PD-NOMA and SCMA combined resource allocation method based on converter network and bitter fish optimization algorithm

Abstract

The invention discloses a PD-NOMA and SCMA combined resource allocation method based on a converter network and a bitter-fish optimization algorithm, which comprises the following steps: firstly, synchronously acquiring a power request, a code word cluster and a channel gain in a 0.2 ms period, and dynamically generating a safe resource corridor based on a cell load and adjacent cell interference; after joint interference representation is output through a double-domain self-attention converter, a coupling strength graph is constructed, priority embedding is applied, and a collision-free initial solution group is quickly generated through step-by-step replacement of an index grading power grid and a code word; and then global refinement is carried out by adopting a bitter fish optimization algorithm, the coupling strength is mapped into a water flow resistance coefficient, the step length is adaptively adjusted by multi-target fitness, and optimal power-code word mapping is output through aggregation and dispersion screening. According to the method, cross-domain joint optimization is realized at a sub-millisecond level, the throughput and the user acceptance rate of a high-density scene are improved, and power, code words and interference hard constraints are ensured to be satisfied by 100%.

Description

Technical Field

[0001] This invention relates to the field of communications, and specifically to a joint resource allocation method of PD-NOMA and SCMA based on converter networks and the bitter fish optimization algorithm. Background Technology

[0002] To meet the demands of massive connectivity and high throughput, 5G and subsequent networks will combine Power Domain Non-Orthogonal Multiple Access (PD-NOMA) with Sparse Code Multiple Access (SCMA). PD-NOMA allocates differentiated power to different users on the same time-frequency resources, and the receiver uses Sequential Interference Cancellation (SIC) to achieve multi-user detection; SCMA, on the other hand, enables multiple users to reuse the same carrier through multi-dimensional sparse codeword mapping, further improving spectrum utilization.

[0003] Existing technologies generally adopt a "power-first, codeword-later" sequential framework: the base station first performs power layer calculation based on the user channel gain, then maps the resulting power layer to a predefined SCMA code cluster, and finally recovers the data through iterative detection or parallel interference cancellation. Power domain allocation often employs water level filling, pairwise pairing, or simplified quadratic programming to quickly obtain the layer coefficients using closed-form expressions; the codeword domain uses greedy search, simulated annealing, label propagation, or offline trained codeword priority matrices, and completes online allocation through table lookup.

[0004] While the aforementioned sequential architecture can smoothly interface with the existing physical layer, once the power coefficient is determined, it no longer interacts and iterates with codeword collision information. This breaks the cross-domain coupling between power and codeword, making it difficult to simultaneously optimize system throughput and interference suppression. As carrier density and the number of users increase, both the power domain's water level filling model and the codeword domain's greedy / random traversal face search space explosion, requiring sacrifices in real-time performance or the introduction of coarse-grained sub-block partitioning, leading to boundary effects and resource fragmentation. Although the centralized-distributed two-level scheduling proposed in recent years combines central millisecond-level coarse decision-making with local microsecond-level fine-tuning, the central period is still as long as milliseconds, and local nodes only obtain partial information, making it impossible to respond promptly to sub-millisecond burst interference.

[0005] In recent years, intelligent algorithms for joint allocation have rapidly emerged. Fully connected neural networks, convolutional networks, and long short-term memory networks are used to learn historical scheduling trajectories, replacing traditional optimization with inference, and inference latency can be controlled to the order of hundreds of microseconds. Meanwhile, swarm intelligence methods such as particle swarm optimization, genetic algorithms, and deep reinforcement learning are used for global search. They typically compress capacity, energy consumption, and interference into multi-objective fitness, outputting feasible solutions within finite iterations through step decay and elite retention mechanisms, and demonstrating good adaptability to peak load scenarios in simulation environments. However, deep learning models are mainly trained offline based on historical trajectories. When the load structure or environmental statistics deviate significantly from the training set, inference accuracy decreases, and their output often requires post-heuristic correction to meet power or interference hard constraints, further increasing link latency. Swarm intelligence optimization can provide global search capabilities, but relies on multiple intra-generation iterations to converge, which easily conflicts with scheduling windows in peak scenarios. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems existing in the related art.

[0007] The purpose of this invention is to provide a PD-NOMA and SCMA joint resource allocation method based on converter networks and the bitter fish optimization algorithm. This method can simultaneously complete the joint decision-making of power stacking and sparse codewords in the sub-millisecond scale, and ensure that carrier-level and cell-level interference does not exceed the safety threshold in dynamic high-density scenarios, while maximizing system throughput and user acceptance rate.

[0008] To achieve the above objectives, the present invention provides a joint resource allocation method based on converter networks, the PD-NOMA optimization algorithm, and SCMA, comprising the following steps:

[0009] S1. Simultaneously acquire the power request, target sparse codeword cluster and real-time channel gain of each user on the base station side, and dynamically generate a power-bandwidth security resource corridor based on the cell load coefficient and neighboring cell interference spectrum. Eliminate candidate combinations that exceed the security resource corridor and simultaneously trigger the power domain high conflict threshold and codeword domain high conflict threshold to obtain a quadrature sparse tensor containing service priority, security power range, candidate codeword set and neighboring cell interference level.

[0010] S2. The quaternary sparse tensor is encoded into a feature matrix by a millisecond-level time window and input into a dual-domain self-attention transformer network. Power domain self-attention, codeword domain self-attention and gated fusion are performed in sequence to obtain a joint interference representation. The joint interference representation is used to characterize the coupling interference strength between the power domain and the codeword domain, and the burst interference features are highlighted by short-time memory gates.

[0011] S3. Based on the joint interference representation, construct a power-codeword mapping table, calculate the product of user power level and codeword occupancy intensity to obtain a coupling strength map, apply neighboring cell interference weights to the coupling strength map and perform dynamic depth adjustment, and then multiply it with the service priority embedding to generate differentiated interference embedding.

[0012] S4. Based on the decreasing order of the interference value of the differential interference embedding, an exponential hierarchical strategy is used to construct a multi-layer power grid within the power range. After mapping the multi-layer power grid to the physical carrier, sparse codewords are selected from the candidate codeword set. If a codeword collision is detected, power differential gain correction and codeword step replacement are performed until the collision is eliminated, forming a power-codeword coupled resource table and verifying the carrier interference margin.

[0013] S5. Using the coupling resource table as the initial solution group, the bitter fish optimization algorithm is used for global search: the coupling strength map is mapped to the water flow resistance coefficient, a multi-objective fitness model including system throughput, user acceptance rate and residual interference is constructed, the fish swarm step size is adaptively adjusted according to the energy gradient and the swarming and dispersing screening is performed, and when the swarm energy increase is lower than the set threshold, the final power-codeword mapping and interference margin table are output and the search ends.

[0014] S6. Send the final power-codeword mapping and interference margin table to the terminal, continuously monitor the physical channel measurement report, and trigger local power or codeword fine-tuning when the normalized deviation between the real-time interference and the predicted value exceeds the deviation threshold, and broadcast the update with a differential command until the normalized deviation is lower than the deviation threshold.

[0015] A further preferred embodiment of the present invention is that step S1 specifically comprises:

[0016] S11. Deploy dual-mode RF-codeword probes in each base station sector and its corresponding edge node, with a sampling period of 0.2ms.

