CN115834306B - Method for directly estimating symbol sequence of multiple communication signals under interference condition - Google Patents
Method for directly estimating symbol sequence of multiple communication signals under interference condition Download PDFInfo
- Publication number
- CN115834306B CN115834306B CN202211377436.3A CN202211377436A CN115834306B CN 115834306 B CN115834306 B CN 115834306B CN 202211377436 A CN202211377436 A CN 202211377436A CN 115834306 B CN115834306 B CN 115834306B
- Authority
- CN
- China
- Prior art keywords
- signal
- symbol
- symbol sequence
- sequence
- fpsse
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The application discloses a method for directly estimating a multi-communication signal symbol sequence under an interference condition, which comprises the steps of extracting parallel symbol sequence estimation characteristics FPSSE based on single-channel observation to obtain characteristic sequences respectively containing symbol information of each target signal; and (3) performing sequence labeling based on the offline trained cyclic neural network to obtain a symbol sequence estimation result, wherein the characteristic sequence is used as input when the offline training is performed on the cyclic neural network. The method adopts the thought of parallel independent estimation of the multi-signal symbol sequence, but not the joint estimation of the multi-signal symbol sequence, so that the calculation complexity of the method is not increased along with the exponential growth of the target signal number. Meanwhile, the method is applicable to scenes with the same or different symbol rates of all target signals, and the practicability is remarkably improved. Because the interference among the target signals is restrained through FPSSE extraction, the estimation performance of the method has obvious advantages compared with a deep convolution network demodulator under the condition that the power difference of the target signals is obvious.
Description
Technical Field
The application belongs to the technical field of digital signal processing, and particularly relates to a method for directly estimating a symbol sequence of a multi-communication signal under an interference condition.
Background
In current cooperative/non-cooperative communications, single-channel receiving devices are more widely adopted for cost, volume, and other reasons, which also makes the time-frequency aliased digital communication signal problem often faced in complex electromagnetic environments more challenging. In order to realize signal analysis and information extraction based on single-channel aliasing observation, a target signal must be separated from the single-channel aliasing observation. Blind signal separation (Blind Signal Separation, BSS) techniques are intended to acquire the signal of interest contained therein based solely on aliased vision separation, where single channel BSS techniques are more intended to achieve this based on single channel vision. Therefore, research on single-channel BSS technology will have important theoretical and practical significance in improving the ability of existing cooperative/non-cooperative communication systems to cope with time-frequency aliased digital communication signals.
The multi-signal symbol sequence estimation is a special single-channel blind signal separation BSS task when a target signal is a digital communication signal, and the existing single-channel BSS algorithm for realizing the multi-signal symbol sequence estimation mainly has the following problems: firstly, due to the adoption of a mode of joint estimation of multiple signal symbol sequences, the calculated amount of a mainstream model-based algorithm based on the technologies of particle filtering PF, processing PSP by survivor paths and the like increases exponentially with the number of target signals and the modulation order of the target signals, and particularly the calculated amount of the algorithm based on the particle filtering PF is too high. Secondly, the PSP which is processed by the survivor path serving as a main stream algorithm can only be applied to the problem scene that the symbol rates of all target signals are equal at present, and the practicability is limited. Thirdly, for the existing data driving algorithm represented by the deep convolutional network demodulator DCND, the adaptive capacity of the existing data driving algorithm to the power level difference of each target signal is not strong because the existing data driving algorithm is not effectively preprocessed, and when each signal symbol sequence is estimated, the existing data driving algorithm is more easily interfered by other signals.
Disclosure of Invention
The application aims to provide a direct estimation method of a multi-communication signal symbol sequence under an interference condition, which solves the problems of high computational complexity and insufficient adaptability to target signal parameters of the existing multi-signal symbol sequence estimation algorithm based on a model and adopting joint estimation.
In order to achieve the above object, the present application provides a method for directly estimating symbol sequences of multiple communication signals under interference conditions, including:
extracting parallel symbol sequence estimation characteristics FPSSE based on single-channel observation to obtain characteristic sequences respectively containing symbol information of each target signal;
and performing sequence labeling based on the offline trained cyclic neural network to obtain a symbol sequence estimation result, wherein the characteristic sequence is used as input when the offline training is performed on the cyclic neural network.
Optionally, before extracting the symbol sequence estimation feature, a single-channel aliasing observation needs to be constructed, which specifically includes:
assuming that a single antenna receives I spectrally overlapping digital communication signals simultaneously, the single-channel aliasing observation can be expressed as:
wherein s is i (n-m i ) Representing a transmission delay of m i The ith target signal of the sampling interval, a i V (N) represents AWGN for its corresponding channel gain, N being the total number of sampling points;
for a typical amplitude/phase modulated digital communication signal s i (n) is further expressed as
Wherein E is signal power;corresponds to the q-th of the transmission i Symbols, Q i Is the total number of symbols; g i (. Cndot.) represents an equivalent pulse shaping function, T, which comprehensively considers the pulse shaping at the transmitting end and the physical transmission channel s Is->Sampling interval and symbol oversampling rate respectively; />Is the residual carrier frequency.
