this embodiment specifically includes:

[0064] S1. Collect channel context information during the operation of the wireless link, input the channel context information into the pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU end configuration management and dynamic coefficient switching module.

[0065] S1 specifically includes:

[0066] Obtain context parameters and standardize them. The context parameters include: link quality parameters, spectrum usage parameters, and transmit configuration parameters.

[0067] The preprocessed context parameters are input into a preset lightweight DNN model, and the optimal FIR filter coefficient set is output. The preset lightweight DNN model construction process includes:

[0068] Network structure: An MLP structure consisting of an input layer, hidden layers, and an output layer is adopted. The dimension of the input layer is determined by the context parameters. The hidden layer contains a number of nodes and the activation function is ReLU. The number of nodes in the output layer is consistent with the order of the FIR filter. The lightweight DNN model structure is as follows:

[0069] ;

[0070] in, ∈R^d is the input feature vector, where d is the input dimension and contains channel context information; ∈R^N represents the output FIR filter coefficient set, with length N consistent with the filter order. For a feedforward neural network (MLP) function with parameter set θ;

[0071] Model training: Based on the target passband ripple, stopband attenuation, and transition band indices, optimal filter coefficients are generated as the supervision target. The mean square error between the predicted coefficients and the target coefficients is calculated as the loss function. The model parameters are adjusted using the backpropagation algorithm. The loss function used is as follows:

[0072] ;

[0073] This represents the coefficients of the i-th target FIR filter, which are continuously adjusted during the training phase using the backpropagation algorithm to make the FIR coefficient vector output by the model... Minimize the MSE error with respect to the target coefficient.

[0074] S2. Receive the optimal FIR filter coefficient set, convert the format of the optimal FIR filter coefficient set based on the preset FR1 full carrier bandwidth initial filter coefficient set and store it, and obtain the corresponding target coefficient set from the optimal FIR filter coefficient set based on the obtained configuration command and the preset mapping table.

[0075] The method for constructing the initial filter coefficient set of the preset FR1 full carrier bandwidth includes:

[0076] Calculate the transmission bandwidth, sampling rate, and normalized transition band ratio for each carrier bandwidth and subcarrier spacing combination within the FR1 band, and generate a normalized transition band ratio table.

[0077] Select the carrier bandwidth and subcarrier spacing combination corresponding to the minimum normalized transition band ratio from the normalized transition band ratio table as the design constraint.

[0078] Based on the design constraints and combined with the preset passband ripple and stopband attenuation indices, the unified filter order that meets the requirements of the entire FR1 frequency band is calculated using filter design tools.

[0079] For each carrier bandwidth, the combination with the smallest normalized transition band ratio among all its subcarrier spacings is selected as the design condition. The corresponding FIR filter coefficients are generated using digital filter design tools, and all coefficients adopt the unified order mentioned above.

[0080] The generated bandwidth coefficient groups are stored in a preset format to form an initial filter coefficient group covering the entire FR1 frequency band.

[0081] The acquisition of the target coefficient set from the optimal FIR filter coefficient set based on the obtained configuration instructions and the preset mapping table specifically includes:

[0082] The carrier bandwidth and subcarrier spacing within the FR1 band are grouped and assigned unique coefficient numbers;

[0083] In the FPGA, a memory block is allocated for each coefficient group to record the starting address and length;

[0084] Establish a mapping relationship between associated carrier bandwidth, subcarrier spacing, coefficient sequence number, and storage block number.

[0085] S3. After loading the target coefficient group into the FIR structure, perform spectrum shaping and filtering processing on the downlink CP-OFDM signal.

[0086] S3 specifically includes:

[0087] The target coefficient set is loaded into the FIR IP core of the FPGA;

[0088] Based on the bandwidth index or AI prediction ID in the configuration instructions, the storage address of the target coefficient group is determined through the mapping table, and the active filter coefficients are switched in real time.