[0017] At any moment Probe synchronously captures users Power request Target sparse code cluster and real-time channel gain And encapsulate them into raw access streams in sequence. ;

[0018] S12. Using the IEEE-1588V3 clock synchronization mechanism, the system clock error of the probe is suppressed, and a frame number is written into the header of each sampled data. With beam number Generate header bytes Header bytes The generation follows the following mapping rules:

[0019] ;

[0020] in, This is the absolute timestamp after synchronization. As the system reference time, The duration of a single frame. This is the actual beam pointing angle. As the sector reference angle, Beam spacing;

[0021] S13, Edge servers in the same frame The system performs high-speed fusion of all probe samples, first generating a power domain occupancy matrix. row index Indicates physical stack-up number and column index. Indicates user identifier;

[0022] Subsequently, a codeword domain conflict matrix is ​​generated. row index Indicates the codeword number;

[0023] Two matrices are jointly used to record the PD-NOMA stack sequence number. With SCMA codeword density This allows the overlay-codeword crossover conflict relationship to be explicitly displayed at the data level;

[0024] S14, Edge server calculates cell load factor in real time. and neighboring cell interference spectrum According to the dynamic tightening or relaxation of the power limit between the two, With bandwidth segment threshold To obtain a feasible resource corridor , is represented as:

[0025] ;

[0026] in, and This is the static programming value of the system. This is the linear adjustment coefficient; Indicates the user's candidate transmit power in the frame. This indicates the occupancy of the corresponding candidate bandwidth segment. The minimum bandwidth allocation granularity allowed by the system is used to ensure that the lower limit of the corridor is schedulable and to avoid the generation of unallocable fragments;

[0027] S15, For all candidates Combining detection methods will lead to feasible resource corridors. In addition, any configurations that trigger high conflict thresholds simultaneously in the power domain and codeword domain should be added to the blacklist. ;

[0028] The server also records the trigger frequency. This provides data support for abnormal rollbacks and regular resource cleaning;

[0029] S16. Perform window smoothing and gapping on samples not included in the blacklist, weakening only random noise while preserving weak burst disturbance features, ultimately generating a quaternary sparse tensor. ,in For business priority, For users The safe power range For its candidate codeword set, The level of interference from neighboring cells.

[0030] Preferably, in step S2, the quaternary sparse tensor is encoded into a feature matrix using a millisecond-level time window, specifically as follows:

[0031] S21. Receive the quaternary sparse tensor generated in each frame in step S1. According to millisecond time windows Perform sharding, and categorize all users within the same class according to their user index. Arrange them into row vectors and generate a consistent input matrix on the server side. For each row vector, a joint encoding method combining normalization, density mapping, and priority mapping is used, represented as follows:

[0032] ;

[0033] in, For time windows Number of users within; These are the lower and upper bounds of the global power in the system planning, respectively. This is the total size of the code word set; The highest known neighboring cell interference level; It is the highest business priority;

[0034] S22. Set a cross-domain interference aggregation operator at the network input end. First, the cross-layer feature matrix The column-wise mapping is divided into power security columns, codeword occupancy columns, adjacent interference columns, and priority columns, i.e., the cross-layer feature matrix is ​​split by operators. The four columns of features are encoded into power security vectors. Codeword occupancy vector Neighbor interference vector With priority vector Then, they are concatenated along the channel dimension to form an aggregate vector. ;

[0035] S23, For aggregated vectors Apply channel normalization and position marking, where the normalization operation ensures a balanced gradient distribution across domains, and the position marking... Based on user index and instantaneous load calculation, the input of the dual-domain self-attention transformer network is obtained after addition. .

[0036] Preferably, in step S2, the feature matrix is ​​input into a dual-domain self-attention transformer network, and power domain self-attention, codeword domain self-attention, and gated fusion are performed sequentially to obtain a joint interference representation. Short-time memory gates are then used to highlight the characteristics of sudden interference. Specifically:

[0037] S24, with the first self-attention layer Perform power domain self-attention operations, focusing on power domain interactions. Capture the stacked interference intensity between different power users by querying a one-click-one-value mapping, and output a power relationship graph. ;

[0038] S25, with the second self-attention layer Performing self-attention operations on the codeword domain, focusing on codeword domain interactions, this layer utilizes sparse attention heads to extract collision patterns among users sharing codewords, forming a codeword relationship graph. Subsequently, the power relationship diagram Gated fusion is performed, and the gating function injects codeword collision semantics while maintaining power sensitivity, as shown below:

[0039] ;

[0040] in, For trainable weight matrix, For bias vectors, Activated for element-wise Sigmoid;

[0041] Based on the above formula, a nonlinear mapping of dual-domain coupled information is performed at the matrix level to obtain the fusion relationship graph. ;

[0042] S26. Merge the relationship diagram Input cross-layer association unit Mapping to a unified interference space yields a joint interference representation. Cross-layer association units, through tensor compression and multi-scale convolution enhancement, explicitly characterize the power-codeword dual-domain coupling law, forming a scene-specific joint high-dimensional representation;

[0043] S27. Add a short-time memory gate to the output of the dual-domain self-attention converter network. The short-time memory gate performs a difference operation on the joint representation of continuous time windows, as shown below:

[0044] ;

[0045] And on Adaptive amplification is used to highlight the surge in interference caused by sudden access.

[0046] As a preferred option, step S3 specifically involves:

[0047] S31. Represent the joint interference obtained in step S2. Write to cross-layer scheduling buffer ;

[0048] S32, Control thread follows user sequence number Line-by-line scanning to extract power levels from the power components of the vector. Then search for the codeword occupancy bits given by S2 in the same row. In the index space Construct a mapping table ;

[0049] S33. Insert cross-layer association operators in the intermediate layer of the network. This cross-layer correlation operator With mapping table Interference weight vector As input, a power-codeword coupling strength map is generated through matrix-vector double multiplication. , is represented as:

[0050] ;

[0051] in, The total number of codewords, Indicates user Interference weight from neighboring cells within the same residential area;

[0052] S34. Power-codeword coupling strength diagram Perform short-term gradient monitoring and calculate the normalized increase of adjacent time windows. The calculation formula is as follows:

[0053] ;

[0054] in, It is the Frobenius norm. To prevent zero constant; when At that time, immediately at the current depth Add an attention-convolutional composite module to the existing layer to update the network depth to [the desired depth]. Used to capture fine-grained interference; when At that time, reduce one layer and simultaneously reclaim parameters to avoid overfitting; threshold It is set by an offline validation set and performs periodic self-checks during runtime;

[0055] S35. After each depth adjustment cycle, evaluate the validation loss of each layer's output in real time. The scheduler uses a weighted hierarchy that prioritizes details over depth and maintains a global perspective over shallow depth. Dynamically enable or disable cross-layer jump connections; if the deep validation loss exceeds the weighted average... Immediately activate cross-layer shortcuts to shorten the back gradient path; otherwise, disable shortcuts to maintain steady state.

[0056] S36, Priority gating embedded in the final aggregation unit , to increase coupling strength With user priority Perform element-wise multiplication to generate differential perturbation embeddings, where the first... Individual users in the time window Differential interference embedding as The differential interference embeddings of all users within the same time window are then combined to form a differential interference embedding vector. The identified differential interference is embedded into the vector. Push to resource mapping interface R.

[0057] As a preferred option, step S4 specifically involves:

[0058] S41, Scheduling core receives differentiated interference embedding vector Based on the different interference embedding of each user The values ​​are sorted stably from largest to smallest, resulting in a user index sequence with decreasing priority. ;

[0059] S42, along the user index sequence Top-down filling of multiple power grids ;

[0060] For the ranking users In the power range The internal application of the index hierarchical strategy calculates its target transmit power. :

[0061] ;

[0062] in As a graded index, This further widens the power gap between high and low priority levels;

[0063] S43, Power grid via carrier mapping function Mapped to One physical carrier, generating an initial power allocation matrix. The data is pushed to the SCMA encoding engine in real time, and the mapping process follows the frequency layering principle of prioritizing low-power carriers and high-power carriers.

[0064] S44, encoding engine for matrices Each user in Candidate codeword set Internal selection of sparse code characters ;If the same carrier If a codeword collision occurs, then in the collision indicator matrix... Set the corresponding element to 1 and send it back to the scheduling core;

[0065] S45, Scheduling Core Read For the set of all users whose flag is 1 Applying differential gain correction, the update is performed simultaneously in both the power dimension and the codeword dimension, as follows:

[0066] ;

[0067] in, This is the power differential gain factor. Replace step size for code words. Indicates in The next codeword is selected by sequential circular shifting within the inner sequence; ensuring that each iteration only locally updates conflicting elements, until... Zeroing means that the collision of code words within the carrier has been eliminated;

[0068] S46. Correct the power-codeword pair Write to the coupled resource table And calculate the interference margin for each carrier. ,in To ensure a safe interference threshold, For actual interference estimation; if If the value is positive, the carrier is marked as safe; otherwise, return to step S45 to trigger a secondary correction to ensure that all carriers are within the safe threshold.

