Channel impulse response peak clustering using neural networks

Abstract

This paper introduces an approach to process channel sounder data acquired from Channel Impulse Response (CIR) of 60GHz and 80GHz channel sounder systems, through the integration of Long Short-Term Memory (LSTM) Neural Network (NN) and Fully Connected Neural Network (FCNN). The primary goal is to enhance and automate cluster detection within peaks from noised CIR data. The study initially compares the performance of LSTM NN and FCNN across different input sequence lengths. Notably, LSTM surpasses FCNN due to its incorporation of memory cells, which prove beneficial for handling longer series.Additionally, the paper investigates the robustness of LSTM NN through various architectural configurations. The findings suggest that robust neural networks tend to closely mimic the input function, whereas smaller neural networks are better at generalizing trends in time series data, which is desirable for anomaly detection, where function peaks are regarded as anomalies.Finally, the selected LSTM NN is compared with traditional signal filters, including Butterworth, Savitzky-Golay, Bessel/Thomson, and median filters. Visual observations indicate that the most effective methods for peak detection within channel impulse response data are either the LSTM NN or median filter, as they yield similar results.

Full text

The paper has been presented at the 2023 International Conference on Advanced Communication Technologies and Networking (CommNet), Rabat, Maroco,
December 11-13, 2023, https://doi.org/10.1109/CommNet60167.2023.10365265
This research was funded in part by the National Science Center (NCN), Poland, grant no. 2021/43/I/ST7/03294 (MubaMilWave). For this purpose of Open
Access, the author has applied a CC-BY public copyright license to any Author Accepted Manuscript (AAM) version arising from this submission.
Channel Impulse Response Peak Clustering Using
Neural Networks
Petr Horky1, Ales Prokes1, Radek Zavorka1, Josef Vychodil1, Jan Marcin Kelner2, Cezary Henryk Ziołkowski 2, Aniruddha Chandra3
1 Department of Radio Electronics, Brno University of Technology, Brno, Czech Republic
2 Institute of Communications Systems, Faculty of Electronics, Military University of Technology, 00-908 Warsaw, Poland
3 National Institute of Technology, Durgapur, India
email: 244016@vutbr.cz
Abstract—This paper introduces an approach to process
channel sounder data acquired from Channel Impulse Response
(CIR) of 60GHz and 80GHz channel sounder systems, through the
integration of Long Short-Term Memory (LSTM) Neural
Network (NN) and Fully Connected Neural Network (FCNN). The
primary goal is to enhance and automate cluster detection within
peaks from noised CIR data. The study initially compares the
performance of LSTM NN and FCNN across different input
sequence lengths. Notably, LSTM surpasses FCNN due to its
incorporation of memory cells, which prove beneficial for
handling longer series.
Additionally, the paper investigates the robustness of LSTM
NN through various architectural configurations. The findings
suggest that robust neural networks tend to closely mimic the
input function, whereas smaller neural networks are better at
generalizing trends in time series data, which is desirable for
anomaly detection, where function peaks are regarded as
anomalies.
Finally, the selected LSTM NN is compared with traditional
signal filters, including Butterworth, Savitzky-Golay,
Bessel/Thomson, and median filters. Visual observations indicate
that the most effective methods for peak detection within channel
impulse response data are either the LSTM NN or median filter,
as they yield similar results.
Keywords— LSTM, FCNN, DBSCAN, anomaly detection,
clusters, peak detection, channel impulse response
I. INTRODUCTION
In recent times, the concept of Artificial Intelligence (AI) has
gained significant prominence, with continual exploration of
novel domains for its application and utilization. AI is a very
broad term and one of the branches of AI is Machine Learning
(ML). Using machine learning algorithms can lead to very
convenient results and it is a useful tool to be used in algorithms
because the process of evaluating data can be described
mathematically. A specific group of ML algorithms is Neural
Networks, which if they reach a certain complexity are called
Deep Learning (DL) [1]. The main issue associated with Neural
Networks and Deep Learning is that they are black boxes. In
other words, the results are not known before the NN is trained
and evaluated. Hence, experimenting with neural network
algorithms in various scenarios is important, as it could produce
promising results.
In this particular case, the objective is to automate the
process of detecting peaks in a CIR obtained from a 60GHz
channel sounder, with 5GHz bandwidth, used for the intra-car
channel measurement[2]. The output of the channel
measurement is denoted as relative power, representing the ratio
of the reflected signal power to the noise, expressed in decibels.
To conduct a comprehensive evaluation of the methods and
test their applicability in various measurements an 80GHz
channel sounder, which operates within a bandwidth of 1GHz
and captures data in the time domain, is also employed. The
measurement scenario in this case involved monitoring
reflections from a frozen field. The reflections from objects with
amplitudes above the noise background can be considered as
anomalies. The insights can subsequently be applied in other
channel measurement applications such as detection [3].
This paper explores the potential of contemporary Deep
Learning techniques for anomaly detection. To achieve this
objective, FCNN and LSTM neural networks are employed. The
paper is structured in two main sections. The first section
outlines the neural network architecture for peak detection,
while the second section compares the neural network results
with results achieved using filters such as Butterworth,
Savitzky-Golay, Bessel/Thomson, and the median filter.
II. DETECTION OF ANOMALIES USING NEURAL
NETWORKS
A. Basic principles
The core concept of utilizing NNs for anomaly detection is
based on training the network to learn the underlying pattern or
curve without relying on mathematical models. In this particular
case, an autoencoder is used. An autoencoder is a special type of
neural network that is trained to copy its input to its output. For
example, given an image of a handwritten digit, an autoencoder
first encodes the image into a lower dimensional latent
representation, then decodes the latent representation back to an
image. An autoencoder learns to compress the data while
minimizing the reconstruction error [4]. Through the use of the
autoencoder, it becomes possible to mitigate the effects of a
noise environment. Moreover, the autoencoder does not respond
to sudden, rapid changes; rather, it captures the overall trend of
the entire function.

