CN111970218A - Method for carrying out communication automatic modulation recognition based on deep multi-hop network - Google Patents

Method for carrying out communication automatic modulation recognition based on deep multi-hop network Download PDF

Info

Publication number
CN111970218A
CN111970218A CN202010878480.7A CN202010878480A CN111970218A CN 111970218 A CN111970218 A CN 111970218A CN 202010878480 A CN202010878480 A CN 202010878480A CN 111970218 A CN111970218 A CN 111970218A
Authority
CN
China
Prior art keywords
network
layer
signal
hop
deep
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.)
Granted
Application number
CN202010878480.7A
Other languages
Chinese (zh)
Other versions
CN111970218B (en
Inventor
王岩
肖辉春
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Taishan University
Original Assignee
Taishan University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Taishan University filed Critical Taishan University
Priority to CN202010878480.7A priority Critical patent/CN111970218B/en
Publication of CN111970218A publication Critical patent/CN111970218A/en
Application granted granted Critical
Publication of CN111970218B publication Critical patent/CN111970218B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00—Modulated-carrier systems
    • H04L27/0012—Modulated-carrier systems arrangements for identifying the type of modulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention discloses a method for carrying out communication automatic modulation recognition based on a deep multi-hop network, which comprises the following steps of firstly receiving communication signal data; then, the collected signals are received and processed; then inputting the processed signal into a deep multi-hop network for modulation mode identification; and finally, outputting the recognition result. The deep multi-hop network provided by the invention can be used for realizing the extraction and analysis of variable input signal length and key identification threshold values, and accurately classifying various modulation modes by more effectively transmitting the learned underwater acoustic signal characteristics among multiple layers. Compared with the traditional deep neural network, the method can overcome the influence of a wireless channel and reasonably identify various commonly used communication modulation modes.