[0089] Using FPGA as the hardware platform and employing two's complement fixed-point data format, the downlink CP-OFDM signal is subjected to spectrum shaping and filtering, with the processing clock frequency set to an integer multiple covering the highest sampling rate of FR1.

[0090] like Figure 3 The diagram shown illustrates the working principle of a 5G FR1 shaping filter according to an embodiment of the present invention. The specific process for operating the 5G FR1 shaping filter designed according to this embodiment is as follows:

[0091] Step 1: Design of initial filter coefficients of uniform order

[0092] Step 1.1: Analyze the filter requirements under the combination of carrier bandwidth and subcarrier spacing in the 5G FR1 band, as follows:

[0093] Based on 3GPP TS 38.104 Table 5.3.2-1 and 3GPP TS 38.104 Table B.5.2-1 / -2 / -3 / -4, the transmission bandwidth, sampling rate, and normalized transition band ratio under all bandwidth and subcarrier spacing conditions of 5G FR1 are obtained according to the following calculation steps. The results are summarized in Table 1.

[0094] Table 1. Normalized transition band ratios for FR1 with different bandwidths and subcarrier combinations.

[0095]

[0096] The meanings of each parameter in Table 1 are as follows:

[0097] TB is an abbreviation for Transmission Bandwidth.

[0098] ;

[0099] N RB It is an abbreviation for Number of Resource Blocks.

[0100] RB is an abbreviation for Resource Block.

[0101] SCS is an abbreviation for Sub Carrier Space.

[0102] CB is an abbreviation for Channel Bandwidth.

[0103] FR1 N for each CB and SCS RB Refer to 3GPP TS 38.104 Table 5.3.2-1;

[0104] SR = FFT Size * SCS / 1000;

[0105] SR is an abbreviation for Sample Rate. In the formula, SCS is in kHz, and SR is in MSPS. The calculation method is as follows: ;

[0106] For the FFT size of each CB and SCS in FR1, refer to 3GPP TS 38.104 Table B.5.2-1 / -2 / -3 / -4;

[0107] TR is an abbreviation for normalized transition ratio, which is the normalized ratio of the transition band relative to the sampling rate. TR = (CB / 2-TB / 2) / (SR / 2).

[0108] Step 1.2: Analyze and determine the uniform order values ​​of the coefficients, as follows:

[0109] Analyze Table 1 to find the bandwidth and subcarrier combination corresponding to the minimum ratio. This combination requires the steepest filter transition band and also requires the highest filter coefficient order.

[0110] Using the combined channel bandwidth, transmission bandwidth, and sampling rate as input parameters for filter coefficient design, and combining the filter passband ripple index and stopband minimum attenuation index requirements, the order of the filter coefficients that meet the requirements can be obtained based on the firpmord function in MATLAB. This order can satisfy the unified order of all bandwidths and subcarriers of FR1.

[0111] Step 1.3: Generate the initial filter coefficient set corresponding to the bandwidth using MATLAB tools, and design the unified filter coefficients that satisfy all subcarrier combinations with the same bandwidth of FR1 as follows:

[0112] Further analysis of the relative sampling rate normalization ratio of the transition band of each group of bandwidths and subcarriers in step 1, comparing and finding the minimum TR value among all subcarrier intervals (15kHz / 30kHz / 60kHz) under each bandwidth. For example, the transition band is steepest when the 20M bandwidth is 15kHz subcarrier. When designing filter coefficients, a unified filter coefficient is designed for the 20M bandwidth carrier according to this combination of conditions, that is, the coefficients are generated only according to the different bandwidths.

[0113] The initial filter coefficients are designed so that all RB resources for each bandwidth are scheduled for use.

[0114] Step 2: Design, Training and Deployment of Intelligent Prediction Model

[0115] A lightweight DNN (MLP structure) is used to model the downlink channel state of 5G base stations and predict the optimal FIR channel filter coefficients. The specific design is as follows:

[0116] Step 2.1: Construct feature vectors:

[0117] The model's input is a set of channel context parameters collected in real time by the BBU, which are divided into three categories:

[0118] Link quality metrics: CQI, BLER, path loss, UE distance;

[0119] Spectrum usage categories: PRB utilization, system load, center frequency;

[0120] Transmit configuration: transmit power, modulation scheme (MCS), carrier bandwidth, subcarrier spacing.