[0069] S47, when right When established, the resource table will be coupled. Snapshots are written to the cache. This serves as the initial solution group for the subsequent global search of the bitter fish.

[0070] As a preferred option, step S5 specifically involves:

[0071] S51, from cache Load the coupling resource table For each entry By power Codeword Index With carrier index Establish three-dimensional coordinates and with A uniform subset as the first school of fish This ensures that the search starting point falls within the feasible solution domain, reducing initialization drift;

[0072] S52, refers to each individual fish in the school. Binding interference coupling weights from step S3 ,Will Abstracted as the water flow resistance coefficient, the greater the resistance, the more the individual is in a high conflict zone;

[0073] S53. Design scenario-based multi-objective fitness and convert it into a single energy value. The evaluation function comprehensively considers system throughput. User acceptance rate With residual interference :

[0074] ;

[0075] in, The weights are non-negative and satisfy the following conditions: ; carrier bandwidth; For individuals On carrier Signal-to-interference-plus-noise ratio; This represents the maximum acceptance rate within the window. To correspond to the remaining interference; For safety threshold;

[0076] S54, Based on energy gradient Dynamically adjust individual step size High-energy fish employ small-step, precise searches, while low-energy fish utilize large-step, global exploration. The update formula is as follows:

[0077] ;

[0078] in, radius random vectors within, radius Random vectors within;

[0079] S55, If the updated coordinates fall into Red area This immediately triggers a penalty jump, migrating the object along the negative gradient direction to the disturbance trough. And record the jump vector in the water flow field. In the middle, it provides a basis for calculating the flow of population;

[0080] S56. After each round of search, perform clustering and dispersion filtering, sorting by energy from high to low, then calculating the cosine distance matrix to check path differences, retaining those with high energy and differences greater than a threshold. Individual fish that are homogeneous and have low energy are removed to maintain population diversity while maintaining convergent tension.

[0081] S57, When the overall energy of the group increases Below the threshold Or search window Exhausted, freeze the current peak individual Output its corresponding final power-codeword mapping and full-carrier interference margin table Immediately deliver the control plane for real-time distribution.

[0082] As a preferred option, step S6 specifically involves:

[0083] S61, Scheduling backend read power - codeword triplet With interference margin Format it as A five-element structure is created and written to a dedicated rendering cache. This forms the initial version of the scheduling entries;

[0084] S62. Aggregate the rendering cache by carrier index, call the lightweight Web-GPU component, and draw the power ladder diagram and codeword heat map in real time; for each carrier... and coding Calculate heat value ,in, carrier Up typing The allocated power, This is the maximum power allowed by the system.

[0085] S63, Synchronous decimation of differential gain factor and ,in accordance with The absolute value is used to construct the scheduling priority weight. The execution queue is obtained by sorting all users in descending order. ;

[0086] S64. Integrating the quinary structure with priority Align 3GPPFAPI fields and encapsulate them into a standard resource scheduling table. ;

[0087] S65. After the distribution is completed, the system continuously monitors the physical channel measurement report and monitors real-time interference. Compared with model predictions The normalized deviation is calculated using the following formula:

[0088] ;

[0089] in, For carrier-by-carrier offset threshold, For index functions; when Immediately invoke the cross-layer embedding generated in step S3. Re-estimate the local power-codeword coupling strength to trigger a fast correction process;

[0090] S66. Based on the new estimated intensity, the affected user set... Perform fine-tuning: If Then reduce the power for the corresponding user. If the probability of codeword collision increases, then... Inward right-hand rotation step selects the next sparse codeword ;

[0091] S67. After all adjustments are completed, write the latest power-codeword-interference triplet back to the visualization panel, and... Time-series statistics are entered into the historical database. .

[0092] As a preferred approach, after the preset evaluation window ends, the system capacity, interference suppression rate, and number of converged frames are statistically analyzed. If any of these indicators is below the threshold, fast sampling is initiated and the search radius for the bitter fish is reset to perform incremental search until the real-time indicator requirements are met. The search parameters are then written into the long-term memory pool to complete cross-cycle adaptive learning.

[0093] Preferably, after the preset evaluation window ends, the system capacity, interference suppression rate, and convergence frame count are statistically analyzed. If any indicator falls below a threshold, rapid sampling is initiated, and the search radius for the "bitter fish" (a metaphor for a type of fish) is reset to perform incremental search until the real-time indicator requirements are met. The search parameters are then written into the long-term memory pool to complete cross-cycle adaptive learning. Specifically:

[0094] S71. Retrieve the transmission receipt logs returned by the base station and summarize the data for each carrier within the evaluation window. Real-time throughput within Instantaneous rate of each user and remaining interference Write the three types of data into the evaluation cache. ;

[0095] S72, in When finished, the statistics module is called to calculate the system capacity. Average speed Interference suppression rate With convergence frame count And generate a baseline report, which is represented as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] in, For carrier bandwidth, carrier exist The average signal-to-interference-plus-noise ratio within the range, For active users, For users Average speed, carrier Average actual interference, To ensure a safe interference threshold, For the first Frame power-codeword coupling strength diagram To converge the decision threshold;

[0101] S73, will With preset threshold vector Perform element-level comparison; if it exists If the indicator is not met, the current assessment window is marked as a high-risk period and the adaptive process is triggered; otherwise, the original parameters are maintained and the process continues.

[0102] S74. Initiate rapid sampling during high-risk periods to capture the latest user access graph. With real-time channel matrix And expand the input window of the cross-layer feature extractor to ;

[0103] S75. Reset the global exploration radius of the improved bitter fish search based on the updated cross-layer features. With local fine-tuning radius , is represented as:

[0104] ;

[0105] ;

[0106] in, and The radius of the previous cycle, The maximum positive interference deviation within the current window. For the minimum rate gap, This is the step size scaling factor;

[0107] S76. Perform an incremental search for bitter fish and output the corrected power-codeword mapping based on the new radius. The affected carriers and users are updated using a partial distribution mechanism; the statistics are recalculated immediately after the distribution is completed. And calculate the improvement vector ;

[0108] S77, if Each component is no less than the corresponding minimum improvement threshold. Then freeze the current search parameters. And written to the long-term memory pool If any component fails to meet the standard, the evaluation window will be shortened to [a shorter timeframe]. The process continues to execute steps S74 to S76 repeatedly until the real-time indicator requirements are met; after completing the cycle, the process enters the next regular evaluation cycle.

[0109] In another aspect, the present invention provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute the above-described resource allocation method based on converter networks and the bitter fish optimization algorithm PD-NOMA and SCMA.

[0110] In another aspect, the present invention provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls logical instructions in the memory to execute the above-mentioned resource allocation method based on converter network and the bitter fish optimization algorithm PD-NOMA and SCMA.

[0111] In another aspect, the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer executes the above-described resource allocation method based on converter networks and the bitter fish optimization algorithm PD-NOMA and SCMA.

[0112] Beneficial effects: The joint resource allocation method based on converter network and bitter fish optimization algorithm PD-NOMA and SCMA of the present invention can simultaneously consider the joint resource allocation mechanism of power stacking and sparse codeword selection in the sub-millisecond scale, so that the base station can still ensure that the carrier level and cell level interference do not exceed the safety threshold in dynamic and high-density access environment, while maximizing system throughput and user acceptance rate.

[0113] Compared with the prior art, the advantages of the present invention are as follows:

[0114] Achieving sub-millisecond-level cross-domain joint optimization: Through 0.2ms probe sampling, millisecond-level time window coding, and a dual-domain self-attention transformer network, the coupling interference between the power domain and the codeword domain is quantized in real time at the sub-millisecond scale; combined with the adaptive step size mechanism of fast initial ungroup generation and the bitter fish optimization algorithm, the resource allocation decision latency meets the 5G NR sub-millisecond-level scheduling window requirements, solving the coupling fragmentation and response lag problems caused by sequential allocation.

