A signal recognition and demodulation method and system
By integrating signal modulation type identification and demodulation functions through a signal identification and demodulation model based on the complex domain Transformer architecture, the problems of low efficiency and limited performance in existing technologies are solved, and efficient and flexible signal identification and demodulation in complex environments are realized.
Patent Information
- Application Number
- CN202411837049.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-11-11
- Estimated Expiration
- 2044-12-13
AI Technical Summary
In existing technologies, signal identification and demodulation are processed independently, resulting in low efficiency and limited performance. In particular, it is difficult to accurately identify signal types and perform efficient demodulation under non-ideal propagation environments or high dynamic conditions. Furthermore, traditional demodulators require prior knowledge of signal modulation parameters and a fixed sampling rate.
A signal recognition and demodulation model based on the complex domain Transformer architecture is adopted. By segmenting the signal into fixed-length segments in the time domain and inserting modulation classification labels, the signal modulation feature is extracted and demodulated using a complex value encoder and decoder. An integrated deep learning model is constructed without the need for prior knowledge of the signal type or precise synchronization information.
It achieves efficient integration of signal recognition and demodulation, has a high degree of adaptability, can flexibly cope with different sampling rates and complex parameter conditions, enhances the system's versatility and robustness, and is suitable for dynamically changing communication environments.
Smart Images
Figure CN119814506B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of wireless communication and deep learning technology, and specifically relates to a signal recognition and demodulation method and system. Background Technology
[0002] Currently, in wireless communication and signal processing applications, the ever-increasing data transmission rates and signal complexity pose significant challenges to traditional signal identification and demodulation techniques. In particular, accurately identifying signal types and efficiently demodulating them in non-ideal propagation environments or high-dynamic conditions has become a critical technical issue.
[0003] In existing technologies, signal identification and demodulation are typically separated into two independent processes, i.e., signal identification and demodulation are designed separately. Specifically, traditional solutions often rely on feature extraction and classification algorithms to identify the signal type, and then select the appropriate demodulator based on the identification results during the demodulation stage. This undoubtedly increases the complexity and latency of the system. Moreover, traditional demodulators require prior knowledge of the signal modulation parameters and require fixed sampling rate and symbol period ratio, etc. SPS (SamplesperSymbol, the number of samples per symbol) refers to the number of samples contained in each symbol in a digital communication system.
[0004] In summary, while existing technologies can identify signal types and demodulate them, their discrete design results in low efficiency for both signal identification and demodulation. Furthermore, for situations involving varying signal modulation parameters such as wireless communication, radar monitoring, and electronic warfare, it is difficult to maintain a fixed sampling rate and symbol period ratio. In other words, existing technologies are only suitable for scenarios with fixed signal modulation parameters, such as cooperative communication, and their performance is limited in non-cooperative communication scenarios. Summary of the Invention
[0005] This invention provides a signal identification and demodulation method and system, aiming to solve the problems of low efficiency and limited performance caused by the independent processing of signal identification and demodulation in the prior art.
[0006] This invention provides a signal identification and demodulation method, the method comprising:
[0007] Receive target signals transmitted via wireless channels, wherein the target signals are wireless signals to be identified and demodulated;
[0008] The target signal is divided into signal segments of fixed length in the time domain, and a modulation classification tag is inserted into the signal token header of each signal segment to generate a signal token sequence.
[0009] The signal token sequence is input into a pre-built signal identification and demodulation model, which is a model built based on a complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain the modulation type feature and extract the signal modulation feature of the target signal in the complex domain.
[0010] The modulation type feature is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and the modulation type corresponding to the maximum probability value is determined as the modulation type of the target signal.
[0011] The complex-valued decoder in the complex-domain Transformer architecture receives the signal modulation features output by the complex-valued encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
[0012] Preferably, before the step of inputting the signal token sequence into the pre-built signal identification and demodulation model, the method further includes:
[0013] Determine the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges;
[0014] Based on the modulation parameter range, a complex training signal corresponding to the training signal is generated through signal simulation.
[0015] Generate signal data and signal parameter information corresponding to the complex training signal, and determine the optimal sampling point position of each signal parameter in the signal parameter information to obtain the optimal sampling point position sequence;
[0016] Based on the signal data, the modulation type, the optimal sampling point position sequence, and the preset decoding symbol sequence, a training sample is constructed. The training sample is divided into multiple signal segments in the time domain, and the segmented signal segments are transformed into signal token sequences. The preset decoding symbol sequence includes modulation symbols corresponding to biphasic phase shift modulation and modulation symbols corresponding to quadrature phase shift keying, and corresponds to the modulation type of the target signal.
[0017] The optimal sampling point supervision signal is generated based on the optimal sampling point position sequence, and the signal token sequence is trained based on the optimal sampling point supervision signal to obtain a basic model for signal recognition and demodulation based on the complex domain Transformer architecture.