Description

Method for carrying out communication automatic modulation recognition based on deep multi-hop network
Technical Field
The invention relates to the technical field of neural networks, in particular to a deep neural network learning method based on signal modulation type identification of communication.
Background
Automatic modulation and identification (AMC) is a research hotspot in intelligent communication systems, and plays an important role in both military and civil fields. In the military field, modulation identification of communication signals is a prerequisite for hindering enemy communications, and by transmitting higher power similar modulation signals in the same frequency band, the communication process can be effectively disturbed, which requires the modulation classifier to detect the correct modulation class to generate the jamming signal. In the civil field, the link adaptive system can obtain the channel quality by measuring the wireless transmission environment, and select a proper modulation mode to obtain better communication efficiency, and under the condition that the radio frequency spectrum resources are in shortage, the link adaptive system further improves the capacity of the communication system by researching an innovative method of AMC.
Wireless channels are typically subject to severe interference, such as multipath effects, doppler shift, time delay, and noise. These interferences can have a significant impact on the communication data transmission. AMC is a challenging task given the particularities of the wireless environment. A typical AMC generally comprises two steps: pre-processing the signal and classifying the processed signal, and the first step is the core of AMC. Currently, there are two main methods of modulation identification: Likelihood-Based (LB) and Feature-Based (FB) methods; unlike LB and FB techniques, which require manual decision thresholds, Deep Learning Methods (DLM) can adaptively search for and determine thresholds. Conventional AMC algorithms require real-time calculations to produce suitable results, not only are the computational complexity higher, but also the latency is longer, which is very disadvantageous for configuring the corresponding algorithms in practical communication systems. DLM can effectively solve these problems. Although training the network requires a significant amount of time, the trained network can perform recognition tasks in near real-time, similar to the validation process of a trained deep learning model.
At present, a deep learning method makes certain progress in the application of signal modulation mode identification, but further research and exploration are still needed.
Disclosure of Invention
The invention aims to provide a method for identifying communication automatic modulation based on a deep multi-hop network so as to make up for the defects of the prior art.
Since the important signal features for distinguishing the modulation modes are interfered by channel factors to cause the key distinguishing information to be covered, when a plurality of modulation features of a signal data set are learned, a deeper neural network easily causes difficulty in training, which is more obvious in CNN and RNN structures, which is mainly caused by network degradation caused by the deeper network structure; however, a shallow network structure cannot extract the basic features hidden in the signal data set comprehensively, so that various modulation modes cannot be classified effectively.
In order to solve the problems and achieve the purposes, the invention adopts the following specific technical scheme:
a method for carrying out communication automatic modulation recognition based on a deep multi-hop network comprises the following steps:
s1: receiving communication signal data;
s2: performing reception processing on the signal collected in S1;
s3: inputting the signal processed by the S2 into a deep multi-hop network for modulation mode identification;
s4: and outputting the recognition result.
Further, in said S1, the received signal is assumed to be separated into the lth multipath component, which is related to τ in the time interval, and arrives at the receiver, and is denoted as
Figure BDA0002653363560000021
Where s (t) denotes the transmitted signal, cl(τ; t) represents the ith path of the equivalent channel impulse response, L ranges from 1 to L, n (t) represents additive noise, and r (t) represents the received signal;
the form of the channel impulse response in equation (9) is written as
cl(τ；t)＝βl(t)(t-τl(t)) (10)
Wherein, betal(t) represents the possible time-varying attenuation on the l-th multipath propagation path, (t- τ)l(t)) is the time t-taul(t) impulse response, τl(t) is the l-th multipath delay, all paths having the same Doppler scaling factor γ, τl(t)≈τl-γt；
Substituting equation (10) into equation (9) can represent the received signal as
Figure BDA0002653363560000022
Further, in S2, in order to fully utilize the classification information of the signal to improve the network identification capability, the signal sequence may be divided into several lengths, each length is input into the deep multi-hop network, which is beneficial to obtain the hidden features of the signal modulation information, and the signal patterns of various lengths are as follows
oq(k)＝c(q+θk) (7)
Where the signal sequence has the total length of M, q is the starting position of the signal sequence, q is 1, 2, …, M-1, representing the maximum fixed length K of the signal sequence, 1, 2, …, K, where K is the regular fixed signal length, θ is the percentage of the selected variable length, c (·) is a function of the selected length that fits the network input, the network receiving oq(k) As an input pattern for modulation classification.
Further, the deep multi-hop network in S3 sequentially includes an input layer, a preprocessing layer, a multi-hop layer, and an output layer.
Further, the input layer represents a wireless communication signal data set, including various signal modulation modes; said Pre-processing layer (Pre-process) comprises Conv, which stands for convolutional network layer, and BN, which stands for batch normalization, the main purpose of BN being to normalize the distribution of signal data during each forward and backward propagation, BN ensuring that the values in the training process have the same order, thus ensuring a more stable signal feature extraction; the Multi-hop layer (Multi-hop) includes a ReLU (linear correction unit) representing a non-linear activation function layer, Conv, Dropout representing an operation to improve the network generalization capability, and Concatenate (in series) representing the aggregation of multiple short connections, which are temporarily discarded according to a limiting probability to prevent the over-fitting phenomenon.