Optionally, performing parallel symbol sequence estimation feature FPSSE extraction based on single channel observations includes:
constructing P single-channel aliasing observations according to the step (1), wherein the modulation pattern of I digital communication signals contained in each aliasing observation is selected from a set omega with the size of C;
for the p (p.epsilon.1, p) th aliasing observation, I FPSSE subsequences are extracted therefrom, each of which may constitute a training sample with the known symbol sequence of its corresponding signal.
Optionally, extracting the FPSSE subsequence comprises: let g i (. Cndot.) has an FIR structure, and is derived fromLasting to->Therefore, receive s i Q of (n) i The sampling point range of the influence of each symbol is as follows:
for the firstSymbol->Has the following components
When (when)Time s i The first order differential sequence of (n) is expressed as:
assuming adequate samplingAnd g i (. Cndot.) is a symmetric function, where:
is->
Then equation (14) is approximated as:
when (when)When d i (n) is
Synthesizing the formulas (1), (15) and (16) to obtain the first-order difference of x (n) as follows:
wherein phi is i Representing the set of the following sampling points:
optionally, extracting the FPSSE subsequence further comprises: for the target signal s i (n) subsequences thereofFrom { d (n) |n ε [1, N]The extraction in } is as follows:
after division according to each signal, { d (n) |n ε [1, N ]]Will be divided into sub-sequence setsWherein each sub-sequence corresponds to a target signal.
Optionally, extracting the FPSSE subsequence further comprises:
for i.e. [1I ]]Further based onThe following normalized amplitude and phase sequences were calculated and respectively noted as
Wherein, I.I represents the amplitude of the complex sequence,representation->The mean value of (4), the angle (·) represents the complex phase, k i ∈[1，Q i -2L i +1]Will->As signal s in the received signal i (n) the corresponding FPSSE sequence.
Optionally, in the off-line training of the cyclic neural network, in order to demodulate signals with different possible modulation patterns, a cyclic neural network set is constructed, and the structures and parameters of the cyclic neural networks in the set are the same, so that the symbol sequence estimation of the target signals with different modulation patterns is trained, and each cyclic neural network comprises two BLSTM layers and one fully-connected classification layer.
Optionally, the output dimension of the classification layer is determined by the debugging order of the signal to which the classification layer is directed.
The application has the technical effects that: the method constructs a single-input-multiple-output symbol sequence estimator capable of carrying out parallel estimation of multiple signal symbol sequences by modeling symbol sequence estimation as a sequence labeling problem and realizing sequence labeling based on a cyclic neural network. Unlike DCND, which takes single-channel aliasing observation as input directly, the method firstly carries out parallel symbol sequence estimation feature extraction based on single-channel observation, and mainly comprises differential operation and specific moment extraction so as to obtain feature sequences respectively containing symbol information of each target signal. The offline training of the RNN-based symbol sequence estimator is a network input that uses the signature as an estimate of each target signal symbol sequence. Different from the existing multi-signal symbol sequence estimation algorithm based on the model, the method adopts the thought of parallel independent estimation of the multi-signal symbol sequence instead of joint estimation of the multi-signal symbol sequence, so that the calculation complexity of the method is not increased exponentially along with the number of target signals. Meanwhile, the method breaks through the limitation of the PSP algorithm on the symbol rate relation of the target signals, can be suitable for scenes with the same or different symbol rates of the target signals, and has the advantage of remarkably improving the practicability. Because the interference among the target signals is restrained through FPSSE extraction, the estimation performance is obviously superior to that of DCND under the condition that the power difference of the target signals is obvious. The simulation experiment result shows that the symbol sequence estimator provided by the method has the capability of simultaneously and accurately estimating a plurality of target signal symbol sequences under different signal power ratios, and has good adaptability to different target signals with larger power level differences. In the generalization capability test, the data driving method provided by the method still shows good estimation performance under the test condition which is not seen in training.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:
fig. 1 is a flow chart of a method for directly estimating symbol sequences of multiple communication signals under interference conditions provided by an embodiment of the present application;
FIG. 2 is a diagram of an RNN-based multi-signal symbol sequence estimator according to an embodiment of the present application;
fig. 3 is a graph of parallel symbol sequence estimation BER for two target signals at different SINR provided by an embodiment of the present application;
fig. 4 is a graph showing BER comparisons of different SIMO symbol sequence estimation algorithms provided by an embodiment of the present application.