[0121] The continuous values ​​of the above context parameters are standardized, and the discrete categories are encoded using one-hot encoding. The final feature vector dimension is 15~30.

[0122] Step 2.2, Network Structure Design

[0123] Employing a lightweight DNN (MLP architecture), it includes:

[0124] Input layer dimension: d∈[15,30], selected based on specific features.

[0125] Hidden layer 1: h1 (e.g., 64) nodes, with ReLU activation function;

[0126] Hidden layer 2: h2 (e.g., 32) nodes, with ReLU activation function;

[0127] Output layer: N nodes, corresponding to N-order FIR coefficients, with activation functions of Linear or Tanh.

[0128] The model function takes the following form:

[0129] ;

[0130] in: ∈R^d: is the input feature vector, where d is the input dimension, containing channel context information such as CQI, BLER, path loss, PRB occupancy, transmit power, etc., in the form after encoding and normalization; ∈R^N represents the output FIR filter coefficient set, with the length N being the same as the filter order; This is a feedforward neural network (MLP) function with parameter set θ.

[0131] The parameter set θ includes the following: First hidden layer parameters (weight matrix and bias term): W(1) ∈ R^(h1×d), b(1) ∈ R^h1; Second hidden layer parameters (weight matrix and bias term): W(2) ∈ R^(h2×h1), b(2) ∈ R^h2; Output layer parameters (weight matrix and bias term): W(3) ∈ R^(N×h2), b(3) ∈ R^N; That is: θ = {W(1),b(1), W(2), b(2), W(3), b(3)};

[0132] Step 2.3: Train the model, as follows:

[0133] During the training phase, offline design tools (such as MATLAB) are used to generate optimal filter coefficients for each set of input channel context parameters based on target passband ripple, stopband attenuation, and transition band indices, using the firpm or remez algorithm as the supervision target for model training.

[0134] Loss function design: Mean squared error (MSE) between the predicted and target coefficients;

[0135] ;

[0136] The above model is continuously adjusted during the training phase using backpropagation algorithms (such as Adam or SGD) to optimize the FIR coefficient vector output by the model. Minimize the MSE error with the target coefficients to optimize the parameter set θ.

[0137] Step 2.4, Online Inference and Interface Output

[0138] After the model is trained, it is deployed on the BBU side and invoked by the scheduling module.

[0139] Real-time reading of current channel context parameters 实时 ;

[0140] Perform forward inference and output filter coefficients. ;

[0141] After compression and quantization, the coefficient set is sent to the RRU via a fronthaul interface (e.g., eCPRI interface).

[0142] Step 3: Design and implementation of the FPGA channel FIR filter module, as detailed below:

[0143] Step 3.1: Generate a coefficient file containing the full bandwidth coefficients.

[0144] Based on the initial filter coefficient set generated in step 1 with the full carrier bandwidth of FR1, the data format of the coefficient set is converted and saved one by one to a coefficient file (such as .coe format) supported by the FPGA FIR IP core. The coefficient sets are sorted according to their order in the coefficient file; assuming there are N sets of coefficients, the corresponding coefficient numbers are 0 to (N-1). A specific set of coefficients can be selected later by its coefficient number, and the coefficient number corresponds one-to-one with the carrier bandwidth. A coefficient mapping table is established to record the mapping relationship between the coefficient number, the coefficient set, and the carrier bandwidth.

