Intelligent hierarchical modulation and demodulation method based on automatic encoder

Abstract

The invention belongs to the technical field of wireless communication, and particularly relates to an intelligent hierarchical modulation and demodulation method based on an automatic encoder, which is used for solving the problem that the design of constellation points is limited in the construction process of an existing hierarchical modulation constellation diagram. On the basis of a communication system structure of an end-to-end automatic encoder (Autoencoder), a layered weighted loss mechanism is creatively introduced, and differential optimization can be performed according to the importance degree of data information, so that hierarchical protection of multi-level information is realized in a model training stage, and the reliability of the model training stage is improved. The mechanism effectively improves the adaptability of the system to data with different priorities under a dynamic channel condition, and overcomes the problem of insufficient flexibility and robustness when a traditional modulation scheme is used for processing multi-service-level transmission.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wireless communication, and specifically relates to an intelligent hierarchical modulation and demodulation method based on an auto-encoder. BACKGROUND

[0002] In recent years, with the increasing requirements for transmission reliability and spectrum utilization in wireless communication systems, the traditional modulation and demodulation method gradually shows problems such as high implementation complexity, insufficient flexibility, and slow response to channel environment changes. With the development and evolution of the fifth generation (5G) and sixth generation (6G) communication technologies, global communication systems are facing multiple challenges such as greater capacity, lower latency, and more service demands.

[0003] To effectively respond to the above challenges, hierarchical modulation (HM) has become an important modulation means to enhance the adaptability of communication systems, which has the ability to dynamically allocate resources according to channel quality and user service level. The basic idea of hierarchical modulation is to divide the bits inside each modulation symbol into multiple levels, each level carries data of different categories or different service levels, and a differentiated protection mechanism is applied to realize a layered transmission strategy. Hierarchical modulation can realize parallel transmission of multiple data streams under limited spectrum conditions, and by superimposing multiple data sub-streams of different levels in a single signal carrier, high-priority information can obtain stronger protection, while low-priority data shares the remaining resources, thereby adapting to changing channel conditions and differentiated user needs.

[0004] Although hierarchical modulation shows certain potential in improving spectrum efficiency and supporting differentiated services, however, traditional hierarchical modulation methods usually rely on static constellation mapping and fixed coding structure, which is difficult to maintain good performance under complex time-varying channel conditions. In addition, traditional hierarchical modulation methods are also difficult to flexibly respond to multi-level quality of service regulation requirements. In this context, deep learning (DL) technology, with its excellent performance in complex nonlinear feature extraction and superior performance in data-driven modeling, is gradually becoming a key new tool for communication system design and optimization. In the face of modulation and demodulation, the physical layer modeling method based on end-to-end auto-encoder (AE) provides a new technical path for hierarchical modulation; such a model can jointly optimize the encoding, modulation and decoding process of the signal through training and learning, significantly improving the robustness of the system under non-ideal channels. However, current related researches mostly focus on single-layer modulation or transmission tasks of static priority data, and have not fully considered the multi-level, adjustable quality of service transmission architecture, which forms obvious application limitations in dynamic resource scheduling and service diversification scenarios. In view of this problem, the present application provides an intelligent hierarchical modulation and demodulation method based on an auto-encoder. SUMMARY

[0005] In view of the problems of traditional hierarchical modulation in the field of wireless communication, an intelligent modulation and demodulation method combining the idea of hierarchical modulation in traditional communication is proposed from the perspective of deep learning to solve the problem of limited constellation point design in the constellation construction process of existing hierarchical modulation. Based on the end-to-end autoencoder (Autoencoder) communication system structure, a hierarchical weighted loss mechanism is creatively introduced, which can optimize different levels of information according to the importance of data information, so as to realize hierarchical protection of multi-level information in the model training stage. This mechanism effectively improves the adaptability of the system to different priority data under dynamic channel conditions, and overcomes the problems of insufficient flexibility and robustness of traditional modulation schemes in processing multi-service level transmission.