Fig. 5 is a generalized performance test chart of the symbol sequence improving estimator according to the embodiment of the present application.
Detailed Description
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The application will be described in detail below with reference to the drawings in connection with embodiments.
As shown in fig. 1, the present embodiment provides a method for directly estimating symbol sequences of multiple communication signals under an interference condition, which includes the following steps:
s1, performing parallel symbol sequence estimation feature FPSSE extraction based on single-channel observation to obtain feature sequences respectively containing symbol information of each target signal;
s2, performing sequence labeling based on the offline trained cyclic neural network to obtain a symbol sequence estimation result, wherein the characteristic sequence is used as input when the offline training is performed on the cyclic neural network.
The cyclic neural network in this embodiment integrates the bidirectional RNN, i.e. BRNN structure, with the LSTM structure to form a BLSTM structure, where each hidden layer includes paired forward and backward LSTM nodes, and adjacent layers are connected by a tanh activation function. The following describes the operation performed by the forward LSTM node in the hidden layer at each time instant:
s t ＝g t ⊙i t +s t-1 ⊙f t (9)
wherein, the liquid crystal display device comprises a liquid crystal display device,and->Respectively representing the output of each forward node at the time t-1 and the time t; g t ，i t ，f t ，o t The method comprises the steps of inputting nodes, inputting gates, forgetting gates and outputting activation function values corresponding to the gates at the moment t respectively; s is(s) t-1 And s t Memory cell states at times t-1 and t, respectively; w (W) gx ，W ix ，W fx ，W ox Respectively inputting a traditional weight matrix between each corresponding unit; w (W) gh ，W ih ，W fh ，W oh Respectively a 'cyclic' weight matrix of the forward hidden layer node and the forward hidden layer node at adjacent time; b g ，b i ，b f ，b o Biasing the vectors for each cell; phi (&) and +.. Here, the vector represents the value of the corresponding part of all the forward nodes in the whole hidden layer. For example, s is composed of the state values of all forward node memory cells in the whole hidden layer.
While the rest of the computation is similar for the backward hidden layer node of the BLSTM except for the direction reversal of the circular connection. Specifically, to calculate the output of each backward node at time t, it is necessary to calculate the cell state of each backward node at time t+1 based on the input at time t. The same pair of BLSTM hidden layers share inputs, and the outputs of the two are sent to the next network layer after being integrated.
The time back propagation technique BPTT or a variant thereof, i.e. the truncated time back propagation technique TBPTT, is used. Unlike standard back propagation techniques, when training based on BPTT, it is assumed that a single RNN layer is spread out in time, treated as a multi-layer network sharing parameters between different time steps, and network parameters are optimized in both directions while taking into account both the longitudinal propagation of gradients between adjacent network layers and the transverse propagation over adjacent time steps. In TBPTT, however, the number of time steps for back propagation is limited to improve the processing efficiency for longer sequences. By training that includes a sufficient number of iterations, network training may be deemed complete when the loss function value on the training set is below a certain threshold or the error rate of the sequence annotation is below a certain level. In the online test of sequence labeling, an input sequence is directly sent to a trained network, and the network output is the sequence labeling result of the current test sample.
Further optimizing scheme, before extracting the estimated characteristics of the symbol sequence, a single-channel aliasing observation needs to be constructed:
assuming that a single antenna receives I spectrally overlapping digital communication signals simultaneously, the single-channel aliasing observation can be expressed as:
wherein s is i (n-m i ) Representing a transmission delay of m i The ith target signal of the sampling interval, a i V (N) represents AWGN for its corresponding channel gain, N being the total number of sampling points;
for a typical amplitude/phase modulated digital communication signal s i (n) is further expressed as
Wherein E is signal power;corresponds to the q-th of the transmission i Symbols, Q i Is the total number of symbols; g i (. Cndot.) represents an equivalent pulse shaping function, T, which comprehensively considers the pulse shaping at the transmitting end and the physical transmission channel s Is->Sampling interval and symbol oversampling rate respectively; />Is the residual carrier frequency.
Further optimizing scheme, parallel symbol sequence estimation based on single-channel observationThe feature FPSSE extraction includes: in order to construct a training sample set, P single-channel aliasing observations are constructed according to the step (1), and the modulation pattern of I digital communication signals contained in each aliasing observation is selected from a set omega with the size of C; for p (p.epsilon.1, P]) And (3) performing aliasing observation, and extracting I FPSSE subsequences from the aliasing observation, wherein each subsequence and a known symbol sequence of a corresponding signal can form a training sample. For example, a training sample constructed based on the ith target signal in the p-th aliasing observation may be expressed as:wherein->From [ L ] of signal i i ，Q i -L i ]And the coded symbols. After initial generation and extraction, all PI training samples in S are divided into C subsets according to their corresponding signal modulation patterns. Thus, there are a total of C training sample subsets { S } d |d∈[1，C]Each training sample subset is generated specifically for RNN training of a certain modulation pattern signal.
