Automatic Recognition of Analog and Digital Modulation Signals using DoE Filter

Minah Kang, Cheolsoo Lee, Jeungmin Joo

Electronic Warfare PEO, Agency for Defense Development, Deajon, 305-600, Korea Tel: +82-42-821-3531, Fax: +82-42-823-3400 E-mail: minah@add.re.kr, dsiqueen@icu.ac.kr, gangsang@add.re.kr
Abstract An algorithm for recognition of analogue and digital modulation types, utilizing the decision-theoretic approach, is developed with the novel key features. The proposed 3 novel key features, the each peak number of the phase and the amplitude component obtained by the DoE (Difference of Estimator) filter and the phase gamma max, has robust properties of sensitive with modulation types and insensitive with SNR variation. This paper describes the algorithm which automatically identifies the modulation types of received signals without prior information with these some new and some old key features. The computer simulation is performed. We investigate the performance of the proposed classifier for classifying 9 modulated signals, and compare with that of the conventional decision tree classifier. Results indicated good performance (i.e. the average probability of correct classification (Pcc) of 99.5 %) at the SNR of 10 dB. Comparing with that of the conventional decision tree classifier (i.e. the average Pcc of 95.4 %), we proved that the performance of the proposed classifier is superior to that of the conventional algorithm. Specially, at the SNR as low as 7 dB, the recognition performance of digital signals shows a noticeable improvement.

I.

INTRODUCTION

In military and civilian communications applications, the system which automatically identifies the modulation signals has been researched for monitoring and recognizing each communication signal with various modulation types in wide band. General wireless communication signals are transmitted with various modulation types in wide frequency band. Specially, in military environment, modulation recognition system is useful to collect information by interception and to interfere in an electronic wave by surveillance and signal confirmation like electronic warfare. It is reason that if we know the modulation type of signal and apply the proper demodulator, the important information can be completely from a incoming signal as well as it helps determining an appropriate jamming waveform and perceiving a threat. Modulation recognition algorithm involves two main steps: the key features extraction for classification; and the modulation type decision of the received signal. In the first step, we extract the key features which are need for modulation recognition. To extract features, there are various techniques such as wavelet transform, likelihood probability and HMM. In the second step, many methods have been also reported in literature for modulation classification. It has been used a variety of techniques such as decision tree classifier,

minimum distance classifier, support vector machine, and the artificiality neural network. According to the tendency to move towards using digital modulation schemes, most of the algorithms deal with the problem of digital modulation classification. Nevertheless, the modulation classification of both analog and digital signals is of interest, because some existing systems still use analog communication techniques. The aim of this paper is to recognize simultaneously analog and digital signals modulation types. The conventional modulation classification algorithm is mainly achieved by the decision tree technique. It is considered as the most practical algorithm because of the requirement of fast processing in reconfigurable software radio. It uses 7 common features for 9 modulation types including both analog and digital modulation without a priori information. These 7 key features are gamma max (max), symmetry of frequency (P), sigma of direct phase (dp), sigma of absolute frequency (df), sigma of absolute phase (ap), frequency compactness (f42) and Occupied bandwidth (oBW) [2]. But, it leave something to desired in using an amplitude gradient as a first classification factor and using only frequency kurtosis as an factor for separating the 4FSK signals from the FM signals. In this paper, we propose the new modulation classification algorithm to enhance recognition rates with the novel robust key features. Because the prototype of modulation classifier was implemented with decision tree approach, we developed a new algorithm of decision tree method with 3 novel features and 3 existing features. The old 3 key features are symmetry of frequency (P), sigma of absolute phase (ap), frequency compactness (f42). The proposed 3 key features are the phase gamma max (max-p), the peak number of the phase component passing the DoE filter (Np), the peak number of the amplitude component passing the DoE filter (Na). The new features are made from differentiated signals by DoE filter. The DoE filter differentiates an amplitude and phase signals and extracts key classification features. Because these features have the values which are clearly divided into each modulation type, we can classify the signal completely. The organization is as follows. Section 2 introduces the DoE filter and describes the 6 classification features including the 3 novel features. Section 3 presents the new classification algorithm using our proposed features. Simulation results and conclusions are given in Section 4 and Section 5, respectively.
B B B B B B B B P PB B B B P PB B B B B B B B

978-1-4244-4522-6/09/$25.00 2009 IEEE

609

ISCIT 2009

1) Symmetry (P) The first key feature, P, measures the spectrum symmetry around the carrier frequency based on the spectral powers for the lower and upper sidebands. It is useful in the discrimination of LSB and USB [1].

