FR2713799A1 - Automatic recognition of modulation type of intercepted signal - Google Patents

Abstract

The modulation recognition involves transforming (2) the intercepted digitally modulated signal to extract digital magnitudes, and comparing (3) these to known magnitudes associated with different types of modulation. The digital magnitudes are determined from the statistical moments of orders 1 to 4 on the amplitude, the phase and the instantaneous frequency of the signal. The relation between these characteristics of the signal and the known signal characteristics is determined by an open loop sigmoid neural net which has previously been apprenticed with examples of the major types of digital modulation.

Description

Method for automatically recognizing modulation
of an intercepted signal
The invention relates to a method for recognizing the type of digital modulation used for transmitting an intercepted and a priori unknown signal, with a view to decoding it, in which this intercepted signal is transformed into a digital form on the basis of which one extracts digital quantities used to discriminate different types of digital modulations and a correspondence is established between these digital quantities and a particular known type of digital modulation.

 There is an important need, in the military but also in the civilian field, to monitor radioelectric activity.

The decoding of the intercepted signals goes through a succession of processing steps including the recognition of the modulation used to transmit this signal. Current data communication techniques make extensive use of digital modulations and in this case it is a matter of recognizing a particular digital modulation among those most frequently used. The most widely used digital modulations in the military field are FSK, PSK, MSK, QAM and OQPSK modulations.

 Recognizing the digital modulation of an intercepted and apparently unknown signal is a complex problem.

 It is known that certain digital modulations have characteristic eye and constellation diagrams. It is therefore possible to recognize the digital modulation used to emit a signal from observation of the eye diagrams and constellation of this signal. This process is however limited to intercepted signals with little or no noise. Its use remains punctual and manual.

 Attempts have already been made to automate the recognition of the digital modulation of an intercepted signal. The most remarkable are based on pattern recognition. In these attempts, the correspondence between the digital quantities and a particular type of digital modulation is established using linear, quadratic classifiers or decision trees. But here again, these known methods remain limited to low noise signals, of the order of 25 to 50 dB.

Furthermore, the material means which they employ are very complicated. They make it possible to recognize a small number of digital modulations, typically 2 or 3 different digital modulations.

 The object of the invention is to propose a method making it possible to recognize the digital modulation used to transmit an intercepted signal and a priori unknown, even highly noisy, among ten or so most frequently used digital modulations.

 To this end, the subject of the invention is such a method, characterized in that, as digital quantities, the first statistical moments of order 1 to 4 are used on the instantaneous amplitude, phase and frequency of this signal, and a correspondence is established between these digital quantities and a type of digital modulation using a non-looped sigmoid neural network which has undergone prior learning to classify said digital quantities on outputs representative of the modulation classes following digital: FSK2, FSK4, FSK8, PSK2, PSK4, PSK8, QAM16, QAM64, OQPSK, MSK.

 Compared to the known state of the art, the characteristics of the method according to the invention contribute to increasing the range of modulations recognized automatically while simplifying the means used to carry out this recognition automatically by implementing a neural network. with a constrained and light architecture which allows faster processing than a fully connected neural network. In addition, such a method makes it possible to obtain performances, in terms of good recognition rate, of the order of 90% even for very noisy intercepted signals (signal / noise ratio which can reach OdB) and almost 100% on less noisy signals.

 The invention is described in detail below with reference to the drawings.

 FIG. 1 is a block diagram illustrating the method according to the invention.

 FIGS. 2 to 7 illustrate calculation steps for the extraction of the digital quantities according to the invention.

 FIG. 8 shows the architecture of the sigmoid neural network used in the method according to the invention.

 In Fig. 1, the intercepted signal S has been filtered and digitized in order to recognize the digital modulation used for its transmission. The complex digital samples obtained at the output of the analog / digital converter (not shown) are first passed through a module 1 for extracting statistical moments of order 1 to 4 on the amplitude, phase and frequency of the signal S .

 These statistical moments are calculated as follows.

 A vector of 2048 digital samples, generally successive, from the sampler are kept in memory for the purpose of calculating the module, the argument and the phase shift of each digital sample comprising a real part of amplitude a and an imaginary part d 'amplitude b.

 The module Mi of each sample i (having a real part ai and an imaginary part bi) is first calculated by applying the relationship shown in Figure 2.

 The argument Pi of each sample i is then calculated by applying the relationship shown in Figure 3.

 The instantaneous frequency Fi in each sample i is finally calculated by applying the relation shown in FIG. 4. The value F2048 is arbitrarily fixed equal to the value F2047.

 We therefore recover three vectors of 2048 elements each containing modulus values, argument values and instantaneous frequency values respectively.

 Statistical moments of order 1 to 4 on the instantaneous amplitude, phase and frequency of the intercepted signal are calculated from these three vectors, which leads to 12 statistical moments M1A M4A, M1P-M4P, M1F -M4F. M1A denotes the mean of the amplitude vector. M2A denotes the variance of the amplitude vector.

