A neural network assisted code division multiplexing over-the-air aggregation reception method

Abstract

The application discloses a neural network assisted code division multiplexing air aggregation receiving method. In the method, a plurality of sending ends spread pilot symbols and data symbols by using respective spreading codes, and transmit on the same time-frequency resource. The receiving end estimates the channel and frequency offset according to the codebook and the pilot, reconstructs the candidate multiplexing chip vector, and combines and matches the distance according to the mapping relationship between the candidate symbol combination and the target aggregation value. On this basis, the neural network outputs the parameter residual, the aggregation receiving metric residual and the gating coefficient, which are subjected to limited correction, physical metric recalculation and reliability gating fusion to generate the aggregation symbol soft information corresponding to each target aggregation value. The application can improve the aggregation receiving reliability under the conditions of channel mismatch, residual frequency offset and inter-code interference without taking the independent data of each sending end as the final output.

Description

Technical Field

[0001] This invention relates to the fields of multiplexing communication and over-the-air aggregation computing technology, specifically to a neural network-assisted code division multiplexing over-the-air aggregation receiving method. Background Technology

[0002] Code division multiplexing (CDM) assigns corresponding spreading codes to different transmitters or different data paths, enabling multiple signals to be superimposed and transmitted on the same time and frequency resources. Over-the-air aggregation computation utilizes the physical superposition characteristics of multiple transmitter signals in a shared radio multiple access channel, allowing the receiver to directly obtain target aggregation values ​​such as symbol summation, weighted summation, state counts, voting results, or quantization model updates. This reduces the processing overhead required for path-by-path detection, path-by-path decoding, and subsequent digital aggregation.

[0003] Under ideal conditions, orthogonal spreading codes can reduce mutual interference between different code channels. However, in actual communication links, different transmitters typically have independent channel responses and frequency deviations, and may also be affected by factors such as chip timing deviations, sampling clock deviations, frequency-selective channels, and inter-carrier interference. These factors introduce amplitude and phase distortions in the chip dimension, causing the equivalent code vector observed by the receiver to deviate from the ideal orthogonality relationship, thus generating inter-symbol interference and leading to mismatches between channel estimation results, candidate signal reconstruction results, and received metrics and the actual received signal.

[0004] Existing code division multiplexing (CDM) reception methods typically aim to recover the data carried by each code channel separately, and then digitally aggregate the detection results of each channel. When the receiver only focuses on the target aggregated value corresponding to multiple transmitter data, the above processing method still requires performing channel-by-channel detection or decoding, which not only increases the complexity of reception processing, but also the single-channel decision error may accumulate or propagate in subsequent aggregation processes.

[0005] Existing neural network-based reception methods can learn reception mappings under complex channel conditions using training data. However, end-to-end neural networks typically replace traditional reception processing directly, and their outputs lack a clear correspondence with the physical signal model. When actual channel conditions deviate from the training data distribution, or when the reception metrics obtained based on the physical model are already highly reliable, the neural network may still output a large correction, affecting the stability and reliability of the reception results. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] The purpose of this invention is to provide a neural network-assisted code division multiplexing (CDM) air aggregation reception method to solve the problems in existing CDM air aggregation reception processes, such as increased equivalent code correlation, inter-code interference, and aggregation reception metric mismatch caused by factors such as residual frequency offset, relative chip timing deviation between transmitters, and frequency-selective channels; high processing complexity and decision error propagation caused by detection and aggregation of each path; and the potential for overcorrection by the neural network even when the physical reception metric reliability is high.

[0008] (II) Technical Solution

[0009] This invention includes the following steps:

[0010] Step 1: Obtain the multiplexed codebook and establish the candidate symbol combination set, target aggregation value set, and aggregation mapping relationship. Receive code division multiplexing pilot signals and data signals transmitted from multiple transmitters on the same time-frequency resources and superimposed through a shared radio multiple access channel to form the received data chip vector.

[0011] Step 2: Using the multiplexed codebook and pilot signals, estimate the equivalent channel parameters and frequency offset of each code channel. Obtain the initial channel estimate, initial frequency offset estimate, pilot reconstruction residual, and relevant information of the valid code.

[0012] Step 3: Based on the multiplexed codebook and the initial estimates of the channel and frequency offset, reconstruct the candidate multiplexed chip vectors and calculate the combination distance. Merge the combination distances according to the aggregation mapping relationship to obtain the initial physical code domain aggregation receive metric, initial aggregation confidence information, and data chip matching residual information.

[0013] Step 4: Combine the initial estimate, reconstructed residual, valid code-related information, data chip matching residual information, and initial aggregation confidence information. Normalize or standardize the combined information to form fixed-dimensional code domain reception features.

[0014] Step 5: Input the fixed-dimensional code domain received features into the neural network-assisted correction model. Obtain the equivalent channel residual correction, frequency offset residual correction, aggregated receiver metric residual correction, and network output gating coefficients.

