Data Analysis for the Automatic

Modulation Classification (AMC)

IQ Samples have been collected using the following devices:
• BB60C, AirT – data capturing devices.
• VSG60A, SMBV100B VSG (R&S) – Data generation devices.
• UWB2, UWB4 antennas.

The following modulations were generated using the VSG60A in two modes:
• 1. Injection Mode.
• 2. Radiation Mode.

Required Modulations:
1. AM
2. FM
3. 2ASK
4. BPSK
5. QPSK
6. 8PSK
7. QAM16
8. QAM64
9. 2FSK
10. 4FSK
11. 8FSK
The Constellation diagrams of these modulation in injection mode are as follows:

Modulation Constellation diagram

AM

FM

2ASK

BPSK
QPSK

8PSK

16QAM

64QAM

2FSK
 4FSK

8FSK

Table 1: Modulation and its constellation diagram

For deep learning classification of the modulation signals there is a prime importance for the
presence of the all the symbols in a given sample length of a digitally modulated RF signal.

Generally, presence of modulation symbols is calibrated using the parameters symbol rate (or) baud
rate, sampling frequency, sample length.

Let “L” be the sample length of the RF signal, “Fs” be the sampling frequency, “Rs” be the
symbols per second (symbol rate).

Number of symbols (M) in a given sample length (L) is given by

 M = L / (Fs / Rs) .... (1)
Fs / Rs --> indicates No of samples per symbol (NSps)
Eq (1) becomes: M= L / NSps .... (2)
As per discussion with DLRL Srinivas sir, the NSps can be around 8 – 10 for constructing a good
constellation of the signal.
The Number of symbols (M) varies with each modulation, and in real time / practical purposes, we
get three cases.
1. If smaller sample length is chosen, then we might not get all the symbols of the modulation for
the model to classify.
2. If longer sample length is chosen, but if the transmitter is sending partial number / repeated
number of symbols for the whole transmission without sending all the unique symbols at least once,
then we can never infer / classify a modulation scheme.
3. A uniform distribution over all the possibilities, this is the case where we assume all the unique
symbols come at least once.

First 2 cases cannot be addressed, but case 3 can be addressed using the classic solution for the
“coupon collector problem”

This problem addresses how many numbers of symbols are required to have all the unique symbols
at least once in our RF signal of a given sample length.
Again, we are assuming uniform distribution and the Expected value is calculated using the
formulae,
M ~ n * ln(n) + 0.5 + gamma * n
where gamma = 0.5772156649 (Eulers-Mascheroni constant)

Table 2: Table indicating the minimum required number of symbols to get the corresponding
probable unique number of symbols for each modulation.

Modulation Unique Expected number Expected Expected number of

type Number of of symbols in a number of symbols in a given
symbols given sample length symbols in a sample length (~99.9%)
(~80-90%) given sample
length (~90-
95%)
2ASK 2 4 8 16
BPSK 2 4 8 16
QPSK 4 6 18 33
8PSK 8 22 41 72
 QAM16 16 55 93 155
QAM64 64 304 458 709
2FSK 2 4 8 16
4FSK 4 6 18 33
8FSK 8 22 41 72

Table 3: Table indicating the Real-world symbol rates per each Modulation type as per web
sources.