[0018] After training is completed, the basic model for signal recognition and demodulation is evaluated and predicted to determine the optimal training model as the signal recognition and demodulation model.
[0019] Preferably, the step of determining the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges, includes:
[0020] Determine the number of modulation types for the training signal;
[0021] Determine the shaping filter for the training signal and the filtering parameters of the filter;
[0022] Determine at least one of the following: the number of samples per symbol parameter or a range of the number of samples per symbol parameter for the signal; and,
[0023] Determine at least one of the following: signal length parameter or a range of signal length parameters.
[0024] Preferably, the step of generating the complex training signal corresponding to the training signal through signal simulation based on the modulation parameter range includes:
[0025] For each of the N modulation types, generate M independent samples to obtain a training sample set containing N×M training samples;
[0026] According to the preset signal simulation rules, the simulation signal of the training sample set is generated;
[0027] The simulation signal is subjected to amplitude normalization processing to obtain the complex training signal corresponding to the training signal.
[0028] Preferably, the step of generating the optimal sampling point supervision signal based on the optimal sampling point position sequence includes:
[0029] The monitoring signal at the optimal sampling point is initialized as a sequence of all "0"s of a preset length;
[0030] Based on each value in the optimal sampling point position sequence, an optimal sampling point monitoring signal is generated.
[0031] Preferably, the step of training the training samples based on the supervision signal at the optimal sampling point to obtain a basic model for signal recognition and demodulation based on the complex domain Transformer architecture includes:
[0032] The complex training signal is input into the complex domain Transformer architecture preprocessor to generate a token sequence;
[0033] The token sequence output by the preprocessor is input into the complex value encoder of the complex domain Transformer architecture to obtain modulation category identification, optimal sampling point and signal modulation characteristics;
[0034] The loss functions of the modulation category identification, the optimal sampling point, and the signal modulation features are weighted and summed, and the backpropagation algorithm is used to train a basic signal identification and demodulation model with a complex domain Transformer architecture.
[0035] Preferably, before the step of generating the optimal sampling point supervision signal based on the optimal sampling point position sequence, the method further includes:
[0036] The time delay of decimal point samples is simulated using random time delay, and the training samples are processed with random time delay.
[0037] The training samples after random delay processing are padded with several "0"s at preset positions to simulate the delay of integer sample points;
[0038] Random noise is injected into the training samples after time delay processing of integer sample points to simulate the situation where the signal is affected by noise.
[0039] This invention provides a signal identification and demodulation system, the system comprising:
[0040] A signal receiving module is used to receive target signals transmitted through a wireless channel, wherein the target signals are wireless signals to be identified and demodulated.
[0041] The token sequence generation module is used to divide the target signal into signal segments of fixed length in the time domain, and insert a modulation classification label into the head of the signal token of each signal segment to generate a signal token sequence.
[0042] The feature extraction module is used to input the signal token sequence into a pre-built signal identification and demodulation model. The signal identification and demodulation model is a model built based on the complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain the modulation type feature and extract the signal modulation feature of the target signal in the complex domain.
[0043] The signal identification and demodulation module is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and determine the modulation type corresponding to the maximum probability value as the modulation type of the target signal; and the complex value decoder in the complex domain Transformer architecture receives the signal modulation features output by the complex value encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
[0044] Preferably, the system further includes: a model building module, used for:
[0045] Determine the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges;
[0046] Based on the modulation parameter range, a complex training signal corresponding to the training signal is generated through signal simulation.
[0047] Generate signal data and signal parameter information corresponding to the complex training signal, and determine the optimal sampling point position of each signal parameter in the signal parameter information to obtain the optimal sampling point position sequence;
[0048] Based on the signal data, the modulation type, the optimal sampling point position sequence, and the preset decoding symbol sequence, a training sample is constructed. The training sample is divided into multiple signal segments in the time domain, and the segmented signal segments are transformed into signal token sequences. The preset decoding symbol sequence includes modulation symbols corresponding to biphasic phase shift modulation and modulation symbols corresponding to quadrature phase shift keying, and corresponds to the modulation type of the target signal.
[0049] The optimal sampling point supervision signal is generated based on the optimal sampling point position sequence, and the signal token sequence is trained based on the optimal sampling point supervision signal to obtain a basic model for signal recognition and demodulation based on the complex domain Transformer architecture.
[0050] After training is completed, the basic model for signal recognition and demodulation is evaluated and predicted to determine the optimal training model as the signal recognition and demodulation model.
[0051] Preferably, the model building module is specifically used for:
[0052] Determine the number of modulation types for the training signal;
[0053] Determine the shaping filter for the training signal and the filtering parameters of the filter;
[0054] Determine at least one of the following: the number of samples per symbol parameter or a range of the number of samples per symbol parameter for the signal; and,
[0055] Determine at least one of the signal length parameters or a range of signal length parameters.