P=
Fig. 1. Difference of Estimator filter technique
fcn

PL PU PL + PU
f cn

(3)
2

PL = X C (i ) , PU = X C (i + f cn + 1)
i =1 i =1

(4) (5)

II. A.

FEATURES EXTRACTION BY DOE FILTER

f cn =

Design of DoE filter Before a received baseband signal come in a classifier, it should be pre-processed. In the first, an amplitude signal and a phase signal are obtained from a received signal. Then they are normalized by a general normalization technique. After this pre-processing stage, an amplitude signal and a phase signal get in the DoE filter. The DoE filter differentiates input signals. To differentiate signals, the method which calculates the difference of current sample and previous sample has been mostly used. However, our algorithm uses the DoE filter for signal differentiation. When the signal samples such as .n-2, n-1, n, n+1, n+2. are inputted, the DoE filter apply an estimator which has a window (W1, W2) with constant length in left and right on current processing sample(n) as shown in Fig.1. The difference of left and right estimator becomes the output of DoE filter. Then, the estimator applied symmetrically in left and right uses the average filter which outputs an average of input samples inside a window, as in (1) and (2).

fc N S 1 fs
B B

where fc is the center frequency, and fs is the sampling frequency. NS means the sample number of one segment and XC(i) means the ith frequency response by the Fourier Transform of corresponding segment. The amplitude modulated signals can be discriminated by checking the value of P. In AM, P approximates zero because the lower and upper sidebands are symmetry. The LSB and the USB with the single sideband are discriminated by checking the sign of P. The value of P is near +1 in the LSB and -1 in the USB.
B B B B B B

2) Sigma of absolute phase (ap)

B B

The second key feature, ap, is the standard deviation of the absolute value of the centered non-linear component of instantaneous phase [1]. It is defined by
B B

ap

f L Average (n) = f R Average (n) =

1 W

W 1 i =0

1 2 NL (i ) 1 = C C An (i )>at

NL (i) An ( i )> at

(6)

x(n i )

(1) (2)

1 W 1 x( n + i ) W i =0

where fL-Average(n) is the output of a left estimator with the window size W and fR-Average(n) is the output of a right estimator.
B B B B

where NL(i) is the value of the nonlinear component of the instantaneous phase at time instants t = i/fs , An(i) is the amplitude of the ith sample. at is an amplitude threshold and C is the number of samples in {(i)} for which An(i)>at . The threshold is used to discriminate between the signals that have the variation of absolute phase component (ap(i)> t ap ) and
B B B B B B B B B B B B B B

have the constant absolute value of nonlinear instantaneous phase component (ap(i) t ap ). The 2PSK signal after
B B

Key Features Extraction In this paper, we selected 6 key features for modulation classification. The 3 key features among them were used in [2] and the novel robust 3 key feature are proposed in this paper. The conventional 3 key features are the symmetry of frequency (P), the sigma of absolute phase (ap), the frequency compactness (f42). The proposed 3 key features are the phase gamma max (max-p), the peak number of the phase component passing the DoE filter (Np), the peak number of the amplitude component passing the DoE filter (Na). B.
B B P PB B B B B B B B

centering has no the absolute phase information because it has only two level polarities to the contrary and the 4PSK signal has the absolute phase information. Thus, this key feature is used for the discrimination of the 2PSK and the 4PSK signals.