Modulation type Practical symbols rates in real world

2ASK Low speed: 300 – 1k (e.g., early modems like
Bell 103)
Industrial / IOT: 10k – 100k (e.g., short-range
telemetry).
High speed: up to 1M in niche RF systems with
wide bandwidth.
BPSK Satellite: 1M – 50M (e.g., GPS uses ~50 baud
for C/A code, but data channels can hit 1
Mbaud).
Wireless: 100k – 10M (e.g., IEEE 802.15.4 at
250 kbps = 250 kbaud).
Deep Space: 10 – 1k (low rates for reliability,
e.g., Voyager at ~160 baud).
QPSK Satellite: 1M – 50M (e.g., DVB-S at 27.5
Mbaud).
Cellular: 100k – 10M (e.g., LTE uses QPSK at
~1–5 Mbaud per subcarrier).
Wi-Fi: 1M – 20M (e.g., 802.11a/g at 12 Mbaud
with OFDM).
8PSK Satellite: 1M – 30M (e.g., DVB-S2 with 8-PSK
at 20–30 Mbaud).
Military / Telemetry: 100k – 10M (e.g., tactical
comms).
Cellular: 1M - 5M (e.g., EDGE in GSM at ~270
kbaud).
QAM16 Wi-Fi: 1M – 54 M (e.g., 802.11n at 20–40
Mbaud with OFDM).
Cellular: 100k – 15M (e.g., LTE at 1–15 Mbaud
per subcarrier).
cable-modems: 1M - 7M (e.g., DOCSIS 2.0 at
5.056941 Mbaud).
QAM64 Wi-Fi: 1M – 54 M (e.g., 802.11ac at 40–80
Mbaud with OFDM).
Cellular: 100k – 15M (e.g., 5G NR at 1–15
Mbaud per subcarrier).
Cable / Satellite: 1M - 30M (e.g., 5G NR at 1–
15 Mbaud per subcarrier).
 2FSK IoT/Radio: 1k – 100k (e.g., Zigbee at 20–40
kbaud).
Legacy Systems: 300 – 10k (e.g., pagers,
amateur radio).
Telemetry: Up to 1 Mbaud in wideband
systems.
4FSK IoT: 10k – 250k (e.g., LoRa at 20–100 kbaud).
Industrial: 1k – 100k (e.g., SCADA systems).
High-speed: Up to 1 Mbaud in specialized RF
links.
8FSK Telemetry: 10k – 200k (e.g., aeronautical
telemetry).
Wireless: 1k - 100k (e.g., custom IoT or M-Ary
FSK systems).
Rare High speed: Up to 500 kbaud in wideband
applications.
Summary Table

Modulation Bits/Symbol Typical Symbol Rate Range (baud) Example Applications

2-ASK 1 300 – 1M Modems, RFID

BPSK 1 10 – 10M Satellite, Bluetooth

QPSK 2 100k – 50M LTE, DVB-S

8-PSK 3 100k – 30M EDGE, DVB-S2

16-QAM 4 100k – 54M Wi-Fi, LTE

64-QAM 6 100k – 54M 5G, DVB-C

2-FSK 1 300 – 1M IoT, Pagers

4-FSK 2 1k – 250k LoRa, SCADA

8-FSK 3 1k – 500k Telemetry

Notes on Practicality
B is bandwidth, α is roll-off factor, B ≈ Rs (1+ α)
1. Bandwidth Limits: Symbol rate is constrained by channel bandwidth. For
α=0.25, 𝐵=1.25*𝑅𝑠. A 1 MHz channel supports 𝑅𝑠 ≈ 800 kbaud.
2. Standards: Rates are often set by protocols (e.g., LTE’s 15 kHz subcarrier spacing scales to
Mbaud with many subcarriers).
3. Trade-offs: Higher-order modulations (e.g., 64-QAM) need lower symbol rates in noisy channels
to maintain error rates, while FSK tolerates noise but uses more bandwidth.
As per discussion with DLRL Srinivas sir, if 64 QAM demodulation is applied across all the lower
modulations and drew a constellation after demodulating the signal then there might be a chance for
classification using CNNs. This requires R&D.

Elsec Modulation Data Analysis:

The Following modulation data is collected by setting the below parameters

Sampling frequency (Fs) = 125e6 Hz
sample length (L) = 16384 samples
Symbol rates (Rs) = 100k, 500k, 1M, 2M, 3M, 5M, 8M
freq_deviations for FSK: 8Mhz, 5Mhz, 4Mhz, 3Mhz, 2.5Mhz, 2Mhz, 1.5Mhz, 1Mhz 500khz,
250khz, 100khz, 50khz
freq_deviations for FM: 3khz, 5khz, 10khz
gammas used for APSK: 2.53, 2.57, 2.84, 3.15
Modulation depth (%) for AM: 25%, 50%, 100%
Filters used: Root cosine for all modulations,
Gaussian fsk filter for fsk modulations

For each symbol rate 500 files are collected using AirT, where each file contains 16384 samples.
APSK, AM, FM don’t use symbol rates as parameters, rather
APSK uses gamma,
AM uses modulation depth (%), and
FM uses frequency deviations for transferring data.