[0056] Preferably, the model building module is specifically used for:
[0057] For each of the N modulation types, generate M independent samples to obtain a training sample set containing N×M training samples;
[0058] According to the preset signal simulation rules, the simulation signal of the training sample set is generated;
[0059] The simulation signal is subjected to amplitude normalization processing to obtain the complex training signal corresponding to the training signal.
[0060] Preferably, the model building module is specifically used for:
[0061] The monitoring signal at the optimal sampling point is initialized as a sequence of all "0"s of a preset length;
[0062] Based on each value in the optimal sampling point position sequence, an optimal sampling point monitoring signal is generated.
[0063] Preferably, the model building module is specifically used for:
[0064] The complex training signal is input into the complex domain Transformer architecture preprocessor to generate a token sequence;
[0065] The token sequence output by the preprocessor is input into the complex value encoder of the complex domain Transformer architecture to obtain modulation category identification, optimal sampling point and signal modulation characteristics;
[0066] The loss functions of the modulation category identification, the optimal sampling point, and the signal modulation features are weighted and summed, and the backpropagation algorithm is used to train a basic signal identification and demodulation model with a complex domain Transformer architecture.
[0067] Preferably, the model building module is further used for:
[0068] The time delay of decimal point samples is simulated using random time delay, and the training samples are processed with random time delay.
[0069] The training samples after random delay processing are padded with several "0"s at preset positions to simulate the delay of integer sample points;
[0070] Random noise is injected into the training samples after time delay processing of integer sample points to simulate the situation where the signal is affected by noise.
[0071] The present invention provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the above-described method.
[0072] The present invention provides a computer-readable storage medium on which a computer program is stored, and the computer program, when executed by a processor, implements the above-described method.
[0073] As can be seen from the above scheme, this invention deeply integrates deep learning and signal processing technologies, providing an integrated deep learning model. By integrating the signal modulation type identification and demodulation functions into a unified complex Transformer model, it can simultaneously solve the signal identification and demodulation problems, effectively overcoming the drawbacks of low efficiency and limited performance caused by separate designs. It also has a high degree of adaptability, requiring no prior knowledge of the signal type or precise synchronization information, and can flexibly cope with complex parameter conditions such as different sampling rates. This feature enhances the system's versatility and robustness, maintaining stable performance even in dynamically changing communication environments. This provides an efficient, flexible, and adaptive solution for applications such as wireless communication, radar monitoring, and electronic warfare. Attached Figure Description
[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0075] Figure 1 A flowchart of the signal identification and demodulation method provided by the present invention;
[0076] Figure 2 A structural diagram of a signal recognition and demodulation model based on a complex Transformer architecture provided by the present invention;
[0077] Figure 3 A schematic diagram of the optimal sampling point monitoring signal provided by the present invention;
[0078] Figure 4 This is a schematic diagram of a signal identification and demodulation model provided by the present invention;
[0079] Figure 5 The utilization provided by the present invention Figure 4 A schematic diagram illustrating signal recognition and demodulation performed by the constructed signal recognition model;
[0080] Figure 6 This is an architecture diagram of the signal identification and demodulation system provided by the present invention;
[0081] Figure 7 This is another architecture diagram of the signal identification and demodulation system provided by the present invention;
[0082] Figure 8 A structural diagram of an electronic device provided by the present invention. Detailed Implementation
[0083] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0084] To address the low efficiency and limited performance caused by the independent processing of signal identification and demodulation in existing technologies, this invention provides a signal identification and demodulation method and system. The signal identification and demodulation method and system provided by this invention will be described in detail below.
[0085] Example 1
[0086] Please see Figure 1 This is a flowchart of a signal identification and demodulation method provided by the present invention. Specifically, the method may include the following steps:
[0087] Step S101: Receive the target signal transmitted through the wireless channel, wherein the target signal is the wireless signal to be identified and demodulated.
[0088] Step S102: As Figure 2 As shown, the target signal is divided into signal segments of fixed length in the time domain, and a modulation classification label is inserted into the signal token header of each signal segment to generate a signal token sequence.
[0089] It should be noted that this application segments the target signal in the time domain. Preferably, the segmentation can be based on a fixed length to obtain several signal segments. For the last signal segment, its length is likely to be less than the fixed length. In this case, to facilitate subsequent processing, the insufficient length portion of the signal segment needs to be padded with zeros to make up the length, thereby forming a signal token. Of course, zero-padding is only a preferred method provided by this invention, and other implementation methods are also possible.
[0090] For each signal token corresponding to a signal segment, preferably, a modulation classification tag can be inserted into the header of the signal token, for example, a special tag can be inserted. <start>The token serves as a modulation classification label.
[0091] Step S103: Input the signal token sequence into a pre-built signal identification and demodulation model. The signal identification and demodulation model is a model built based on the complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain the modulation type feature and extract the signal modulation feature of the target signal in the complex domain.