[0006] To achieve the above object, the technical scheme adopted by the present application is:

[0007] An intelligent hierarchical modulation and demodulation method based on an autoencoder, the autoencoder model comprising: an encoder module and a decoder module; original bit data is input into the encoder module, and modulation symbols are output by the encoder module; the modulation symbols are transmitted to the decoder module through a channel, and reconstructed bit data is output by the decoder module; characterized in that the original bit data is divided into K logical levels, and each logical level is preset with a weight parameter w1, w2…w K ; the encoder module, the decoder module and the channel model constitute an intelligent hierarchical modulation and demodulation model to complete training, and during the training process, hierarchical loss weight coefficients are set:

[0008]

[0009] wherein w i represents the weight parameter of the i-th logical level, a i represents the loss weight coefficient of the i-th logical level, and w1 represents the reference weight parameter of the highest priority logical level.

[0010] Based on the hierarchical loss weight coefficient setting, the loss function Loss is:

[0011]

[0012] wherein Loss i represents the independent loss term of the i-th logical level.

[0013] Further, the independent loss term of each logical level adopts a binary cross-entropy loss function.

[0014] Further, in the encoder module, firstly, a convolutional layer with a kernel length of k is adopted to perform preliminary feature extraction on the input original bit data, k is the number of modulation bits; then three same convolutional layers are adopted to perform high-dimensional feature extraction, the output features are projected to a preset two-dimensional signal space via a fully connected linear mapping layer, and finally, normalization processing is performed to obtain the modulation symbol.

[0015] Further, the decoder module corresponds to the encoder module, and the received signal is first subjected to feature extraction by three convolutional layers, the output features are converted to an output dimension consistent with the number of modulation bits via a fully connected linear mapping layer, and finally, the reconstructed bit data is obtained through a flattening operation.

[0016] Based on the above technical solutions, the present application has the following advantages:

[0017] Compared with the traditional quadrature amplitude modulation, the hierarchical modulation realizes the hierarchical transmission strategy according to the importance of data by embedding data bits from different data types or different priorities in a single modulation symbol and implementing differentiated protection mechanisms for each layer of data.

[0018] The present application draws on the core idea of hierarchical modulation and proposes an intelligent hierarchical modulation and demodulation method based on an autoencoder. The intelligent hierarchical modulation and demodulation method based on deep learning mechanism realizes hierarchical learning and differentiated protection of different service level data bits by introducing a hierarchical weighted loss function on the end-to-end autoencoder system, thereby improving the expression ability of the system for multi-level information and the anti-interference performance. Unlike the traditional global loss function with uniform weighting, the hierarchical weighted loss function assigns different penalty coefficients to different levels of data, thereby guiding the network to prioritize the reconstruction accuracy of key levels and preferentially protecting important data during model training, thus realizing the result of hierarchical modulation.

[0019] In summary, the present application effectively overcomes the performance bottleneck caused by the averaging of the optimization function objective of the traditional autoencoder in multi-level data optimization, and is particularly suitable for signal transmission in multi-priority communication scenarios. Compared with the traditional hierarchical modulation or quadrature amplitude modulation (QAM) strategy, the deep learning driven hierarchical modulation and demodulation architecture proposed by the present application has significant advantages in the balance of the modulation constellation structure and the bit error rate (BER) performance, thereby verifying its practical application potential in multi-level service transmission scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1It is a structure schematic diagram of a wireless communication transmission system based on an auto-encoder (AE) simulation in the application.

[0021] Figure 2 It is a schematic diagram of an intelligent hierarchical modulation and demodulation model training process in the application.

[0022] Figure 3 It is a structure schematic diagram of an auto-encoder model in the application.

[0023] Figure 4 It is a constellation comparison diagram of an intelligent hierarchical modulation and demodulation model under different signal-to-noise ratios in the application.

[0024] Figure 5 It is a performance comparison diagram of an intelligent hierarchical modulation and demodulation model under different signal-to-noise ratios in the application.

[0025] Figure 6 It is a constellation comparison diagram of an intelligent hierarchical modulation and demodulation model under different weight coefficients in the application.

[0026] Figure 7 It is a performance comparison diagram of high-bit of an intelligent hierarchical modulation and demodulation model under different weight coefficients in the application.

[0027] Figure 8 It is a performance comparison diagram of low-bit of an intelligent hierarchical modulation and demodulation model under different weight coefficients in the application. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical scheme and beneficial effects of the application clearer, the application will be further described in detail below in combination with the drawings and examples.

