IEEE Consumer Electronics Magazine - January/February 2022 - 54

Special Section on Security and Sustainability Technologies for Smart Cities
Figure 4. Distance matrix for Simulated CAN Data.
then corresponds to a probability distribution,
and EM-GMM tries to discover their parameters
(mean and covariance). Although this algorithm
requires prior knowledge of the number of clusters,
we chose to use it because it can help to
investigate if a particular CAN ID belongs with
multiple vehicular functions.
EVALUATION
We implemented the four clustering approaches
in Python with the numpy and scipy
libraries, and with the PyClustering tool.10 We
evaluate our approach using 2 CAN data sets: simulatedCAN
data and real J1939.
Simulated CAN Data
We conducted a feasibility study using a simulated
CAN data log with 10 different CAN IDs, simulating
an acceleration mode and a braking mode.
This data log hadmessages with different frequencies
and message intervals to simulate real data.
There were five IDs specifically related to acceleration
and another four IDs related to braking, with
one message that was transmitted during both
modes. This data was used to verify and validate
our hypothesis that clustering can be effective.
GiventhatrawCANdataisdependent on the vehicle
it was captured from, our simulated data gives
us a baseline to understand the workings of our
algorithms when applied toward raw captured
CAN data. Simulated data was generated using
Vector CANoe and has complete information
about the relationship between messages and
vehicle functions.
Figure 4 shows the distance matrix for this
data. We passed this matrix into the AH clustering
algorithm as well as a dendrogram creator. A
dendrogram visualizes hierarchical clustering in
a tree-like structure that records the sequences
of merges of the leaves, in this case, the CAN
messages. The dendrogram (see Figure 5) is
54
Figure 5. Dendrogram for Simulated CAN Data. The
two major groupings for the CAN IDs are easily
distinguished visually.
created from the leaves first to create the cluster
tree. It shows that there are two clusters of the
CAN IDs, which is expected since we simulated
two separate driving modes of braking and accelerating.
Our clustering algorithm confirmed the
dendrogram results. CAN IDs 1-5 and 8 were designated
as accelerating and were clustered into
one group, and CAN IDs 6, 7, 9, and 10 were designated
as braking and clustered into another
group. The simulated data was used to confirm
our hypothesis and approach before applying
clustering to real CAN data.
Real J1939 Data
To address the chicken-and-egg problem of
incomplete information about the ground truth
that our clustering algorithms are trying to discover,
we obtained CAN data that followed the
J1939 standard to examine the performance of our
approach. The Society of Automotive Engineers
standard SAE J1939 is the vehicle bus standard
based on CAN that has been widely adopted by
diesel engine manufacturers for use in large tractors
and trucks.11 This standard defines that all
J1939 packets, except for the request packet,
should contain eight bytes of data and a standard
header which contains a Parameter Group Number
(PGN), that is embedded in the message's 29bit
identifier. A PGN identifies a message's function
and associated data. J1939 attempts to define
standardPGNs to encompass a wide range of automotive
and other vehicle purposes. PGNs define
the message functions so we validate the clustering
algorithm results by using the standard as a
basis for ground truth.
The J1939 data set was captured from a John
Deere tractor as it performed tillage operations
at different sites. This data set is raw, has no
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - January/February 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2022