M3A denotes the statistical moment of order 3 of the amplitude vector. M4A designates the moment of order 4 of the amplitude vector. MlP-M4P respectively designate the moments of order 1 to 4 for the phase vector. M1F-M4F respectively designate the moments of order 1 to 4 for the frequency vector.

 The different statistical moments are calculated from the relationships shown in Figures 5 to 7 which use the terminology defined above.

 The power of the received signal is normalized by setting MlA = 1 and the quantities M2A-M4A are recalculated from the following relationships.

M2A = M2A / M1A2
M3A = M3A / M1A3
M4A = M4A / M1A4
11 useful quantities M2A-M4F are therefore obtained.

 The 11 quantities M2A-M4F are applied at the input of the sigmoid neural network 2 shown in FIG. 1 and which is described below with reference to FIG. 8.

 This sigmoid neural network 2 comprises a first input layer of 11 neurons Nl.l-Nl.ll, a first hidden layer of 20 neurons N2.1-N2.20, a second hidden layer of 13 neurons N3.1- N3.13 and an output layer of 10 neurons N4.1-N4.10.

 This architecture of the neural network 2 contributes to a particularly simple embodiment both in the form of a program and in the form of a wired circuit. The document "LANGAGES & SYSTEMES-INFOPC N 71; The artificial neural networks of Henri Lemberg" describes the architecture and the functioning of a neural network. Indeed, the proposed network contains half as many connections as the same fully connected network. This is due to the fact that this network is a juxtaposition of simpler networks, having undergone independent learning, on sub-problems of the general problem.

 The sigmoid neural network of FIG. 8 is used to classify the 11 quantities M2A-M4F, which prove to be a good summary of information for quasiperiodic signals, in one of the following 10 digital modulation classes: FSK2, FSK4, FSK8, PSK2, PSK4, PSK8, QAM16, QAM64, MSK.

 It should be understood here, that each neuron of the output layer, in response to the digital quantities presented at the input of the network, releases a generally decimal value comprised in the interval [-1,1], that each neuron of the output layer represents a particular modulation class so that a measure of the probability of correspondence between the quantities M2A-M4F and the modulation class represented by the neuron is directly given by the output value of this neuron.

 The connections between neurons are established as follows.

 The 11 neurons N1.1-N1.11 are each connected to the first 5 neurons N2.1-N2.5 of the first hidden layer.

The neurons N2.1-N2.5 are each connected to the first 3 neurons N3.1-N3.3 of the second hidden layer. These neurons form a first neural sub-network with 11 inputs and 3 outputs.

 The 7 neurons Nl.l, N1.2, Nl.3, N1.5, N1.7, Nl.9, N1.11 are each connected to the 5 neurons N2.6-N2.10 of the first hidden layer. The neurons N2.6-N2.10 are each connected to the 5 neurons N3.4-N3.8 of the second hidden layer. These neurons form a second neural sub-network with 7 inputs and 5 outputs.

 The 7 neurons Nl.l, N1.2, N1.3, N1.5, N1.7, Nl.9, N1.11 are each still connected to the 5 neurons N2.11-N2.15 of the first hidden layer. The neurons N2.11-N2.15 are each connected to the 3 neurons N3.9-N3.11 of the second hidden layer. These neurons form a third neural sub-network with 7 inputs and 3 outputs.

 The 7 neurons Nl.l, N1.2, N1.3, N1.5, N1.7, Nl.9, N1.11 are each still connected to the 5 neurons N2.16-N2.20 of the first hidden layer. The neurons N2.16-N2.20 are each connected to the neurons N3.12-N3.13 of the third hidden layer. These neurons form a fourth neural sub-network with 7 inputs and 2 outputs.

 The first neural sub-network is provided so as to classify the input quantities M2A-M4F in one of the 3 metaclasses of digital modulation, respectively (FSK2, FSK4, FSK8, OQPSK, MSK); (PSK2, PSK4, PSK8); (QAM16, QAM64) respectively represented by its output neurons.

In other words, the output of neuron N3.1 gives a measure of the probability of correspondence between a metaclass comprising the digital modulations FSK2, FSK4, FSK8, OQPSK and MSK and the input quantities. The output of neuron N3.2 gives a measure of the probability of correspondence between the metaclass including the digital modulations
PSK2, PSK4 and PSK8 and the input quantities. The output of neuron N3.3 gives a measure of the probability of correspondence between the metaclass comprising the digital modulations QAM16 and QAM64 and the input quantities.

In the same way, the second neural sub-network is provided so as to classify the quantities of M2A inputs
M4F in one of the 5 digital modulation classes, respectively FSK2, FSK4, FSK8, OQPSK and MSK. Neurons
N3.4-N3.8 respectively represent the modulations
FSK2, FSK4, FSK8, OQPSK and MSK.