B. Long short-term memory
Long short-term memory networks are a special kind of
RNN that can learn long-term dependencies [5]. They find
application in various domains, such as predicting time series in
medical contexts, as seen in ECG signal analysis [6], [7],
forecasting stock market trends [8], or interpreting data outputs
from sensors [9].
C. Fully connected Neural network
Additionally, FCNN has been used to compare its results
with the LSTM neural network. A similar FCNN has been
previously used in an intra-vehicle scenario [10]. Unlike the
LSTM, the FCNN does not consider prior values, which may
lead to slightly less accurate outputs. But this attribute
paradoxically contributes to anomaly detection, because the
anomalies are determined by subtracting the predicted values
from the input values.
III. MEASUREMENT
The channel impulse response from the channel sounder
consists of IQ samples [2]. The data is initially transformed into
the time domain, which then serves as the input dataset. To
prevent the gradient explosion while training the NN, data are
scaled with a standard scaler, described in a formula:
 󰇛󰇜
󰇛󰇜 ()
where Xstandardized is the standardized version of the feature X, X
is the original feature, mean(X) is the mean of the feature X, and
std(X) is the standard deviation of the feature X.
In this case, a supervised learning method is employed,
meaning that the samples and targets are known. The data is then
divided into segments, and a sliding window is used to create the
sample dataset and target dataset. The method is described with
the formula below:
  󰇟     󰇠
   (2)
where i is the position of sliding windows and k is the number
of input samples.
To provide a more accurate description of this process,
Figure 1 illustrates an example of the sliding window with 4
input samples.
Fig. 1. Example of sliding window with k = 4
The input dataset is afterwards divided into testing sets, each
with different sequence lengths of 2, 20, 50, and 100 samples.
A. Input sequence length evaluation
To analyse the impact of the input sequence length on
channel modelling, neural networks with architectures shown in
TABLE I and TABLE II are used. The FCNN architecture
consists of dense layers, rectified linear activation functions, and
dropout functions, which are implemented to prevent overfitting
by deactivating a percentage of neurons during training. To
properly assess LSTM, all parameters are kept identical to those
in FCNN, except for the substitution of dense layers with LSTM
layers. Both neural networks share an autoencoder architecture,
featuring higher-dimensional layers at the start and end, with a
bottleneck in the middle.
TABLE I. FCNN
Neural
network
architecture
dense 64, relu, dropout 0.2
dense 32, relu, dropout 0.2
dense 16, relu,
dense 32, relu, dropout 0.2
dense 64, relu, dropout 0.2
time distributed dense 1
TABLE II. LSTM
Neural
network
architecture
lstm 64, relu, dropout 0.2
lstm 32, relu, dropout 0.2
flatten
repeat vector 16
lstm 32, relu, dropout 0.2
lstm 64, relu, dropout 0.2
time distributed dense 1
The primary distinction between LSTM and FCNN is that
FCNN lacks memory cells. Notably, the FCNN yielded good
results only with the shortest input sequences, failing to produce
satisfactory outcomes with longer input sequences as shown in
Figure 2.
FCNN with varying sequence length
In contrast, the LSTM NN consistently delivered good
results with all input lengths as shown in Figure 3. For shorter
input sequences, the predicted series closely matches the input
sequence. With longer input sequences, the results tend to
replicate only the trend of the function, which is desirable for
identifying anomalies in the series.

