An Adaptive Intelligent Demodulation Method for Communication Signals in Complex Environments

Abstract

This invention relates to an adaptive intelligent demodulation method for communication signals in complex environments, belonging to the field of wireless communication technology. The method includes: constructing a training dataset containing OFDM received signals with various modulation schemes; using the time-domain IQ sequence and frequency-domain IQ sequence of the received signal as input samples; and using the original bit sequence with added start and end symbols as the output bitstream label; training a signal recovery neural network using the training dataset, consisting of a multi-scale convolutional embedding layer, an encoder employing a bidirectional cross-attention mechanism, and a decoder performing autoregressive decoding based on a mask self-attention mechanism; and using the trained signal recovery neural network to perform autoregressive decoding on the OFDM received signal, recovering the original bitstream end-to-end without prior modulation information. This invention enables end-to-end learning of the mapping from received signal to information bits and recovers the bitstream through autoregression, effectively improving signal recovery capabilities in complex environments.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an adaptive intelligent demodulation method for communication signals in complex environments. Background Technology

[0002] Wireless communication systems operating in complex environments such as plateaus, oceans, urban canyons, and dense forests are simultaneously subjected to the dual effects of natural environments (such as low-temperature freeze-thaw cycles, high humidity and salinity, and building obstructions) and electromagnetic environments (such as dynamic multipath propagation, strong noise, and spectrum congestion). These external conditions not only weaken the stability of wireless links but also further affect the terminal radio frequency links through environmental stresses (such as rapid temperature changes and power supply ripple fluctuations), making the physical layer signals exhibit stronger time-varying and randomness.

[0003] To balance spectral efficiency and reliability in dynamic channels, communication systems typically select different modulation schemes based on channel conditions. For example, higher-order modulation is used to improve throughput when channel quality is good, while lower-order modulation is used to reduce the bit error rate (BER) when channel quality is poor. However, terminal hardware capabilities are limited, and its channel estimation, interference suppression, and decision-making performance are easily affected by noise and non-ideal factors, making it difficult to accurately select modulation schemes and execute the corresponding demodulation process. Especially in scenarios with rapidly changing channels or unstable reception conditions, the receiver struggles to identify the current modulation scheme in a timely and accurate manner, leading to incorrect symbol resolution and significantly reducing the overall communication performance of the system.

[0004] Existing adaptive modulation mechanisms typically rely on fixed algorithms or independent modulation identification modules. Their core processes often assume relatively slow channel changes and that modulation schemes can be identified individually. However, in real-world systems, modulation identification is frequently affected by factors such as signal-to-noise ratio fluctuations, frequency offset interference, and multipath structure changes, making it difficult to guarantee accurate identification. Furthermore, traditional receiver architectures separate modulation identification, channel estimation, equalization, and demodulation, lacking overall coordination. This results in weak adaptability of each module to non-stationary channels, making it difficult to effectively resist complex interference caused by the superposition of various distortions. Summary of the Invention

[0005] Based on the above analysis, this invention aims to disclose an adaptive intelligent demodulation method for communication signals in complex environments; it is a joint reception method that can automatically sense changes in modulation scheme and simultaneously recover signals without prior information. The method can be directly deployed in wireless communication devices, reducing the bit error rate (BER) to improve communication reliability and system efficiency under rapidly changing channel conditions.

[0006] This invention discloses an adaptive intelligent demodulation method for communication signals in complex environments, comprising:

[0007] A training dataset is constructed, which contains OFDM received signals with various modulation schemes. The time-domain IQ sequence and frequency-domain IQ sequence of the received signal are used as input samples, and the original bit sequence with start and end symbols added is used as the output bitstream label.

[0008] A signal recovery neural network is trained using the training dataset, the neural network comprising:

[0009] Multi-scale convolutional embedding layers are used to extract time-frequency domain features;

[0010] The encoder employs a bidirectional cross-attention mechanism to align and fuse time-frequency domain features, generating an overall feature vector of the signal.

[0011] The decoder, triggered by the start symbol, performs autoregressive decoding based on a mask self-attention mechanism, outputting bit probabilities bit by bit until an end symbol is generated, thus restoring the original bit stream.

[0012] The trained signal recovery neural network is used to perform autoregressive decoding on the OFDM received signal to recover the original bit stream end-to-end without prior modulation information.

[0013] Furthermore, constructing the training dataset includes:

[0014] The original bit information is converted into IQ data and mapped into a modulation symbol sequence according to different modulation methods. After serial-to-parallel conversion, pilot insertion, inverse discrete Fourier transform, parallel-to-serial conversion and cyclic prefix insertion, the transmitted signal is generated. After channel transmission and noise superposition, the received signal is obtained.

[0015] The time-domain received signal is converted into a frequency-domain received signal using the Discrete Fourier Transform.

[0016] Set the maximum modulation order and the maximum sampling length of the received signal, and use zero padding to compensate the original bit sequence, time-domain IQ sequence and frequency-domain IQ sequence into equal-length sequences respectively;

[0017] The IQ data of the time-domain IQ sequence and the frequency-domain IQ sequence are converted into time-domain data sequences and frequency-domain data sequences composed of real and imaginary data, respectively. Start and end symbols are added to the beginning and end of the original bit sequence after zero padding as labels for the output bit stream.