[0029] As shown in Figure 1 It is a wireless communication transmission system based on an auto-encoder (AE) simulation, which includes a sending end, a channel and a receiving end. The sending end encodes original bit information into a transmittable signal, which is sent to the receiving end through the wireless channel. The receiver decodes and recovers the received signal.

[0030] In the automatic encoder-based modulation and demodulation method, the original bit data is first input into the encoding module at the sending end, the input information is extracted and nonlinearly mapped by using a multi-layer neural network structure, and finally a complex domain signal meeting the modulation symbol structure requirement is output, realizing the automatic modulation process of the signal. Thus, the modulation signal generated by using the automatic encoder can be used to optimize two problems in the communication sending end process: first, the modulation signal is used to construct a two-dimensional space constellation, thereby realizing the visual evaluation of the signal distribution characteristics and the system modulation performance, and can be directly compared with the traditional modulation technology; second, the modulation signal is input into different types of channel models, thereby simulating the transmission characteristics and distortion behavior in various channel environments. At the receiving end of the automatic encoder, the damaged signal is recovered and reconstructed by the neural network, thereby completing the complete communication process from encoding, modulation, channel transmission to decoding.

[0031] Based on this, the embodiment provides an intelligent hierarchical modulation and demodulation method based on an automatic encoder. The automatic encoder model comprises an encoder module and a decoder module. The original bit data is input into the encoder module, and the modulation symbol is output by the encoder module. The modulation symbol is transmitted to the decoder module through a channel, and the reconstructed bit data is output by the decoder module. The original bit data is divided into K logical levels, and each logical level is sequentially preset with a weight parameter w1, w2…w K The encoder module, the decoder module and the channel model constitute an intelligent hierarchical modulation and demodulation model to complete training. In the training process, a hierarchical loss weight coefficient is set:

[0032]

[0033] wherein w i represents the weight parameter of the i-th logical level, a i represents the loss weight coefficient of the i-th logical level, and w1 represents the reference weight parameter of the highest priority logical level.

[0034] Based on the hierarchical loss weight coefficient, the loss function Loss is set as:

[0035]

[0036] wherein Loss i represents the independent loss term of the i-th logical level.

[0037] As shown in Figure 2 , the embodiment introduces a weighted loss function mechanism, which allocates corresponding loss weight coefficients to different logical levels corresponding to each bit in the modulation symbol, thereby simulating the mechanism of differentiating protection of data with different priorities in hierarchical modulation. More specifically:

[0038] In the traditional hierarchical modulation technology, in order to improve the anti-interference ability of high priority information, the way of mapping it to the center area of the modulation constellation diagram is often used, and low priority information is allocated to the peripheral position of the constellation diagram; This design strategy makes it possible to maintain high demodulation accuracy of high priority information carried in the center area even if the peripheral symbol is misjudged under low SNR conditions; Usually, each symbol in the modulation constellation diagram represents a signal point, containing data bits from different priorities, and the transmission performance of high and low priority information is optimized by adjusting the geometric layout and Euclidean distance between symbols. In the traditional QAM structure, the hierarchical modulation constellation diagram is usually composed of a pre-defined fixed structure, and the symbol spacing and layout rules are determined by the pre-set hierarchical modulation coefficient. The size of the modulation coefficient directly affects the sparsity and protection ability of the constellation points in space: if it is set too large, the constellation points are too concentrated, which is beneficial to the protection of the high priority layer, but the low priority layer will be severely limited under high SNR conditions, and cannot effectively utilize the spectrum resources; If the coefficient is set too small, it is difficult to effectively isolate and enhance the protection of the high priority layer under low SNR conditions, resulting in a decline in overall communication performance. Due to the lack of flexibility in this traditional design, the constellation pattern is difficult to adjust dynamically according to real-time channel conditions, resulting in the problem of insufficient adaptability in a variable channel environment.

[0039] To overcome the above shortcomings, the embodiment proposes an intelligent hierarchical modulation method based on deep learning, which is based on an end-to-end autoencoder architecture, relies on the adaptive modeling capability of deep neural networks, and automatically learns the optimal constellation structure and hierarchical coefficient matching the channel conditions through the training process; The core idea is to assign different loss weights to the logical layers corresponding to each bit in the modulation symbol, and to simulate the differential protection mechanism for different priority data by introducing a weighted loss function, thereby realizing the joint optimization of the constellation structure and the hierarchical modulation strategy. During the model training process, the system not only learns the mapping relationship from the input bit sequence to the modulation symbol, but also dynamically adjusts the penalty weight of different logical layers in the loss function according to their importance. This process enables the model to adaptively optimize the geometric structure of the constellation diagram and its hierarchical strategy at the symbol mapping level and the weight level according to different channel conditions, thereby balancing and improving the transmission performance of different priority data. As shown in Figure 2 , which is significantly different from the traditional fixed hierarchical modulation method, and has the advantages of dynamic learning, automatic optimization and scene adaptation, and is particularly suitable for multi-level quality of service communication systems in complex and time-varying channel environments.

[0040] In summary, through the above mechanism, during the model training process, the high priority layer is given a higher α i to obtain stronger optimization tilt, thereby ensuring its reconstruction quality and robustness; while the low priority layer appropriately reduces the weight αi , to allow a moderate sacrifice of accuracy under high noise conditions, so as to provide more space resources in the constellation for key bits, to improve the reliability and efficiency of the overall communication system. Therefore, the model trained based on the weighted loss function can automatically learn the optimal modulation constellation distribution according to different channel states, especially in low SNR conditions, it can enhance the transmission robustness of high priority information, and in high SNR conditions, it can release the capacity potential corresponding to low priority information, to realize efficient resource utilization; finally, the wireless communication system can effectively control the bit error rate of high priority bits while significantly improving the error performance of low priority data and the overall spectrum utilization efficiency.

[0041] Further, as Figure 3 The structure diagram of the autoencoder model in the embodiment is shown, wherein m represents the number of bits of the information to be transmitted, and k represents the number of modulation bits. In the encoder part, a convolution kernel with a length of k is first used to perform preliminary information extraction on the input original bit information string; then three layers of convolution kernels with a length of 1 are used for high-dimensional feature extraction. Different from the first convolution, the subsequent three convolutions are used to further mine deeper information representation on the basis of the preliminary local features, so as to enhance the learning ability of the model for complex mapping relationship; next, the convolution layer output features are projected to the preset two-dimensional signal space through a fully connected linear mapping layer, wherein the output dimension is m x 2, corresponding to the in-phase component and the quadrature component of the m symbols respectively; finally, in order to meet the power or amplitude constraint requirements of the wireless communication physical layer for the transmitted signal, the signal needs to be normalized before being sent to the channel. The normalization layer normalizes the encoder output to control the signal energy or amplitude not to exceed the set threshold, so as to ensure that the final output meets the transmission power specification; thus, the automatic mapping from the input bit sequence to the symbol sequence conforming to the IQ modulation format is completed, that is, the function implementation of the transmitting end of the communication system. The decoder basically corresponds to the encoder. The received signal is first extracted by a three-layer convolution kernel with a length of 1, and then converted to an output dimension consistent with the number of modulation bits through a fully connected linear mapping layer; finally, through the flatten operation, the multi-dimensional output is transformed into a one-dimensional sequence, which is convenient for performing bit-level decision tasks, and the original bit sequence is reconstructed. A channel model is provided between the encoder and the decoder. The module does not contain trainable parameters, and its role is to simulate the interference and degradation suffered by the signal during transmission in the actual physical channel.

[0042] In the embodiment, additive white Gaussian noise (AWGN) is used as the channel model. This model is the most classical and widely used in communication theory, and its mathematical expression form is:

[0043] Y = X + N

[0044] Wherein, X represents the transmitting signal of the sending end, Y is the receiving signal of the receiving end, and N is a random noise term with Gaussian distribution; the channel model is characterized by having a constant power spectral density in the entire frequency band range, and each sample is independent and has no correlation.

[0045] The beneficial effects of the present application will be described in detail below in combination with simulation tests.