[0092] The Transformer model architecture, proposed by Google in its 2017 paper "Attentions is All you need," uses a self-attention mechanism instead of the RNN (Recurrent Neural Network) network structure commonly used in NLP (Natural Language Processing) tasks. Compared to RNN networks, the biggest advantage of the Transformer architecture is its parallel computation capability. See also... Figure 2 The Transformer model architecture is essentially an encoder-decoder architecture. The Transformer can be divided into two parts: an encoding component and a decoding component. The encoding component, referred to as the "complex encoder" below, is used to extract the signal modulation type and the sampling point supervision signal, while the decoding component, referred to as the "complex decoder" below, is used to predict the real-time symbol probability distribution based on historical demodulated symbols.
[0093] The complex Transformer architecture used in this invention can be obtained by extending the real Transformer architecture in various ways. Furthermore, the extended Transformer architecture can directly process complex signals without converting them to real numbers. For example, it can be obtained by extending the real Transformer architecture as follows:
[0094] (1) Complex-valued linear layer extension
[0095] The complex weight matrix is multiplied by its conjugate transpose, i.e., y = x W. H , where W is the weight matrix in complex form.
[0096] (2) Complex value multi-head self-attention extension
[0097] The complex-valued attention matrix is converted to a real value using a specific function f(·), followed by probability normalization.
[0098]
[0099] Where Q, K, and V are the query vector, key vector, and value vector, respectively, and d is the dimension of the key vector.
[0100] (3) Complex-valued activation function extension: Activating the real and imaginary parts of the signal independently can be expressed as follows:
[0101]
[0102] in and Let g(·) represent taking the real part and taking the imaginary function, respectively, and g(·) is the activation function in the real number field.
[0103] (4) Complex-valued layer normalization extension: This involves calculating the mean and variance of complex data and then standardizing them.
[0104]
[0105] Var[·] represents the expectation and variance operations, and γ and β are trainable complex parameters.
[0106] It should be noted that the above methods (1) to (4) are preferred methods provided by the present invention and should not be construed as limiting the present invention. Other methods can also be used in actual applications, and the present invention does not limit them.
[0107] For ease of reading, the construction process of the signal identification and demodulation model will be explained in detail below.
[0108] Step S104: The modulation type feature is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and the modulation type corresponding to the maximum probability value is determined as the modulation type of the target signal.
[0109] Specifically, <start>The modulation type feature corresponding to the token is transmitted via a fully connected layer (see [link to documentation]). Figure 2 The probability distribution of modulation types is calculated, and the modulation type corresponding to the maximum probability value is determined as the modulation type of the target signal.
[0110] Step S105: The complex-valued decoder in the complex-domain Transformer architecture receives the signal modulation features output by the complex-valued encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
[0111] Please see Figure 5 As shown, during symbol decoding, the complex-valued decoder of the complex-domain Transformer architecture uses the output features of the complex-valued encoder as initial input. Specifically, for the first decoding step, the special... <sos>A token is input into the model, marking the start of sequence demodulation. At each decoding time step, the complex-valued decoder uses the previously identified symbols to predict the symbol probability distribution for the current time step; that is, the complex-valued decoder uses the previous k-1 symbols to predict the symbol probability distribution for the current k-th time step. The symbol with the highest probability is selected as the current decoding output and used as the input for the next time step. This process is repeated until the model predicts... <eos>Demodulation ends when the token or the preset maximum sequence length is reached.
[0112] Furthermore, after obtaining the identified and demodulated information data, the results can be output. Specifically, the decoded symbol sequence obtained in the above steps can be organized and output to form the final identified and demodulated information data.
[0113] Therefore, this invention deeply integrates deep learning and signal processing technologies, providing an integrated deep learning model. By integrating the identification and demodulation functions of signal modulation type into a unified complex Transformer model, it can simultaneously solve the signal identification and demodulation problems, effectively overcoming the drawbacks of low efficiency and limited performance caused by separate designs. It also has a high degree of adaptability, requiring no prior knowledge of the signal type or precise synchronization information, and can flexibly cope with complex parameter conditions such as different sampling rates. This feature enhances the system's versatility and robustness, maintaining stable performance even in dynamically changing communication environments. This provides an efficient, flexible, and adaptive solution for applications such as wireless communication, radar monitoring, and electronic warfare.
[0114] As mentioned earlier, before processing the signal token sequence, this invention pre-constructs a model based on the complex domain Transformer architecture, thereby integrating the signal modulation type and demodulation function into this model. This is one of the key inventive points of this invention. Below, we will describe in detail the construction process of the "signal identification and demodulation model." Please refer to [link to relevant documentation]. Figures 2 to 5 .
[0115] Specifically, the signal identification and demodulation model can be constructed by following these steps:
[0116] Step 1: Determine the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain. The modulation parameter information includes at least one of the modulation parameters and modulation parameter ranges.