[0145] Step 3.2: Instantiate FPGA FIR IP cores based on design parameters

[0146] The channel filter is an FIR filter processing module implemented using an FIR IP core. The FIR IP core is an integrated module on an FPGA device, not the subject of this invention. It implements FIR filtering based on externally input coefficients. When generating an instance of the IP core, it needs to be parameterized, specifically including two points:

[0147] Import the coefficient file (e.g., .coe format) containing all bandwidth coefficients, and store them sequentially in the internal coefficient storage space of the IP core according to the order of the file content. Each group of coefficients occupies the same size storage space block. Number these storage space blocks in ascending order of address. The corresponding storage space block number can be determined according to the coefficient sequence number. Add the mapping relationship between coefficient sequence number and coefficient storage space block number to the coefficient mapping table.

[0148] The input clock frequency is designed to be an integer power of 2 (e.g., 4 times the highest sampling rate) of the total FR1 carrier bandwidth to support all carrier sampling rates.

[0149] Step 4: FPGA configuration module design and implementation, as detailed below:

[0150] The configuration module supports two operations: update and switch. The switching process does not reconstruct the filter structure, thus achieving real-time adaptation under multiple bandwidths. This configuration module implements the following:

[0151] The system receives control commands or AI prediction results from the BBU through the configuration interface, and parses out the configuration content, including: input carrier bandwidth, operation type, and new coefficient data.

[0152] Design a lookup table based on the coefficient mapping table. Using the lookup table, input the bandwidth to determine the selected coefficient index and the selected coefficient storage space block.

[0153] If it is a coefficient update command, the configuration module writes the new coefficient data to the selected coefficient storage space block through the update interface; if it is a coefficient switching command, the configuration module controls the coefficient storage module to output the coefficient corresponding to the selected coefficient number through the switching interface.

[0154] The embodiments of the present invention have the following beneficial effects:

[0155] This invention constructs a filter structure with a unified order, covering all typical bandwidths within the FR1 band. It is uniformly loaded into an FPGA, avoiding structural redundancy and resource waste, and improving the consistency and maintainability of the system design. This invention innovatively introduces a lightweight DNN (such as an MLP) model to perceive the link's operating status in real time. Combined with a pre-trained model, it outputs the optimal FIR coefficients, enabling the filter to optimize according to environmental changes. This dynamic optimization mechanism significantly improves the system's spectrum shaping accuracy and interference control capabilities, far superior to traditional fixed-coefficient filters or static coefficient table loading schemes. It supports online switching and updating of filter coefficients within the same structure through a configuration module. This mechanism not only simplifies hardware resource design but also enhances the system's flexible deployment capabilities in public / private networks and industry-specific scenarios. This invention features a deep coupling design between the filter algorithm, AI model, and FPGA hardware platform, including detailed optimizations such as unified order, coefficient format quantization, and model output encoding, ensuring end-to-end real-time performance from model prediction to hardware updates. Simultaneously, the decoupled hardware-software architecture enhances the system's scalability and engineering feasibility, facilitating practical commercial deployment. By combining AI model prediction at the BBU end with configuration module linkage at the RRU end and dynamic coefficient injection mechanism, this invention realizes a closed-loop control process for intelligent spectrum shaping, which has good algorithm scalability and feedback iteration capability, and can continuously optimize model accuracy as the actual network operation effect changes.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A 5G FR1 shaped filter system, characterized in that, include: The BBU-side intelligent prediction module, the RRU-side configuration management and dynamic coefficient switching module, and the RRU-side unified structure channel filter module; The BBU-end intelligent prediction module is used to collect channel context information during the operation of the wireless link, input the channel context information into a pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU-end configuration management and dynamic coefficient switching module. The RRU end configuration management and dynamic coefficient switching module is used to receive the optimal FIR filter coefficient group, convert the format of the optimal FIR filter coefficient group based on the preset FR1 full carrier bandwidth initial filter coefficient group and store it, and obtain the corresponding target coefficient group from the optimal FIR filter coefficient group based on the obtained configuration command and the preset mapping table. The unified structure channel filter module at the RRU end performs spectrum shaping and filtering on the downlink CP-OFDM signal after loading the target coefficient group into the FIR structure. The BBU-side intelligent prediction module is specifically used for: Obtain context parameters and standardize them. The context parameters include: link quality parameters, spectrum usage parameters, and transmit configuration parameters. The preprocessed context parameters are input into a preset lightweight DNN model, and the optimal FIR filter coefficient set is output. The preset lightweight DNN model construction process includes: Network structure: An MLP structure consisting of an input layer, hidden layers, and an output layer is adopted. The dimension of the input layer is determined by the context parameters. The hidden layer contains a number of nodes and the activation function is ReLU. The number of nodes in the output layer is consistent with the order of the FIR filter. The lightweight DNN model structure is as follows: ； in, ∈R^d is the input feature vector, where d is the input dimension and contains channel context information; ∈R^N represents the output FIR filter coefficient set, with length N consistent with the filter order. For a feedforward neural network (MLP) function with parameter set θ; Model training: Based on the target passband ripple, stopband attenuation, and transition band indices, optimal filter coefficients are generated as the supervision target. The mean square error between the output FIR filter coefficient set and the target coefficients is calculated as the loss function. The model parameters are adjusted using the backpropagation algorithm. The loss function used is as follows: ； in, This represents the coefficients of the i-th target FIR filter, which are continuously adjusted during the training phase using the backpropagation algorithm, so that... Minimize the mean square error between the target coefficient and the target coefficient.