Figure 2 shows the RNN-based SIMO multi-signal symbol sequence estimator proposed by the present method. Here, single input refers to outputting multiple signal symbol sequence estimation results based on single channel aliasing observations only. The symbol sequence estimator is first trained offline before online testing. In offline training, C RNNs are respectively in { S } d |d∈[1，C]Training to complete sequence labeling task. In the on-line test, first, I FPSSE subsequences are extracted from the aliasing observation to be tested, and then are respectively sent into the trained RNNs corresponding to the modulation patterns, and finally, the outputs of the RNNs are respectively estimated symbol sequences of the I signals.
Further optimization, conventional symbol sequence estimators (e.g., coherent demodulators) typically perform symbol estimation directly based on decision making on the result of synchronous decimation of the down-converted received signal, where multiple communication signals are simultaneously included in the received signal, due to the signalsThe mutual interference will greatly increase the error rate. Therefore, when estimating the aliasing multiple signal symbol sequence, a new feature extraction mode needs to be designed, and the key point is to avoid mutual interference of each signal in the symbol sequence estimation process, and then the multiple signal symbol sequence estimation can be respectively completed in a sequence labeling mode. In formula (2), g i (. Cndot.) has an FIR structure, and is derived fromLasting to->Therefore, receive s i Q of (n) i The sampling point range of the influence of each symbol is as follows:
for the firstSymbol->The method comprises the following steps:
when (when)Time s i The first order differential sequence of (n) is expressed as:
assuming adequate samplingAnd g i (. Cndot.) is a symmetric function, where:
is->
Then equation (14) can be approximated as:
when (when)When d i (n) is
Formula (15) and (16) show that:
1)implicit in relation to->Is->Is a piece of information of (a). In general { s } can be said i (n)|i∈[1，I]Sign information of } is included in that it is located +.>In a first order differential sequence at the point.
2){s i (n)|i∈1，I]The first order differential sequence of is sparse in nature, i.e., there will be more significant values at only a small fraction of points. The positions of the significant value points are determined by the symbol oversampling rate of each target signal.The point is exactly the target signal s i The symbol switching time of (n) will be different from the symbol sub-sequence affecting the values of the two adjacent sampling points, so that the difference value will contain the information of the front and rear symbols (as shown in the formula (15)) and the amplitude will be larger when the difference operation is performed. And at the target signal s i The non-symbol switching time of (n) has the symbol subsequence affecting the value of the adjacent sampling point kept unchanged, so the differential value will be small (as shown in formula (16), the differential value at this time is basically only g i (. Cndot.) the change in value between adjacent samples is determined, and under conditions where the sample is sufficient, the change is weaker).
Synthesizing the formulas (1), (15) and (16) to obtain the first-order difference of x (n) as follows:
wherein phi is i Representing the set of the following sampling points:
from equation (17), signal s i The sign information of (n) is contained inWhile wherein interference from other target signals is effectively suppressed. Because of->And->All related, therefore->Is structured in nature. Therefore, this requirement is based on +.>Element pair signal s in (a) i The symbols of (n) are inferred taking into account the context information in the input sequence. Here, it is considered to use RNNs with a BLSTM structure to accomplish this labeling task.
Further optimizing scheme, since the oversampling rate of each target signal symbol may be different in practice and the transmission delay of each signal is random, it can be assumed thatWith such orthogonality, the differential sequence of the received signal may be actually decomposed into a plurality of differential subsequences corresponding to the respective target signals. Thus, extracting the FPSSE subsequence further comprises: for the target signal s i (n) its subsequence->From { d (n) |n ε [1, N]The extraction in } is as follows:
note that due toAnd has the formula (16), the other signals are for +.>The influence of (c) will be effectively suppressed.
According to the signalsAfter division, { d (n) |n ε [1, N]Will be divided into sub-sequence setsWherein each sub-sequence corresponds to a target signal.
Further optimization, it is desirable that the input of a symbol sequence estimator is related to the transmitted symbols only, in an ideal case. This is to avoid the influence of other symbol independent signals and channel parameters on the symbol estimation result. And as can be seen from equation (17), in addition to the transmission symbol,the amplitude and phase of (a) are still respectively subject to +.>A kind of electronic device with high-pressure air-conditioning systemIs a function of (a) and (b). Thus, extracting the FPSSE subsequence further comprises: for i.e. [1I ]]Further based on->The following normalized amplitude and phase sequences were calculated, respectively designated +.>
Wherein, I.I represents the amplitude of the complex sequence,representation->The mean value of (4), the angle (·) represents the complex phase, k i ∈[1，Q i -2L i +1]Will->As signal s in the received signal i (n) the corresponding FPSSE sequence.