 The third neural sub-network is designed to classify the input quantities M2A-M4F in one of the 3 digital modulation classes, PSK2, PSK4 and PSK8, respectively. The neurons N3.9-N3.11 respectively represent the PSK2, PSK4 and PSK8 modulations.

 The fourth neural sub-network is designed to classify the input quantities M2A-M4F in one of the 2 digital modulation classes, QAM16 and QAM64 respectively. The neurons N3.12-N3.13 respectively represent the QAM16 and QAM64 modulations.

 This division of the a priori complex neural network 2 into four neural sub-networks of much simpler architecture contributes to simplifying the learning phase for obtaining weighting coefficients for the neurons and lightening the structure of the final network. From the above, it appears that the 11 quantities only serve the problem of metaclasses. The other 3 subnets are content with 7 input quantities. the structure of neural network 2 is thus lightened between the first two layers by deleting connections.

 These weights are obtained as follows.

 A database is constructed comprising 3000 sets of input quantities with 300 sets for each type of modulation among the 10 types of modulation mentioned above. The signal / noise ratio of each game varies between 0 and + 50dB.

Each neural subnetwork undergoes learning on half of a base of examples (i.e. 1500 sets of quantities randomly drawn from the base per program).

 The weighting coefficients assigned to the input connections of each neuron of each neural sub-network are calculated by the back propagation algorithm of the stochastic gradient. In the course of this algorithm, the weighting coefficients are initialized at random in the interval [-0.1; 0.1]. The step of the stochastic gradient is equal to 0.01. The base is passed 500 times on each subnet.

Learning the neural network 2 as described above is only made possible by the fact that the neural subnets are independent (at the level of their hidden layers)
In this neural network, the neurons of the output layer are very simply connected to the neurons of the hidden second layer. The connections between the neurons of the output layer and those of the second hidden layer are given by the correspondence table shown in appendix I in which the output of a neuron of the output layer corresponds to a logical AND of the outputs. of two hidden second layer neurons.

The outputs of neurons N4.1-N4.10 correspond respectively to the digital modulation classes FSK2,
FSK4, FSK8, PSK2, PSK4, PSK8, QAM16, QAM64, OQPSK, MSK.

The neural network 2 can be followed by a processing module 3 for the outputs of neurons N4.1-N4.10. The function of this processing module is to determine the smallest Euclidean distance between the output of the neural network, for example the vector [0.6; -0.5; ....- 0.2], and 10 coded neural network outputs in the form of 10 vectors each having 10 elements. In each vector, one of the elements is equal to 1 and identifies by its rank in the vector a particular class of digital modulation, the other elements being equal to -1. Annex II shows these 10 vectors in correspondence with a digital modulation class. The smallest Euclidean distance which has been determined can be compared with an adjustable threshold so that the decision D at the output of the processing module identifies or not a recognized digital modulation with a quantifiable certainty factor. This threshold allows rejection of the new modulation type presented to network 2 (apart from the 10 types considered above)
APPENDIX I
N4.1 <- N3.1 AND N3.4
N4.2 <- N3.1 AND N3.5
N4.3 <- N3.1 AND N3.6
N4.4 <- N3.2 AND N3.9
N4.5 <- N3.2 AND N3.10
N4.6 <- N3.2 and N3.11
N4.7 <- N3.3 AND N3.12
N4.8 <- N3.3 and N3.13
N4.9 <- N3.1 AND N3.7
N4.10 <- N3.1 AND N3.8
ANNEX II [1, -1, -1, -1, -1, -1, -1, -1, -1, -1] -> FSK2 [-1, -1,1, -1, -1, -1, -1, -1, -1, -1] -> FSK4 [-1, -1, -1, -1,1, -1, -1, -1, -1, -1] -> PSK4 [-1, -1, -1, -1, -1, -1,1, -1, -1, -1] -> QAM16 [-1, -1, -1, -1, -1, -1, -1, -1,1, -1] -> OQPSK

Claims (2)

 1.) A method for recognizing the type of digital modulation used to transmit an intercepted and a priori unknown signal, with a view to decoding it, in which one transforms (2) this intercepted signal into a digital form on the basis of which extract digital quantities used to discriminate different types of digital modulations and establish (3) a correspondence between these digital quantities and a particular known type of digital modulation, characterized in that, as digital quantities, the first statistical moments of order 1 to 4 on the instantaneous amplitude, phase and frequency of this signal, and a correspondence is established between these digital quantities and a type of digital modulation using a sigmoid neural network not looped which has undergone a prior learning to classify said digital quantities on outputs representative of the digital modulation classes e following: FSK2, FSK4, FSK8, PSK2, PSK4, PSK8, QAM16, QAM64, OQPSK, MSK.  2.) A neural network (2) for implementing the method according to claim 1, characterized in that it consists of four neural sub-networks having undergone independent learning, these four sub-networks being part in a structure with two hidden layers each having a plurality of neurons, the neurons (N2.1-N2.20) of a hidden layer not being completely connected to the neurons (N3.1-N3.13) of the another hidden layer.