[0115] Improve system performance in high-density scenarios: Dynamically capture power-codeword superposition effect using coupling strength graph, and prioritize the embedding of differentiated interference to ensure the safe power range of high-priority users. Compared with traditional water level filling and greedy traversal, it significantly improves system throughput and user acceptance rate under the same carrier density and user scale, and controls the interference exceedance rate to within 5%.

[0116] To ensure that hard constraints are met: the safe resource corridor pre-eliminates power-bandwidth combinations that exceed limits, the codeword collision detection closed-loop executes power differential gain correction and step replacement until the conflict is eliminated, and the bitter fish optimization algorithm uses the remaining interference as the negative objective function term to ensure that the final power-codeword mapping always meets the triple hard constraints of power upper limit, codeword collision and interference threshold.

[0117] Enhanced environmental adaptability: Short-time memory gates amplify sudden disturbance gradients in real time, coupling strength maps dynamically adjust network depth and cross-layer jump connections, the bitter fish algorithm adaptively balances local refinement and global exploration based on energy gradients, and the evaluation window feedback mechanism dynamically adjusts the search radius, enabling the system to converge faster and maintain stable optimality when there are sudden changes in load structure or neighboring area disturbances.

[0118] Supports smooth deployment and low-overhead operation: Edge servers and base station probes use IEEE-1588V3 synchronization to achieve cross-domain data alignment, 3GPP FAPI zero-copy delivery and differential command broadcasting avoid full table retransmission, and Web-GPU visualization panel renders resource allocation status in real time; the whole solution does not require modification of the existing physical layer implementation and is compatible with the 5G NR protocol stack. Attached Figure Description

[0119] Figure 1 This is an overall flowchart of the resource allocation method based on converter networks and the bitter fish optimization algorithm PD-NOMA and SCMA of the present invention;

[0120] Figure 2 This is a flowchart of the dual-domain self-attention transformer network encoding process in this invention;

[0121] Figure 3 This is a flowchart of the global optimization process based on the bitter fish algorithm in this invention. Detailed Implementation

[0122] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0123] The following is combined with Figures 1-3 This invention describes a joint resource allocation method based on converter networks and the bitter fish optimization algorithm PD-NOMA and SCMA.

[0124] Example: This example provides a resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms, such as... Figure 1 As shown, it includes the following steps:

[0125] S1. Simultaneously acquire the power request, target sparse codeword cluster and real-time channel gain of each user on the base station side, and dynamically generate a power-bandwidth security resource corridor based on the cell load coefficient and neighboring cell interference spectrum. Eliminate candidate combinations that exceed the security resource corridor and simultaneously trigger the power domain high conflict threshold and codeword domain high conflict threshold to obtain a quadrature sparse tensor containing service priority, security power range, candidate codeword set and neighboring cell interference level.

[0126] S11. Deploy dual-mode RF-codeword probes in each base station sector and its corresponding edge node, with a sampling period of 0.2ms.

[0127] At any moment Probe synchronously captures users Power request Target sparse code cluster and real-time channel gain And encapsulate them into raw access streams in sequence. .

[0128] S12. Using the IEEE-1588V3 clock synchronization mechanism, the system clock error of the probe is suppressed, and a frame number is written into the header of each sampled data. With beam number Generate header bytes Header bytes The generation follows the following mapping rules:

[0129] ;

[0130] in, This is the absolute timestamp after synchronization. As the system reference time, The duration of a single frame. This is the actual beam pointing angle. As the sector reference angle, Beam spacing;

[0131] Based on this encoding, probe data from different cells and different beams are aligned to the same time slot. Within this framework, cross-domain information is subsequently mapped at the matrix level.

[0132] S13, Edge servers in the same frame The system performs high-speed fusion of all probe samples, first generating a power domain occupancy matrix. row index Indicates physical stack-up number and column index. Indicates user identifier;

[0133] Subsequently, a codeword domain conflict matrix is ​​generated. row index Indicates the codeword number;

[0134] Two matrices are jointly used to record the PD-NOMA stack sequence number. With SCMA codeword density This allows the overlay-codeword crossover conflict relationship to be explicitly displayed at the data level.

[0135] S14, Edge server calculates cell load factor in real time. and neighboring cell interference spectrum According to the dynamic tightening or relaxation of the power limit between the two, With bandwidth segment threshold To obtain a feasible resource corridor , is represented as:

[0136] ;

[0137] in, and This is the static programming value of the system. This is the linear adjustment coefficient; Indicates the user's candidate transmit power in the frame. This indicates the occupancy of the corresponding candidate bandwidth segment. The minimum bandwidth allocation granularity allowed by the system is used to ensure that the lower limit of the corridor is schedulable and to avoid the generation of unallocable fragments;

[0138] The corridor directly reflects the dual safety redundancy requirements for power and bandwidth during peak periods, serving as a hard constraint boundary for mapping and optimization.

[0139] S15, For all candidates Combining detection methods will lead to feasible resource corridors. In addition, any configurations that trigger high conflict thresholds simultaneously in the power domain and codeword domain should be added to the blacklist. ;

[0140] The server also records the trigger frequency. This provides data support for abnormal rollbacks and regular resource cleaning.

[0141] S16. Perform window smoothing and gapping on samples not included in the blacklist, weakening only random noise while preserving weak burst disturbance features, ultimately generating a quaternary sparse tensor. ,in For business priority, For users The safe power range For its candidate codeword set, The level of interference from neighboring cells.

[0142] S2. The quaternion sparse tensor is encoded into a feature matrix using a millisecond-level time window and input into a dual-domain self-attention transformer network. Power domain self-attention, codeword domain self-attention, and gated fusion are then performed sequentially to obtain a joint interference representation. This joint interference representation characterizes the coupling interference strength between the power domain and the codeword domain, and highlights burst interference features through a short-time memory gate. Specific steps are as follows: Figure 2 As shown, it includes:

[0143] S21. Receive the quaternary sparse tensor generated in each frame in step S1. According to millisecond time windows Perform sharding, and categorize all users within the same class according to their user index. Arrange them into row vectors and generate a consistent input matrix on the server side. For each row vector, a joint encoding method combining normalization, density mapping, and priority mapping is used, represented as follows:

[0144]

[0145] in, For time windows Number of users within; These are the lower and upper bounds of the global power in the system planning, respectively. This is the total size of the code word set; The highest known neighboring cell interference level; It is the highest business priority;

[0146] The matrix maps quaternion information uniformly to The range ensures amplitude consistency and semantic integrity of the scene.

[0147] S22. Set a cross-domain interference aggregation operator at the network input end. First, the cross-layer feature matrix The column-wise mapping is divided into power security columns, codeword occupancy columns, adjacent interference columns, and priority columns, i.e., the cross-layer feature matrix is ​​split by operators. The four columns of features are encoded into power security vectors. Codeword occupancy vector Neighbor interference vector With priority vector Then, they are concatenated along the channel dimension to form an aggregate vector. The aggregated vectors retain domain details and provide a unified entry point for subsequent multi-head attention in the network.

[0148] S23, For aggregated vectors Apply channel normalization and position marking, where the normalization operation ensures a balanced gradient distribution across domains, and the position marking... Based on user index and instantaneous load calculation, the input of the dual-domain self-attention transformer network is obtained after addition. This enables the network to synchronously identify load bursts caused by power levels, codeword sparsity, and user arrival order.

[0149] S24, with the first self-attention layer Perform power domain self-attention operations, focusing on power domain interactions. Capture the stacked interference intensity between different power users by querying a one-click-one-value mapping, and output a power relationship graph. The relationship graph provides an explicit power weight benchmark for subsequent codeword domain attention layers, reducing the probability of misjudging sparse features under high load conditions.

[0150] S25, with the second self-attention layer Performing self-attention operations on the codeword domain, focusing on codeword domain interactions, this layer utilizes sparse attention heads to extract collision patterns among users sharing codewords, forming a codeword relationship graph. Subsequently, the power relationship diagram Gated fusion is performed, and the gating function injects codeword collision semantics while maintaining power sensitivity, as shown below:

[0151] ;

[0152] in, For trainable weight matrix, For bias vectors, Activated for element-wise Sigmoid;

[0153] Based on the above formula, a nonlinear mapping of dual-domain coupled information is performed at the matrix level to obtain the fusion relationship graph. .