Furthermore, the multi-hop layer adopts a basic depth identity mapping network structure, and then different layers in the structure are reduced or expanded to obtain deeper signal classification representation, so that the identification effect is further improved.
In addition, although the identity mapping network structure can improve the identification performance by deepening the network structure, when the depth reaches the layer number, the effect cannot be further improved, and by adding direct connection among layers in the network, the transmitted learning value can be gradually optimized, so that more high-dimensional signal classification features are collected.
The direct connection is a network structure formed by linking different layers inside, and the multi-hop connection rate is a mode of controlling the number of connections among multiple layers so as to improve the transmission efficiency of signal characteristics learned by a deep network.
As mentioned above, the multi-hop layer uses an identical mapping form of cross-layer connection, which is equivalent to a very smooth way to realize the transmission process of parameters during the modulation class learning period, αi+1And the previous layer of alphaiThe relationship between them is a purely linear superposition relationship, expressed as
αi+1＝αi+G(αi) (1)
Where G (-) is a direct access function that passes learned information between deep network layers.
In addition, if the output expression of the subsequent layer is further derived, as shown below
αi+2＝αi+1+G(αi+1) (2)
αi+2＝αi+G(αi)+G(αi+1) (3)
…
Figure BDA0002653363560000041
The vector of any subsequent layer will have a portion of the linear contribution from the previous layer, J is the number of layers after the ith layer, J is the total number of layers from the ith layer to the ith layer, and αIIs the output of layer I.
Moreover, the layer output expression of equation (4) indicates that the inverse identity mapping propagation is also a similar smoothing process, and can be expressed as identity mapping definition in equation (2) above, corresponding to the identity mapping definition in equation (2)
Figure BDA0002653363560000042
αhaRepresenting an ideal vector value, αIRepresenting real vector values, the chain rule here can be expressed directly as
Figure BDA0002653363560000043
Output alphaIAn identity mapping between any network layers is generated and can be passed back to alpha at any previous layeriThe transfer process is very fast and direct, along with the superposition of layers, the learned weight does not have too much transmission loss, and the formula (6) is linear superposition operation instead of multiplication operation; thus, the estimated gradients are properly composed by the neurons in each layer and do not degrade the parameter delivery efficiency of learning due to transmission between deep network layers, which explains why the network structure that allows identity mapping is so deep.
Further, in the multi-hop layers, each layer takes as input the output of some previous layer, where the total number of cross-layer connections is defined as U, the learned signal characteristics have some redundancy with dense connections between layers, and the learned collective knowledge is shared between network layers of the same densely connected block to fully utilize redundant signal information, formulated as
Figure BDA0002653363560000044
Where ρ is the multi-hop connectivity rate, indicating how many short connections are in the dense connection layer, λ represents the total number of interconnections in the internal layer of the multi-hop network, duRepresenting a selected connectivity layer within the internal layers of the multi-hop network. With the deepening of the network structure, the recognition effect can be effectively obtainedAnd (4) improving. In addition, dense short connections also have a regularizing effect, which has a suppressing effect on network performance degradation. Moreover, the network bandwidth is narrower and the learned weight transfer flows more reasonably between cross-layer connections. The number of feature maps is small in the output of each layer. Other network forms such as ResNet have hundreds of output layers, which can burden the training network. The dense connection form enables gradient transmission to be more effective, and enables the network to mine more deep signal features, thereby improving the classification effect.
Furthermore, a receiving domain receptive field range is also arranged in the multi-hop layer, which determines how many extracted features are used for final classification and identification in each layer, which is helpful for the network to obtain more modulation classification features and enhance the final identification effect.
Furthermore, the receiving domain receptive field range can adjust how much new signal information in each layer contributes to the global classification state, when the network is deep, the receiving domain receptive field range is connected across layers instead of summing, which can generate a large amount of input signal characteristic information, and the signal information is transmitted through multi-hop connection, thereby providing more identification information for the deep multi-hop network, and the size of the receiving receptive field is enlarged at each layer due to the collection of the signal characteristic map; if the deep multi-hop network deepens each time to generate phi feature maps, the value of the reception receptive field in the V-th layer can be expressed as
Figure BDA0002653363560000051
Wherein phi is0Is the value of the initial reception field, phivIs the value of the varying receptive field corresponding to the V-th layer, V is 1, 2, …, V, which means that each composite layer of the deep multihop network generates a Φ feature map, which is treated as the extracted information of the network layer, and each layer can access its previous signal feature map and obtain more collective modulation classification knowledge.
The deep multi-hop network is applied to communication modulation mode identification.
Furthermore, the deep multi-hop network is especially used for underwater communication modulation mode identification.
The invention has the advantages and beneficial effects that:
the deep multi-hop network provided by the invention can be used for realizing the extraction and analysis of variable input signal length and key recognition threshold values, and accurately classifying various modulation modes by more effectively transmitting the learned underwater acoustic signal characteristics among multiple layers, belongs to a moderately sparse network structure, and has higher training speed and shorter training time. Compared with the traditional deep neural network, the deep multi-hop network can overcome the influence of a wireless channel and reasonably identify various common communication modulation modes.