3) Frequency compactness (f42) The frequency modulated signals can be distinguished by the instantaneous frequency distribution. The key feature, f42 defined by (7), is the kurtosis of the normalized-centered instantaneous frequency [1].
P PB B P PB B

610

f 42 =

{ } {E{f (t )}}
4 E f N (t ) 2 N

(7)
B B

PDoE (i ) = f L p (i ) f R p (i )

(13)

where fN is the normalized-centered instantaneous frequency based on zero. The instantaneous frequency distribution of the FM signal is compacter than that of the 2/4FSK signals. That is, the frequency kurtosis of the FM signal is larger than that of the 2/4FSK signal. Thus, the FM and 2/4FSK signals can be distinguished by comparing this key features with suitable threshold th f .
B B

where PDoE(i) is the output of the DOE filter corresponding to the phase components and fL-p(i) and fR-p(i) are each left and right estimator outputs. W is the window size and Pcn(i) is the normalized-centered instantaneous phase signal. Then, to find the peaks of the output signal, we subtract the previous sample value from the current sample value, as in (14).
B B B B B B

42

S p (i ) = PDoE (i ) PDoE (i 1)
B B B B

(14)

4) Phase gamma max (max-p) The envelope variation of instantaneous phase can be another key feature for the discrimination of the FM and the 2/4FSK signals. Compared with the 2/4FSK signal, the phase envelop of the FM signal is largely fluctuated. This key feature, max-p, is defined by
B B B B

where Sp(i) is the differentiated value of PDoE(i), ( i = 2, 3, ., N). In the peak point of the DoE filter output signal, the sign of Sp(i) and Sp(i+1) is different, that is, Sp(i) is positive and Sp(i+1) is negative. Then, among these peaks, we selects the peaks which PDoE(i) > t N p with a suitable threshold. The number of
B B B B B B B B B B

max p = max

DFT ( Pcn (i )) NS

selected peaks is a novel key feature, Np, defined by (15).

B B

(8)

N p = n i | S p (i ) > 0, S p (i + 1) < 0, PDoE (i ) > t N P

(15)

where Pcn(i) is the value of the normalized-centered instantaneous phase at time instants t = i/fs , (i = 1, 2,., NS) and it is defined by
B B B B B B

Because this key feature is robust in SNR variation and have the discriminating values on each modulation types, it is contributed to enhance the recognition rate.

Pcn (i ) = pn (i ) 1
B B B B B B

(9)

where Pcn(i) = P(i)/mp and where mp is the average value of the instantaneous phase over one frame; that is,

mp =

1 N P(i) N i =1

(10)

6) The peak number of the amplitude component passing the DOE filter (Na) This novel feature, Na, is obtained by the same process with Np except for using the amplitude information instead of the phase information. It is the peak number of the amplitude component derived from the DoE filter. Firstly, we calculate the amplitude envelope using the Hilbert transform. Then, the amplitude envelope is normalized and passes the DoE filter. The output of DoE filter is defined by (18).
B B B B B B

The phase gamma max is similar to gamma max in [2] but they are different in using the phase information instead of the amplitude information.

f L a (i ) =

1 W 1 W

W 1 m =0

acn (i m)

(16)

5) The peak number of the phase component passing the DoE filter (Np) The peak number of the phase component obtained by the DoE filter, Np, is the most important key feature of the proposed algorithm. After the normalized instantaneous phase signal passed the DoE filter, the output is calculated by the difference of two estimators and is defined by (13).
B B B B

f R a (i ) =

W 1 m =0

acn (i + m)

(17)

ADoE (i ) = f L a (i ) f R a (i )
B B

(18)

f L p (i) =

1 W 1 pcn (i m) W m =0 1 W 1 pcn (i + m) W m =0

(11)

f R p (i ) =

(12)

where ADoE(i) is the output of the DoE filter corresponding to the amplitude components and fL-a(i) and fR-a(i) are each left and right estimator outputs. W is the window size and acn(i) is the normalized-centered amplitude envelope. Then, to find the peaks of the output signal, we subtract the previous sample value from the current sample value, as in (19).
B B B B B B

S a (i ) = ADoE (i ) ADoE (i 1)

(19)

611

where Sa(i) is the differentiated value of ADoE(i), ( i = 2, 3, ., N). In the peak point of the DoE filter output signal, the sign of Sa(i) and Sa(i+1) is different, that is, Sa(i) is positive and Sa(i+1) is negative. Then, among these peaks, we selects the peaks which ADoE(i) > t N a with a suitable threshold. The number of
B B B B B B B B B B B B B B

selected peaks is a novel key feature, Na, defined by (20).