Fig. 2. LSTM with varying sequence length
To conclude the results, the LSTM NN achieved consistent
results and therefore it will be used for further analysis of
choosing the correct network architecture.
B. LSTM architecture
For the upcoming tests, LSTM NN with input sequence of
100 samples is utilized. The sequence of 100 samples yielded
the best results in predicting trends rather than replicating the
function. The NN is initially trained with architecture in TABLE
III and then for each measurement, all of the layers are halved to
explore how the network's robustness impacts its performance.
Smaller neural networks often exhibit good data generalization
capabilities but may struggle to represent the data accurately.
Conversely, robust neural networks can effectively learn
underlying patterns, although this can sometimes result in
overfitting.
TABLE III. LSTM
Neural
network
architecture
lstm 64, relu, dropout 0.2
lstm 32, relu, dropout 0.2
flatten
repeat vector 16
lstm 32, relu, dropout 0.2
lstm 64, relu, dropout 0.2
time distributed dense 1
Fig. 3. LSTM of 60GHz channel sounder, with varying NN architecture
The results from Figure 4 indicate that when the size of the
neural network decreases, the output values follow the trend of
the function rather than replicating the function itself. To further
observe the behavior of the NN architecture, a second test was
conducted using an 80GHz channel sounder system for
measuring a frozen field scenario. The results shown in Figure 5
demonstrate that chosen LSTM architercure is convenient for
processing data from different CIR measurements
Fig. 4. LSTM of 80GHz channel sounder, with varying NN architecture
The drawback of using neural networks is that they need to
be trained for each signal impulse response. The training process
consisted of 20 epochs, with each epoch requiring 2 seconds for
training.
IV. ANOMALY DETECTION
A. Input-Output analysis
To detect anomalies, we calculate the distance between each
two points in a graph. Using mean square error is not suitable
for this application because it includes negative peaks. The
solution for this calculation is a simple subtraction of the
predicted function from the input function.
  󰇛󰇜 󰇛󰇜 (3)
The graph of the distances between the points is shown in
Figure 6. In this picture, the peaks are clearly visible. To fully
automate the process of identifying peaks or even entire clusters,
a thresholding method needs to be applied.
B. DBSCAN
To separate anomalies from the noise a clustering method
Density-Based Spatial Clustering of Applications with Noise
(DBSCAN) from scikit learn library [11] with default
parameters will be used. DBSCAN algorithm defines clusters as
continuous regions of high density. For each instance, the
algorithm counts how many instances are located within a small
distance from it [1]. In the following figures, the graphs of the
original and predicted data are displayed. Additionally, there is
a figure showing the calculated distances between the functions.
The distinguishing colors in the graph below represent
anomalies obtained from the default settings of the DBSCAN
function, with the min_samples value set to two.