Drawings
Fig. 1 is a schematic diagram of an identity mapping network element according to embodiment 1;
fig. 2 is a schematic structural diagram of a depth identity mapping network according to embodiment 1;
fig. 3 is a schematic diagram of the general structure of a deep multihop network according to embodiment 1;
fig. 4 is a schematic internal structure diagram of a deep multihop network according to embodiment 1;
fig. 5 is a schematic diagram illustrating comparison of the recognition performance of the deep multi-hop network according to embodiment 1 under different signal sequence lengths;
fig. 6 is a schematic diagram illustrating comparison of the recognition performance of the deep multi-hop network according to embodiment 1 under different extended receptive field ranges;
fig. 7 is a schematic diagram illustrating comparison of the identification performance of the deep multi-hop network according to the method for comparing network identification and connection rate mapping according to different identity maps in embodiment 1.
Detailed Description
The invention will be further explained and illustrated by means of specific embodiments and with reference to the drawings.
Example 1:
a method for carrying out communication automatic modulation recognition based on a deep multi-hop network comprises the following steps:
s1: receiving communication signal data;
s2: performing reception processing on the signal collected in S1;
s3: inputting the signal processed by the S2 into a deep multi-hop network for modulation mode identification;
s4: and outputting the recognition result.
Further, in said S1, the received signal is assumed to be separated into the lth multipath component, which is related to τ in the time interval, and arrives at the receiver, and is denoted as
Figure BDA0002653363560000061
Where s (t) denotes the transmitted signal, cl(τ; t) represents the ith path of the equivalent channel impulse response, L ranges from 1 to L, n (t) represents additive noise, and r (t) represents the received signal;
the form of the channel impulse response in equation (9) is written as
cl(τ；t)＝βl(t)(t-τl(t)) (10)
Wherein, betal(t) represents the possible time-varying attenuation on the l-th multipath propagation path, (t- τ)l(t)) is the time t-taul(t) impulse response, τl(t) is the l-th multipath delay, all paths having the same Doppler scaling factor γ, τl(t)≈τl-γt；
Substituting equation (10) into equation (9) can represent the received signal as
Figure BDA0002653363560000071
Further, in S2, in order to fully utilize the classification information of the signal to improve the network identification capability, the signal sequence may be divided into several lengths, each length is input into the deep multi-hop network, which is beneficial to obtain the hidden features of the signal modulation information, and the signal patterns of various lengths are as follows
oq(k)＝c(q+θk) (7)
Where the signal sequence has the total length of M, q is the starting position of the signal sequence, q is 1, 2, …, M-1, representing the maximum fixed length K of the signal sequence, 1, 2, …, K, where K is the regular fixed signal length, θ is the percentage of the selected variable length, c (·) is a function of the selected length that fits the network input, the network receiving oq(k) As an input pattern for modulation classification.
Further, the deep multi-hop network in S3 includes an input layer, a preprocessing layer, a multi-hop layer, and an output layer in sequence, as shown in fig. 3.
Further, the input layer represents a wireless communication signal data set, including various signal modulation modes; said Pre-processing layer (Pre-process) comprises Conv, which stands for convolutional network layer, and BN, which stands for batch normalization, the main purpose of BN being to normalize the distribution of signal data during each forward and backward propagation, BN ensuring that the values in the training process have the same order, thus ensuring a more stable signal feature extraction; the multi-hop layer (as shown in fig. 4) includes ReLU (linear correction unit) representing a non-linear activation function layer, Conv, Dropout representing an operation to improve the network generalization capability, and Concatenate (in series) representing aggregation of multiple short connections, which are temporarily discarded according to a limiting probability to prevent an over-fitting phenomenon.
The multi-hop layer adopts a basic depth identity mapping network structure and is obtained by reducing or expanding different layers in the structure. Experiments show that the classification precision of the network is continuously improved along with the continuous increase of the structure depth. When the number of layers reaches a certain amount, the recognition accuracy is rapidly reduced and the network becomes more difficult to train. The main reason is the network degradation problem, and the phenomenon can be relieved by establishing an identity mapping in the network. When the deeper network structure reaches the saturation of extracted data features, the connection is realized by adding some identity mapping, namely the form of network front layer output equal to rear layer input. By the method, the network structure can be further deepened, so that a higher-level data form certificate is extracted to obtain a better classification result. Meanwhile, the identification performance is not seriously reduced in the verification process. A single network element consisting of a typical identity mapping network is shown in fig. 1, which directly passes the output of the previous layer to the next layer in order to pass the learned data features, which are essential for improving the classification accuracy, inside the network.
In addition, although the identity mapping network structure (as shown in fig. 2) can improve the recognition performance by deepening the network structure, when the depth reaches the number of layers, the effect cannot be further improved, and by adding direct connection between the layers inside the network, the learning value of transmission can be gradually optimized, so that more high-dimensional signal classification features are collected.
The direct connection is a network structure formed by linking different layers inside, and the multi-hop connection rate is a mode of controlling the number of connections among multiple layers so as to improve the transmission efficiency of signal characteristics learned by a deep network.
As mentioned above, the multi-hop layer adopts an identical mapping form of cross-layer connection, which is equivalent to a very smooth way to realize the transmission process of the parameters during the modulation type learning, and the relationship between α i +1 and the previous layer α i is a pure linear superposition relationship (as shown in fig. 1), which is expressed as a pure linear superposition relationship (as shown in fig. 1)