[0117] In one implementation, the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain can be determined in the following manner, specifically including: determining the number of modulation types of the training signal; determining the shaping filter of the training signal and the filtering parameters of the filter (for example, if a root-raised cosine filter is used, the roll-off coefficient parameter or the range of the roll-off coefficient parameter needs to be determined); determining at least one of the parameters of the number of sampling points per symbol of the signal or the range of the number of sampling points per symbol parameter; and determining at least one of the parameters of the signal length or the range of the signal length parameter.
[0118] Step 2: Based on the modulation parameter range, generate a complex training signal corresponding to the training signal through signal simulation.
[0119] In one scenario, for each of the N modulation types, M independent samples are generated to obtain a training sample set containing N×M training samples; a simulation signal of the training sample set is generated according to a preset signal simulation rule; the simulation signal is subjected to amplitude normalization processing to obtain the complex training signal corresponding to the training signal.
[0120] For example, for a modulated signal within a parameter range, values can be randomly and uniformly selected within a specified range for signal generation.
[0121] It should be noted that, based on a custom parameter range, simulation software can be used to generate complex training signals of various types and under various conditions, thereby constructing a simulation dataset. Furthermore, existing technologies disclose specific methods for generating training signals based on simulation signals, which will not be elaborated upon here.
[0122] Step 3: Generate the signal data and signal parameter information corresponding to the complex training signal, and determine the optimal sampling point position for each signal parameter in the signal parameter information to obtain the optimal sampling point position sequence. (See [link to previous step]). Figure 3 .
[0123] In one scenario, the optimal sampling point monitoring signal can be generated as follows: the optimal sampling point monitoring signal is initialized as a sequence of all "0"s of a preset length; the optimal sampling point monitoring signal is generated based on each value in the optimal sampling point position sequence.
[0124] For example, the optimal sampling point position sequence can be generated as follows: Consider an undelayed signal data (signal index starting at "0"), with a sampling point parameter of s per symbol, and the signal data containing k symbols. The optimal sampling point position sequence is then:
[0125] In one specific implementation, such as Figure 3 As shown, an optimal sampling point monitoring signal can be generated based on the optimal sampling point position sequence. Let the optimal sampling point position sequence of the signal be S, and for each value S[i], i=0,…,k-1 in the optimal sampling point position sequence S, it is processed according to the following two cases:
[0126] The first type, if S[i] is an integer value, the assignment expression for the optimal sampling point monitoring signal T is: T[S[i]]=1;
[0127] The second approach, if S[i] is a non-integer value, is as follows:
[0128]
[0129] Where T represents the optimal sampling point monitoring signal, and the operator... Indicates rounding down, operator This indicates rounding up to the nearest integer.
[0130] Step 4: Based on the signal data, the modulation type, the optimal sampling point position sequence, and the preset decoding symbol sequence, a training sample is constructed. The training sample is divided into multiple signal segments in the time domain, and the segmented signal segments are transformed into signal token sequences. The preset decoding symbol sequence includes modulation symbols corresponding to biphasic phase shift modulation and modulation symbols corresponding to quadrature phase shift keying, and corresponds to the modulation type of the target signal.
[0131] In one implementation, the training sample symbol sequence includes: two modulation symbols, "0" and "1," corresponding to BPSK (Binary Phase Shift Keying); and four modulation symbols, "00," "01," "10," and "11," corresponding to QPSK (Quadrature Phase Shift Keying). Here, symbols of different modulation types are considered different symbols. It should be noted that BPSK and QPSK mentioned here are only two preferred modulation methods and are not intended to limit the invention; of course, other modulation types are also possible, and those skilled in the art can make reasonable settings according to actual needs.
[0132] Step 5: Generate an optimal sampling point supervision signal based on the optimal sampling point position sequence, and train the signal token sequence based on the optimal sampling point supervision signal to obtain a basic model for signal recognition and demodulation based on the complex domain Transformer architecture.
[0133] Step 6: After training is completed, the basic model for signal recognition and demodulation is evaluated and predicted to determine the optimal training model as the signal recognition and demodulation model.
[0134] In one specific implementation, the complex training signal can be input into the complex-domain Transformer architecture preprocessor to generate a token sequence; the token sequence output by the preprocessor is input into the complex-valued encoder of the complex-domain Transformer architecture to obtain modulation category identification, optimal sampling point, and signal modulation features; the loss functions of the modulation category identification, the optimal sampling point, and the signal modulation features are weighted and summed, and the backpropagation algorithm is used to train the basic model of signal identification and demodulation of the complex-domain Transformer architecture.
[0135] Specifically, the complex domain signal is first uniformly divided into a series of fixed-length segments, with any segments shorter than the required length padded with zeros; each segment is then passed through a fully connected layer to form a signal token. A special [feature / method] is then inserted at the head of the signal token. <start>Finally, the above token sequence is added to the corresponding embedded vector.
[0136] It should be noted that, <start>The token corresponds to the modulation type feature output of the signal, used for modulation category identification. A fully connected layer calculates the probability distribution of different modulation types (e.g., BPSK, QPSK), and the prediction error is minimized by a cross-entropy loss function compared to the true modulation category label. The signal token corresponds to the signal feature encoding output, used for the prediction of the optimal sampling point, extracted by another fully connected layer. The optimal sampling point prediction is compared with the optimal sampling point supervision signal T, and a binary cross-entropy loss is used to optimize the estimation accuracy.