[0015] Step Six: Scaling and limiting the equivalent channel residual correction and frequency offset residual correction, and correcting the corresponding initial estimates. Reconstruct the candidate multiplexing chip vector based on the corrected parameters, and recalculate the combination distance, physical code domain aggregation receiver metric, and aggregation confidence information.

[0016] Step 7: Determine the actual fusion gating coefficients based on the recalculated aggregation confidence information and network output gating coefficients. Perform gating fusion between the processed aggregation receive metric residual correction and the recalculated physical code domain aggregation receive metric to output aggregation symbol soft information.

[0017] (III) Beneficial Effects

[0018] The advantages of this invention are as follows: Compared with the prior art, this invention has the following beneficial effects: First, it directly constructs the physical code domain aggregation reception metric for the target aggregation value, reducing the processing overhead of path-by-path detection, path-by-path decoding, and subsequent digital aggregation, and reducing the impact of single-path decision error propagation on the aggregation result; Second, it uses a neural network to perform limited correction on the equivalent channel parameters, frequency offset, and residual errors of the aggregation reception metric, and uses the corrected parameters to recalculate the physical code domain aggregation reception metric, thereby improving the aggregation reception reliability when there are parameter estimation biases, inter-symbol interference, and model mismatch; Third, it performs reliability gating on the neural network correction strength based on the recalculated aggregation confidence information, reducing the risk of over-correction when the physical reception metric reliability is high, while taking into account physical interpretability, reception stability, and engineering feasibility. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall framework for code division multiplexing over-the-air aggregation transmission and reception.

[0020] Figure 2 This is a schematic diagram of the code division multiplexing air aggregation receiving method;

[0021] Figure 3 This is a schematic diagram of a neural network-assisted restricted correction and reliability gating structure. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention and not to limit the scope of protection of the present invention; in the absence of conflict, the technical features of the various embodiments can be combined with each other.

[0023] like Figure 1 As shown, the communication scenario of this invention includes multiple transmitters and one receiver. The multiple transmitters respectively use terminal-specific spreading codes to chip-spread pilot symbols and data symbols. The expanded multi-chip signals are superimposed on a shared radio multiple access channel using the same time and frequency resources. The receiver does not aim to output independent data streams from each transmitter as its final goal, but rather constructs a target-oriented aggregation value-based reception metric based on the code division multiplexing physical model, and uses a neural network to perform limited corrections to the remaining mismatch in the physical model. For example... Figure 2As shown, the present invention sequentially performs fixed configuration and signal reception, initial code domain estimation, physical code domain aggregation reception metric construction, fixed-dimensional feature generation, neural network-assisted residual output, restricted parameter correction and physical metric recalculation, and reliability-gated fusion and aggregation output. The specific implementation of these steps is as follows.

[0024] Step 1: Fixed configuration and code division multiplexing signal reception.

[0025] Let the number of transmitters be... The spreading code length is The multiplexing codebook is ,in Indicates assignment to the first Each transmitter uses a terminal-specific spreading code. To ensure comparability of chip energy under different code lengths and codebook configurations, the unit energy normalization expression for each spreading code is as follows:

[0026]

[0027] For orthogonal codebooks, different spreading codes satisfy... For low cross-correlation codebooks, different spreading codes satisfy... The spreading code can be a real digit or a complex digit, specifically a Walsh-Hadamard code, orthogonal binary code, or a pre-designed low cross-correlation code. This indicates the preset maximum number of cross-correlation coefficients.

[0028] Taking binary phase shift keying as an example, the first The first sender in the... The modulation symbols for each data position are The corresponding spreading chip vector expression is:

[0029]

[0030] Pilot symbols are chip-expanded using the same spreading code or a pilot spreading code with a predetermined correspondence to the spreading code, enabling the receiver to establish a correspondence between pilot observations and data observations using a unified codebook. Within a fixed configuration period, the number of transmitters, spreading code length, codebook, modulation symbol set, candidate symbol combination set, mapping relationship between candidate symbol combinations and target aggregation values, and neural network input / output dimensions remain unchanged, so as to pre-generate candidate combination lookup tables and aggregation mapping lookup tables.

[0031] The receiver is at the first The length of each data location is [value]. The received chip vector, its unified model expression is:

[0032]

[0033] in, For the first The code channel in the first Channel distortion matrix of chip dimension for each data location The residual frequency offset phase rotation matrix, This is the composite vector of noise, residual interference, and unmodeled errors. Under flat fading conditions, we can let... Under frequency-selective channel or orthogonal frequency division multiplexing bearer conditions, It can be a diagonal matrix composed of the responses of common subcarriers, or an equivalent matrix obtained by mapping the channel taps. Indicates the first The complex channel coefficients of each code channel, where I represents the identity matrix matching the chip dimension.