Fig. 5. CIR 1 - Detected anomalies after applying NN and DBSCAN
Fig. 6. CIR 2 - Detected anomalies after applying NN and DBSCAN
Fig. 7. CIR 3 - Detected anomalies after applying NN and DBSCAN
The results using LSTM NN and DBSCAN show that this
method is effective for detecting peaks and assigning clusters in
a noise environment, however, this method experiences notable
errors within the first 100 samples of the graph as shown in
Figure 6 and Figure 8. These discrepancies might be caused by
the initial lack of context at the beginning of the graph
prediction.
V. COMPARISON TO SIGNAL FILTERS
To further validate its effectiveness, the results from LSTM
NN presented in TABLE IV are compared to commonly used
filters from the SciPy library [12]. In this case, Butterworth,
Savitzky-Golay, Bessel/Thomson, and median filters were
verified.
TABLE IV. LSTM ARCHITECTURE
Neural
network
architecture
lstm 8, relu, dropout 0.2
lstm 4, relu, dropout 0.2
flatten
repeat vector 2
lstm 4, relu, dropout 0.2
lstm 8, relu, dropout 0.2
time distributed dense 1
To achieve desirable results, it is necessary to correctly
configure filter parameters. As there is no automated method for
parameter configuration, they were chosen manually, beginning
with conservative values, and subsequently fine-tuned to
visually obtain the best results. This manual selection of
parameters contrasts with the objective of automation. The
selected parameters are listed in TABLE V. The Following
figures illustrate the impact of various filters on peak cluster
detection.

TABLE V. FILTER PARAMETERS
Savitzky-Golay
Window length = 100, Polynomial order = 5
Butterworth
Order of the filter = 10, Cutoff frequency = 0.04*Fs,
Filter type = Lowpass
Bessel/Thomson
Order of the filter = 4, Cutoff frequency = 0.1*Fs,
Filter type = Lowpass
Median filter
Input array length = 100
Fig. 8. CIR 1 - Comparison of LSTM NN and other filters
Fig. 9. CIR 2 - Comparison of LSTM NN and other filters

Fig. 10. CIR 3 - Comparison of LSTM NN and other filters
The results show that the highest precision is achieved when
using DBSCAN with LSTM NN and Median filter. Although
some filters produce favourable outcomes, they are inconsistent
in their results. It is also notable, that LSTM NN results vary
from the previous measurement, despite the same neural
network architecture. This is due to the fact that NN can achieve
different results with each training cycle These findings
indicate that while neural networks may be suitable for this
application, further testing is required. The primary area for
improvement lies in integrating more input data or exploring
combinations of digital filters and neural networks.
VI. ACKNOWLEDGEMENT
The research was conducted within the framework of a
research grant project funded by the Czech Science Foundation,
under Lead Agency Project No. 22-04304L, Mutli-band
prediction of millimetre-wave propagation effects for dynamic
and fixed scenarios in rugged time-varying environments.
VII. CONCLUSION
This study deals with the application of Deep Learning
techniques, specifically Long Short-Term Memory (LSTM) and
Fully Connected Neural Networks (FCNN), for the purpose of
anomaly detection in millimeter wave channel impulse
response. The core principle involves training neural networks
to learn trends of the reflected signals impulse response and to
automate the process of peak detection.
The research reveals that LSTM neural networks
outperformed FCNN in scenarios of predicting signal impulse
response. The LSTM demonstrated robustness in predicting
trends. In contrast, FCNN exhibited limitations with longer
input sequences, leading to less accurate results.
The study also employs the DBSCAN clustering method to
further distinguish anomalies from noise, showing that the
combination of LSTM and DBSCAN is effective in detecting
peaks in noise environments.
Furthermore, the results given by the LSTM neural network
and DBSCAN are compared with traditional signal filters like
Butterworth, Savitzky-Golay, Bessel/Thomson, and median
filters. The deployment of these filters comes with a manual
selection of parameters such as cutoff frequency, filter order,
and window length, which contradicts the goal of automation.
Based on the visual comparison, the most effective methods for
peak detection are either median filtering or the use of LSTM
neural networks. However, it's important to note that a drawback
of neural networks is the necessity to train them for each impulse
response, which consumes approximately 40 seconds.
Additionally, due to the training process, there might be slight
variations in the results with each new training session.
The further scope of research could divide into two groups.
firstly, exploring the application of the latest neural network
architectures or creating ensemble models that combine neural
networks with machine learning models to achieve improved
output trend representations and secondly, examining novel
methods for evaluating the anomalies from the CIR input and its
calculated trend function.