αi+1＝αi+G(αi) (1)
Where G (-) is a direct access function that passes learned information between deep network layers.
In addition, if the output expression of the subsequent layer is further derived, as shown below
αi+2＝αi+1+G(αi+1) (2)
αi+2＝αi+G(αi)+G(αi+1) (3)
…
Figure BDA0002653363560000081
Any subsequent layer's vectors will have a portion fromLinear contribution of previous layers, J is the number of layers after the ith layer, J is the total number of layers from the ith to the ith layer, and αIIs the output of layer I.
Moreover, the layer output expression of equation (4) indicates that the inverse identity mapping propagation is also a similar smoothing process, and can be expressed as identity mapping definition in equation (2) above, corresponding to the identity mapping definition in equation (2)
Figure BDA0002653363560000082
αhaRepresenting an ideal vector value, αIRepresenting real vector values, the chain rule here can be expressed directly as
Figure BDA0002653363560000083
Figure BDA0002653363560000091
Output alphaIAn identity mapping between any network layers is generated and can be passed back to alpha at any previous layeriThe transfer process is very fast and direct, along with the superposition of layers, the learned weight does not have too much transmission loss, and the formula (6) is linear superposition operation instead of multiplication operation; thus, the estimated gradients are properly composed by the neurons in each layer and do not degrade the parameter delivery efficiency of learning due to transmission between deep network layers, which explains why the network structure that allows identity mapping is so deep.
Further, in the multi-hop layers, each layer takes as input the output of some previous layer, where the total number of cross-layer connections is defined as U, the learned signal characteristics have some redundancy with dense connections between layers, and the learned collective knowledge is shared between network layers of the same densely connected block to fully utilize redundant signal information, formulated as
Figure BDA0002653363560000092
Where ρ is the multi-hop connection rate, indicating how many short connections are in the dense connection layer. With the deepening of the network structure, the recognition effect can be effectively improved actually. In addition, dense short connections also have a regularizing effect, which has a suppressing effect on network performance degradation. Moreover, the network bandwidth is narrower and the learned weight transfer flows more reasonably between cross-layer connections. The number of feature maps is small in the output of each layer. Other network forms such as ResNet have hundreds of output layers, which can burden the training network. The dense connection form enables gradient transmission to be more effective, and enables the network to mine more deep signal features, thereby improving the classification effect.
Furthermore, a receiving domain receptive field range is also arranged in the multi-hop layer, which determines how many extracted features are used for final classification and identification in each layer, which is helpful for the network to obtain more modulation classification features and enhance the final identification effect.
Furthermore, the receiving domain receptive field range can adjust how much new signal information in each layer contributes to the global classification state, when the network is deep, the receiving domain receptive field range is connected across layers instead of summing, which can generate a large amount of input signal characteristic information, and the signal information is transmitted through multi-hop connection, thereby providing more identification information for the deep multi-hop network, and the size of the receiving receptive field is enlarged at each layer due to the collection of the signal characteristic map; if the deep multi-hop network deepens each time to generate phi feature maps, the value of the reception receptive field in the V-th layer can be expressed as
Figure BDA0002653363560000101
Wherein phi is0Is the value of the initial reception field, phivIs the value of the varying receptive field corresponding to the v-th layer, v being 1, 2…, V, which means that each composite layer of a deep multi-hop network generates a Φ signature, treating the signature as an extracted information for the network layers, each layer can access its previous signal signature and obtain more collective modulation classification knowledge.
Example 2:
this example presents a specific validation and comparative test of the method presented in example 1.
In the verification experiment, the method of setting the signal data set from the public signal data and the true communication state is as follows: the modulated signal data set consists of ten Modulation patterns, respectively Single-side band (SSB), Frequency Modulation (FM). Constant Phase-Shift Keying (CPFSK), Gaussian FSK (Gaussian FSK, GFSK), Pulse Amplitude Modulation (PAM), 16Quadrature Amplitude Modulation (16 QAM), 64QAM, Binary Phase-Shift Keying (BPSK), Quadrature PSK (Quadrature PSK, QPSK), and 8 PSK. The signal length is divided into 32, 64, 128 and 256. The temporal fading model uses rayleigh distribution. The number of training and verification vectors of each modulation mode is 10000, and the range of Signal-to-Noise Ratio (SNR) is-20 dB to 20 dB. The added noise is assumed to be zero-mean white gaussian noise with limited bandwidth. The noise varies with SNR, with a standard deviation of 10-SNR/10Is calculated by the formula (2). Maximum carrier frequency offset of 0.5-3Hz, the random number generator seed of the noise source is set to 4919. In the frequency selective fading simulation, the number of sine waves is 8, and the range profile of the signal filter for the pluggable time delay is 10. The roll-off coefficient of the raised cosine pulse shaping filter is 0.35.
During the training verification process, an appropriate gradient descent optimizer must be selected to ensure that the network can eventually converge. The wireless communication signal has a number of local minima in the data set, which minima are derived from the channel characteristics. Adam can solve the fusion problem during training. The network completes the training work and needs to be verified by a verification data set, so that a standard for measuring the training effect of the network needs to be provided. The cross entropy function may be used as a measure of the criterion. The number of layers in the multihop network is set to 88 in view of harvesting more signal features and training efficiency.