[0018] The time-domain data sequence, frequency-domain data sequence, and corresponding output bitstream labels are used to construct the training dataset.

[0019] Furthermore, the multi-scale convolutional embedding layer includes a temporal processing branch and a frequency domain processing branch, each branch consisting of a sequentially connected multi-scale convolutional module and a spatial attention module; wherein,

[0020] The multi-scale convolution module contains multiple parallel temporal convolution sub-modules. Each sub-module uses a convolution kernel of different size and a configurable dilation rate to capture channel characteristic signals at different scales and stack them to output multi-scale features.

[0021] The spatial attention module performs average pooling along both the signal length and feature dimensions, generates a spatial weight matrix through two-dimensional convolution and Sigmoid activation, weights and fuses multi-scale features and compresses them to a specified dimension, and outputs a time-domain feature vector or a frequency-domain feature vector.

[0022] Furthermore, in the encoder, the bidirectional cross-attention mechanism performs two independent attention calculations:

[0023] The first attention calculation is performed using the time-domain feature vector as the query and the frequency-domain feature vector as the key and value.

[0024] A second attention calculation is performed using the frequency domain feature vector as the query and the time domain feature vector as the key and value. The results of the two attention calculations are added and fused to obtain the overall feature vector of the signal.

[0025] Furthermore, the encoder employs two parallel processing branches, each performing independent attention calculations; specifically including:

[0026] The first processing branch consists of a time-domain → frequency-domain attention submodule, a first residual normalization module, a first feedforward neural network, and a second residual normalization module connected in sequence.

[0027] The second processing branch consists of a frequency domain → time domain attention submodule, a third residual normalization module, a second feedforward neural network, and a fourth residual normalization module connected in sequence.

[0028] And element-wise adders and ReLU activation units connected to the outputs of the two branches;

[0029] The first processing branch is used to calculate attention weights with temporal features with added position encoding as queries and frequency features with added position encoding as keys and values, and output temporally aligned frequency-domain enhanced features;

[0030] The second processing branch is used to calculate attention weights with frequency domain features with added position coding as queries and time domain features with added position coding as keys and values, and outputs frequency domain aligned time domain enhanced features;

[0031] The element-wise adder and ReLU activation unit are used to fuse the enhanced features of the two branches to generate the overall feature vector of the signal.

[0032] Furthermore, the decoder employs an autoregressive approach, using the start symbol as a trigger and the overall feature vector of the signal as a condition to generate a bit sequence bit by bit until the end symbol is output, at which point decoding stops, resulting in the final predicted bit stream sequence.

[0033] Furthermore, the decoder employs a multi-layer stacked architecture, with each decoding layer containing the following sub-modules executed sequentially:

[0034] The first-level multi-head self-attention module receives the bit sequence mapped by the linear embedding layer, and uses a causal mask to shield future positions and padding bits to ensure that the current position depends only on historical information. After residual connection and layer normalization, it outputs a bit feature representation that incorporates context information.

[0035] The second-level multi-head cross-attention module uses the overall feature vector of the signal output by the encoder as the key and value, and the above bit feature representation as the query to calculate the cross-attention weight, realize the mapping between the received signal and the bit stream, and then performs residual connection and layer normalization again.

[0036] The feedforward neural network performs a non-linear transformation on the cross-attention output and outputs the enhanced features of the decoding layer through residual connections and layer normalization.

[0037] Multiple decoding layers are repeatedly stacked in a multi-layer stacked architecture to deepen feature representation;

[0038] Finally, the features are mapped to the output space through a linear layer and a softmax function, and transformed into bit-by-bit classification probabilities.

[0039] Furthermore, a mask is used to predict the training signal to recover the neural network model;

[0040] Add padding, start, and end markers to the output bitstream labels to expand the label space;

[0041] A mask is generated based on the padding position and sequence order, so that the model training process does not focus on future positions and padding positions; the cross-entropy loss function is used to compare the predicted bit probability with the label probability, and the network parameters are updated with the goal of minimizing the bit error rate.

[0042] Furthermore, the signal recovery problem is constructed during training as The five-ary classification task, where 0 and 1 represent bit values, 2 represents padding bits, 3 represents the start symbol, and 4 represents the end symbol.

[0043] With the aim of minimizing the bit error rate, the neural network parameters are optimized based on the output bitstream label and the predicted bit probability.

[0044] The decoder ultimately maps features to bit classification probabilities using the Softmax function, achieving bit-by-bit decision output.

[0045] Furthermore, the trained signal recovery neural network is used to perform autoregressive decoding on the OFDM received signal, the process of which includes:

[0046] 1) Input the time-domain data sequence and frequency-domain data sequence of the OFDM received signal into the multi-scale convolutional embedding layer of the signal recovery neural network to perform time-domain and frequency-domain feature extraction;

[0047] 2) Input the time-domain and frequency-domain features into the encoder, align and fuse the two feature vectors, and map the input sequence to a continuously represented vector. ;

[0048] ;

[0049] in, Indicates encoding operation; Representing time-domain features and Represents frequency domain characteristics;

[0050] 3) The decoder uses an autoregressive approach, based on vectors. The first bit is generated by combining the start symbol with the newly generated bit and the existing output sequence to predict the next bit.

[0051] The prediction update of the bitstream sequence is represented as:

[0052] ;

[0053] in, Indicates a decoding operation. , These represent the predicted sequence results at the previous and current time steps, respectively. This process is iterated until a specific end symbol is output; the final predicted bitstream sequence is then output. .

[0054] This invention can achieve one of the following beneficial effects:

[0055] The adaptive intelligent demodulation method for communication signals in complex environments disclosed in this invention can learn the mapping from received signals to information bits end-to-end and recover the bit stream through autoregression, effectively improving the signal recovery capability of adaptive modulation communication systems in complex environments.

[0056] In this invention, multi-scale convolutional embedding layers combined with spatial attention mechanisms are used to simultaneously extract time-domain and frequency-domain features, effectively capturing the deep structure of signals under multipath fading and rapidly time-varying channels, thereby improving robustness to complex environments.

[0057] By employing bidirectional cross-attention (time domain ↔ frequency domain cross-checking) to achieve time-frequency feature alignment and fusion, it can more effectively suppress inter-symbol interference (ISI) and inter-carrier interference (ICI) compared to traditional unidirectional processing.

[0058] The decoder generates data bit by bit through an autoregressive method and combines causal masking to avoid future information leakage. In scenarios where the channel changes rapidly and there is no prior modulation information, it achieves dynamic and real-time signal recovery and reduces the bit error rate (BER). Attached Figure Description

[0059] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0060] Figure 1 This is a flowchart of the adaptive intelligent demodulation method for communication signals in complex environments according to an embodiment of the present invention;

[0061] Figure 2 This is a comparison chart of bit error rates using different deep learning networks under BPSK modulation in this embodiment of the invention;

[0062] Figure 3 This is a comparison chart of bit error rates using different deep learning networks under QPSK modulation in this embodiment of the invention;

[0063] Figure 4 This is a comparison chart of bit error rates using different deep learning networks under 8PSK modulation in an embodiment of the present invention;

[0064] Figure 5 This is a comparison chart of bit error rates using different deep learning networks under 16QAM modulation in an embodiment of the present invention.

[0065] Figure 6 The recognition accuracy of the DL-AMR network in this embodiment of the invention under different modulation methods;

[0066] Figure 7 The recognition accuracy of the Seq-Rec network in this embodiment of the invention under different modulation schemes;

[0067] Figure 8 The bit error rate of the DL-AMR network in this embodiment of the invention under different modulation schemes;

[0068] Figure 9 The bit error rate of the Seq-Rec network in this embodiment of the invention under different modulation schemes;

[0069] Figure 10 This refers to the clustering feature distribution extracted by the encoder in the Seq-Rec configuration in this embodiment of the invention.

[0070] Figure 11 This refers to the clustering feature distribution extracted by the encoder in the Seq-Rec-C configuration in this embodiment of the invention.

[0071] Figure 12 This refers to the clustering feature distribution extracted by the encoder in the Seq-Rec-F configuration in this embodiment of the invention.

[0072] Figure 13 This represents the clustering feature distribution extracted by the encoder in the Seq-Rec-T configuration in this embodiment of the invention. Detailed Implementation

[0073] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and, together with the embodiments of the present invention, serve to illustrate the principles of the present invention.

[0074] One embodiment of the present invention discloses an adaptive intelligent demodulation method for communication signals in complex environments, such as... Figure 1 As shown, it includes:

[0075] S1. Construct a training dataset, which contains OFDM received signals with various modulation schemes. Use the time-domain IQ sequence and frequency-domain IQ sequence of the received signals as input samples, and use the original bit sequence with start and end symbols added as the output bitstream label.

[0076] S2. Train a signal recovery neural network using the training dataset, the neural network comprising:

[0077] Multi-scale convolutional embedding layers are used to extract time-frequency domain features;

[0078] The encoder employs a bidirectional cross-attention mechanism to align and fuse time-frequency domain features, generating an overall feature vector of the signal.

[0079] The decoder, triggered by the start symbol, performs autoregressive decoding based on a mask self-attention mechanism, outputting bit probabilities bit by bit until an end symbol is generated, thus restoring the original bit stream.

[0080] S3. Use the trained signal recovery neural network to perform autoregressive decoding on the OFDM received signal to recover the original bit stream end-to-end without prior modulation information.

[0081] Specifically, in S1, constructing the training dataset includes:

[0082] S101. Based on different modulation methods, the original bit information is converted into IQ data and mapped into a modulation symbol sequence. After sequentially undergoing serial-to-parallel conversion (S / P), pilot insertion, inverse discrete Fourier transform (IDFT), parallel-to-serial conversion (P / S), and cyclic prefix insertion, a transmit signal is generated. The received signal is obtained after being transmitted through the channel and superimposed with noise. ;

[0083] S102. Transform the received signal in the time domain using Discrete Fourier Transform (DFT). Converted to frequency domain received signal ;

[0084] S103. Set the maximum modulation order and the maximum sampling length of the received signal, and use zero padding to compensate the original bit sequence, time domain IQ sequence and frequency domain IQ sequence into equal-length sequences respectively.

[0085] S104. Convert the IQ data of the time-domain IQ sequence and the frequency-domain IQ sequence into time-domain data sequences and frequency-domain data sequences composed of real and imaginary parts, respectively, and add start symbols to the beginning and end of the original bit sequences after zero padding. <start>"and end symbol" <end>"As a label for the output bitstream;

[0086] The IQ data of the time-domain and frequency-domain sequences are decomposed into real and imaginary parts, respectively, and denoted as the time-domain data sequence. and frequency domain data sequences It is expressed as follows:

[0087] ;

[0088] ;

[0089] in, and These represent functions for extracting the real and imaginary parts of a sequence, respectively.

[0090] S105. Construct a training dataset by combining the time-domain data sequence, the frequency-domain data sequence, and the corresponding output bitstream labels.

[0091] The training dataset is constructed as follows ,in, It is the input signal data. It is the output bitstream label. This is the number of samples in the training dataset.

[0092] In the signal recovery neural network of this embodiment, the multi-scale convolutional embedding layer includes a time-domain processing branch and a frequency-domain processing branch. Each branch consists of a sequentially connected multi-scale convolutional module and a spatial attention module.

[0093] The multi-scale convolution module contains multiple parallel temporal convolution sub-modules. Each sub-module uses a convolution kernel of different size and a configurable dilation rate to capture channel characteristic signals at different scales and stack them to output multi-scale features.

[0094] In each temporal convolutional submodule, the convolution dilation rate is... The kernel size is The input vector is , Indicates the number of processes currently being processed. The bit vector, after passing through a non-causal dilated convolution layer, is the first... Bit vector The update is as follows:

[0095] ;

[0096] in, The trainable weights of the convolution kernel are represented. For bias, Represents the ReLU activation function. This indicates a weight normalization operation;

[0097] The outputs of each temporal convolutional submodule are stacked, and the shape of the feature vector output by the multi-scale temporal convolutional module is as follows: ,in, , , These represent the signal length dimension, feature dimension, and multi-scale dimension, respectively.

[0098] The spatial attention module performs average pooling along both the signal length and feature dimensions, generates a spatial weight matrix through two-dimensional convolution and Sigmoid activation, weights and fuses multi-scale features and compresses them to a specified dimension, and outputs a time-domain feature vector or a frequency-domain feature vector.

[0099] Spatial attention modules along the dimensions and Perform average pooling, dividing the dimension into... The vectors are respectively converted into shapes of and The feature map is then concatenated with the pooling results from both directions, followed by convolution, batch normalization, and non-linear activation.

[0100] Two-dimensional convolution and sigmoid activation are applied to the feature maps in two directions respectively to obtain the corresponding spatial weight matrices. The weight matrices and the original three-dimensional vectors are then fused through matrix multiplication to achieve weighted allocation of importance for each spatial location.

[0101] The weighted vector is then flattened and convolved to compress it to the specified dimension, and the feature vector is output.

[0102] After extracting time-domain and frequency-domain feature vectors from the signal using multi-scale convolutional embedding layers, the encoder aligns and fuses these features. To capture the sequential characteristics of the feature vectors, sinusoidal position coding is applied to add position information to the input sequence, resulting in a time-domain and frequency-domain feature sequence containing position information. The calculation method is as follows:

[0103] ;

[0104] ;

[0105] in, Indicates the first line, number Data indexed by column position. Indicates the feature dimension.

[0106] Specifically, in the signal recovery neural network, the encoder uses a bidirectional cross-attention mechanism to perform two independent attention calculations, aligning and fusing the temporal and frequency domain features extracted by the multi-scale convolutional embedding layer.

[0107] The first attention calculation is performed using the time-domain feature vector as the query and the frequency-domain feature vector as the key and value.

[0108] A second attention calculation is performed using the frequency domain feature vector as the query and the time domain feature vector as the key and value.

[0109] The results of the two attention calculations are added together and fused to obtain the overall feature vector of the signal.

[0110] In one specific implementation, the encoder employs two parallel processing branches, each performing independent attention calculations; specifically including:

[0111] The first processing branch consists of a time-domain → frequency-domain attention submodule, a first residual normalization module, a first feedforward neural network, and a second residual normalization module connected in sequence.

[0112] The second processing branch consists of a frequency domain → time domain attention submodule, a third residual normalization module, a second feedforward neural network, and a fourth residual normalization module connected in sequence.

[0113] And element-wise adders and ReLU activation units connected to the outputs of the two branches;

[0114] The first processing branch is used to calculate attention weights with temporal features with added position encoding as queries and frequency features with added position encoding as keys and values, and output temporally aligned frequency-domain enhanced features;

[0115] The second processing branch is used to calculate attention weights with frequency domain features with added position coding as queries and time domain features with added position coding as keys and values, and outputs frequency domain aligned time domain enhanced features;

[0116] The element-wise adder and ReLU activation unit are used to fuse the enhanced features of the two branches to generate the overall feature vector of the signal.

[0117] In each attention calculation,

[0118] For the input feature sequence The query vector Q, key vector K, and value vector V are obtained by linear transformation:

[0119] ;

[0120] in, , , It is a learnable weight matrix. The query vector and key vector are multiplied by a dot product and then scaled using a softmax activation function to obtain the weight matrix. Finally, the weight matrix and value vector are fused using matrix multiplication to output the weighted features. The attention calculation is represented as follows:

[0121] ;

[0122] The outputs of multiple attention heads are concatenated, and then nonlinear mapping is performed through a feedforward neural network to enhance feature representation. After iteration, the time-domain and frequency-domain features are added and activated by the ReLU function to obtain a vector representing the overall features of the signal. .

[0123] Specifically, the decoder uses an autoregressive approach, generating bit sequences bit by bit with the start symbol as the trigger and the overall feature vector of the signal as the condition, until the end symbol is output and decoding stops, thus obtaining the final predicted bitstream sequence.

[0124] More specifically, the decoder adopts a multi-layer stacked architecture, and each decoding layer contains the following sub-modules executed in sequence:

[0125] The first-level multi-head self-attention module receives the bit sequence mapped by the linear embedding layer, and uses a causal mask to shield future positions and padding bits to ensure that the current position depends only on historical information. After residual connection and layer normalization, it outputs a bit feature representation that incorporates context information.

[0126] The second-level multi-head cross-attention module uses the overall feature vector of the signal output by the encoder as the key and value, and the above bit feature representation as the query to calculate the cross-attention weight, realize the mapping between the received signal and the bit stream, and then performs residual connection and layer normalization again.

[0127] The feedforward neural network performs a non-linear transformation on the cross-attention output and outputs the enhanced features of the decoding layer through residual connections and layer normalization.

[0128] Multiple decoding layers are repeatedly stacked in a multi-layer stacked architecture to deepen feature representation;

[0129] Finally, the features are mapped to the output space through a linear layer and a softmax function, and transformed into bit-by-bit classification probabilities.

[0130] Specifically, in S2, a mask is used to predict the training signal to recover the neural network model;

[0131] In the original bitstream labels 0 and 1, adding the number 2 indicates a padding bit, and adding 3 indicates a start symbol. <start>", 4 indicates the end symbol" <end>", expand the label for signal recovery to The corresponding bitstream padding and sequence order generation mask ensures that the model neither sees future positions nor focuses on padding positions when processing bit sequences of different modulated signals.

[0132] The signal recovery problem is structured as a classification task. The goal is to minimize the BER (bitrate error) based on the bitstream labels. and predicting bitstream The network optimization objective is defined as follows:

[0133] ;

[0134] in, These are the parameters of the neural network model.

[0135] The cross-entropy function is used as the loss function, defined by comparing the predicted probability of each bit with the actual probability of the label:

[0136] ;

[0137] in, Indicates the bitstream label in the first... The first Classification probability, Indicates the prediction result in the first... The first Classification probability.

[0138] In S3, the trained signal recovery neural network is used to perform autoregressive decoding on the OFDM received signal. During this process...

[0139] 1) Time-domain data sequence of OFDM received signal and frequency domain data sequences The input is fed into the multi-scale convolutional embedding layer of the signal recovery neural network to perform time-domain and frequency-domain feature extraction;

[0140] Functional operations representing multi-scale convolutional embedding layers, temporal features and frequency domain features The extraction process can be represented as:

[0141] ;

[0142] ;

[0143] 2) Time-domain features and frequency domain features The input encoder aligns and fuses two feature vectors, mapping the input sequence to a continuously represented vector. ; ,in, Indicates encoding operation;

[0144] 3) The decoder uses an autoregressive approach, based on vectors. and a specific start symbol <start>"Generate the first bit, combine the newly generated bit with the existing output sequence, and predict the next bit;

[0145] The prediction update of the bitstream sequence can be expressed as: ;in, Indicates a decoding operation. , These represent the predicted sequence results at the previous time step and the current time step, respectively. This process is iterated until a specific end symbol is output. <end>Output the final predicted bitstream sequence. .

[0146] Preferably, the reasoning process employs a greedy search strategy, where the model selects the word with the highest probability as the output.

[0147] This embodiment further includes performance verification testing to evaluate the demodulation performance of the signal recovery neural network under actual channel conditions:

[0148] Test environment configuration: The number of subcarriers was set to 64, using a comb pilot structure. The transmitted signal consisted of two consecutive OFDM symbols, with symbol mapping using a hybrid modulation scheme of BPSK / QPSK / 8PSK / 16QAM. Channel simulation was based on the 3GPP TDL model, where TDL-A / E were used for the validation set (corresponding to NLOS and LOS scenarios respectively), and TDL-B / C / D were used for the training set. This test performed bit error rate (BER) statistics in 1dB steps within the SNR range of -5dB to 20dB.

[0149] Comparison benchmark settings: To verify the technical effect of this embodiment, three types of prior art were selected for comparison:

[0150] Per-CSI: The theoretical upper bound of performance assuming the receiver has perfect channel state information;

[0151] DeepRx / DL-AMR / Comm-Transformer: A receiver solution based on deep learning;

[0152] Seq-Rec: A scheme based on the signal recovery neural network in this embodiment;

[0153] Seq-Rec-T / Seq-Rec-F: A single-domain input variant of this embodiment, used to verify the contribution of dual-domain complementarity;

[0154] Seq-Rec-C: A single-scale convolution variant in this embodiment, used to verify multi-scale convolution contributions.

[0155] Test steps: After the autoregressive decoding described in S3 is completed, statistically predict the bitstream sequence. With the original bit sequence The BER performance. The number of test sets for each signal-to-noise ratio point is 2000.

[0156] Results analysis:

[0157] 1. Demodulation performance analysis under known modulation schemes:

[0158] As attached Figures 2 to 5 As shown, the bit error rate (BER) performance of the proposed Seq-Rec model exhibits the following characteristics under various modulation modes: The Per-CSI method, based on ideal channel state information, is primarily limited by the decrease in signal-to-noise ratio (SNR) and the increase in modulation order, while possessing natural immunity to channel fading. In the low SNR region (SNR≤5dB), the BER curve of Seq-Rec closely approximates the Per-CSI benchmark, and this convergence becomes increasingly significant with increasing modulation order, indicating that Seq-Rec can still approach the theoretical performance boundary under strong noise interference. However, in the high SNR region (SNR≥15dB), limited by the residual effects of multipath delay spread and Doppler shift, Seq-Rec cannot reproduce the zero-BER plateau of Per-CSI, especially under higher-order modulations such as 16QAM, where the BER tends to... Despite reaching a saturation point, Seq-Rec demonstrates a systematic performance advantage over existing deep learning methods (such as DeepRx, DL-AMR, and Comm-Transformer) across the entire signal-to-noise ratio range, revealing its stronger generalization ability to represent channel non-stationary characteristics through a multi-scale time-frequency feature fusion mechanism. Furthermore, the validation dataset employs a dynamic cyclic prefix configuration to simulate real-world channel delay spread variations: when the cyclic prefix duration is less than the maximum channel delay spread, traditional methods, unable to adaptively compensate for inter-symbol interference (ISI), exhibit significantly lower performance ceilings than Seq-Rec, further confirming the robustness of this invention to complex channel environments.

[0159] 2. Modulation identification and demodulation performance analysis under unknown modulation schemes

[0160] As attached Figure 6 and Figure 7 As shown, given that the Per-CSI, DeepRx, and Comm-Transformer architectures lack modulation scheme recognition capabilities, this validation focuses solely on a horizontal comparison with DL-AMR. The training dataset employs a hybrid modulation configuration to support joint recognition and demodulation tasks across modulation types. While the modulation recognition accuracy of both models exhibits similar trends, Seq-Rec demonstrates faster convergence with increasing signal-to-noise ratio in both QPSK and 8PSK modes. Within the dBSNR range, it can stably achieve 100% recognition accuracy without fluctuations. Conversely, DL-AMR's sustained recognition accuracy is insufficient, making it difficult to maintain 100% accuracy across the entire range, resulting in a theoretical upper limit to its bit error rate performance. In adaptive modulation communication systems, the transmitter dynamically selects the modulation strategy based on real-time channel quality to balance spectral efficiency and transmission reliability—low-order modulation ensures robust transmission under low signal-to-noise ratio (SNR), while high-order modulation improves spectral efficiency under high SNR conditions. Seq-Rec can achieve 100% recognition accuracy at the SNR operating point adapted to each modulation mode, thus better meeting the engineering application requirements of adaptive modulation systems.

[0161] In comparison, as shown in the appendix Figure 8 and Figure 9 As shown, the Seq-Rec model achieves systematic performance advantages across the entire modulation range. Under high signal-to-noise ratio conditions, the bit error rate (BER) of Seq-Rec can stably reach [value missing] for each modulation scheme. The demodulation accuracy of Seq-Rec is significantly better than that of DL-AMR. Notably, Seq-Rec's performance improvement for higher-order 16QAM modulation is more pronounced than that of QPSK and 8PSK, revealing its enhanced symbol discrimination capability through deep fusion of time-frequency features. Overall, DL-AMR is usable in low-order modulation scenarios with favorable channel conditions, while Seq-Rec, with its stable generalization capability across modulation types and its advantages in sequence autoregressive modeling, is more suitable for adaptive reception scenarios with multiple modulation modes coexisting in complex electromagnetic environments.

[0162] This embodiment further includes performance comparison verification experiments to illustrate the synergistic effect of multi-scale convolutional embedding layers and dual-domain input:

[0163] Under the same channel conditions as S1-S3, three comparative models were constructed to quantitatively analyze the contribution of key modules:

[0164] Comparison Model 1 (Seq-Rec-T): Remove the frequency domain data sequence input in S1, retain only the time domain data sequence processed by the multi-scale convolutional embedding layer, and keep the encoder and decoder structures unchanged from this embodiment;

[0165] Comparison Model 2 (Seq-Rec-F): Remove the time-domain data sequence input in S1, retain only the frequency-domain data sequence processed by the multi-scale convolutional embedding layer, and keep the encoder and decoder structures unchanged from this embodiment;

[0166] Comparison Model 3 (Seq-Rec-C): Keep the dual-domain input format of S1 unchanged, replace the multi-scale convolutional embedding layer with a single-scale convolutional structure (the convolutional kernel size is fixed at 3, and the dilation rate d=1), and keep the encoder and decoder structures unchanged from this embodiment.

[0167] Experimental procedure: Observe the results at a signal-to-noise ratio of 10dB. Generate 2000 test samples for each of the QPSK / 16QAM / 64QAM mixed modulation signals. Reduce the overall feature vector of the encoder output signal to a two-dimensional space using t-SNE and observe the clustering distribution.

[0168] Experimental results: such as Figures 10-13 As shown, the complementary information from two domains and multi-scale feature extraction significantly enhance noise robustness and inter-class discriminative ability. Conversely, Seq-Rec-C exhibits weaker clustering at the same signal-to-noise ratio because single convolution cannot capture multi-resolution details. Seq-Rec-T and Seq-Rec-F, due to insufficient single-domain information, show significantly higher feature entanglement at all signal-to-noise ratios. These results confirm that two-domain inputs provide effective complementary information, and multi-scale convolution is crucial for optimizing signal feature perception. Therefore, Seq-Rec achieves optimal feature representation in a fully noisy environment.

[0169] In summary, the adaptive intelligent demodulation method for communication signals in complex environments disclosed in this invention can learn the mapping from received signals to information bits end-to-end and recover the bit stream through autoregression, effectively improving the signal recovery capability of adaptive modulation communication systems in complex environments.

[0170] In this scheme, multi-scale convolutional embedding layers are combined with spatial attention mechanisms to simultaneously extract time-domain and frequency-domain features, effectively capturing the deep structure of signals under multipath fading and rapidly time-varying channels, and improving robustness to complex environments.

[0171] By employing bidirectional cross-attention (time domain ↔ frequency domain cross-checking) to achieve time-frequency feature alignment and fusion, it can more effectively suppress inter-symbol interference (ISI) and inter-carrier interference (ICI) compared to traditional unidirectional processing.

[0172] The decoder generates data bit by bit through an autoregressive method and combines causal masking to avoid future information leakage. In scenarios where the channel changes rapidly and there is no prior modulation information, it achieves dynamic and real-time signal recovery and reduces the bit error rate (BER).

[0173] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.< / end> < / start> < / end> < / start> < / end> < / start>

Claims

1. An adaptive intelligent demodulation method for communication signals in complex environments, characterized in that, include: A training dataset is constructed, which contains OFDM received signals with various modulation schemes. The time-domain IQ sequence and frequency-domain IQ sequence of the received signal are used as input samples, and the original bit sequence with start and end symbols added is used as the output bitstream label. A signal recovery neural network is trained using the training dataset, the neural network comprising: Multi-scale convolutional embedding layers are used to extract time-frequency domain features; The encoder employs a bidirectional cross-attention mechanism to align and fuse time-frequency domain features, generating an overall feature vector of the signal. The decoder, triggered by the start symbol, performs autoregressive decoding based on a mask self-attention mechanism, outputting bit probabilities bit by bit until an end symbol is generated, thus restoring the original bit stream. The trained signal recovery neural network is used to perform autoregressive decoding on the OFDM received signal to recover the original bit stream end-to-end without prior modulation information. In the encoder, the bidirectional cross-attention mechanism performs two independent attention calculations: The first attention calculation is performed using the time-domain feature vector as the query and the frequency-domain feature vector as the key and value. A second attention calculation is performed using the frequency domain feature vector as the query and the time domain feature vector as the key and value. The results of the two attention calculations are added and fused to obtain the overall feature vector of the signal.

2. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 1, characterized in that, Constructing the training dataset includes: The original bit information is converted into IQ data and mapped into a modulation symbol sequence according to different modulation methods. After serial-to-parallel conversion, pilot insertion, inverse discrete Fourier transform, parallel-to-serial conversion and cyclic prefix insertion, the transmitted signal is generated. After channel transmission and noise superposition, the received signal is obtained. The time-domain received signal is converted into a frequency-domain received signal using the Discrete Fourier Transform. Set the maximum modulation order and the maximum sampling length of the received signal, and use zero padding to compensate the original bit sequence, time-domain IQ sequence and frequency-domain IQ sequence into equal-length sequences respectively; The IQ data of the time-domain IQ sequence and the frequency-domain IQ sequence are converted into time-domain data sequences and frequency-domain data sequences composed of real and imaginary data, respectively. Start and end symbols are added to the beginning and end of the original bit sequence after zero padding as labels for the output bit stream. The time-domain data sequence, frequency-domain data sequence, and corresponding output bitstream labels are used to construct the training dataset.

3. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 1, characterized in that, The multi-scale convolutional embedding layer includes a temporal processing branch and a frequency domain processing branch. Each branch consists of sequentially connected multi-scale convolutional modules and spatial attention modules. The multi-scale convolution module contains multiple parallel temporal convolution sub-modules. Each sub-module uses a convolution kernel of different size and a configurable dilation rate to capture channel characteristic signals at different scales and stack them to output multi-scale features. The spatial attention module performs average pooling along both the signal length and feature dimensions, generates a spatial weight matrix through two-dimensional convolution and Sigmoid activation, weights and fuses multi-scale features and compresses them to a specified dimension, and outputs a time-domain feature vector or a frequency-domain feature vector.

4. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 1, characterized in that, The encoder employs two parallel processing branches, each performing independent attention calculations; specifically including: The first processing branch consists of a time-domain → frequency-domain attention submodule, a first residual normalization module, a first feedforward neural network, and a second residual normalization module connected in sequence. The second processing branch consists of a frequency domain → time domain attention submodule, a third residual normalization module, a second feedforward neural network, and a fourth residual normalization module connected in sequence. And element-wise adders and ReLU activation units connected to the outputs of the two branches; The first processing branch is used to calculate attention weights with temporal features with added position encoding as queries and frequency features with added position encoding as keys and values, and output temporally aligned frequency-domain enhanced features; The second processing branch is used to calculate attention weights with frequency domain features with added position coding as queries and time domain features with added position coding as keys and values, and outputs frequency domain aligned time domain enhanced features; The element-wise adder and ReLU activation unit are used to fuse the enhanced features of the two branches to generate the overall feature vector of the signal.

5. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 1, characterized in that, The decoder uses an autoregressive approach, generating bit sequences bit by bit based on the start symbol and the overall feature vector of the signal, until the end symbol is output and decoding stops, thus obtaining the final predicted bitstream sequence.

6. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 5, characterized in that, The decoder employs a multi-layer stacked architecture, with each decoding layer containing the following sub-modules executed in sequence: The first-level multi-head self-attention module receives the bit sequence mapped by the linear embedding layer, and uses a causal mask to shield future positions and padding bits to ensure that the current position depends only on historical information. After residual connection and layer normalization, it outputs a bit feature representation that incorporates context information. The second-level multi-head cross-attention module uses the overall feature vector of the signal output by the encoder as the key and value, and the above bit feature representation as the query to calculate the cross-attention weight, realize the mapping between the received signal and the bit stream, and then performs residual connection and layer normalization again. The feedforward neural network performs a non-linear transformation on the cross-attention output and outputs the enhanced features of the decoding layer through residual connections and layer normalization. Multiple decoding layers are repeatedly stacked in a multi-layer stacked architecture to deepen feature representation; Finally, the features are mapped to the output space through a linear layer and a softmax function, and transformed into bit-by-bit classification probabilities.

7. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 1, characterized in that, The neural network model is recovered by using masking to predict the training signal; Add padding, start, and end markers to the output bitstream labels to expand the label space; A mask is generated based on the padding position and sequence order, so that the model training process does not focus on future positions and padding positions; the cross-entropy loss function is used to compare the predicted bit probability with the label probability, and the network parameters are updated with the goal of minimizing the bit error rate.

8. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 7, characterized in that, The signal recovery problem is constructed during training as The five-ary classification task, where 0 and 1 represent bit values, 2 represents padding bits, 3 represents the start symbol, and 4 represents the end symbol; With the aim of minimizing the bit error rate, the neural network parameters are optimized based on the output bitstream label and the predicted bit probability. The decoder ultimately maps features to bit classification probabilities using the Softmax function, achieving bit-by-bit decision output.

9. The adaptive intelligent demodulation method for communication signals in complex environments according to claim 8, characterized in that, The trained signal recovery neural network is used to perform autoregressive decoding on the OFDM received signal. The process includes: 1) Input the time-domain data sequence and frequency-domain data sequence of the OFDM received signal into the multi-scale convolutional embedding layer of the signal recovery neural network to perform time-domain and frequency-domain feature extraction; 2) Input the time-domain and frequency-domain features into the encoder, align and fuse the two feature vectors, and map the input sequence to a continuously represented vector. ; ； in, Indicates encoding operation; Representing time-domain features and Represents frequency domain characteristics; 3) The decoder uses an autoregressive approach, based on vectors. The first bit is generated by combining the start symbol with the newly generated bit and the existing output sequence to predict the next bit. The prediction update of the bitstream sequence is represented as: ； in, Indicates a decoding operation. , These represent the predicted sequence results at the previous and current time steps, respectively. This process is iterated until a specific end symbol is output; the final predicted bitstream sequence is then output. .

Images