[0034] The expression for the frequency offset phase rotation matrix is:

[0035]

[0036] in, For the first Normalized residual frequency offset of each transmitter For the first The data location in the th data position The normalized time coordinates of each chip. In the single-carrier implementation, the normalized frequency offset is based on the chip rate; in the orthogonal frequency division multiplexing implementation, the normalized frequency offset is based on the subcarrier spacing or sampling rate, and a unified reference is used within the same system.

[0037] In a single-carrier implementation, the transmitting end continuously transmits a spread spectrum chip sequence at a predetermined chip rate. The receiving end, after down-conversion, sampling, timing synchronization, and coarse frequency offset compensation, transmits the sequence in a continuous... Each chip constitutes a received chip vector. In the orthogonal frequency division multiplexing implementation, each transmitter first completes chip spreading, and then maps the spread spectrum chips to the same set of common subcarriers; when the number of common subcarriers is... At that time, the common subcarrier is divided into A length of Each chip subcarrier group carries a spreading chip vector. Each transmitter does not occupy exclusive, non-overlapping subcarriers; instead, they transmit by superimposing subcarriers on the same chip subcarrier group.

[0038] Step 2: Initial estimation of the code domain.

[0039] Code division multiplexing pilot signals include at least training pilots positioned at the beginning of a frame or transport block. In one implementation, tracking pilots are also inserted at predetermined intervals within the data interval to update slowly changing channel and frequency offset. To improve the identifiability of multiple independent frequency offset parameters, different transmitters may employ terminal-specific pilot symbol sequences, terminal-specific pilot positions, or a combination of both.

[0040] The receiver performs correlation, matched filtering, despreading, or projection on the superimposed pilots based on the multiplexed codebook. For an ideal orthogonal codebook, the inner product of the received pilot chip and the terminal-specific spreading code can be performed; for a low cross-correlation codebook, linear projection or regularized projection can be performed based on the code correlation matrix to suppress leakage from other code channels.

[0041] In one coarse frequency offset estimation method, at least two training pilot blocks with known time intervals are set. The receiver obtains code domain observations of the same code channel on different training pilot blocks and calculates a coarse frequency offset estimate based on the correlation phase between the code domain observations. This coarse estimate is used to limit the subsequent candidate search interval, reducing the computational cost of multi-dimensional joint search.

[0042] The received chips from all pilot positions are arranged in a predetermined order to form a pilot observation vector. Given a frequency offset vector Under the given conditions, construct the pilot matrix based on the known pilot symbols, codebook, and chip time coordinates. The pilot observation vector expression is:

[0043]

[0044] in, This is a vector formed by concatenating the equivalent channel parameters used to parameterize the distortion matrix of each code channel in a predetermined order. In flat fading, each code channel corresponds to one complex channel parameter; in frequency-selective channels, each code channel can correspond to multiple subcarrier response parameters or multiple channel tap parameters.

[0045] Given the frequency offset assumption, the equivalent channel parameters are estimated using regularized least squares:

[0046]

[0047] in, This is the regularization coefficient. The receiver constructs a candidate set near the coarse frequency offset estimate. For each candidate frequency offset vector, calculate the channel estimation and pilot reconstruction error, and select the parameter combination with the minimum pilot reconstruction error:

[0048]

[0049] When there are a large number of transmitters, the current frequency offset estimate of other code channels can be fixed, and a one-dimensional candidate search can be performed on each code channel in turn. After each update, all equivalent channel parameters are recalculated. This process is repeated until the pilot reconstruction error decreases by less than a predetermined threshold or the predetermined number of iterations is reached.

[0050] The receiver also calculates effective code-related information based on the current channel and frequency offset estimation. Let the... The equivalent code vector of each code channel is Then the normalized effective code correlation coefficient is:

[0051]

[0052] In actual calculations, the unknown true value is replaced with the corresponding estimated value, and a positive lower limit is set for the energy in the denominator. When When the normalized correlation coefficient is 1, the correlation coefficient is 1; when At that time, its modulus reflects the effective cross-correlation caused by the non-ideality of the spreading code, residual frequency offset, chip timing deviation, and chip-dimensional channel distortion. The off-diagonal elements, off-diagonal element energies, maximum cross-correlation values, or inter-symbol interference indices calculated from the effective code correlation matrix can be used as subsequent network input features.

[0053] Step 3: Construction of physical code domain aggregation reception metric.

[0054] The receiver establishes a candidate symbol combination set for a fixed number of transmitters. ,in For binary phase shift keying, the total number of candidate symbol combinations is Let the aggregation mapping be... The target aggregate value set is Target aggregate value The corresponding candidate combination set expression is:

[0055]

[0056] For the Combine each data location with any candidate symbol to reconstruct the candidate multiplexing chip vector based on the initial channel and frequency offset estimation:

[0057]

[0058] The receiver calculates the combined distance between the actual received chip vector and the candidate multiplexed chip vector. Generally, the Mahalanobis distance relative to the noise and residual interference covariance matrix is ​​used.