In the further optimization scheme, in the offline training of the cyclic neural network, in order to demodulate signals with different possible modulation patterns, a cyclic neural network set is constructed, the structures and parameters of the cyclic neural networks in the set are the same, the training is performed for symbol sequence estimation of target signals with different modulation patterns, and each cyclic neural network comprises two BLSTM layers and one fully-connected classification layer. The activation function of the classification layer is softmax.
Further optimizing scheme, the output dimension of the classification layer is determined by the debugging order of the signal facing the classification layer.
Performance and test condition index
The SIMO multi-signal symbol sequence estimator based on RNN obtained by the method is tested under different conditions and compared with the existing multi-signal symbol sequence estimation algorithm based on single-channel observation.
The simulation experiment 1 tests the estimation performance of the symbol sequence estimator under different target signal power ratios, and compares the RNN structure selected in the method with other RNN network structures. The simulation experiment 2 compares the symbol sequence estimator provided by the method with the existing symbol sequence estimation algorithm based on model and data driving in two aspects of symbol sequence estimation precision and time complexity. Considering the importance of generalization capability to data-driven class methods, simulation experiment 3 tests the test performance of the sequence estimator proposed by the method under the condition that the sequence estimator is not found during training.
The Bit Error Rate (BER) of each target signal is used as an index. BER corresponding to the i-th target signal may be defined as follows:
wherein, the liquid crystal display device comprises a liquid crystal display device,the bit number inconsistent with the actual bit sequence after the estimated symbol sequence is converted into the corresponding bit sequence; b (B) i For the total length of the target signal bit sequence.
When estimating a certain target signal symbol sequence, the error bit is mainly derived from disturbance of other target signals and environmental noise, so that the performance of the symbol sequence estimator under different disturbance intensities is mainly examined in the simulation experiment of the method, and the signal-to-interference-and-noise ratio SINR is used as a measure of the disturbance intensity in the aliasing observation. It should be noted that the target signal and the interfering signal are relative concepts, and when a symbol sequence estimation is performed on a certain target signal, all other target signals are regarded as interference to the current signal, and vice versa. Thus, for target signal i (i e [1, i ]), its SINR in the aliasing observation is defined as:
simulation experiment
The basic training parameters of the symbol sequence estimator are set as follows: RNN is implemented based on the pyrench0.4.0 platform, and training/validation and test data is generated based on MatlabR2018a on Intel (R) Core (TM) i7-6500UCPU@2.50GHz processor. Training/verification/testing of a single RNN is done based on 4 x 105/1 x 105/4 x 106 modulation symbols, respectively. Different channel gains were simulated during the training sample generation, with SNR set at 25dB. The number of forward and backward nodes of each hidden layer is 128. The training optimizer and learning rate were set to classical Adam and 0.001, respectively. The training batch size was 100 and the training round number was 200.
Simulation experiment 1 the performance test and analysis of the symbol sequence estimator provided by the method under different target signal power ratios.
In the present simulation experiment, two target signals, BPSK and 2PAM signals, respectively, are considered. To examine the applicability of the proposed symbol sequence estimator to different scenarios of symbol rates of the target signals, the symbol rates of the two target signals were set to 0.112KB (Baud) and 0.1KB, respectively. Fig. 3 shows BER of the symbol sequence estimator of the present method after parallel symbol sequence estimation of two target signals at different power ratios. Wherein (a) BPSK (b) 2PAM. Snr=25 dB. L (L) 1 ＝L 2 =1. When calculating the SINR of a certain signal, another target signal is considered as interference. Also shown are BER results of a conventional coherent demodulator and a symbol sequence estimator constructed based on an original RNN (in the original RNN symbol sequence estimator, the BLSTM layer is replaced with an original RNN layer having the same number of neurons. In early experiments, a symbol sequence estimator based on a fully connected layer, i.e., a fully connected layer, was also constructed. It can be seen that the symbol sequence estimator provided by the method successfully performs parallel estimation on two target signal symbol sequences at different SINR in the case that coherent demodulation (i.e. conventional symbol sequence estimation flow) is almost completely disabled. Even when the interfering signal is significantly stronger than the current target signal, the symbol sequence estimator provided by the method still completes accurate estimation of the target signal symbol sequence (e.g., when SINR = -10dB, both target signal BER is below 10-2), at which point the conventional coherent demodulator has substantially failed. This benefits mainly from the FPSSE extraction performed. As can be seen from formulas (16) (17), inThe interference from other aliased digital communication signals has been substantially suppressed, making the estimation of the current target signal symbol sequence less susceptible to interference from other target signals. On the other hand, since the long-term and bi-directional context information in the input/output can be utilized, the results of the BLSTM network can be seen to be superior to those of the original RNN network. This also demonstrates the necessity of utilizing both bi-directional context information and long term context information in the sequence annotation problem modeled by the present method.