B B

N a = n i | S a (i ) > 0, S a (i + 1) < 0, ADoE (i ) > t N a

(20)

This key feature is useful to separate the CW signal from the amplitude modulated signals such as the LSB, the USB and the AM signals. III. PROPOSED MODULATION CLASSIFICATION ALGORITHM

The proposed algorithm for the modulation classification is based on the decision-theoretic approach. The classifier with this decision tree technique is globally used for analog and digitally modulated signals recognition because it is easy to implement and has low complexity. A general procedure of the decision tree classifier is constructed by setting the best key feature for discriminating the modulated signals at each node of the tree. Because it is progressed orderly, different tree classifiers are obtained by changing the order the variables are analyzed. So the optimal classifier should be found by thorough search and select the suitable threshold which has the minimal classification error. The conventional decision tree classifier separates the 9 modulation types with the proper threshold values corresponding to each key feature as shown in Fig.2 [2]. Because the decision tree algorithm orderly goes on, the

classification error in the primary stage results in the performance deterioration. Thus, it is necessary to use the key feature which can be robust in SNR variation and clearly discriminate the modulated signals in the first classification stage. In the conventional classifier, the gamma max, which identifies the amplitude envelope variations, is used as the first key feature. This feature separates the constant envelope signals (e.g., CW, FM, 2/4FSK, 2/4PSK) from the amplitude modulated signals (e.g., AM, LSB, USB). But, it is defective for the first classification because it results in some misclassified signals. Also, the conventional classifier is weak in discriminating of the 4FSK and the FM signals by using only frequency kurtosis. In this paper, to make up for these defects, the new algorithm with novel features is proposed. It uses the peak number of the phase component obtained by the DoE filter, Np, as the first separating feature. The feature causes the comparatively complete classification which separates the 9 modulated types into the 4 regions at a time. Among the four set, the first set consists of the 4FSK and the FM signals and the second set has only the 2FSK signal. The third set consists of the 2PSK and the 4PSK signals and the last set consists of the CW, the AM, the LSB and the USB signals. This first separating stage requires three proper threshold values of Np such as tNp1, tNp2 and tNp3 and is performed by following procedure.
B B B B B B B B B B

N p > t N p1

1 set : 4FSK, FM

(21) (22) (23) (24)

t N p 2 < N p < t N p 1 2 set : 2FSK t N p 3 < N p < t N p 2 3 set : 2PSK, 4PSK N p < tNp3 4 set : CW, AM, LSB, USB

In the second separating stage, the CW signals is discriminated from the amplitude modulated signals (e.g., AM, LSB, USB) using the peak number of the amplitude component obtained by the DoE filter, Na. The remaining signals (e.g., AM, LSB, USB) is classified by the feature, P, which measures the spectrum symmetry around the carrier frequency. The 2PSK signals and the 4PSK signals is completely separated by using only the standard deviation of the normalized absolute phase, ap. Lastly, the frequency kurtosis, f42, and the envelope variation of instantaneous phase, max-p, are used in the separation of the 4FSK signals and the FM signals. Because the discrimination of these signals is difficult, the classification by only one feature has much probability of errors. So we separate them with two features, f42 and max-p, by following 3 steps.
B B B B P PB B B B P PB B B B

1 step:
f 42 < t f 1 4FSK
42

(25) (26)

Fig. 2.

Functional flow chart in conventional decision tree classifier

f 42 > t f 2 FM
42

612

Fig. 3.