The verification results are as follows:
in fig. 5, the case of signal modulation identification of different lengths is discussed. The input length is from 32 to 256 and the recognition rate increases as the length increases. The proposed network can adapt to changes in length to obtain efficient results. In the range of-20 dB to 4dB, the length of 256 is respectively 14%, 8% and 5% higher than the lengths of 32, 64 and 128, and the increase of the signal length is beneficial to the improvement of the recognition rate; the main reason is that longer signal lengths provide richer signal characteristics for multi-hop networks. When the SNR is more than 4dB, the signal sequence with the length of 128 has better identification effect than 256, the average improvement is about 1.5%, which mainly comes from the signal sequence with the length of 128, and for a multi-hop network, the learning of the signal modulation classification characteristic is more beneficial at high SNR.
Fig. 6 shows the recognition rate of the deep multi-hop network in different receiving domain extended receptive field ranges. From-20 dB to-10 dB, the six extended receptive field ranges have similar recognition performance, while the 10 extended range is 2% less than the other ranges. With increasing SNR, the extended receptive field range of 80 increased by about 3% on average over the extended receptive field ranges of 10, 20, and 40, in the range of-10 dB to 0 dB. Under the condition of higher signal-to-noise ratio, the identification rate is continuously improved along with the expansion of the receiving receptive field range. The extended receptive field range of 80 is approximately 6%, 5% and 3% higher than the extended receptive field ranges of 10, 20 and 40, respectively, and also approximately 2% higher than the extended ranges of 60 and 100. This indicates that the improvement in recognition effectiveness is correlated with the extended range of the receptive field observed by the multihop network. This is mainly due to the larger field of view, and the wider signal modulation information available for recognition. This also indicates that when the extension range reaches a certain range, the effective modulation classification information provided to the multi-hop network does not continue to increase, and the recognition performance does not improve further.
In fig. 7, the deep multi-hop network has an identification performance at different connection rates (connection rates), and is also compared with evolgdnn, resnet 50, respompenet, and ressignnet at the same time. Resnet takes the form of 50 layers. Deeper network structures are more prone to overfitting phenomena, so resinet is represented in layer 34, while ressignnet uses relu and learnelu as active activation functions. At SNR ≦ 5dB, the recognition gap was less than 1% for 4 changes in ligation rates. The connection rate 50% has the highest recognition effect when the signal-to-noise ratio is increased, and SNR ranges from 2dB to 20dB, being about 1% and 2% higher than connection rates of 25% and 100%, respectively. The connection rate is 100%, which has similar recognition effect, and the difference between the two is about 1% on average. The result shows that the proper establishment of cross-layer connection among the multiple layers of the multi-hop network provided by the invention is beneficial to establishing the identification result. It should also be noted that too many short connections do not continuously improve the recognition result. The 50% connection rate is more than 5% higher than the 12.5% connection rate. This means that good recognition results can be achieved as long as sufficient signal characteristics are transmitted between the layers.
All networks have similar low recognition results in the low SNR range of-20 dB to-15 dB. When the SNR rises to-10 dB, the network with a connection rate of 12.5% has the best recognition effect, 14.72%, 2.18%, 11.06% and 17.95% better than evolgdnn, respnet 50, respnet and resignnet with leakrelu, respectively. Within the SNR range, the resnet with relu is slightly higher than the multi-hop network by 2.15%, the identification results of the two networks are approximate, the identification rates are both less than 45%, and the effect of identifying various AMCs is limited. From-10 dB to 0dB, the multi-hop network (connection rate 12.5%) is 30.67%, 12.62%, 14.81%, 9.96% and 37.73% higher than evolgdnn, respet 50, respontet, resignet with relu and resignet with leakrelu, respectively. At SNR > 0dB, the networks used for comparison were reduced by 27.58%, 4.29%, 8.07%, 5.52% and 28.24%, respectively, compared to the multihop network at a connection rate of 12.5%. The situation shows that the multi-hop network provided by the invention has better identification performance than the traditional network, can better transmit signal characteristics in an identity mapping connection network structure, and can become an advanced signal characteristic extractor for distinguishing various modulation modes.
And comparing the training time and parameter quantity results of the deep multi-hop network under the conditions that the hardware condition is that the CPU is i5, the GPU is 2080ti, the system version is ubuntu 18.04 and the deep learning framework is tensoflow version 1.12. As the connection rate becomes greater, both the time and the overall parameters in table 1 continue to increase. The connection rate of 100% is 10 times higher than the connection rate of 25% and the total parameter is approximately 13 times higher than the connection rate of 25%. However, the recognition effect of the 25% connection rate is slightly better than that of the 100% connection rate, which proves the high efficiency of the multi-hop network structure provided by the invention, and the network form with small parameters is more suitable for being deployed in the wireless communication scene with limited resources.
Table 1: single training time and parameter size in multi-hop networks
Rate of connection Single training time (unit: second) Total parameter
12.5％ 4 319,770
25％ 6 821,370
50％ 20 2,887,770
100％ 70 10,783,770
The result proves that the deep multi-hop network method provided by the invention can effectively identify the communication modulation mode, and the increase of the signal length is beneficial to the improvement of the identification rate; under the condition of higher signal-to-noise ratio, the identification rate is continuously improved along with the expansion of the receiving receptive field range; the establishment of cross-layer connection between multiple layers of the deep multi-hop network is favorable for establishing a recognition result. In addition, the recognition effect of the method provided by the invention is obviously superior to that of other conventional deep learning methods.