[0137] Please see Figure 4 , <start>The token corresponds to the modulation feature output of the signal, and the token corresponds to the signal feature encoded output, which serves as the decoder input. Simultaneously, the decoder inputs the preceding symbol to predict the probability distribution of the next symbol. When predicting the first symbol, the decoder inputs a special... <sos>The token, indicating the start of demodulation, is used to output the probability distribution of the first symbol type after passing through the encoder and fully connected layer. For any subsequent decoding step, the decoder takes the ground truth labels of previously decoded symbols as input, predicts the symbol probability distribution at the current time step, and uses the cross-entropy loss function to minimize the prediction error compared to the ground truth symbol list labels. When the decoding step reaches the end of the sequence, its supervision label is... <eos>This indicates the end of demodulation.
[0138] Furthermore, in a preferred embodiment of the present invention, data augmentation can be performed on the training samples in step 3, before generating the optimal sampling point supervision signal based on the optimal sampling point position sequence, thereby improving the performance of the constructed model. Specific methods may include: simulating decimal point sample delay using random time delay and performing random time delay processing on the training samples; adding several "0"s at preset positions to the training samples after random time delay processing to simulate integer sample point delay; and injecting random noise into the training samples after integer sample point delay processing to simulate the situation where the signal is affected by noise.
[0139] In one implementation, data augmentation can be performed in the following ways, specifically including:
[0140] (1) Random delay: The signal is subjected to random time delay processing, with the delay length being a random value between [0,1] to simulate the time delay of fractional sample points, and the time delay length is increased for each value in the optimal sampling point position sequence. For example, if the original signal is delayed by p sampling points to form a new signal, then each value in the optimal sampling point position sequence of the new signal is increased by p, i.e., S+p, where S is the optimal sampling point position sequence of the original signal.
[0141] (2) Random padding: Randomly pad the signal with a number of "0"s before and after it to simulate the delay of integer sample points. The number of zeros to be padded is a random value between [0, n], where n is a preset zero-padding parameter. The number of zeros padded before the signal is increased by the number of zeros padded before the signal in each value of the optimal sampling point position sequence. For example, consider padding the signal with m and n zeros before and after it respectively to form a new signal. Then the value of each value in the optimal sampling point position sequence of the new signal is increased by m, i.e., S+m, where S is the optimal sampling point position sequence of the original signal.
[0142] (3) Noise injection: Injecting noise of random magnitude into the signal to simulate the effect of noise on the signal, where the noise magnitude ensures that the signal-to-noise ratio satisfies [n l n h dB range.
[0143] Of course, the specific methods listed here are for illustrative purposes only and should not be construed as limiting the invention. In practical applications, the methods can be reasonably set according to specific circumstances. It should also be noted that after data augmentation of the training samples, the training samples in subsequent steps such as steps 3 to 6 are all data-augmented training samples.
[0144] It should be noted that steps 1 to 6 above constitute the complete construction process of the identification and demodulation model provided by this invention. As can be seen, this invention deeply integrates deep learning and signal processing technologies. By integrating the identification and demodulation functions of signal modulation types into a unified complex Transformer model, it constructs an integrated deep learning model that can simultaneously solve the signal identification and demodulation problems, effectively overcoming the drawbacks of low efficiency and limited performance caused by separate designs. It also possesses high adaptability, requiring no prior knowledge of the signal type or precise synchronization information, and can flexibly cope with complex parameter conditions such as different sampling rates. This feature enhances the system's versatility and robustness.
[0145] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions involved are not necessarily required by this application.
[0146] After introducing the signal identification and demodulation method provided by the present invention, the signal identification and demodulation system provided by the present invention will be described.
[0147] Example 2
[0148] Please refer to Figure 6 This is an architecture diagram of a signal recognition and demodulation system provided by the present invention. The system includes the following modules: a signal receiving module 210, a token sequence generation module 220, a feature extraction module 230, and a signal recognition and demodulation module 240.
[0149] The signal receiving module 210 is used to receive the target signal transmitted through the wireless channel, wherein the target signal is the wireless signal to be identified and demodulated.
[0150] The token sequence generation module 220 is used to divide the target signal into signal segments of fixed length in the time domain, and insert a modulation classification label into the head of the signal token of each signal segment to generate a signal token sequence.
[0151] The feature extraction module 230 is used to input the signal token sequence into a pre-built signal identification and demodulation model. The signal identification and demodulation model is a model built based on the complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain the modulation type feature and extract the signal modulation feature of the target signal in the complex domain.