[0154] S26. Merge the relationship diagram Input cross-layer association unit Mapping to a unified interference space yields a joint interference representation. Cross-layer association units, through tensor compression and multi-scale convolution enhancement, explicitly characterize the power-codeword dual-domain coupling pattern, forming a scene-specific joint high-dimensional representation; output It combines hierarchical dependencies with details of local conflicts, providing a differentiable feature basis for the resource mapping module.

[0155] S27. Add a short-time memory gate to the output of the dual-domain self-attention converter network. The short-time memory gate performs a difference operation on the joint representation of continuous time windows, as shown below:

[0156] ;

[0157] And on Adaptive amplification is performed to highlight the surge in interference caused by sudden access. The amplified result, along with the current frame representation, is output to the resource mapping stage to ensure that subsequent power-codeword joint scheduling can perceive and respond to drastic changes in network status on a sub-millisecond scale.

[0158] S3. Based on the joint interference representation, construct a power-codeword mapping table, calculate the product of user power level and codeword occupancy strength to obtain a coupling strength map, apply neighboring cell interference weights to the coupling strength map and perform dynamic depth adjustment, and then multiply it with the service priority embedding to generate differentiated interference embedding.

[0159] S31. Represent the joint interference obtained in step S2. Write to cross-layer scheduling buffer .

[0160] S32, Control thread follows user sequence number Line-by-line scanning to extract power levels from the power components of the vector. Then search for the codeword occupancy bits given by S2 in the same row. In the index space Construct a mapping table .

[0161] S33. Insert cross-layer association operators in the intermediate layer of the network. This cross-layer correlation operator With mapping table Interference weight vector As input, a power-codeword coupling strength map is generated through matrix-vector double multiplication. , is represented as:

[0162] ;

[0163] in, The total number of codewords, Indicates user Interference weight from neighboring cells within the same residential area;

[0164] Based on the above formula, the power interaction term is accumulated in the shared codeword dimension and the neighboring cell interference adjustment coefficient is applied to depict the absolute intensity of the dual-domain superposition effect in real time.

[0165] S34. Power-codeword coupling strength diagram Perform short-term gradient monitoring and calculate the normalized increase of adjacent time windows. The calculation formula is as follows:

[0166] ;

[0167] in, It is the Frobenius norm. To prevent zero constant; when At that time, immediately at the current depth Add an attention-convolutional composite module to the existing layer to update the network depth to [the desired depth]. Used to capture fine-grained interference; when At that time, reduce one layer and simultaneously reclaim parameters to avoid overfitting; threshold It is set by an offline validation set and performs periodic self-checks during runtime.

[0168] S35. After each depth adjustment cycle, evaluate the validation loss of each layer's output in real time. The scheduler uses a weighted hierarchy that prioritizes details over depth and maintains a global perspective over shallow depth. Dynamically enable or disable cross-layer jump connections; if the deep validation loss exceeds the weighted average... Immediately activate cross-layer shortcuts to shorten the back gradient path; conversely, close shortcuts to maintain steady state, ensuring that the network computation path adapts to the optimal scenario throughout the process.

[0169] S36, Priority gating embedded in the final aggregation unit , to increase coupling strength With user priority Perform element-wise multiplication to generate differential perturbation embeddings, where the first... Individual users in the time window Differential interference embedding as The differential interference embeddings of all users within the same time window are then combined to form a differential interference embedding vector. The identified differential interference is embedded into the vector. Push to resource mapping interface R.

[0170] S37. Embed the labeled vector stream Push to resource mapping interface .

[0171] S4. Based on the decreasing order of the interference value of the differential interference embedding, an exponential hierarchical strategy is used to construct a multi-layer power grid within the power range. After mapping the multi-layer power grid to the physical carrier, sparse codewords are selected from the candidate codeword set. If a codeword collision is detected, power differential gain correction and codeword step replacement are performed until the collision is eliminated, forming a power-codeword coupled resource table and verifying the carrier interference margin.

[0172] S41, Scheduling core receives differentiated interference embedding vector Based on the different interference embedding of each user The values ​​are sorted stably from largest to smallest, resulting in a user index sequence with decreasing priority. The sequence records the user with the highest power sensitivity first, ensuring that high-risk users get the safe power window first when the raster is filled later.

[0173] S42, along the user index sequence Top-down filling of multiple power grids ;

[0174] For the ranking users In the power range The internal application of the index hierarchical strategy calculates its target transmit power. :

[0175] ;

[0176] in As a graded index, This further widens the power gap between high and low priority levels.

[0177] Based on the above formula, while ensuring that high-risk users are in the low-power seat belt, adjustable levels are provided for subsequent users. Sequential write The multi-layer power grid is then constructed.

[0178] S43, Power grid via carrier mapping function Mapped to One physical carrier, generating an initial power allocation matrix. The data is pushed to the SCMA encoding engine in real time. During the mapping process, the frequency layering principle of prioritizing low-power carriers and high-power carriers is followed to reduce cross-carrier co-channel interference.

[0179] S44, encoding engine for matrices Each user in Candidate codeword set Internal selection of sparse code characters ;If the same carrier If a codeword collision occurs, then in the collision indicator matrix... Set the corresponding element to 1 and send it back to the scheduling core.

[0180] S45, Scheduling Core Read For the set of all users whose flag is 1 Applying differential gain correction, the update is performed simultaneously in both the power dimension and the codeword dimension, as follows:

[0181] ;

[0182] in, This is the power differential gain factor. Replace step size for code words. Indicates in The next codeword is selected by sequential circular shifting within the inner sequence; ensuring that each iteration only locally updates conflicting elements, until... Zeroing means that the collision of code words within the carrier has been eliminated.

[0183] S46. Correct the power-codeword pair Write to the coupled resource table And calculate the interference margin for each carrier. ,in To ensure a safe interference threshold, For actual interference estimation; if If the result is positive, the carrier is marked as safe; otherwise, return to step S45 to trigger a secondary correction to ensure that all carriers are within the safe threshold.

[0184] S47, when right When established, the resource table will be coupled. Snapshots are written to the cache. This serves as the initial solution group for the subsequent global search of the bitter fish. Because... Having passed the conflict self-correction check in this step, duplicate detection can be skipped during the search phase, significantly shortening the convergence time.

[0185] This step completes the rapid generation and security verification of power-codeword coupling resources, leaving ample time for intelligent optimization.

[0186] S5. Using the coupled resource table as the initial solution group, a global search is performed using the bitter fish optimization algorithm: the coupling strength map is mapped to the water flow resistance coefficient, and a multi-objective fitness model is constructed, including system throughput, user acceptance rate, and residual interference. The fish swarm step size is adaptively adjusted according to the energy gradient, and aggregation and dispersion screening are performed. When the swarm energy increase is lower than a set threshold, the final power-codeword mapping and interference margin table are output, and the search ends. Specific steps are as follows: Figure 3 As shown, it includes:

[0187] S51, from cache Load the coupling resource table For each entry By power Codeword Index With carrier index Establish three-dimensional coordinates and with A uniform subset as the first school of fish This ensures that the search starting point falls within the feasible solution domain, reducing initialization drift;

[0188] S52, refers to each individual fish in the school. Binding interference coupling weights from step S3 ,Will Abstracted as the water flow resistance coefficient, the greater the resistance, the more the individual is in a high-conflict zone; when updating coordinates, the fish automatically moves away from the high-resistance area, forming a self-driven avoidance based on physical meaning, thereby shortening the ineffective exploration distance.

[0189] S53. Design scenario-based multi-objective fitness and convert it into a single energy value. The evaluation function comprehensively considers system throughput. User acceptance rate With residual interference :

[0190] ;

[0191] in, The weights are non-negative and satisfy the following conditions: ; carrier bandwidth; For individuals On carrier Signal-to-interference-plus-noise ratio; This represents the maximum acceptance rate within the window. To correspond to the remaining interference; For safety threshold;

[0192] By using dimensionless normalization and differential weighting, multiple objectives are compressed into a single scalar energy. The higher the energy value, the more beneficial the power-codeword mapping is to the overall system benefits.

[0193] S54, Based on energy gradient Dynamically adjust individual step size High-energy fish employ small-step, precise searches, while low-energy fish utilize large-step, global exploration. The update formula is as follows:

[0194] ;

[0195] in, radius random vectors within, radius The random vector within; the Sigmoid coefficient continuously schedules the step size, enabling the algorithm to adaptively balance local refinement and global exploration in the same iteration.

[0196] S55, If the updated coordinates fall into Red area This immediately triggers a penalty jump, migrating the object along the negative gradient direction to the disturbance trough. And record the jump vector in the water flow field. In the middle, it provides a basis for calculating the flow status of the group.

[0197] S56. After each round of search, perform clustering and dispersion filtering, sorting by energy from high to low, then calculating the cosine distance matrix to check path differences, retaining those with high energy and differences greater than a threshold. Individual fish are eliminated by removing those that are homogeneous and have low energy levels, thus maintaining population diversity while preserving convergent tension.

[0198] S57, When the overall energy of the group increases Below the threshold Or search window Exhausted, freeze the current peak individual Output its corresponding final power-codeword mapping and full-carrier interference margin table The data is immediately delivered to the control plane for real-time distribution. At this point, the power-codeword joint allocation has completed global refinement, maximizing system throughput and acceptance efficiency while ensuring safety thresholds.

[0199] S6. Send the final power-codeword mapping and interference margin table to the terminal, continuously monitor the physical channel measurement report, and trigger local power or codeword fine-tuning when the normalized deviation between the real-time interference and the predicted value exceeds the deviation threshold, and broadcast the update with a differential command until the normalized deviation is lower than the deviation threshold.

[0200] S61, Scheduling backend read power - codeword triplet With interference margin Format it as A five-element structure is created and written to a dedicated rendering cache. This forms the first version of the scheduling entries, providing a unified data base for visualization and interface encapsulation.

[0201] S62. Aggregate the rendering cache by carrier index, call the lightweight Web-GPU component, and draw the power ladder diagram and codeword heat map in real time; for each carrier... and coding Calculate heat value ,in, carrier Up typing The allocated power, The maximum allowable power of the system; output under the coupling of power normalization and interference margin. The temperature range is mapped to the color channel using a linear gradient, making the remaining interference margin visible and presenting the temperature distribution.

[0202] S63, Synchronous decimation of differential gain factor and ,in accordance with The absolute value is used to construct the scheduling priority weight. The execution queue is obtained by sorting all users in descending order. The queue reserves time slices for multi-threaded tasks of the base station to enable rapid connection for high-risk users.

[0203] S64. Integrating the quinary structure with priority Align 3GPPFAPI fields and encapsulate them into a standard resource scheduling table. The scheduling table is pushed to the edge control plane in one go using the zero-copy MMIO method, avoiding redundant copying and shortening the air interface transmission latency, achieving millisecond-level configuration refresh.

[0204] S65. After the distribution is completed, the system continuously monitors the physical channel measurement report and monitors real-time interference. Compared with model predictions The normalized deviation is calculated using the following formula:

[0205] ;

[0206] in, For carrier-by-carrier offset threshold, For index functions; when Immediately invoke the cross-layer embedding generated in step S3. Re-estimate the local power-codeword coupling strength to trigger a fast correction process.

[0207] S66. Based on the new estimated intensity, the affected user set... Perform fine-tuning: If Then reduce the power for the corresponding user. If the probability of codeword collision increases, then... Inward right-hand rotation step selects the next sparse codeword The above changes are encapsulated as differential instructions. It only covers local fields and broadcasts them via the uncompressed PDCP link, avoiding resending the entire table.

[0208] S67. After all adjustments are completed, write the latest power-codeword-interference triplet back to the visualization panel, and... Time-series statistics are entered into the historical database. The statistics are used as dynamic priors for improving the initial population of the bitter fish algorithm in the next scheduling cycle, realizing cross-cycle knowledge accumulation. At this point, the closed-loop correction and visualization update are all completed.

[0209] S7. After the preset evaluation window ends, the system capacity, interference suppression rate and convergence frame count are statistically analyzed. If any indicator is lower than the threshold, fast sampling is started and the search radius of the bitter fish is reset to perform incremental search until the real-time indicator requirements are met. The search parameters are written into the long-term memory pool to complete cross-cycle adaptive learning.

[0210] S71. Retrieve the transmission receipt logs returned by the base station and summarize the data for each carrier within the evaluation window. Real-time throughput within Instantaneous rate of each user and remaining interference Write the three types of data into the evaluation cache. This ensures consistent sampling for subsequent statistical analysis.

[0211] S72, in When finished, the statistics module is called to calculate the system capacity. Average speed Interference suppression rate With convergence frame count And generate a baseline report, which is represented as follows:

[0212] ;

[0213] ;

[0214] ;

[0215] ;

[0216] in, For carrier bandwidth, carrier exist The average signal-to-interference-plus-noise ratio within the range, For active users, For users Average speed, carrier Average actual interference, To ensure a safe interference threshold, For the first Frame power-codeword coupling strength diagram To convergence decision threshold.

[0217] S73, will With preset threshold vector Perform element-level comparison; if it exists If the indicator is not met, the current evaluation window is marked as a high-risk period and the adaptive process is triggered; otherwise, the original parameters are maintained and the process continues.

[0218] S74. Initiate rapid sampling during high-risk periods to capture the latest user access graph. With real-time channel matrix And expand the input window of the cross-layer feature extractor to This ensures that the search algorithm has the latest prior knowledge regarding mutation interference and capacity bottlenecks.

[0219] S75. Reset the global exploration radius of the improved bitter fish search based on the updated cross-layer features. With local fine-tuning radius , is represented as:

[0220] ;

[0221] ;

[0222] in, and The radius of the previous cycle, The maximum positive interference deviation within the current window. For the minimum rate gap, This is the step size scaling factor; by simultaneously adjusting the global boundary and local step size through dual feedback of interference and rate, the search focus is automatically biased towards global exploration in high-interference domains and fine mining by users with capacity bottlenecks.

[0223] S76. Perform an incremental search for bitter fish and output the corrected power-codeword mapping based on the new radius. The affected carriers and users are updated using a partial distribution mechanism; the statistics are recalculated immediately after the distribution is completed. And calculate the improvement vector .

[0224] S77, if Each component is no less than the corresponding minimum improvement threshold. Then freeze the current search parameters. And written to the long-term memory pool If any component fails to meet the standard, the evaluation window will be shortened to [a shorter timeframe]. The system then repeats steps S74 to S76 until the real-time performance requirements are met. After the cycle is completed, the system enters the next regular evaluation cycle, thus achieving cross-cycle adaptive learning and continuous performance improvement of our overall technical solution.