Simulation experiment 2: the symbol sequence estimator provided by the method is compared and analyzed with the performance of the existing symbol sequence estimation algorithm.
In the simulation experiment, the SIMO multi-signal symbol sequence estimator provided by the method is compared with the existing SIMO symbol sequence estimation algorithm based on model and data driving. In order to make a fair comparison, two aliased BPSK signals with the same symbol rate are considered here, subject to the limitations of the applicable range of the PSP algorithm. The relative delay between the two signals is 0.1 times the symbol duration. For the PSP algorithm, the step size of its LMS iteration is set to 0.005. The number of convolution kernels included in the two convolution layers of DCND is 128 and 64, respectively. Fig. 4 shows the test BER of different symbol sequence estimation algorithms (the two target signal results are shown separately). Wherein (a) bpsk#1, (b) bpsk#2. It can be seen that the symbol sequence estimator proposed by the present method significantly reduces BER compared to existing model-based and data-driven symbol sequence estimators, especially when there is a significant difference in the power of the two target signals. This is mainly because the inputs of existing symbol sequence estimation algorithms are indistinguishable for different target signals (joint demodulation in PSP based on symbol rate extraction results of aliasing observations and direct aliasing observations in DCND), i.e. symbol sequence estimation is performed for different target signals based on the same inputs, which makes interference between different target signals unavoidable. In contrast, in the symbol sequence estimator provided by the method, respective FPSSE sub-sequences are extracted for different target signals respectively, and the sparsity of the FPSSE sub-sequences determines that the mutual interference of the extracted target signals is weak, so that the symbol sequence estimator can be well adapted to scenes with significant power differences of the target signals. Meanwhile, the result also verifies that the symbol sequence estimator provided by the method has good adaptability to the scene when the symbol rates of the target signals are the same, so that the symbol sequence estimator is a practical algorithm in this aspect compared with the PSP. On the other hand, table 1 lists the number of trainable parameters and the average required run time in the symbol sequence estimator and the existing symbol sequence estimation algorithm proposed by the present method (the number of trainable parameters is only for the data driven class method). It can be seen that under the test condition of the simulation experiment, the time complexity of the estimator provided by the method is similar to that of the PSP algorithm. However, it should be noted that the current simulation experiment is a test performed under the condition that the number of target signals is only 2 and the modulation orders of the target signals are all 2. The estimator adopts a parallel estimation mode, namely, symbol sequence estimation is synchronously and independently carried out on a plurality of target signals, and the calculation complexity of the estimator does not increase exponentially with the modulation order of the target signals. In contrast, if the number of target signals increases or the modulation order of the target signals increases, the time complexity of the PSP algorithm increases exponentially, and it is expected that the average operating time increases significantly.
TABLE 1 different SIMO symbol sequence estimation algorithms trainable parameter numbers and average estimation times (every 1000 symbols)
Simulation experiment 3: the generalization capability test and analysis of the symbol sequence estimator provided by the method.
Generalization capability is always a more important issue for data driven class approaches. Because the training process cannot traverse all possible data, whether the training process has the capability of coping with the data which is not found during the training is one of the keys for determining whether the data-driven method has practicability. In the simulation experiment, the performance of the symbol sequence estimator provided by the method under the generalization condition is tested. Here, the test under the generalization condition mainly means that the signal/channel parameters and the like employed in generating the test data do not appear in the training data generation. In the first case in fig. 5, the test SNR is reduced from 25dB to 15dB. In the second case, the relative delay of the two target signals is set to be randomly generated between (0.1-0.9) times the symbol duration, and then the average BER is calculated over a number of tests. It can be seen that although the symbol sequence estimator proposed by the method is trained offline only under certain conditions, it can still cope well with generalized test conditions that are not seen in training. Even though the SNR decreases by up to 10dB, the resulting BER increase is still within acceptable limits. It should also be noted that the performance of the symbol sequence estimator proposed by the present method at lower test SNR will predictably be improved if different SNRs are considered during training. This can be achieved simply by generating training data over a wider range of SNR.
Finally, it should be noted that: the above examples are only specific embodiments of the present application, and are not intended to limit the scope of the present application, but it should be understood by those skilled in the art that the present application is not limited thereto, and that the present application is described in detail with reference to the foregoing examples: any person skilled in the art may modify or easily conceive of the technical solution described in the foregoing embodiments, or perform equivalent substitution of some of the technical features, while remaining within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the corresponding technical solutions. Are intended to be encompassed within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.