[0059]

[0060] in, Let be the covariance matrix of noise and residual interference. For The minimum eigenvalue is set with a positive lower bound to avoid underestimation of covariance leading to abnormally amplified distances. When the combined distance is expressed, it is:

[0061]

[0062] For the target aggregation value Consider the prior probabilities of candidate combinations The physical code domain aggregation reception metric is

[0063]

[0064] When all candidate combinations have equal probability The common bias introduced by this constant for each target aggregate value can be omitted. To reduce computational complexity, a maximum logarithm approximation can also be used, expressed as:

[0065]

[0066] When the candidate combinations are equally probable, this formula is the same as... The only difference is a constant that is the same for all target aggregation values.

[0067] The metrics for all target aggregate values ​​are normalized. To improve numerical stability, the largest metric is subtracted first, and the aggregate posterior probability expression is:

[0068]

[0069] The expression for the normalized posterior entropy, calculated based on the aggregated posterior probability, is as follows:

[0070]

[0071] Therefore, in From time to time The maximum aggregate posterior probability, the normalized posterior entropy, and the difference between the two largest aggregate posterior probabilities together constitute the aggregate confidence information.

[0072] Step 4: Generation of fixed-dimensional code field reception features.

[0073] For each data location, the receiver generates fixed-dimensional code domain reception features based on the initial code domain estimate, candidate multiplexing chip matching results, and aggregate confidence information. These features include at least the real and imaginary parts of the equivalent channel initial estimate, the initial frequency offset estimate, the normalized pilot reconstruction residual, the minimum matching distance of the data chips, effective code correlation features, and initial aggregate confidence information. They may further include noise covariance features and received power. The expression for the fixed-dimensional code domain reception features is:

[0074]

[0075] in, To normalize the pilot reconstruction residual, To match residuals for data chips, For valid code-related features, For fixed-dimensional noise features extracted from the noise covariance matrix, For received power, This is the aggregated posterior probability vector arranged in a predetermined order according to the target aggregate values. The maximum aggregate posterior probability corresponding to the nth data position. It is the difference between the posterior probabilities of the first two aggregations.

[0076] Continuous features are standardized using the mean and standard deviation obtained from the training data, and a predetermined cutoff range can be set. The feature order, mean, standard deviation, and cutoff range determined during the training phase remain unchanged during the deployment phase. For posterior probabilities and normalized posterior entropies located in the zero-to-one interval, they can be directly input or subjected to a fixed linear transformation.

[0077] When the channel and frequency offset remain approximately constant within a pilot update interval, the parameter-type features of multiple data locations within that interval can be averaged, weighted, or attention-based to form interval-level features; the data matching residual, aggregated posterior probability, metric residual, and gating coefficient still change per data location.

[0078] Step 5: Neural network-assisted residual output.

[0079] like Figure 3 As shown, the neural network-assisted correction model includes a shared feature extraction layer, a parameter residual output branch, an aggregated receiver metric residual output branch, and a gated output branch. The shared feature extraction layer is used to extract a joint representation of channel state, frequency offset, inter-symbol interference, reconstruction residual, and aggregated uncertainty.

[0080] In one feasible network architecture, the shared feature extraction layer comprises two or three fully connected networks. Each fully connected network is followed by a nonlinear activation function, and batch normalization, layer normalization, or random deactivation modules can be configured. The parameter residual output branch updates the interval-level features based on the pilot signals and outputs equivalent channel residual corrections and frequency offset residual corrections. The aggregated receiver metric residual output branch outputs the same number of metric residuals as the target aggregated values ​​per data location. The gating output branch outputs network gating coefficients between zero and one per data location using a bounded activation function. .

[0081] For the pilot update interval Pooling the features within the interval The network input-output relationship expression is:

[0082]

[0083]

[0084] in, This represents the mapping function corresponding to the parameter residual output branch. This represents the data location-level mapping function corresponding to the aggregated receive metric residual output branch and the gated output branch. Represents the set of network parameters for a neural network-assisted correction model, indicated by the superscript. Indicates that the corresponding variable belongs to the first One pilot update interval.

[0085] The neural network does not directly output independent data symbols from each transmitter, nor does it directly replace the physical code domain aggregation receiver metric. The parameter residuals are only used to correct the physical parameters after scaling and clipping, and the aggregation receiver metric residuals are only fused with the recalculated physical metrics after centering, clipping, and reliability gating.

[0086] Training samples can be generated from simulated channels, calibrated channels, or actual acquired signals, and cover a predetermined signal-to-noise ratio range, code channel amplitude and phase range, frequency offset range, chip timing deviation, channel selectivity, spreading code correlation, inter-symbol interference intensity, and transmitter power differences. Training samples for orthogonal frequency division multiplexing (OFDM) implementations can also cover frequency-selective channels with a maximum multipath delay not exceeding the cyclic prefix guard interval, responses of different common subcarriers, residual inter-carrier interference, and Doppler spread.