[0152] The signal identification and demodulation module 240 is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and determine the modulation type corresponding to the maximum probability value as the modulation type of the target signal; and the complex value decoder in the complex domain Transformer architecture receives the signal modulation features output by the complex value encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
[0153] Furthermore, such as Figure 7 As shown, the system may further include a model building module 200, used to determine the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges; based on the modulation parameter range, generating a complex training signal corresponding to the training signal through signal simulation; generating signal data and signal parameter information corresponding to the complex training signal, and determining the optimal sampling point position of each signal parameter in the signal parameter information to obtain an optimal sampling point position sequence; constructing training samples based on the signal data, the modulation type, the optimal sampling point position sequence, and a preset decoding symbol sequence. The sample is segmented into multiple signal segments in the time domain, and the segmented signal segments are transformed into signal token sequences. The preset decoding symbol sequence includes modulation symbols corresponding to biphasic phase-shift modulation and quadrature phase-shift keying, and corresponds to the modulation type of the target signal. An optimal sampling point supervision signal is generated based on the optimal sampling point position sequence, and the signal token sequence is trained based on the optimal sampling point supervision signal to obtain a basic signal recognition and demodulation model based on a complex-domain Transformer architecture. After training, the basic signal recognition and demodulation model is evaluated and predicted to determine the optimal training model as the signal recognition and demodulation model.
[0154] It should be noted that since the construction process of the signal identification and demodulation model has been described in detail in the method embodiments, the relevant parts can be found in the description of the method embodiments section, and will not be repeated here.
[0155] In one implementation, the model building module 200 is specifically used to determine the number of modulation types of the training signal; determine the shaping filter of the training signal and the filtering parameters of the filter; determine at least one of the parameters of the number of sampling points per symbol or the range of parameters of the number of sampling points per symbol; and determine at least one of the parameters of the signal length or the range of parameters of the signal length.
[0156] In one scenario, the model building module 200 is specifically used to generate M independent samples for each of the N modulation types, thereby obtaining a training sample set containing N×M training samples; generate a simulation signal for the training sample set according to a preset signal simulation rule; and perform amplitude normalization processing on the simulation signal to obtain the complex training signal corresponding to the training signal.
[0157] In another scenario, the model building module 200 is specifically used to initialize the optimal sampling point supervision signal as a preset length of all "0" signal sequence; and generate the optimal sampling point supervision signal according to each value in the optimal sampling point position sequence.
[0158] In another scenario, the model building module 200 is specifically used to input the complex training signal into the complex domain Transformer architecture preprocessor to generate a token sequence; input the token sequence output by the preprocessor into the complex value encoder of the complex domain Transformer architecture to obtain modulation category identification, optimal sampling point, and signal modulation features; perform a weighted summation of the loss functions of the modulation category identification, the optimal sampling point, and the signal modulation features, and use the backpropagation algorithm to train the basic model of signal identification and demodulation of the complex domain Transformer architecture.
[0159] Furthermore, the model building module 200 is also used to simulate the decimal point sample delay using random time delay and to perform random time delay processing on the training samples; to add several "0"s at preset positions to the training samples after random time delay processing to simulate the integer sample point delay; and to inject random noise into the training samples after integer sample point delay processing to simulate the situation where the signal is affected by noise.
[0160] It should be noted that the system implementation is basically similar to the method implementation, so the description is relatively simple. For relevant details, please refer to the description in the method implementation section.
[0161] Example 3
[0162] To address the above problems, the present invention provides an electronic device, such as... Figure 8 As shown, it includes a memory 310, a processor 320, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method applied to the electronic device.
[0163] It should be noted that the electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, the processor 320 and the memory 310. Those skilled in the art will understand that... Figure 8 This is merely an example of an electronic device and does not constitute a limitation on the electronic device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.
[0164] The processor 320 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.
[0165] The memory 310 can be an internal storage unit of the electronic device, such as a hard drive or memory. The memory 310 can also be an external storage device of the electronic device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard. Furthermore, the memory 310 can include both internal and external storage units. The memory 310 is used to store the computer program and other programs and data required by the electronic device. The memory 310 can also be used to temporarily store data that has been output or will be output.
[0166] Example 4
[0167] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into the commuter broadcast control device. The computer-readable storage medium stores one or more computer programs, which, when executed by a processor, implement the methods described above.
[0168] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.