Claims (5)
1. A method for directly estimating a symbol sequence of a multi-communication signal under an interference condition, comprising:
extracting parallel symbol sequence estimation characteristics FPSSE based on single-channel observation to obtain characteristic sequences respectively containing symbol information of each target signal;
performing sequence labeling based on the offline trained cyclic neural network to obtain a symbol sequence estimation result, wherein the characteristic sequence is used as input when the offline training is performed on the cyclic neural network;
the extracting of the parallel symbol sequence estimation feature FPSSE based on single-channel observation comprises the following steps:
constructing P single-channel aliasing observations: assuming that a single antenna receives I spectrally overlapping digital communication signals simultaneously, the single-channel aliasing observation is expressed as:
wherein s is i (n-m i ) Representing a transmission delay of m i The ith target signal of the sampling interval, a i V (N) represents AWGN for its corresponding channel gain, N being the total number of sampling points;
for a typical amplitude/phase modulated digital communication signal s i (n) is further expressed as
Wherein E is i Is the signal power;corresponds to the q-th of the transmission i Symbols, Q i Is the total number of symbols; g i (. Cndot.) represents an equivalent pulse shaping function, T, which comprehensively considers the pulse shaping at the transmitting end and the physical transmission channel s Is->Sampling interval and symbol oversampling rate respectively; />Is the residual carrier frequency;
the modulation pattern of the I digital communication signals contained in each aliasing observation is selected from a set omega with the size of C;
for p (p.epsilon.1, P]) Performing aliasing observation, namely extracting 1 FPSSE subsequence from the aliasing observation, wherein each subsequence and a known symbol sequence of a corresponding signal form a training sample; wherein extracting the FPSSE subsequence comprises: let g i (. Cndot.) has finite length unit impulse response FIR structure and lasts 2L i -1 symbol, and fromLasting to->Therefore, receive s i Q of (n) i The sampling point range of the influence of each symbol is as follows: />
For the firstSymbol->Has the following components
Wherein E is i For the signal power to be high,corresponds to the q-th of the transmission i The symbols g i (. Cndot.) represents an equivalent pulse shaping function, T, which comprehensively considers the pulse shaping at the transmitting end and the physical transmission channel s Is->Sampling interval and symbol oversampling rate respectively;
when (when)Time s i The first order differential sequence of (n) is expressed as:
wherein Q is i Is the total number of symbols;
assuming adequate samplingAnd g i (. Cndot.) is a symmetric function, where:
is->
Then equation (14) is approximated as:
when (when)When d i (n) is:
wherein n represents time, a i Representation transmissionTime delay of m i The ith target signal of the sampling interval corresponds to the channel gain,is the residual carrier frequency;
the first order difference d (n) of x (n) is obtained by combining equations (1) (15) and (16):
wherein phi is i Representing the set of the following sampling points:
2. the method of claim 1, wherein extracting the FPSSE subsequence further comprises: for the target signal s i (n) subsequences thereofFrom { d (n) |n ε [1, N]The extraction in } is as follows:
after division according to each signal, { d (n) |n ε [1, N ]]Will be divided into sub-sequence setsWherein each sub-sequence corresponds to a target signal.
3. The method of claim 2, wherein extracting the FPSSE subsequence further comprises:
for i.epsilon.1, I]Further based onThe following normalized amplitude and phase sequences were calculated, respectively designated +.>
Wherein, the | represents the amplitude of the complex sequence,representation->The mean value of (4), the angle (·) represents the complex phase, k i ∈[1，Q i -2L i +1]Will->As signal s in the received signal i (n) the corresponding FPSSE sequence.
4. The method of claim 1, wherein in the off-line training of the recurrent neural network, in order to demodulate signals of different modulation patterns that may exist, a set of recurrent neural networks is constructed, and each recurrent neural network in the set has the same structure and parameters, and is trained for symbol sequence estimation of target signals of different modulation patterns, and each recurrent neural network includes two BLSTM layers and one fully connected classification layer.
