IEEE - Aerospace and Electronic Systems - June 2022 - 10

Using MLSTM and Multioutput Convolutional LSTM Algorithms for Detecting Anomalous Patterns in Streamed Data of Unmanned Aerial
with trainable filters to produce feature maps in the first clayer.
A layer ofconnection weights is included in each filter.
Typically, every four pixels in the feature map form a group.
These pixels produce additional feature maps in the first slayer
when passed through a sigmoid function. This procedure
carries on in the consequent c-layers and s-layers. The
feature-mapping layers share the weights and form a plane. It
should also be noted that the filters in this model are used to
connect a series ofoverlapping receptive fields and transform
the 2-D raster batch input into a single unit in the output.
Figure 5 shows a conceptual schematic diagram ofa standard
CNN (adapted from Liu et al. [14]).
Training the CNN is similar to training a standard neuFigure
4.
Reshaping the data for ConvLSTM as a 3-D tensor.
CONVOLUTIONALLSTM
The ConvLSTM is a combination of two techniques: 1) convolutional
neural networks (CNN) and 2) LSTM. CNN is a
type of neural networks with a grid-like topology. It is used
to process 2-D data, such as images and videos. In CNN, the
core operation is the convolution operation. Convolution is a
mathematical operation on two functions ofreal-valued arguments,
anditisusedtoproduceathird function. Theconvolution
operation replaces matrix multiplication, which is timeconsuming,
causing the number of weights to decrease and
consequently reducing the complexity of the network (Liu
et al. [14]) (see (10) for a 1-D convolution)
stðÞ ¼
X1
t¼1
Ht ¼ ot tanhCtðÞ(15Þ
CNN consists of two different types of layers, which are
connected interchangeably: convolution layers (c-layers) and
subsampling layers (s-layers). The input matrix is convolved
where Ut is the input, Ct is the Memory cell output,
Ht is the hidden state, and denotes the convolution
(15)
x tðÞwt tðÞ:
(10)
ral network using backpropagation; where in the first stage
of the training, information is propagated in the feed-forward
direction through different layers. Leading features
are obtained by applying digital filters at each layer. The
values of the output are then computed. During the second
stage, the error between the expected and actual values
of the output is calculated. Backpropagation tends to
minimize this error. The weight matrix is further adjusted,
and the network is thus fine-tuned (Liu et al. [14]).
The input for ConvLSTM is a matrix whose elements are
vectors. Each vector contains the current input in addition
to the last past input values. Equations (11)-(15) (adapted
from [21]) are used to meet the automation ofConvLSTM
it ¼ s WxiUt þWhiHt1 þWciCt1 þ biðÞ
(11)
ft ¼ s WxfUt þWhfHt1 þWcfCt1 þ bf
(12)
Ct ¼ ftCt1 þ it tanhWxcUt þWhcHt1 þ bcðÞ
(13)
ot ¼ s WxoUt þWhoHt1 þWcoCt þ boðÞ
(14)
Figure 5.
Conceptual schematic diagram ofa standard CNN (adapted from Liu et al. [14]).
10
IEEE A&E SYSTEMS MAGAZINE
JUNE 2022
IEEE - Aerospace and Electronic Systems - June 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - June 2022