Claims (10)

1. A method for carrying out communication automatic modulation recognition based on a deep multi-hop network is characterized by comprising the following steps:
s1: receiving communication signal data;
s2: performing reception processing on the signal collected in S1;
s3: inputting the signal processed by the S2 into a deep multi-hop network for modulation mode identification;
s4: and outputting the recognition result.
2. The method for automatic modulation recognition in communication of claim 1, wherein in S1, assuming that the received signal is separated into the lth multipath component, which is related to τ in the time interval, the signal arriving at the receiver is represented as
Figure FDA0002653363550000011
Where s (t) denotes the transmitted signal, cl(τ; t) represents the ith path of the equivalent channel impulse response, L ranges from 1 to L, n (t) represents additive noise, and r (t) represents the received signal;
the form of the channel impulse response in equation (9) is written as
cl(τ；t)＝βl(t)(t-τl(t)) (10)
Wherein, betal(t) represents the possible time-varying attenuation on the l-th multipath propagation path, (t- τ)l(t)) is the time t-taul(t) impulse response, τl(t) is the l-th multipath delay, all paths having the same Doppler scaling factor γ, τl(t)≈τl-γt；
Substituting equation (10) into equation (9) represents the received signal as
Figure FDA0002653363550000012
3. The method for automatic modulation recognition of communication according to claim 1, wherein in S2, the signal of each length is received and processed, and the signal patterns of various lengths are as follows
oq(k)＝c(q+θk) (7)
Where the signal sequence has the total length of M, q is the starting position of the signal sequence, q is 1, 2, …, M-1, representing the maximum fixed length K of the signal sequence, 1, 2, …, K, where K is the regular fixed signal length, θ is the percentage of the selected variable length, c (·) is a function of the selected length that fits the network input, the network receiving oq(k) As an input pattern for modulation classification.
4. The method for communication automatic modulation recognition of claim 1, wherein the deep multi-hop network in S3 comprises an input layer, a preprocessing layer, a multi-hop layer, and an output layer in this order.
5. The method of communication automatic modulation recognition of claim 4, wherein the input layer represents a wireless communication signal data set including various signal modulation schemes; the preprocessing layer comprises Conv and BN, wherein Conv represents a convolutional network layer, and BN represents batch normalization; the multi-hop layer comprises a ReLU, a Conv, a Dropout and a Concatenate, wherein the ReLU represents a non-linear activation function layer, the Dropout represents an operation for improving the generalization capability of the network, and the Concatenate represents aggregation of a plurality of short connections.
6. The method of claim 4, wherein the multi-hop layers employ a basic depth identity mapping network structure by reducing or expanding different layers in the structure; the network structure is characterized in that direct connection is added among all layers in the network structure, the direct connection is formed by linking different layers in the network structure, and the multi-hop connection rate is a mode for controlling the number of connections among multiple layers.
7. The method of communication automodulation identification as claimed in claim 4 wherein the multi-hop layers employ an identity mapping across layer connections, αi+1And the previous layer of alphaiThe relationship between them is a purely linear superposition relationship, expressed as
αi+1＝αi+G(αi) (1)
Where G (-) is a direct access function that passes learned information between deep network layers.
8. The method of communication automodulation recognition of claim 4 wherein each of the multi-hop layers takes as input the outputs of some previous layer, defines a total number of cross-layer connections of U, the learned signal characteristics have some redundancy with dense connections between layers, and the learned collective knowledge is shared between network layers of the same densely-connected block to fully utilize redundant signal information, formulated as
Figure FDA0002653363550000021
Where ρ is the multi-hop connection rate, λ represents the total number of interconnections in the internal layers of the multi-hop network, and duRepresented within a multi-hop networkA selected connection layer in the partial layer.
9. The method of claim 4, wherein a reception field range is further provided in the multi-hop layer; if the deep multi-hop network deepens each time to generate phi feature maps, the value of the reception receptive field in the V-th layer can be expressed as
Figure FDA0002653363550000031
Wherein phi is0Is the value of the initial reception field, phivIs the value of the varying receptive field corresponding to the V-th layer, V1, 2, …, V, which means that each composite layer of the deep multihop network generates a Φ signature, which is treated as the extracted information of the network layer, each layer has access to its previous signal signature.
10. Use of a deep multihop network as claimed in claim 1 for identification of a modulation scheme for underwater communication.
CN202010878480.7A 2020-08-27 2020-08-27 Method for carrying out communication automatic modulation recognition based on deep multi-hop network Active CN111970218B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010878480.7A CN111970218B (en) 2020-08-27 2020-08-27 Method for carrying out communication automatic modulation recognition based on deep multi-hop network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010878480.7A CN111970218B (en) 2020-08-27 2020-08-27 Method for carrying out communication automatic modulation recognition based on deep multi-hop network