[0225] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0226] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A resource allocation method based on converter networks, the PD-NOMA optimization algorithm, and SCMA, characterized in that, Includes the following steps: S1. Simultaneously acquire the power request, target sparse codeword cluster and real-time channel gain of each user on the base station side, and dynamically generate a power-bandwidth security resource corridor based on the cell load coefficient and neighboring cell interference spectrum. Eliminate candidate combinations that exceed the security resource corridor and simultaneously trigger the power domain high conflict threshold and codeword domain high conflict threshold to obtain a quadrature sparse tensor containing service priority, security power range, candidate codeword set and neighboring cell interference level. S2. The quaternary sparse tensor is encoded into a feature matrix by a millisecond-level time window and input into a dual-domain self-attention transformer network. Power domain self-attention, codeword domain self-attention and gated fusion are performed in sequence to obtain a joint interference representation. The joint interference representation is used to characterize the coupling interference strength between the power domain and the codeword domain, and the burst interference features are highlighted by short-time memory gates. S3. Based on the joint interference representation, construct a power-codeword mapping table, calculate the product of user power level and codeword occupancy intensity to obtain a coupling strength map, apply neighboring cell interference weights to the coupling strength map and perform dynamic depth adjustment, and then multiply it with the service priority embedding to generate differentiated interference embedding. S4. Based on the decreasing order of the interference value of the differential interference embedding, an exponential hierarchical strategy is used to construct a multi-layer power grid within the power range. After mapping the multi-layer power grid to the physical carrier, sparse codewords are selected from the candidate codeword set. If a codeword collision is detected, power differential gain correction and codeword step replacement are performed until the collision is eliminated, forming a power-codeword coupled resource table and verifying the carrier interference margin. S5. Using the coupling resource table as the initial solution group, the bitter fish optimization algorithm is used for global search: the coupling strength map is mapped to the water flow resistance coefficient, a multi-objective fitness model including system throughput, user acceptance rate and residual interference is constructed, the fish swarm step size is adaptively adjusted according to the energy gradient and the swarming and dispersing screening is performed, and when the swarm energy increase is lower than the set threshold, the final power-codeword mapping and interference margin table are output and the search ends. S6. Send the final power-codeword mapping and interference margin table to the terminal, continuously monitor the physical channel measurement report, and trigger local power or codeword fine-tuning when the normalized deviation between the real-time interference and the predicted value exceeds the deviation threshold, and broadcast the update with a differential command until the normalized deviation is lower than the deviation threshold.

2. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 1, characterized in that, Step S1 is as follows: S11. Deploy dual-mode RF-codeword probes in each base station sector and its corresponding edge node, with a sampling period of 0.2ms. At any moment Probe synchronously captures users Power request Target sparse code cluster and real-time channel gain And encapsulate them into raw access streams in sequence. ; S12. Using the IEEE-1588V3 clock synchronization mechanism, the system clock error of the probe is suppressed, and a frame number is written into the header of each sampled data. With beam number Generate header bytes Header bytes The generation follows the following mapping rules: ； in, This is the absolute timestamp after synchronization. As the system reference time, The duration of a single frame. This is the actual beam pointing angle. As the sector reference angle, Beam spacing; S13, Edge servers in the same frame The system performs high-speed fusion of all probe samples, first generating a power domain occupancy matrix. row index Indicates physical stack-up number and column index. Indicates user identifier; Subsequently, a codeword domain conflict matrix is ​​generated. row index Indicates the codeword number; Two matrices jointly record the PD-NOMA stack sequence number. With SCMA codeword density This allows the overlay-codeword crossover conflict relationship to be explicitly displayed at the data level; S14, Edge server calculates cell load factor in real time. and neighboring cell interference spectrum According to the dynamic tightening or relaxation of the power limit between the two, With bandwidth segment threshold To obtain a feasible resource corridor , is represented as: ； in, and This is the static programming value of the system. This is the linear adjustment coefficient; Indicates the user's candidate transmit power in the frame. This indicates the corresponding candidate bandwidth segment usage. The minimum bandwidth allocation granularity allowed by the system is used to ensure that the lower limit of the corridor is schedulable and to avoid the generation of unallocable fragments; S15, For all candidates Combining detection methods will lead to feasible resource corridors. In addition, any configurations that trigger high conflict thresholds simultaneously in the power domain and codeword domain should be added to the blacklist. ; The server also records the trigger frequency. This provides data support for abnormal rollbacks and regular resource cleaning; S16. Perform window smoothing and gapping on samples not included in the blacklist, weakening only random noise while preserving weak burst disturbance features, ultimately generating a quaternary sparse tensor. ,in For business priority, For users The safe power range For its candidate codeword set, The level of interference from neighboring cells.

3. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 2, characterized in that, In step S2, the quaternion sparse tensor is encoded into a feature matrix using a millisecond-level time window, specifically as follows: S21. Receive the quaternary sparse tensor generated in each frame in step S1. According to millisecond time windows Perform sharding, and categorize all users within the same class according to their user index. Arrange them into row vectors and generate a consistent input matrix on the server side. For each row vector, a joint encoding method combining normalization, density mapping, and priority mapping is used, represented as follows: ； in, For time windows Number of users within; These are the lower and upper bounds of the global power in the system planning, respectively. This is the total size of the code word set; The highest known neighboring cell interference level; It is the highest business priority; S22. Set a cross-domain interference aggregation operator at the network input end. First, the cross-layer feature matrix The column-wise mapping is divided into power security columns, codeword occupancy columns, adjacent interference columns, and priority columns, i.e., the cross-layer feature matrix is ​​split by operators. The four columns of features are encoded into power security vectors. Codeword occupancy vector Neighbor interference vector With priority vector Then, they are concatenated along the channel dimension to form an aggregate vector. ; S23, For aggregated vectors Apply channel normalization and position marking, where the normalization operation ensures a balanced gradient distribution across domains, and the position marking... Based on user index and instantaneous load calculation, the input of the dual-domain self-attention transformer network is obtained after addition. .

4. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 3, characterized in that, In step S2, the feature matrix is ​​input into the dual-domain self-attention transformer network, and power domain self-attention, codeword domain self-attention, and gated fusion are performed sequentially to obtain a joint interference representation. Short-time memory gates are then used to highlight burst interference features. Specifically: S24, with the first self-attention layer Perform power domain self-attention operations, focusing on power domain interactions. Capture the stacked interference intensity between different power users by querying a one-click-one-value mapping, and output a power relationship graph. ; S25, with the second self-attention layer Performing self-attention operations on the codeword domain, focusing on codeword domain interactions, this layer utilizes sparse attention heads to extract collision patterns among users sharing codewords, forming a codeword relationship graph. Subsequently, the power relationship diagram Gated fusion is performed, and the gating function injects codeword collision semantics while maintaining power sensitivity, as shown below: ； in, For trainable weight matrix, For bias vectors, Activated for element-wise Sigmoid; Based on the above formula, a nonlinear mapping of dual-domain coupled information is performed at the matrix level to obtain the fusion relationship graph. ; S26. Merge the relationship diagram Input cross-layer association unit Mapping to a unified interference space yields a joint interference representation. Cross-layer association units, through tensor compression and multi-scale convolution enhancement, explicitly characterize the power-codeword dual-domain coupling law, forming a scene-specific joint high-dimensional representation; S27. Add a short-time memory gate to the output of the dual-domain self-attention converter network. The short-time memory gate performs a difference operation on the joint representation of continuous time windows, as shown below: ； And on Adaptive amplification is used to highlight the surge in interference caused by sudden access.

5. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 4, characterized in that, Step S3 is as follows: S31. Represent the joint interference obtained in step S2. Write to cross-layer scheduling buffer ; S32, Control thread follows user sequence number Line-by-line scanning to extract power levels from the power components of the vector. Then search for the codeword occupancy bits given by S2 in the same row. In the index space Construct a mapping table ; S33. Insert cross-layer association operators in the intermediate layer of the network. This cross-layer correlation operator With mapping table Interference weight vector As input, a power-codeword coupling strength map is generated through matrix-vector double multiplication. , is represented as: ； in, The total number of codewords, Indicates user Interference weight of neighboring cells in the same community; S34. Power-codeword coupling strength diagram Perform short-term gradient monitoring and calculate the normalized increase of adjacent time windows. The calculation formula is as follows: ； in, It is the Frobenius norm. To prevent zero constant; when At that time, immediately at the current depth Add an attention-convolutional composite module to the existing layer to update the network depth to [the desired depth]. Used to capture fine-grained interference; when At that time, reduce one layer and simultaneously reclaim parameters to avoid overfitting; threshold It is set by an offline validation set and performs periodic self-checks during runtime; S35. After each depth adjustment cycle, evaluate the validation loss of each layer's output in real time. The scheduler uses a weighted hierarchy that prioritizes details over depth and maintains a global perspective over shallow depth. Dynamically enable or disable cross-layer jump connections; if the deep validation loss exceeds the weighted average... Immediately activate cross-layer shortcuts to shorten the back gradient path; otherwise, disable shortcuts to maintain steady state. S36, Priority gating embedded in the final aggregation unit , to increase coupling strength With user priority Perform element-wise multiplication to generate differential perturbation embeddings, where the first... Individual users in the time window Differential interference embedding as The differential interference embeddings of all users within the same time window are then combined to form a differential interference embedding vector. The identified differential interference is embedded into the vector. Push to resource mapping interface R.

6. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 5, characterized in that, Step S4 is as follows: S41, Scheduling core receives differentiated interference embedding vector Based on the different interference embedding of each user The values ​​are sorted stably from largest to smallest, resulting in a user index sequence with decreasing priority. ; S42, along the user index sequence Top-down filling of multiple power grids ; For the ranking users In the power range The internal application of an index-based hierarchical strategy calculates its target transmit power. : ； in As a graded index, This further widens the power gap between high and low priority levels; S43, Power grid via carrier mapping function Mapped to One physical carrier, generating an initial power allocation matrix. The data is pushed to the SCMA encoding engine in real time, and the mapping process follows the frequency layering principle of prioritizing low-power carriers and high-power carriers. S44, encoding engine for matrix Each user in Candidate codeword set Internal selection of sparse code characters ; If the same carrier If a codeword collision occurs, then in the collision indicator matrix... Set the corresponding element to 1 and send it back to the scheduling core; S45, Scheduling Core Read For the set of all users whose flag bit is 1 Applying differential gain correction, the update is performed simultaneously in both the power dimension and the codeword dimension, as follows: ； in, This is the power differential gain factor. Replace step size for code words. Indicates in The next codeword is selected by sequential circular shifting within the inner sequence; ensuring that each iteration only locally updates conflicting elements, until... Zeroing means that the collision of code words within the carrier has been eliminated; S46. Correct the power-codeword pair Write to the coupled resource table And calculate the interference margin for each carrier. ,in To ensure a safe interference threshold, For actual interference estimation; if If the value is positive, the carrier is marked as safe; otherwise, return to step S45 to trigger a secondary correction to ensure that all carriers are within the safe threshold. S47, when right When established, the resource table will be coupled. Snapshots are written to the cache. This serves as the initial solution group for the subsequent global search of the bitter fish.

7. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 6, characterized in that, Step S5 is as follows: S51, from cache Load the coupling resource table For each entry By power Codeword Index With carrier index Establish three-dimensional coordinates and with A uniform subset as the first school of fish This ensures that the search starting point falls within the feasible solution domain, reducing initialization drift; S52, refers to each individual fish in the school. Binding interference coupling weights from step S3 ,Will Abstracted as the water flow resistance coefficient, the greater the resistance, the more the individual is in a high conflict zone; S53. Design scenario-based multi-objective fitness and convert it into a single energy value. The evaluation function comprehensively considers system throughput. User acceptance rate With residual interference : ； in, The weights are non-negative and satisfy the following conditions: ; carrier bandwidth; For individuals On carrier Signal-to-interference-plus-noise ratio; This represents the maximum acceptance rate within the window. To correspond to the remaining interference; For safety threshold; S54, Based on energy gradient Dynamically adjust individual step size High-energy fish employ small-step, precise searches, while low-energy fish utilize large-step, global exploration. The update formula is as follows: ； in, radius random vectors within, radius Random vectors within; S55, If the updated coordinates fall into Red area This immediately triggers a penalty jump, migrating the object along the negative gradient direction to the disturbance trough. And record the jump vector in the water flow field. In the middle, it provides a basis for calculating the flow of population; S56. After each round of search, perform clustering and dispersion filtering, sorting by energy from high to low, then calculating the cosine distance matrix to check path differences, retaining those with high energy and differences greater than a threshold. Individual fish that are homogeneous and have low energy are removed to maintain population diversity while maintaining convergent tension. S57, When the overall energy of the group increases Below the threshold Or search window Exhausted, freeze the current peak individual Output its corresponding final power-codeword mapping and full-carrier interference margin table Immediately deliver the control plane for real-time distribution.

8. The resource allocation method based on converter networks and the PD-NOMA and SCMA optimization algorithms according to claim 7, characterized in that, Step S6 is as follows: S61, Scheduling backend read power - codeword triplet With interference margin Format it as A five-element structure is written to a dedicated rendering cache. This forms the initial version of the scheduling entries; S62. Aggregate the rendering cache by carrier index, call the lightweight Web-GPU component, and draw the power ladder diagram and codeword heat map in real time; for each carrier... and coding Calculate heat value ,in, carrier Up typing The allocated power, This is the maximum power allowed by the system; S63, Synchronous decimation of differential gain factor and ,in accordance with The absolute value is used to construct the scheduling priority weight. The execution queue is obtained by sorting all users in descending order. ; S64. Integrating the quinary structure with priority Align 3GPPFAPI fields and encapsulate them into a standard resource scheduling table. ; S65. After the distribution is completed, the system continuously monitors the physical channel measurement report and monitors real-time interference. Compared with model predictions The normalized deviation is calculated using the following formula: ； in, For carrier-by-carrier deviation threshold, For index functions; when Immediately invoke the cross-layer embedding generated in step S3. Re-estimate the local power-codeword coupling strength to trigger a fast correction process; S66. Based on the new estimated intensity, the affected user set... Perform fine-tuning: If Then reduce the power for the corresponding user. If the probability of codeword collision increases, then... Inward right-hand rotation step selects the next sparse codeword ; S67. After all adjustments are completed, write the latest power-codeword-interference triplet back to the visualization panel, and... Time-series statistics are entered into the historical database. .

9. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 8, characterized in that, After the preset evaluation window ends, the system capacity, interference suppression rate and convergence frame count are statistically analyzed. If any indicator is lower than the threshold, fast sampling is started and the search radius of the bitter fish is reset to perform incremental search until the real-time indicator requirements are met. The search parameters are then written into the long-term memory pool to complete cross-cycle adaptive learning.

10. The resource allocation method based on converter networks and the PD-NOMA and SCMA joint optimization algorithms according to claim 9, characterized in that, After the preset evaluation window ends, the system capacity, interference suppression rate and convergence frame number are statistically analyzed. If any indicator is lower than the threshold, fast sampling is started and the bitter fish search radius is reset to perform incremental search until the real-time indicator requirements are met. The search parameters are written into the long-term memory pool to complete cross-cycle adaptive learning. Specifically: S71. Retrieve the transmission receipt logs returned by the base station and summarize the data for each carrier within the evaluation window. Real-time throughput within Instantaneous rate for each user and remaining interference Write the three types of data into the evaluation cache. ; S72, in When finished, the statistics module is called to calculate the system capacity. Average speed Interference suppression rate With convergence frame count And generate a baseline report, which is represented as follows: ； ； ； ； in, For carrier bandwidth, carrier exist The average signal-to-interference-plus-noise ratio within the range, For active users, For users Average speed, carrier Average actual interference, To ensure a safe interference threshold, For the first Frame power-codeword coupling strength diagram To converge the decision threshold; S73, will With preset threshold vector Perform element-level comparison; if it exists If the indicator is not met, the current assessment window is marked as a high-risk period and the adaptive process is triggered; otherwise, the original parameters are maintained and the process continues. S74. Initiate rapid sampling during high-risk periods to capture the latest user access graph. With real-time channel matrix And expand the input window of the cross-layer feature extractor to ; S75. Reset the global exploration radius of the improved bitter fish search based on the updated cross-layer features. With local fine-tuning radius , is represented as: ； ； in, and The radius of the previous cycle, The maximum positive interference deviation within the current window. For the minimum rate gap, This is the step size scaling factor; S76. Perform an incremental search for bitter fish and output the corrected power-codeword mapping based on the new radius. The affected carriers and users are updated using a partial distribution mechanism; the statistics are recalculated immediately after the distribution is completed. And calculate the improvement vector ; S77, if Each component is no less than the corresponding minimum improvement threshold. Then freeze the current search parameters. And written to the long-term memory pool If any component fails to meet the standard, the evaluation window will be shortened to [a shorter timeframe]. The process continues to execute steps S74 to S76 repeatedly until the real-time indicator requirements are met; after completing the cycle, the process enters the next regular evaluation cycle.

Images