[0169] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of protection of this invention shall still be determined by the scope defined in the appended claims.< / eos> < / sos> < / start> < / start> < / start> < / eos> < / sos> < / start> < / start>
Claims
1. A signal identification and demodulation method, characterized in that, The method includes: Receive target signals transmitted via wireless channels, wherein the target signals are wireless signals to be identified and demodulated; The target signal is divided into signal segments of fixed length in the time domain, and a modulation classification tag is inserted into the signal token header of each signal segment to generate a signal token sequence. The signal token sequence is input into a pre-built signal identification and demodulation model, which is a model built based on a complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain modulation type features and extract the signal modulation features of the target signal in the complex domain. The modulation type feature is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and the modulation type corresponding to the maximum probability value is determined as the modulation type of the target signal. The complex-valued decoder in the complex-domain Transformer architecture receives the signal modulation features output by the complex-valued encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
2. The signal identification and demodulation method according to claim 1, characterized in that, Prior to the step of inputting the signal token sequence into the pre-built signal identification demodulation model, the method further includes: Determine the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges; Based on the modulation parameter range, a complex training signal corresponding to the training signal is generated through signal simulation. Generate signal data and signal parameter information corresponding to the complex training signal, and determine the optimal sampling point position of each signal parameter in the signal parameter information to obtain the optimal sampling point position sequence; Based on the signal data, the modulation type, the optimal sampling point position sequence, and the preset decoding symbol sequence, a training sample is constructed. The training sample is divided into multiple signal segments in the time domain, and the segmented signal segments are transformed into signal token sequences. The preset decoding symbol sequence includes modulation symbols corresponding to biphasic phase shift modulation and modulation symbols corresponding to quadrature phase shift keying, and corresponds to the modulation type of the target signal. The optimal sampling point supervision signal is generated based on the optimal sampling point position sequence, and the signal token sequence is trained based on the optimal sampling point supervision signal to obtain a basic model for signal recognition and demodulation based on the complex domain Transformer architecture. After training is completed, the basic model for signal recognition and demodulation is evaluated and predicted to determine the optimal training model as the signal recognition and demodulation model.
3. The signal identification and demodulation method according to claim 2, characterized in that, The step of determining the modulation type and modulation parameter information of the training signal to be identified and demodulated in the time domain, wherein the modulation parameter information includes at least one of modulation parameters and modulation parameter ranges, includes: Determine the number of modulation types for the training signal; Determine the shaping filter for the training signal and the filtering parameters of the filter; Determine at least one of the following: the number of samples per symbol parameter or a range of the number of samples per symbol parameter for the signal; and, Determine at least one of the following: signal length parameter or a range of signal length parameters.
4. The signal identification and demodulation method according to claim 2, characterized in that, The step of generating the complex training signal corresponding to the training signal through signal simulation based on the modulation parameter range includes: For each of the N modulation types, generate M independent samples to obtain a training sample set containing N×M training samples; According to the preset signal simulation rules, the simulation signal of the training sample set is generated; The simulation signal is subjected to amplitude normalization processing to obtain the complex training signal corresponding to the training signal.
5. The signal identification and demodulation method according to claim 2, characterized in that, The step of generating the optimal sampling point supervision signal based on the optimal sampling point position sequence includes: Initialize the monitoring signal at the optimal sampling point to a sequence of all "0"s of a preset length; Based on each value in the optimal sampling point position sequence, an optimal sampling point monitoring signal is generated.
6. The signal identification and demodulation method according to claim 2, characterized in that, The step of training the training samples based on the supervision signal of the optimal sampling point to obtain the basic model for signal recognition and demodulation based on the complex domain Transformer architecture includes: The complex training signal is input into the complex domain Transformer architecture preprocessor to generate a token sequence; The token sequence output by the preprocessor is input into the complex value encoder of the complex domain Transformer architecture to obtain modulation category identification, optimal sampling point and signal modulation characteristics; The loss functions of the modulation category identification, the optimal sampling point, and the signal modulation features are weighted and summed, and the backpropagation algorithm is used to train a basic signal identification and demodulation model with a complex domain Transformer architecture.
7. The signal identification and demodulation method according to any one of claims 2 to 6, characterized in that, Before the step of generating the optimal sampling point supervision signal based on the optimal sampling point position sequence, the method further includes: The time delay of decimal point samples is simulated using random time delay, and the training samples are processed with random time delay. The training samples after random delay processing are padded with a number of "0"s at preset positions to simulate the delay of integer sample points; Random noise is injected into the training samples after time delay processing of integer sample points to simulate the situation where the signal is affected by noise.
8. A signal identification and demodulation system, characterized in that, The system includes: A signal receiving module is used to receive target signals transmitted through a wireless channel, wherein the target signals are wireless signals to be identified and demodulated. The token sequence generation module is used to divide the target signal into signal segments of fixed length in the time domain, and insert a modulation classification label into the head of the signal token of each signal segment to generate a signal token sequence. The feature extraction module is used to input the signal token sequence into a pre-built signal identification and demodulation model. The signal identification and demodulation model is a model built based on the complex domain Transformer architecture. The complex value encoder in the complex domain Transformer architecture processes the input signal token sequence to obtain modulation type features and extract the signal modulation features of the target signal in the complex domain. The signal identification and demodulation module is used to calculate the probability distribution of the modulation type through the fully connected layer in the complex domain Transformer architecture, and determine the modulation type corresponding to the maximum probability value as the modulation type of the target signal; and the complex value decoder in the complex domain Transformer architecture receives the signal modulation features output by the complex value encoder and demodulates the signal modulation features in the time domain to obtain the identified and demodulated information data.
9. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.