Filter Used: Root cosine

Modulations collected No of samples for each modulation in M (10^6)
Symbol Rates
100k 500k 1M 2M 3M 5M 8M
8PSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
8PSK_EDGE 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
16QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
16QAM_EDGE 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
32QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
32QAM_EDGE 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
 64QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
128QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
256QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
512QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
1024QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
2048QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
4096QAM 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
AQPSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
ASK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 16.384 M
BPSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
MSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
OQPSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
pi4_DQPSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
PI8_D8PSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
QPSK 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
QPSK_45_OFFSET 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M
QPSK_EDGE 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M 8.192 M

100k 500k 1M 2M 3M 5M 8M
Frequency deviations 50 100 250 500 500 1 1 2 1.5 3 2.5 5 4 8
kH kH kH kH kH M M M M M M M M M
z z z z z Hz Hz Hz Hz Hz Hz Hz Hz Hz
2FSK - - - - - - - - - - - - 8.19 8.19
2M 2M
4FSK - - - - - - - - - - - - 8.19 8.19
2M 2M
8FSK - - - - - - - - - - - - 8.19 8.19
2M 2M

Filter Used: Gaussian FSK

100k 500k 1M 2M 3M 5M 8M
Frequency deviations 50 100 250 500 500 1 1 2 1.5 3 2.5 5 4 8
kH kH kH kH kH M M M M M M M M M
z z z z z Hz Hz Hz Hz Hz Hz Hz Hz Hz
2FSK 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1
92 92 92 92 92 92 92 92 92 92 92 92 92 92
M M M M M M M M M M M M M M
4FSK 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1
92 92 92 92 92 92 92 92 92 92 92 92 92 92
M M M M M M M M M M M M M M
 8FSK 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1 8.1
92 92 92 92 92 92 92 92 92 92 92 92 92 92
M M M M M M M M M M M M M M

Filter Used: Root cosine

Gamma 2.53 2.57 2.84 3.15
16APSK - 8.192 M - 8.192 M
32APSK 8.192 M - 8.192 M -

Modulation Depth (%) 25 50 100

AM 16.384 M 16.384 M 16.384 M

Frequency deviation 3 kHz 5 kHz 10 kHz

FM 16.384 M 16.384 M 16.384 M
Table 4: Modulation and respective collected samples for each unique parameter.
Each file contains 16384 IQ samples. So, Sample length L = 16384, Fs = 125e6, using Eq (2) ...

NSps (Number of samples per symbol)

Modulation Symbol Rate
type
100k 500k 1M 2M 3M 5M 8M
2ASK 1250 250 125 62 41 25 15
BPSK 1250 250 125 62 41 25 15
QPSK 1250 250 125 62 41 25 15
8PSK 1250 250 125 62 41 25 15
QAM16 1250 250 125 62 41 25 15
QAM64 1250 250 125 62 41 25 15
2FSK 1250 250 125 62 41 25 15
4FSK 1250 250 125 62 41 25 15
8FSK 1250 250 125 62 41 25 15

Table 5: Modulation and the respective Number of samples per symbols for collected data

In reference to Table 2,
M (Number of symbols)
Modulation Symbol Rate
type
100k 500k 1M 2M 3M 5M 8M
 2ASK 13 65 131 264 399 655 1092
BPSK 13 65 131 264 399 655 1092
QPSK 13 65 131 264 399 655 1092
8PSK 13 65 131 264 399 655 1092
QAM16 13 65 131 264 399 655 1092
QAM64 13 65 131 264 399 655 1092
2FSK 13 65 131 264 399 655 1092
4FSK 13 65 131 264 399 655 1092
8FSK 13 65 131 264 399 655 1092
Table 6: Modulation and the respective Number of symbols for collected data

Green indicates that the symbol rates in this modulation satisfies two conditions:

1. Samples per symbols is between 8 – 10.

2. probability of finding all unique symbols in that respective modulation is above 99.9%.

Yellow indicates that the probability of finding all unique symbols in that respective modulation is
between 90% - 95%.

Red indicates that the probability of finding all unique symbols in that respective modulation is
below 90%.

For the red and yellow symbol rates, if more samples are considered then the probability of finding
the unique symbols increases, currently this analysis is for 16384 samples.

Using AIRT, the max number of samples can be collected is 16384 per file, but to combine one or
more files to compensate for the sample length, there are issues like time synchronization and
frequency synchronization between each file.