5. The method of claim 4, wherein the output dimension of the classification layer is determined by the debug order of the signal to which it is directed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377436.3A CN115834306B (en) | 2022-11-04 | 2022-11-04 | Method for directly estimating symbol sequence of multiple communication signals under interference condition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377436.3A CN115834306B (en) | 2022-11-04 | 2022-11-04 | Method for directly estimating symbol sequence of multiple communication signals under interference condition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115834306A CN115834306A (en) | 2023-03-21 |
| CN115834306B true CN115834306B (en) | 2023-09-05 |
Family
ID=85526656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211377436.3A Active CN115834306B (en) | 2022-11-04 | 2022-11-04 | Method for directly estimating symbol sequence of multiple communication signals under interference condition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115834306B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101394388A (en) * | 2008-10-24 | 2009-03-25 | 武汉大学 | Demodulating and modulation processing method for radio digital domain |
| CN106850470A (en) * | 2017-02-17 | 2017-06-13 | 中国人民解放军海军工程大学 | A kind of Interference Cancellation based on affine precoding and the channel estimation methods of two-way cooperation |
| CN108206799A (en) * | 2016-12-16 | 2018-06-26 | 中国航天科工飞航技术研究院 | A kind of carrier frequency bias estimation for burst communication |
| CN110048982A (en) * | 2019-04-04 | 2019-07-23 | 中国人民解放军战略支援部队信息工程大学 | The blind separating method and device of the paired carrier multiple access signal of distinct symbols rate |
| CN113259283A (en) * | 2021-05-13 | 2021-08-13 | 侯小琪 | Single-channel time-frequency aliasing signal blind separation method based on recurrent neural network |
| CN114464206A (en) * | 2022-04-11 | 2022-05-10 | 中国人民解放军空军预警学院 | Single-channel blind source separation method and system |
-
2022
- 2022-11-04 CN CN202211377436.3A patent/CN115834306B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101394388A (en) * | 2008-10-24 | 2009-03-25 | 武汉大学 | Demodulating and modulation processing method for radio digital domain |
| CN108206799A (en) * | 2016-12-16 | 2018-06-26 | 中国航天科工飞航技术研究院 | A kind of carrier frequency bias estimation for burst communication |
| CN106850470A (en) * | 2017-02-17 | 2017-06-13 | 中国人民解放军海军工程大学 | A kind of Interference Cancellation based on affine precoding and the channel estimation methods of two-way cooperation |
| CN110048982A (en) * | 2019-04-04 | 2019-07-23 | 中国人民解放军战略支援部队信息工程大学 | The blind separating method and device of the paired carrier multiple access signal of distinct symbols rate |
| CN113259283A (en) * | 2021-05-13 | 2021-08-13 | 侯小琪 | Single-channel time-frequency aliasing signal blind separation method based on recurrent neural network |
| CN114464206A (en) * | 2022-04-11 | 2022-05-10 | 中国人民解放军空军预警学院 | Single-channel blind source separation method and system |
Non-Patent Citations (1)
| Title |
|---|
| Xin Cai ; Xiang Wang ; Zhitao Huang ; Fenghua Wang.Single-Channel Blind Source Separation of Communication Signals Using Pseudo-MIMO Observations.IEEE Communications Letters.2018,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115834306A (en) | 2023-03-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Reinforcement learning with perturbed rewards | |
| CN109412725B (en) | Radio communication PCMA signal blind demodulation method and device | |
| Bouboulis et al. | The augmented complex kernel LMS | |
| Zhang et al. | The kernel conjugate gradient algorithms | |
| US8295341B2 (en) | Cognitive radio device | |
| CN112115821B (en) | Multi-signal intelligent modulation mode identification method based on wavelet approximate coefficient entropy | |
| CN111884976B (en) | Channel interpolation method based on neural network | |
| CN112347871B (en) | Interference signal modulation identification method for communication carrier monitoring system | |
| Shlezinger et al. | Model-based machine learning for communications | |
| WO2021088465A1 (en) | Fast modulation recognition method using multilayer perceptron, and employing data fusion of multiple distribution tests | |
| CN111431565B (en) | Optical communication MIMO detection method and system | |
| Rahman et al. | Deep learning-based improved cascaded channel estimation and signal detection for reconfigurable intelligent surfaces-assisted MU-MISO systems | |
| Zheng et al. | Deep learning based trainable approximate message passing for massive MIMO detection | |
| CN112422208B (en) | Signal detection method based on antagonistic learning under unknown channel model | |
| CN115834306B (en) | Method for directly estimating symbol sequence of multiple communication signals under interference condition | |
| Wang et al. | Long short-term memory neural equalizer | |
| CN114614920B (en) | Signal detection method based on data and model combined driving of learning factor graph | |
| CN114938232B (en) | LSTM-based simultaneous co-frequency full-duplex digital domain self-interference suppression method | |
| Yarotsky et al. | Machine learning-assisted channel estimation in massive MIMO receiver | |
| CN112821971A (en) | Time-varying channel signal detection method based on countermeasure learning | |
| Brooks et al. | Multichannel adaptive blind deconvolution using the complex cepstrum of higher order cross-spectra | |
| CN115801509B (en) | Method for directly estimating single communication signal symbol sequence under interference condition | |
| Yellepeddi et al. | Efficient system tracking with decomposable graph-structured inputs and application to adaptive equalization with cyclostationary inputs | |
| CN114662523A (en) | Spread spectrum signal detection method integrating cognitive enhanced compressed sensing and artificial neural network | |
| Nguyen et al. | Blind equalization of dispersive fast fading Ricean channels via the EMV algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |