# Decoding of binary data (AIS) from socket

I am currently working on a AIS message decoder written in pure Python. It's just a little project to teach me some things about binary data, decoding, etc. Because I am only programming for about a year or so I am not quite sure, if my approaches is reasonable. The full source code be found here: https://github.com/M0r13n/pyais

Without going too deep into the details, my question is: What is a reasonably fast way to decode binary content in pure Python?

Let's say I have a bit sequence with a length of 168 bits and this bit sequence contains encoded data. The data may not be a multiple of 2 and therefore won't fit into typical data structures such as bytes.

I tried three approaches:

1: Store the bin sequence as a normal string and convert each substring into an int individually:

bit_str = '000001000101011110010111000110100110000000100000000000000001010111001111101011010110110000010101101000100010000010011001100101001111111111110110000010001000100010001110'
d = {
'type': int(bit_vector[0:6], 2),
'repeat': int(bit_vector[6:8], 2),
'mmsi': int(bit_vector[8:38], 2),
'ais_version': int(bit_vector[38:40], 2),
'imo': int(bit_vector[40:70], 2),
'callsign': ascii6(bit_vector[70:112]), # custom method of mine, ignore for now
# ...
}


2: Using bitstring's BitArray and slicing:

b = BitArray(bin=bit_vector)
# access each piece of data like so
type_ = b[0:6].int


3: Using the bitarray module:

-> The Bitarray module does not have a nice way to convert individual parts into ints, so I dropped it.

Approach 1 (my current one) decodes #8000 messages in 1.132967184 seconds and the second one takes nearly 3 seconds.

Overall I am fairly happy with my first idea, but I feel like, I could be missing something. My main concern is readability, but the code should not be overly slow. In C I would have used structs, is the ctypes module worth a consideration?

EDIT: It was suggested, that I include all relevant parts of my code here. So here we go:

The overall flow can be summarized like this. First part is the name of the method and the second part in brackets it the new data type, after the method was called:

Socket data(bytes) -> decode("utf-8")(string) -> decode_ascii6()(string) -> decode_msg_1()(dict)

Data is coming from a TCP socket:

def ais_stream(url="ais.exploratorium.edu", port=80):
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((url, port))
while True:
# there may be multiple messages
for msg in s.recv(4096).decode("utf-8").splitlines():
yield msg


Decoding

Each message is then decoded and it's checksum is computed. Because AIS is using 6bit ASCII and not 8bit like normal, we need convert the payload. The checksum is just a logical XOR of all characters. Also there are different message types and each type requires it's own decoding. I simply created a list of the first 25 types and stored the corresponding function in it. Because functions are objects in python, that should not be a problem. And this is a very clean way to express my intention. The to_int and signed methods are just wrappers around Pythons normal int method, that check of empty strings, etc.

def decode_ascii6(data):
"""
Decode AIS_ASCII_6 encoded data and convert it into binary.
:param data: ASI_ASCII_6 encoded data
:return: a binary string of 0's and 1's, e.g. 011100 011111 100001
"""
binary_string = ''

for c in data:
c = ord(c) - 48
if c > 40:
c -= 8
binary_string += f'{c:06b}'

return binary_string

def checksum(msg):
"""
Compute the checksum of a given message
:param msg: message
:return: hex
"""

c_sum = 0
for c in msg[1::]:
if c == '*':
break
c_sum ^= ord(c)

return c_sum

def decode(msg):
m_typ, n_sentences, sentence_num, seq_id, channel, data, chcksum = msg.split(',')
# convert normal ASCII into AIS_ASCII_6
decoded_data = decode_ascii6(data)
msg_type = int(decoded_data[0:6], 2)

if checksum(msg) != int("0x" + chcksum[2::], 16):
print(f"\x1b[31mInvalid Checksum dropping packet!\x1b[0m")
return None

if n_sentences != '1' or sentence_num != '1':
print(f"\x1b[31mSentencing is not supported yet!\x1b[0m")
return None

if 0 < msg_type < 25:
# DECODE_MSG is a list of functions
return DECODE_MSG[msg_type](decoded_data)

return None


Actual AIS Message Decoding

As I said, each message types is different and so needs to be treated individually. I decided to decode each message and return a dictionary containing all relevant information. Because 90% of the data is numeric, I also convert every chunk of data into an integer. Some information needs additional context. For example the NAVIGATION_STATUS, where each number has a special meaning. Therefore I initialize a Dictionary when the script is called, which serves as a lookup table. The lookup time is O(1) and the code gets very readable. See the decoding of message type 1 as an example:

def decode_msg_1(bit_vector):
"""
AIS Vessel position report using SOTDMA (Self-Organizing Time Division Multiple Access)
Src: https://gpsd.gitlab.io/gpsd/AIVDM.html#_types_1_2_and_3_position_report_class_a
"""
status = to_int(bit_vector[38:42], 2)
maneuver = to_int(bit_vector[143:145], 2)
return {
'type': to_int(bit_vector[0:6], 2),
'repeat': to_int(bit_vector[6:8], 2),
'mmsi': to_int(bit_vector[8:38], 2),
# ...
'lon': signed(bit_vector[61:89]) / 600000.0,
'lat': signed(bit_vector[89:116]) / 600000.0,
'course': to_int(bit_vector[116:128], 2) * 0.1,
# ...
'maneuver': (maneuver, MANEUVER_INDICATOR[maneuver]),
}


I would love to get some feedback and tips from you! :-)

• Did you also have a look at the struct module in the standard library? Commented Oct 6, 2019 at 11:27
• Yes, but it does not seem to offer the possibility to extract data of arbitrary Width, e.g. 6 bits or 17 bits. Commented Oct 6, 2019 at 11:31
• True, but that also seems to be true of both your methods, at least of the method one. Commented Oct 6, 2019 at 11:32
• Or rather, there can be only one variable-length field, the last (with b[123:]). Commented Oct 6, 2019 at 11:33
• Please include all relevant code here.
– TomG
Commented Oct 6, 2019 at 16:43

## Avoid stringly-typed data

Given this sequence of operations:

Socket data(bytes) -> decode("utf-8")(string) -> decode_ascii6()(string) -> decode_msg_1()(dict)


and based on the specification for AIS, UTF-8 is not an appropriate decode format. ASCII should be used instead - it's a fixed-width character format, and from what I read, none of the characters in the format deviate from that.

I discourage your approach in #1. Binary data are binary data; having any concern for performance rules out a string-based representation.

## Generators and message boundaries

You should probably have a more nuanced approach to buffering. You should only yield after ensuring that your buffer ends on the boundary of a message, delimited by an ASCII newline byte. Another 'gotcha' is that recv only returns up to the number of bytes you specify, so you'll need to have logic to measure the returned result and deal with this.

The sanest way to represent this is two different generator functions. The first asks for 4096 bytes, yields chunks of one line at a time, and maintains a local buffer between loop iterations of incomplete lines.

The second generator function iterates over the result of the first generator function and does the actual unpacking into structured message objects.

## Modularize your file

There's enough going on here - you have a bunch of documentation, a bunch of functions, etc. - that it's worth breaking out your file into multiple files in a module. It's still executable from the outside via python -m. I've made a pull request doing only this.

• Thank you for you comments and improvements. I am already thinking and trying around to avoid stringly-typed data. I hope that I can get it to work soon. I don't think that there are two generators needed, because the AIS spec does not allow for a message to be longer than 4096 bytes. At least I did not find any exception. The modularization is definitely nice and will help other to gain a quicker overview! Commented Oct 14, 2019 at 8:07
• The problem is not that a message may be longer than 4096 bytes - the problem is that one message may be split across the buffer boundary if many messages are received at once. You're using TCP, so there is no guarantee of predictable packet size. Commented Oct 14, 2019 at 13:27
• I see. Now that we are working with raw bytes, values can be extracted by bitwise operations like so: raw_bytes[0] & 0x3f to get the first 6 bits. That should be pretty fast, doesnt it? Commented Oct 14, 2019 at 19:18
• I think so :) That being said, I think your concern for performance is a little unbalanced. I ran the code against your default server and the rate of message reception is very low. You should be optimizing for correctness and simplicity first. Commented Oct 14, 2019 at 19:20
• Thats definitly right, but my primary concern is to learn something about working with binary data. Otherwise I would use a c++-based library such as libais. Thanks for your great support! Commented Oct 14, 2019 at 20:28