Ultrasonic Large Scenario Model (ULSM): Vector Embedding

System for Ultrasonic Echo Wave Characteristics

Shafait Azam

, Mashnunul Huq

and Andreas Pech

Department of Computer Science and Engineering, Frankfurt University of Applied Sciences, Frankfurt am Main, Germany

Keywords: Ultrasonic Sensor, Vector Embeddings, Residual Neural Network, Signal Processing, Human Computer

Interaction, Pattern Recognition, Transfer Learning.

Abstract: Ultrasonic sensors emitting ultrasound waves can be effectively used in Human Computer Interaction (HCI)

to assist visually disabled humans. With the embedding of the sensor echoes into assistive tools, real-time

spatial awareness for mobility is enhanced. Moreover, material identification aids object recognition by

detecting different materials through their echo signatures. In this article, we study the use of ultrasonic

sensors in HCI systems focusing on their ability to detect materials by analysing the ultrasonic wave

characteristics. These services aim to improve the autonomy and security of people with visual impairments,

offering a complete assistive solution for daily navigation and interaction processes. We have planned to

create a vector database for storing these embeddings generated from reflected waves of various materials

and objects. In this work, we propose a precise vector embeddings generation framework for ultrasonic

systems using ResNet50 convolutional neural network. In the future, Generative AI will use these embeddings

to serve a range of applications for greater autonomy and safety, providing an assistive travel and interaction

solution for the visually impaired.

1 INTRODUCTION

Ultrasonic sensors are based on ultrasonic waves and

can be used to determine distance, presence, or even

the level between objects. Ultrasonic sensors have the

advantage of working in harsh conditions. Working

principle of ultrasonic Sensor is that it functions like

the basic sound wave propagation (acoustical

wavelength, sound reflection) logic at a frequency in

the ultrasonic region. That’s why ultrasonic waves

show similar characteristics as sound in time domain.

In ultrasonic non-destructive testing (NDT), the

transmission of high-frequency sound waves in a

material is used to determine the characteristics of

that material such as surface information, orientation

etc (Matz, 2006) (Taiju Shibata, 2001) . The reflected

waves from that material can significantly distinguish

the unique properties of that material. Not only that,

Ultrasonic sensors mounted with embedded systems

can also calculate the distance of that material.

https://orcid.org/0009-0005-4161-0184

https://orcid.org/0009-0006-5401-8259

https://orcid.org/0000-0003-4985-3437

In recent studies, NDT was able to figure out the

characteristics of core internal graphite blocks (Taiju

Shibata, 2001). The key idea behind this complex

method is to rely on the reflected waves bounced back

from a surface and examine them precisely. These

reflected waves show different characteristics for

different materials. In porous ceramics the

propagation characteristics of ultrasonic waves is

quite unique because of their porous shape. In these

cases, the relationship between wave velocity and

porosity should be taken into account (Taiju Shibata,

2001). Because of these unique characteristics,

ultrasonic signals or reflected echo signals must be

preprocessed. Ultrasonic signals can be impacted by

several key factors, including the frequency and

bandwidth of the signal, the path and distance of the

inspection, the position and size of the material, and

the properties of the material (Pagodinas, 2002). To

detect various materials and create discrepancies

between them several signal processing techniques

are used. Some of these techniques are implemented

704

Azam, S., Huq, M. and Pech, A.

Ultrasonic Large Scenario Model (ULSM): Vector Embedding System for Ultrasonic Echo Wave Characteristics.

DOI: 10.5220/0013251600003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 704-711

ISBN: 978-989-758-730-6; ISSN: 2184-4313

on the embedded system or the hardware and some of

them are done after post echo generation. Ultrasonic

waves tend to scatter, so the efficacy of these

techniques in detecting reflected echoes in materials

shows high attenuation of waves. Some of the simple

signal processing techniques can be analog signal

filtering, transducer damping, pulse shaping, clipping

the whole signal and adjusting it, amplitude

controlling, applying filters for noise reduction,

improving the spatial resolution of multiple

reflections, etc (Pagodinas, 2002). Some of the signal

preprocessing techniques also includes shifting the

signal from time domain to frequency domain. In our

case we have captured the ultrasound signal and

generated spectrogram from that signal. Because we

have realized that extracting echo wave

characteristics in frequency domain is effective to

generate vector embeddings. Because different

materials have different echo wave characteristics

which can be utilized in pattern recognition.

Figure 1: Ultrasound signal reflected from plastic.

Figure 1, shows the whole ultrasound signal

which is preprocessed where X-AXIS symbolizes

Time and Y-AXIS symbolizes Amplitude. we have

generated spectrogram from it. Spectrogram holds the

frequency properties of a specific material.