[0087] When the training samples include the real channel, the real frequency offset, and the teacher's reception metric, the expression for the parameter residual label is:

[0088]

[0089] The expression for the centralized metric residual label is:

[0090]

[0091] In one gating supervision method, when the aggregate decision of the recalculated physical metric is correct and the reliability condition is met, the gating supervision label is set to zero; when the physical metric is incorrect or does not meet the reliability condition, the gating supervision label is set to one, or a continuous label between zero and one is constructed based on the difference between the physical metric and the teacher's metric.

[0092] The comprehensive training loss expression is:

[0093]

[0094] Among them, each Non-negative weights; The classification loss is the difference between the aggregated soft information and the true aggregated value after fusion. , and Each network output is constrained to closely approximate the training labels. This is used to suppress excessive gate control on highly reliable samples of physical metrics. After training, the network parameters and feature normalization parameters are fixed, and only forward inference is performed during the deployment phase.

[0095] Step 6: Correction of constrained parameters and recalculation of physical metrics.

[0096] The element-wise scaling and amplitude limiting of the complex channel parameter residuals is expressed as follows:

[0097]

[0098] in, This is the channel residual scaling factor. To define the upper limit of the complex amplitude, the expression for the complex amplitude limit function is:

[0099]

[0100] in, This represents a complex scalar whose amplitude needs to be limited. This represents the corresponding complex amplitude limit value; when the input is a vector, the complex amplitude limit function processes each element in the vector separately.

[0101] The frequency offset residual is first scaled and limited to a predetermined residual range, then added to the initial frequency offset estimate, and the sum is limited to the allowable range after the previous coarse frequency offset compensation. The receiver then... Generate the corrected code channel distortion matrix, and reconstruct all candidate multiplexing chip vectors by combining the frequency offset residual with the initial frequency offset estimate. Then, recalculate the combination distance, physical code domain aggregation receiver metric, aggregation posterior probability, and normalized posterior entropy according to step three.

[0102] By reapplying the parameter residuals to the physical signal model, rather than directly superimposing them onto the original aggregate metric, channel and frequency offset corrections can have a consistent impact on the amplitude, phase, and correlation structure of all candidate multiplexed chips. When multiple data locations exist within a pilot update interval, interval-level parameter residuals can be used collectively to recalculate the physical metrics for each data location within that interval.

[0103] Step 7: Reliability-gated fusion and aggregated output.

[0104] The aggregated receiving metric residual is first centered along the dimension of the target aggregated value, and then its magnitude is limited. Its expression is:

[0105]

[0106] Here, clip indicates amplitude limitation. This measures the upper limit of the residual magnitude. Centralization eliminates the non-uniqueness caused by adding the same constant to all aggregate measures.

[0107] Calculate the maximum aggregate posterior probability based on the recalculated physical metrics. and normalized posterior entropy And select the gating limit:

[0108]

[0109] When the reliability of recalculated physical metrics is high, a smaller first gating upper limit is used; otherwise, a larger second gating upper limit is used to reserve more correction space for network metric residuals. The actual fusion gating coefficients and the expressions for the fused aggregated metric are as follows:

[0110]

[0111] in, Aggregate the received metric for the recalculated physical code field. The metric residual after centralization and amplitude limiting, This is the aggregated reception metric after fusion.

[0112] The final expression for the aggregate symbol soft information is:

[0113]

[0114] When a hard decision on aggregation is required, the target aggregation value with the largest final aggregation symbol soft information is selected. When setting channel coding or sequence constraints that match the target aggregation value set, the soft information can also be input into the corresponding decoding or sequence inference module.

[0115] Example 1: Orthogonal code single-carrier air summation reception from two transmitters

[0116] This embodiment employs two transmitters and one receiver, carrying code division multiplexing chips in a single-carrier manner, and using the sum of the binary phase shift keying symbols from the two transmitters as the target aggregation value. Communication, initial estimation, neural network, and gating parameters are shown in Tables 1 and 2, respectively.

[0117] Table 1 Communication and Initial Estimation Parameters for Example 1

[0118] Number of senders 2 Spread code length 4 chips Terminal-specific spreading code The first spreading code is [1,1,1,1]ᵀ / 2; the second spreading code is [1,-1,1,-1]ᵀ / 2. Modulation method Binary phase shift keying Transmission method Single-carrier, overlapping transmission of the same chip time period Initial training pilot Two repeating pilot blocks, each with 8 pilot symbols. Training pilot block guard interval 16 chips Normalized frequency offset search range The coarse estimate of the frequency offset is 0.02 above and below. Frequency offset search step size 0.001 Regularization coefficient 0.01

[0119] Table 2 Neural Network and Gating Parameters in Example 1

[0120] Fixed-dimensional code domain feature dimension 17 Shared feature extraction layer Two fully connected layers, with hidden layer widths of 64 and 32 respectively. Channel residual scaling factor 0.20 Upper limit of channel residual amplitude 0.30 Frequency offset residual scaling factor 0.50 Upper limit of frequency offset residual 0.03 Maximum aggregate posterior probability threshold 0.85 Normalized posterior entropy threshold 0.50 First gate limit 0.15 Second gating limit 0.75 Upper limit of residual amplitude 2.50

[0121] In this embodiment, each training pilot block includes 8 pilot symbols, and each pilot symbol is chip-spread using a spreading code of length 4. Therefore, each training pilot block occupies 32 chips. The 16 chips mentioned in Table 1 refer to the idle chip interval between the end position of the first training pilot block and the start position of the second training pilot block. Therefore, the time interval between the start positions of the two training pilot blocks is 48 chips. The receiver calculates a coarse frequency offset estimate based on the correlation phase between the code domain observations of the same code channel in the two training pilot blocks and according to the start position interval of the 48 chips.

[0122] The receiver loads a codebook containing two terminal-specific spreading codes and establishes a mapping between four candidate symbol combinations and three summation aggregate values. The four candidate symbol combinations... , , and Mapped to aggregate values ​​respectively 0, 0, and 2. The two transmitters transmit at the same carrier frequency and in the same chip time period. The receiver forms a received chip vector every four chips received.

[0123] The receiver performs code correlation or projection processing on the two repeating pilot blocks respectively, obtaining code domain observations of the two code channels on the two pilot blocks, and obtains a coarse frequency offset estimate based on the phase difference between the observations of the same code channel. A candidate set is formed within a range of 0.02 above and below the coarse frequency offset estimate, with a step size of 0.001. For each set of frequency offset candidates, a pilot matrix is ​​constructed, and the two equivalent channel parameters are estimated using a regularized least squares method with a regularization coefficient of 0.01. The frequency offset and channel parameters with the smallest pilot reconstruction error are selected as the initial estimates, and the normalized pilot reconstruction residual and the cross-correlation characteristics between the two effective codes are calculated.

[0124] For each data location, the receiver reconstructs the candidate multiplexing chip vectors corresponding to the four candidate symbol combinations and calculates the distances between the four combinations. (Aggregate value) 2 and 2 each correspond to a candidate combination, and the aggregation value 0 corresponds to two candidate combinations; the distances between the two combinations corresponding to the aggregation value 0 are logarithmically summed and exponentially merged to obtain the three physical code domain aggregation reception metrics and three aggregation posterior probabilities.

[0125] The seventeen-dimensional code domain reception features include: four real and imaginary parts of the two complex channel initial estimates, two frequency offset initial estimates, a normalized pilot reconstruction residual, a minimum matching distance of the data chip, an effective code cross-correlation feature, an equivalent noise variance, a received power, three aggregate posterior probabilities, a maximum aggregate posterior probability, a normalized posterior entropy, and a gap between the first two aggregate posterior probabilities.

[0126] The neural network's channel residual branch outputs four real numbers to represent the real and imaginary parts of the two complex channel residuals; the frequency offset residual branch outputs two real numbers; the aggregated receiver metric residual branch outputs three real numbers; and the gated output branch outputs a real number between zero and one. Constrained corrections and physical metric recalculation are performed using the parameters listed in Table 2. The gate upper limit is set to 0.15 when the recalculated maximum aggregated posterior probability is not less than 0.85 and the normalized posterior entropy is not greater than 0.50; otherwise, it is set to 0.75. The final output is the aggregated value. The soft information of the three-dimensional aggregate symbols corresponding to 0 and 2.

[0127] Example 2: Orthogonal codes and orthogonal frequency division multiplexing bearers at four transmitters

[0128] This embodiment employs four transmitters and one receiver. Four orthogonal codes are selected from an 8-byte Walsh-Hadamard codebook as terminal-specific spreading codes, and orthogonal frequency division multiplexing is used to carry the pre-formed code division multiplexing chips. The communication and frame structure parameters, estimation and neural network parameters are shown in Tables 3 and 4, respectively, and the summation aggregation value and the number of candidate combinations are shown in Table 5.

[0129] Table 3 Communication and Frame Structure Parameters in Example 2

[0130] Number of senders 4 Spread code length 8 chips Codebook type From a Walsh-Hadamard orthogonal codebook of length 8, select 4 codes. Modulation method Binary phase shift keying Fast Fourier Transform Points 64 Cyclic prefix length 16 sampling points Number of common subcarriers The 16 chips are divided into two chip subcarrier groups of length 8, and each transmitter uses the same subcarrier group. Number of data locations per frame 100 Initial training pilot Two repeating pilot blocks, each with 8 pilot symbols. Tracking pilot spacing One tracking pilot is set for every 20 data locations.