Functional flow chart in proposed decision tree classifier

2048 samples. The best key features threshold values are selected such that the optimum probability of correct decisions is obtained from the 400 realizations for each modulation type of interest. And the probability of correct classification (Pcc) is calculated by 1000 runs of our algorithm. The classification result of the conventional classifier at the SNR of 10 dB is provided in the matrix shown in Table I [2]. It was shown the probability of correct classification (Pcc) of 95.4 %. It proved the weakness in the discrimination of the AM signal and the constant envelope signals such as the CW and the 2PSK signals obtained by using the gamma max as the first key feature. Also, we confirmed the ambiguity in the separation of the FM and the 4FSK signals. Table II shows the performance of the proposed modulation classifier. The proposed classifier can achieve the probability of correct classification (Pcc) of 99.5 % at the SNR of 10 dB. With supplying weak points in the conventional algorithm, it remarkably improves the recognition rates of the modulation types such as the AM signals, the FM signals and the 2/4PSK signals. Considering only recognition rates, other classification techniques such as Minimum Distance Classifier (MDC) and
TABLE I CONVENTIONAL CLASSIFICATION RESULT (PCC=95.4%)

f t f 1 < 42 < t f 2 Go to 2 step

42 42

(27)

2 step:

max p < t max p 1

FM Go to 3 step

(28) (29) (30)

max p > t max p 2 4FSK

t max p 1 < max p < t max p 2 3 step:
f 42 < t f 3
42

Actual Modulation CW CW AM LSB USB 100 4

Estimated Modulation Type @ SNR = 10 dB AM 91 100 100 94 2 0.5 100 98 82 3.5 93.5 5 0.5 LSB USB FM 2FSK 4FSK 2PSK 4PSK 5

4FSK FM

(31) (32)

FM 2FSK 4FSK 2PSK 14.5 6.5 4PSK

f 42 > t f 3
42

The above-mentioned classification procedure of the modulated signals is shown in Fig.3. IV. SIMULATION RESULTS The performance of the proposed algorithm is evaluated for 9 modulation types. The numerical simulations, corrupted with band-limited Gaussian noise, were performed for both analog and digital modulation signals. None of many classification techniques work reliably with signals that have low SNR (below 10 dB) and consider a large range of modulation types including both digital and analog. So, our simulation is performed at the SNR of 10 dB. In the detail conditions, the carrier frequency is 150 KHz and the sampling frequency is 1.2 MHz. The message signal bandwidth is 10 KHz and the frequency deviations of the FM and the digital signals are each 20 KHz and 25 KHz, respectively. The segment defined as a minimum unit for analysis of signals has

TABLE II PROPOSED CLASSIFICATION RESULT ( PCC=99.5%) Actual Modulation CW CW AM LSB USB FM 2FSK 4FSK 2PSK 4PSK 0.3 2 100 99.9 0.9 98.9 99.9 99.5 100 98 99.7 100 0.5 0.1 0.2 0.1 Estimated Modulation Type @ SNR = 10 dB AM LSB USB FM 2FSK 4FSK 2PSK 4PSK

613

Support Vector Classifier (SVC) is better than the Decision Tree Classifier (DTC) [2]. However, compared with the best performance of other techniques, the improved DTC of our algorithm is superior to any other methods. Merely, we can look forward to better performance from the implementation of MDC and SVC with our novel features as our future works. Specially, as shown in Table II, the recognition rates of the digital signals are prominently improved. Also, our classifier for digital modulation classification indicates good performance with the average Pcc of 98 % on an AWGN channel, even at the SNR as low as 7 dB. The detailed performance of 4 digital modulation types results in the Pcc of 99.4 % in the 2FSK4, 93.2 % in the 4FSK, 99.4 % in the 2PSK, 99.9% in the 4PSK, respectively. It is necessary to improve the classification of the 4FSK signals and the FM signals at the low SNR. V. CONCLUSION