REFERENCES
[1] A. Géron, Hands-on machine learning with
Scikit-Learn, Keras, and TensorFlow, Second
edition. in concepts, tools, and techniques to
build intelligent systems. Beijing: O’Reilly,
2019.
[2] A. Prokes, J. Vychodil, M. Pospisil, J.
Blumenstein, T. Mikulasek, and A. Chandra,
“Time-domain nonstationary intra-car channel
measurement in 60 GHz band,” in 2016
International Conference on Advanced
Technologies for Communications (ATC),
2016, pp. 1–6. doi:
10.1109/ATC.2016.7764753.
[3] A. Prokes et al., “Multipath Propagation
Analysis for Vehicle-to-Infrastructure
Communication at 60 GHz,” in 2019 IEEE
Vehicular Networking Conference (VNC),
2019, pp. 1–8. doi:
10.1109/VNC48660.2019.9062771.
[4] Martín~Abadi et al., “ TensorFlow: Large-
Scale Machine Learning on Heterogeneous
Systems.” 2015. [Online]. Available:
https://www.tensorflow.org/
[5] S. Hochreiter and J. Schmidhuber, “Long short-
term memory,” Neural Comput, vol. 9, no. 8,
pp. 1735–1780, 1997.
[6] Zabir Al Nazi, Ananna Biswas, Tasnim Azad
Abir, and Md. Abu Rayhan, “Classification of
ECG signals by dot Residual LSTM Network
with data augmentation for anomaly
detection,”
[7] J. Hu, W. L. Goh, and Y. Gao, “Classification
of ECG Anomaly with Dynamically-biased
LSTM for Continuous Cardiac Monitoring,” in
2023 IEEE International Symposium on
Circuits and Systems (ISCAS), IEEE, May
2023, pp. 1–5. doi:
10.1109/ISCAS46773.2023.10181690.
[8] Y. Guo, “Stock Price Prediction Based on
LSTM Neural Network: The Effectiveness of
News Sentiment Analysis,” in Proceedings -
2020 2nd International Conference on
Economic Management and Model
Engineering, ICEMME 2020, Institute of
Electrical and Electronics Engineers Inc., Nov.
2020, pp. 1018–1024. doi:
10.1109/ICEMME51517.2020.00206.
[9] Anran Huang, Yingming Lyu, Zicheng Guo,
and Xiangyang Zhao, “A Temperature and
Humidity Compensation Method for On-board
NOx Sensors with LSTM Network,” IEEE,
2020.
[10] R. Shukla et al., “Deep Learning based Power
Delay Profile Trend Generation: A 60 GHz
Intra-Vehicle Case Study,” in 2022 IEEE
International Symposium on Antennas and
Propagation and USNC-URSI Radio Science
Meeting (AP-S/URSI), 2022, pp. 1–2. doi:
10.1109/AP-S/USNC-
URSI47032.2022.9887316.
[11] F. Pedregosa et al., “Scikit-learn: Machine
Learning in Python,” 2011. [Online].
Available: http://scikit-learn.sourceforge.net.
[12] P. Virtanen et al., “SciPy 1.0: Fundamental
Algorithms for Scientific Computing in
Python,” Nat Methods, vol. 17, pp. 261–272,
2020, doi: 10.1038/s41592-019-0686-2.