[0131] Table 4. Estimation, Neural Network, and Gating Parameters in Example 2

[0132] Normalized frequency offset search range The coarse estimate of the frequency offset is 0.05 above and below. Frequency offset search step size 0.002 Regularization coefficient 0.02 Shared feature extraction layer Two fully connected layers, with hidden layer widths of 128 and 64 respectively. Channel residual scaling factor 0.15 Upper limit of channel residual amplitude 0.25 Frequency offset residual scaling factor 0.40 Upper limit of frequency offset residual 0.05 Maximum aggregate posterior probability threshold 0.80 Normalized posterior entropy threshold 0.80 First gate limit 0.10 Second gating limit 0.70 Upper limit of residual amplitude 3.00

[0133] Table 5. Sum of aggregated values ​​from four sending ends and number of candidate combinations

[0134] -4 1 -2 4 0 6 2 4 4 1

[0135] The four transmitters first use their respective spreading codes to complete the chip expansion of the pilot and data symbols, and then map the spreading chips to the same 16 common subcarriers. These 16 common subcarriers are divided into two chip subcarrier groups of length 8. Each chip subcarrier group carries a spreading chip vector corresponding to a data location; therefore, one orthogonal frequency division multiplexing (OFDM) symbol can carry two data locations. The receiver performs coarse synchronization, cyclic prefix removal, fast Fourier transform, and common subcarrier extraction, and then constructs a received chip vector from each of the 8-length chip subcarrier groups.

[0136] The receiver uses four spreading codes to correlate or project two repeating pilot blocks, obtaining code-domain observations of four code channels. A coarse frequency offset estimate is then obtained based on the phase difference between the repeating pilots. Candidate searches are performed within a 0.002 step size range above and below the coarse estimate, within a range of 0.05. To reduce the complexity of the four-dimensional joint search, a code-channel alternating search is adopted: after updating each frequency offset parameter, the pilot matrix is ​​reconstructed and all channel parameters are updated until the stopping condition is met. The initial channel and frequency offset estimates are updated every 20 data positions of the tracking pilots.

[0137] The four binary phase shift keying transmitters have a total of 16 candidate symbol combinations, each mapped to an aggregate value. , The receiver reconstructs 16 candidate multiplexed chip vectors and calculates 16 combination distances, which are then merged into five physical code domain aggregated reception metrics according to Table 5. An aggregated value of 0 corresponds to six candidate combinations, and the joint contribution of the six candidate combinations is retained through logarithmic summation and exponential merging.

[0138] The code domain reception features include initial estimates of the channel and frequency offset for the four code channels, normalized pilot reconstruction residuals, minimum matching distance of data chips, energy of off-diagonal elements of the effective code correlation matrix, maximum effective cross-correlation value, equivalent noise variance, received power, five aggregated posterior probabilities, maximum aggregated posterior probability, normalized posterior entropy, and the interval between the first two aggregated posterior probabilities. The channel residual output branch outputs eight real numbers, the frequency offset residual output branch outputs four real numbers, the aggregated receiver metric residual output branch outputs five real numbers, and the gated output branch outputs one real number.

[0139] The parameters listed in Table 4 are used to scale and limit the four complex channel residuals and four frequency offset residuals. Based on the corrected parameters, 16 candidate multiplexed chip vectors and five physical code domain aggregation receiver metrics are regenerated. When the recalculated maximum aggregation posterior probability is not less than 0.80 and the normalized posterior entropy is not greater than 0.80, the gating upper limit is set to 0.10; otherwise, it is set to 0.70. Gating fusion is performed on the five centered and amplitude-limited metric residuals to finally output the aggregated value. , The five-dimensional aggregate symbol soft information corresponding to 0, 2 and 4.

[0140] The above embodiments are illustrative of the present invention and not intended to limit its scope of protection. Equivalent substitutions, parameter adjustments, and functional combinations made within the scope of the inventive concept are all applicable to the present invention.

Claims

1. A neural network-assisted code division multiplexing air aggregation receiving method, characterized in that, Includes the following steps: Step 1: Obtain a multiplexed codebook containing multiple terminal-specific spreading codes, and establish a candidate symbol combination set, a target aggregation value set, and their aggregation mapping relationship; receive code division multiplexing pilot signals and code division multiplexing data signals formed by multiple transmitters using their respective terminal-specific spreading codes and superimposing them on the same time-frequency resources through a shared radio multiple access channel; and perform chip grouping on the code division multiplexing data signals to form a received data chip vector; Step 2: Use the multiplexed codebook to perform code domain estimation on the code division multiplexed pilot signal to obtain the initial channel estimate, initial frequency offset estimate, pilot reconstruction residual, and effective code related information for each code channel; Step 3: Based on the multiplexed codebook, initial channel estimate, and initial frequency offset estimate, reconstruct the candidate multiplexed chip vector of each candidate symbol combination and calculate its combination distance with the received data chip vector to obtain data chip matching residual information; merge the combination distances corresponding to the same target aggregation value according to the aggregation mapping relationship to obtain the initial physical code domain aggregation reception metric and initial aggregation confidence information. Step 4: Combine the initial channel estimate, initial frequency offset estimate, pilot reconstruction residual, effective code related information, data chip matching residual information, and initial aggregation confidence information in a predetermined order, and perform normalization or standardization processing to form fixed-dimensional code domain reception features. Step 5: Using the fixed-dimensional code domain reception features as input to the neural network-assisted correction model, output the equivalent channel residual correction, frequency offset residual correction, aggregated reception metric residual correction, and network output gating coefficient. Step 6: Scaling and limiting the equivalent channel residual correction and frequency offset residual correction, and using the processed residual correction to correct the initial channel estimate and initial frequency offset estimate, respectively. Based on the corrected equivalent channel parameters and frequency offset, the candidate multiplexing chip vector is reconstructed, and the combination distance, physical code domain aggregation reception metric and corrected aggregation confidence information are recalculated. Step 7: Determine the actual fusion gating coefficient based on the corrected aggregate confidence information and the network output gating coefficient; The aggregated reception metric residual correction is centered and its amplitude is limited. The processed aggregated reception metric residual correction is then gated and fused with the recalculated physical code domain aggregated reception metric. The fusion result is then normalized, and the aggregated symbol soft information corresponding to each target aggregated value is output.

2. The method according to claim 1, characterized in that, The terminal-specific spreading code is a Walsh-Hadamard code or other orthogonal code or low cross-correlation code, and satisfies the following conditions: ,as well as ,in, Indicates the first Each terminal has its own unique spreading code. and They represent the first The and the first Each terminal has its own unique spreading code. , and All are terminal indexes, and ; superscript This indicates the conjugate transpose. Indicates modulo, This indicates the preset maximum number of cross-correlation coefficients.

3. The method according to claim 1, characterized in that, The code division multiplexed pilot signal and code division multiplexed data signal are carried using orthogonal frequency division multiplexing; when the spreading code length is The number of common subcarriers is At that time, the common subcarrier is divided into A length of Each chip subcarrier group carries a spreading chip vector, and all transmitters use the same chip subcarrier group.

4. The method according to claim 1, characterized in that, The code division multiplexing pilot signal includes at least two training pilot blocks with known time intervals. The receiving end performs correlation, despreading, or projection on the superimposed pilots according to the terminal-specific spreading code. It obtains a coarse frequency offset estimate based on the correlation phase between code domain observations of the same code channel on different training pilot blocks, and performs candidate search or code-channel alternating update near the coarse frequency offset estimate. Under a given frequency offset assumption, it estimates the equivalent channel parameters using least squares or regularized least squares.

5. The method according to claim 1, characterized in that, The code division multiplexing pilot signal also includes tracking pilots set at predetermined intervals; the receiver updates the equivalent channel parameters and frequency offset according to the tracking pilots, and uses the update results for parameter residual correction and physical code domain aggregation receiver metric recalculation in adjacent data intervals.

6. The method according to claim 1, characterized in that, The combined distance is the Mahalanobis distance between the difference between the received data chip vector and the candidate multiplexed chip vector and the noise and residual interference covariance matrix. For candidate symbol combinations mapped to the same target aggregation value, logarithmic summation and exponential merging are performed according to the prior probability of the candidate symbol combination, or the maximum logarithmic approximation is used to obtain the physical code domain aggregation reception metric corresponding to the target aggregation value.

7. The method according to claim 1, characterized in that, The fixed-dimensional code domain received features include the energy of the off-diagonal elements of the normalized effective code correlation matrix, the maximum absolute value of the off-diagonal elements in the normalized effective code correlation matrix, the equivalent noise variance obtained by dividing the trace of the noise and residual interference covariance matrix by the chip dimension, and the received power obtained by dividing the square norm of the received data chip vector by the chip dimension. The neural network-assisted correction model includes a shared feature extraction layer, a parameter residual output branch, an aggregated receiver metric residual output branch, and a gated output branch. The parameter residual output branch outputs the equivalent channel residual correction and frequency offset residual correction at the pilot update interval level, and the aggregated receiver metric residual output branch and the gated output branch output the aggregated receiver metric residual correction and the network output gate coefficient at the data location level, respectively.

8. The method according to claim 1, characterized in that, Complex amplitude limits are applied to the scaled equivalent channel residual correction amount, and interval limits are applied to the scaled frequency offset residual correction amount. The corrected frequency offset is then limited to the allowable frequency offset range after the previous coarse frequency offset compensation. The candidate multiplexing chip vector is reconstructed using the constrained corrected equivalent channel parameters and frequency offset, and the combination distance and physical code domain aggregation reception metric are recalculated.

9. The method according to claim 1, characterized in that, The maximum aggregated posterior probability and normalized posterior entropy are calculated based on the recalculated physical code domain aggregated reception metric. When the maximum aggregated posterior probability is not less than a first threshold and the normalized posterior entropy is not greater than a second threshold, a first gating upper limit is adopted; otherwise, a second gating upper limit greater than the first gating upper limit is adopted. The network output gating coefficient is multiplied by the adopted gating upper limit to obtain the actual fusion gating coefficient.

Images