[0046] Based on the hierarchical modulation and demodulation model and the loss function described above, in the model training process, to ensure the convergence stability and performance optimization effect of the network, the training parameters are set as follows: the total number of training rounds is set to 200, the optimization algorithm uses Adam optimizer, and the initial learning rate is set to 1x10-3; in the model training, the independent loss term of each logical level is a binary cross-entropy loss function (BCE Loss), which is suitable for bit-level binary decision tasks and can effectively guide the model to optimize the bit decision accuracy, thereby improving the end-to-end performance of the overall system.

[0047] In order to evaluate the robustness and performance of the deep learning structure under different channel conditions, this embodiment adjusts the signal-to-noise ratio (SNR), i.e. the ratio of signal power to noise power, to construct channel environments of different intensities, thereby simulating and verifying the performance of the system under the AWGN channel.

[0048] In order to further compare the performance advantages of the intelligent hierarchical modulation and demodulation method based on the auto-encoder, this embodiment tests the AWGN with different signal-to-noise ratios, and the constellation diagram results and performance comparison results obtained by using the deep learning method are as follows: Figure 4 and Figure 5As shown. By observing the constellation diagram under different signal-to-noise ratios, it can be found that when the channel signal-to-noise ratio is low, for example, when the signal-to-noise ratio is 6dB, in order to enhance the transmission protection of the high priority data layer, the intelligent hierarchical modulation model proposed in the present application automatically compresses all the symbol points corresponding to the low priority data layer to the same position in the constellation diagram, thereby forming four constellation points determined only by the high priority layer; the constellation formed by these four points is similar to the traditional 4QAM modulation constellation diagram, which is derived from the fact that the high priority layer contains 2 bits, which corresponds to 4 combinations of QAM modulation and demodulation, and the model effectively improves the noise immunity of the high priority layer. As the signal-to-noise ratio gradually increases to 15dB, the generated constellation diagram begins to expand in structure; at this time, the constellation points are distributed in four obvious partitions in the diagram, and the degree of aggregation of the points in each region is slightly weakened compared to before, the interval between the points gradually widens, and the constellation cluster with clear boundaries is initially formed; but due to the overlapping phenomenon in the central region, the overall constellation diagram has not yet fully expanded. When the signal-to-noise ratio is further increased to 18dB, the constellation points are basically completely separated, forming a complete structure of 64 constellation points; at this time, the distance between the symbols is unevenly distributed, and the distance between the points in the central region is significantly smaller than that in the edge region; and when SNR reaches 21dB, the model finally generates a constellation diagram that highly fits the standard 64QAM format, with only a small range of point disturbance, showing excellent constellation configuration accuracy. From the above phenomena, it can be seen that under low signal-to-noise ratio conditions, the model proposed in the present application improves the robustness of the high priority layer by overlapping the symbol points of the low priority layer and constructing the constellation region division, and the bit error rate is significantly better than that of the traditional hierarchical modulation scheme; when the signal-to-noise ratio is high, the constellation diagram generated by the model presents a more regular geometric distribution, while ensuring that the bit error rate of the high priority layer is controlled within an acceptable range, it fully utilizes the transmission capacity of the low priority layer, and significantly reduces the bit error rate of the low priority layer.

[0049] As Figure 5The bit error rate curves of the high priority layer and the low priority layer under different SNR conditions are shown to quantify the performance difference. When the signal-to-noise ratio is in the interval of 0 dB to 15 dB, the application exhibits obvious error correction capability superior to the conventional scheme in the high priority layer, while the bit error rate of the low priority layer only shows slight difference, and the overall performance advantage is more significant. When the signal-to-noise ratio is improved to the interval of 18 dB to 24 dB, the performance comparison trend reverses. At this stage, the error performance of the model proposed in the low priority layer is superior to the conventional hierarchical modulation method, which reflects the effective utilization ability of low priority data resources. In the high priority layer, the bit error rate of the conventional method decreases rapidly; in contrast, the bit error rate of the high priority layer of the model proposed in the application decreases gently until it gradually approaches the bit error rate level of the low priority layer at a higher signal-to-noise ratio. In summary, the application can preferentially guarantee the reliable transmission of key bits in a low signal-to-noise ratio environment, and further improve the spectral efficiency in a high signal-to-noise ratio scenario, realizing flexible trade-off between different transmission levels, and having strong scene adaptability and communication performance advantage.