This paper has aimed to recognize automatically the types of modulations when it is not known a priori if the signal is analogue or digital. The decision tree approach has been to carry out the classification process in two stages: the first of which extracts the key features from the received signal into one of the nine signal classes and the second stage is to classify the received signals into one of the nine classes and to recognize the modulation type. In this paper, we proposed the 3 novel key features, the peak number of the phase and amplitude component passing the DOE filter and the phase gamma max, which has robust properties of sensitive with modulation type. With these some new and some old ideas from previously published works, our automatic modulation recognition algorithm which has robust properties of insensitive with SNR variation is proposed in this paper. Some results of simulations were shown. As a result, the recognition rate of the proposed classifier is the average Pcc of 99.5 % at the SNR of 10 dB. Compared to previous classifier, it is the noticeably improved result. Also, it indicates good performance with the average Pcc of 98 % at the SNR as low as 7 dB. These good results for signals came from the large effects of robust properties of a novel feature. Specially, the recognition rates of the digital signals are prominently improved. Many classification techniques for digital modulation types have been introduced by the tendency to move towards using digital modulation schemes. Thus, as a future work, we will improve the optimum algorithm for various digital signals with the suggested features obtained by DOE filter or other new features. Also, our future work will include the implementation of MDC and SVC with the proposed features in this paper. REFERENCES
[1] E.E. Azzouz, and A.K.Nandi, Procedure for automatic recognition of analogue and digital modulations, IEE proceedings Communications, vol. 143, no.5, Oct.1996, pp. 259-266

[2] Cheol-Sun Park and Dae Young Kim, A novel robust feature of modulation classification for reconfigurable software radio, IEEE Trans. Consumer Electronics, Nov.2006 (accepted for publication). [3] A.K. Nandi, and E.E. Azzouz, Algorithm for automatic modulation recognition of communication signals, IEEE Trans. Communications, vol. 46, no. 4, pp. 431-436, April 1998. [4] Cheol-Sun Park, Won Jang, Sun-Phil Nah, and Dae Young Kim, Automatic modulation recognition using support vector machine in software radio applications, in Proc. 9th IEEE ICACT, Feb. 2007, pp. 9-12 [5] Cheol-Sun Park, Jun-Ho Choi, Sun-Phil Nah, Won Jang, Automatic modulation recognition of digital signals using wavelet features and SVM, in Proc. 10th IEEE ICACT, Feb. 2008, pp. 387-390 [6] K. C. Ho, W. Prokopiw, and Y. T. Chan, Modulation identification by the Wavelet Transform, in Proc. IEEE MILCOM, Nov. 1995, pp. 886-890 [7] Y. T. Chan and L. Gadbois, Identification of the modulation type of a signal, Signal Processing, vol. 16, no. 2, Feb. 1989. [8] Jiang Yuan, Zhang Zhao-Yang, and Qiu Pei-Liang, Modulation classification of communication signals, in Proc. IEEE MILCOM, Oct. 2004, pp. 1470-1476 [9] Ilan Druckmann et al., Automatic modulation type recognition, in Proc. of IEEE Canadian Conf. on Electrical and Computer Engineering, pp. 65-68, May 1998. [10] Zhilu Wu et al., Automatic digital modulation recognition based on support vector machine, in Proc. of IEEE Conf. on Neural Networks and Brain, pp. 1025-1028, Oct. 2005. [11] Octabvia A. Dobre, Ali Abdi, Yeheskel Bar-Ness and Wei Su, A survey of automatic modulation classification Techniques classical approaches and new trends, IEE Proc. Communication to be published in 2006. [12] Stefan C. Kremer and Joanne Sheils, A testbed for automatic modulation recognition using artificial neural networks, IEEE Canadian Conf. on Electrical and Computer Engineering, pp. 67-70, May 1997. [13] Kyoung Hun Lee, Se Hun Joo, In Ho Hwang, and Jong Tae Lee, Method and apparatus for identifying analog signal or digital signal using DoE filter, in KR1020050046913 A, Feb. 2007.
P P P P

614