Publications (2)

Publication Number Publication Date
CN111970218A true CN111970218A (en) 2020-11-20
CN111970218B CN111970218B (en) 2021-09-14

Family

ID=73399604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010878480.7A Active CN111970218B (en) 2020-08-27 2020-08-27 Method for carrying out communication automatic modulation recognition based on deep multi-hop network

Country Status (1)

Country Link
CN (1) CN111970218B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109547374A (en) * 2018-11-23 2019-03-29 泰山学院 A kind of depth residual error network and system for subsurface communication Modulation Identification

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101064914A (en) * 2006-04-29 2007-10-31 上海贝尔阿尔卡特股份有限公司 Method and apparatus for performing combined relay in wireless communication network
US7797367B1 (en) * 1999-10-06 2010-09-14 Gelvin David C Apparatus for compact internetworked wireless integrated network sensors (WINS)
CN107295453A (en) * 2016-03-31 2017-10-24 扬州大学 A kind of wireless sensor network data fusion method
CN109299697A (en) * 2018-09-30 2019-02-01 泰山学院 Deep neural network system and method based on underwater sound communication Modulation Mode Recognition
CN109361635A (en) * 2018-11-23 2019-02-19 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on depth residual error network
KR20190031832A (en) * 2017-09-18 2019-03-27 국방과학연구소 Automatic modulation recognition method with feature selection for uncertain fading channels
CN111431825A (en) * 2020-02-25 2020-07-17 泰山学院 Signal automatic classification and identification method based on deep multi-flow neural network

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7797367B1 (en) * 1999-10-06 2010-09-14 Gelvin David C Apparatus for compact internetworked wireless integrated network sensors (WINS)
CN101064914A (en) * 2006-04-29 2007-10-31 上海贝尔阿尔卡特股份有限公司 Method and apparatus for performing combined relay in wireless communication network
CN107295453A (en) * 2016-03-31 2017-10-24 扬州大学 A kind of wireless sensor network data fusion method
KR20190031832A (en) * 2017-09-18 2019-03-27 국방과학연구소 Automatic modulation recognition method with feature selection for uncertain fading channels
CN109299697A (en) * 2018-09-30 2019-02-01 泰山学院 Deep neural network system and method based on underwater sound communication Modulation Mode Recognition
CN109361635A (en) * 2018-11-23 2019-02-19 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on depth residual error network
CN111431825A (en) * 2020-02-25 2020-07-17 泰山学院 Signal automatic classification and identification method based on deep multi-flow neural network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
DILEEP P: "Dense Layer Dropout Based CNN Architecture for Automatic Modulation Classification", 《IEEE XPLORE》 *
姚宇晨，彭虎: "基于深度学习的通信信号自动调制识别技术", 《人工智能》 *
王岩等: "基于深度残差网络的水下通信调制信号分类", 《数字海洋与水下攻防》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109547374A (en) * 2018-11-23 2019-03-29 泰山学院 A kind of depth residual error network and system for subsurface communication Modulation Identification
CN109547374B (en) * 2018-11-23 2021-11-23 泰山学院 Depth residual error network and system for underwater communication modulation recognition

Also Published As

Publication number Publication date
CN111970218B (en) 2021-09-14

Similar Documents

Publication Publication Date Title
CN113225102B (en) Low signal-to-noise ratio code capturing method based on random continuous phase modulation signal
CN110086555B (en) Grouping pilot frequency distribution method and device in large-scale MIMO system
CN113630130B (en) End-to-end digital communication demodulation method
CN107135041B (en) RBF neural network channel prediction method based on phase space reconstruction
Zhang et al. A data preprocessing method for automatic modulation classification based on CNN
Wu CNN and RNN-based deep learning methods for digital signal demodulation
Kim et al. Adversarial attacks with multiple antennas against deep learning-based modulation classifiers
CN112307987B (en) Method for identifying communication signal based on deep hybrid routing network
CN113055107B (en) Interference strategy generation method for radio station with unknown communication mode
Wang et al. Adoption of hybrid time series neural network in the underwater acoustic signal modulation identification
Ko et al. Adaptive modulation with diversity combining based on output-threshold MRC
CN111970218B (en) Method for carrying out communication automatic modulation recognition based on deep multi-hop network
Huynh-The et al. MIMO-OFDM modulation classification using three-dimensional convolutional network
CN111464469A (en) Hybrid digital modulation mode identification method based on neural network
CN110519013A (en) A kind of underwater sound communication self-adaptive modulation method based on intensified learning
CN114039825A (en) SIMO-HPO-CDSK communication method based on Rayleigh fading channel
CN113141214B (en) Deep learning-based underwater optical communication misalignment robust blind receiver design method
CN102082619B (en) Transmission adaptive method based on double credible evaluations
Hung et al. An adaptive multistage multiuser detector for MC-CDMA communication systems using evolutionary computation technique
CN106231629B (en) A kind of self-organizing network system and method for realizing rate adaptation and anti-Doppler frequency displacement
Yi et al. 6G intelligent distributed uplink beamforming for transport system in highly dynamic environments
CN109831264B (en) Time sequence underwater sound channel quality prediction method and system based on nearest neighbor regression
Jin et al. On Channel Classification by Using DTMB Signal
CN110190908A (en) A kind of planisphere design method minimizing incoherent extensive SIMO error rate of system in ISI channel
Mathew et al. Semi blind neural network based channel estimation technique for OFDM receivers

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