[0050] In addition to the performance comparison under different signal-to-noise ratios, the embodiment tests the influence of different hierarchical weight coefficients on the constellation structure. The embodiment is tested under the condition of a fixed signal-to-noise ratio of 12 dB, and the results are shown in Figure 6 When α = 2, the constellation points output from the encoder have fallen into four regions, but after being disturbed by channel noise at the receiving end, the symbol point distribution is seriously overlapped, and the region discrimination is ambiguous. With the increase of the weight coefficient, the distance between the symbol points in the same region of the constellation diagram transmitted by the encoder decreases, and the point distribution tends to converge, and the constellation diagram received by the decoder begins to show a clear partition structure. This indicates that the larger the hierarchical weight coefficient is, the more the model tends to aggregate low priority layer symbol points, thereby enhancing the noise resistance of the high priority layer and improving its decision accuracy.

[0051] In addition, as shown in Figure 7 and Figure 8 Further show the trend of BER performance of the high priority layer and the low priority layer under different hierarchical weight coefficients. From Figure 7 and Figure 8It can be observed that when the conventional deep learning modulation model does not introduce a weighting mechanism, the high-priority layer and low-priority layer bit error rate curve distribution has no obvious rules, and fluctuates greatly, basically around the performance curve of standard 64QAM, and it is difficult to achieve effective classification. When the classification weight coefficient is greater than 1, the model generally shows the characteristics that the high-priority layer bit error rate is better than the low-priority layer, verifying the effectiveness of the weighted loss function in realizing hierarchical modulation. Further comparison of experimental results under different weight coefficients shows that as the weight coefficient increases, the high-priority layer obtains higher protection weight, and its bit error rate continues to decrease; but at the same time, the utilization rate of the low-priority layer decreases, and the bit error rate rises. It is worth noting that when the signal-to-noise ratio is improved to a certain extent, the bit error rates of the high-priority layer and the low-priority layer gradually converge, and eventually approach the standard 64QAM modulation performance curve. This shows that in a high signal-to-noise ratio environment, the intelligent hierarchical modulation model proposed in the present application can improve the overall modulation efficiency while ensuring the reliability of high-layer data, and achieve performance comparable to traditional high-order modulation methods.

[0052] The above is only a specific implementation of the present application, any feature disclosed in the specification can be replaced by other equivalent or similar purpose alternative features unless specifically described, and all features disclosed or steps in all methods or processes can be combined in any way except for mutually exclusive features and / or steps.

Claims

1. An intelligent hierarchical modem method based on autoencoder, the autoencoder model comprising: The encoder module and the decoder module; The original bit data is input into the encoder module, and modulation symbols are output by the encoder module; The modulated symbols are transmitted to a decoder module through a channel, and reconstructed bit data is output by the decoder module; characterized in that the original bit data is divided into K logical levels, and each logical level is sequentially provided with a weight parameter w1, w2, …, w K The encoder module, the decoder module and the channel model constitute an intelligent hierarchical modulation and demodulation model to complete training, and a hierarchical loss weight coefficient is set in the training process. wherein w i represents the weight parameter of the i-th logical level, a i represents the loss weight coefficient of the i-th logical level, w1 represents the reference weight parameter of the highest priority logical level. The loss function Loss is set based on the hierarchical loss weight coefficient: wherein Loss i represents the independent loss term of the i-th logical level.

2. The intelligent hierarchical modem method based on autoencoder according to claim 1, wherein, The independent loss item of each logical level adopts a binary cross-entropy loss function. 3.The intelligent hierarchical modem method based on auto-encoder of claim 1, wherein, In the encoder module, a convolution layer with a convolution kernel length of k is first adopted to perform preliminary feature extraction on the input original bit data, and k is the modulation bit number; then three same convolution layers are adopted to perform high-dimensional feature extraction, the output features are projected to a preset two-dimensional signal space through a full connection linear mapping layer, and finally normalized to obtain modulation symbols.

4. The intelligent hierarchical modem method based on autoencoder according to claim 3, characterized in that, The decoder module corresponds to the encoder module, and the received signal is first subjected to feature extraction through three convolution layers, the output features are converted to an output dimension consistent with the modulation bit number through a full connection linear mapping layer, and finally subjected to a flattening operation to obtain reconstructed bit data.

Images