2. The system according to claim 1, characterized in that, The process of constructing the initial filter coefficient set of the preset FR1 full carrier bandwidth includes: Calculate the transmission bandwidth, sampling rate, and normalized transition band ratio for each carrier bandwidth and subcarrier spacing combination within the FR1 band, and generate a normalized transition band ratio table. Select the carrier bandwidth and subcarrier spacing combination corresponding to the minimum normalized transition band ratio from the normalized transition band ratio table as the design constraint. Based on the design constraints and combined with the preset passband ripple and stopband attenuation indices, the unified filter order that meets the requirements of the entire FR1 frequency band is calculated using filter design tools. For each carrier bandwidth, the combination with the smallest normalized transition band ratio among all its subcarrier spacings is selected as the design condition. The corresponding FIR filter coefficients are generated using digital filter design tools, and all coefficients adopt the unified filter order. The generated bandwidth coefficient groups are stored in a preset format to form an initial filter coefficient group covering the entire FR1 frequency band.

3. The system according to claim 1, characterized in that, The preset mapping table construction process includes: The carrier bandwidth and subcarrier spacing within the FR1 band are grouped and assigned unique coefficient numbers; In the FPGA, a memory block is allocated for each coefficient group to record the starting address and length; Establish a mapping relationship between associated carrier bandwidth, subcarrier spacing, coefficient sequence number, and storage block number.

4. The system according to claim 1, characterized in that, The unified structure channel filter module at the RRU end is specifically used for: The target coefficient set is loaded into the FIR IP core of the FPGA; Based on the bandwidth index or AI prediction ID in the configuration instructions, the storage address of the target coefficient group is determined through the mapping table, and the active filter coefficients are switched in real time. Using FPGA as the hardware platform and employing two's complement fixed-point data format, the downlink CP-OFDM signal is subjected to spectrum shaping and filtering, with the processing clock frequency set to an integer multiple covering the highest sampling rate of FR1.

5. A design method for a 5G FR1 shaped filter based on the 5G FR1 shaped filter system according to any one of claims 1-4, characterized in that, include: S1. Collect channel context information during the operation of the wireless link, input the channel context information into the pre-trained lightweight DNN model, output the optimal FIR filter coefficient set, and send the optimal FIR filter coefficient set to the RRU end configuration management and dynamic coefficient switching module. S2. Receive the optimal FIR filter coefficient set, convert the format of the optimal FIR filter coefficient set based on the preset FR1 full carrier bandwidth initial filter coefficient set and store it, and obtain the corresponding target coefficient set from the optimal FIR filter coefficient set based on the obtained configuration command and the preset mapping table. S3. After loading the target coefficient group into the FIR structure, perform spectrum shaping and filtering processing on the downlink CP-OFDM signal.

Images