Based on the properties of ultrasonic signals HCI

systems can be built. To ensure the quality of life for

disabled people smart wheelchairs can be beneficial

(Sanders, 2021). Not only this, by introducing the

power of recent technologies such as Generative AI

any message can be generated and transmitted to the

visually impaired person in correspondence with

his/her surroundings. The existing method to assist a

visually impaired person is called a white cane or a

walking stick. It detects its surroundings with the help

of sensors and tries to alert people (Nayan, 2016).

people have a vision impairment which can be either

partial or full. White cane can help people but it may

get confused in an unfamiliar environment. Our focus

is to build a system that can efficiently serve blind

people and generate alert messages to the visually

impaired so that they can have a proper understanding

of the situation. Human actions include gestures and

movements in an environment. A system for human

activity recognition is meant to recognize these

activities in the context of real-life situations so that

we can understand what lies beneath them. The field

of contextual information detection has been

attracting intense interest among numerous

researchers for decades due to its applications in HCI

(Ghosh, 2023).

In the area of signal processing, Recurrent Neural

Networks (RNN), Long-Short Term Memory

(LSTM) (Zhou, 2022), and transformers are widely

used. Speech, text, audio and music can be generated

with the help of Generative AI. Transformers are the

building blocks of Gen AI consisting of multilayer

encoder and decoder that works with the attention

mechanism (Vaswani, 2017).

This paper initializes and demonstrates the first

phase of the Ultrasonic Large Scenario Model

(ULSM) which is to generate the embedding vectors

from spectrograms via the ResNet50. To generate our

desired vector embeddings we have conducted

research on VGG16, VGG19 and ResNet50

(Agarwal, 2021). But we have chosen the ResNet50

model because of its architecture and residual

connections which is suitable to process

spectrograms. To varify the correctness of these

embeddings we have calculated the similarity scores.

These embeddings are stored in a vector database for

different materials and objects. Our key idea is to

combine the abilities of ultrasonic sensors with Gen

AI systems to solve various real-life problems.

2 RELATED WORKS

Audio signals are within the human audible range for

example 20 Hz to 20 KHz, but ultrasonic signals work

beyond 20 KHz which is above human hearing.

Regardless, both types of signals are used to

communicate and they provide the ability for

information exchange with spectral elliptically

shaped sound waves. So, it can be said that the

methods we use to process audio or sound waves

nowadays can be applied to ultrasonic signals for

processing purposes. Ultrasonic signals can be treated

Ultrasonic Large Scenario Model (ULSM): Vector Embedding System for Ultrasonic Echo Wave Characteristics

705

as time series data with continuous values. The

methods used to represent time series data are the

building blocks to develop time series-based

applications. To represent data adequately and draw

any conclusion from the given time series data, vector

embeddings are necessary. Signal2Vec is a novel

technique that harnesses the capabilities of natural

language processing methodologies to convert

continuous-time series data into a meaningful vector-

based representation. This transformation enables a

diverse array of applications, encompassing time

series classification, prediction and anomaly

detection (Nalmpantis, 2019). The inspiration model

was word2vec which can understand the semantic and

syntactic meaning of words (Ma, 2015).

Another model named Wave2Vec, which is a

semantic learning model to learn deep representations

of medical concepts from Electronic Health Records

(EHRs). This model is capable of handling bio signals

such as Electroencephalogram (EEG),

Electrocardiogram (ECG), and Electromyography

(EMG) (Yuan, 2019). These continuous time series

signals are converted into vectors to extract semantic

meaning. This base model is a combination of two

separate models known as Wave2Vec-sc and

Wave2Vec-so. Wave2Vec-sc is responsible for

extracting the dormant characteristics of bio signals

with the help of a sparse autoencoder (SAE). On the

other hand, Wave2Vec-so can be trained to predict

neighbouring representations with the help of a

SoftMax layer (Yuan, 2019).

We are inspired by the research conducted by the

Facebook AI research team. They have developed a

model known as Wav2Vec 2.0 which can process raw

audio signals efficiently to solve speech processing

problems (Baevski, 2020). This model consists of

several stages such as Feature encoder, Context

Network, Quantization, and Self-Supervised

Learning. It encodes raw speech audio using a multi-

layer convolutional neural network into high-level

continuous feature vectors. These embeddings are

then fed into a Transformer network to create

contextualized representations. During the

pretraining stage, part of the model employs a

quantization module to transform the latent

representations into a limited set of potential

embeddings. These representations are chosen from

multiple codebooks (Baevski, 2020). The Gumbel

SoftMax function is used to choose discrete

codebooks (Gumbel, 1954) (Jang, 2016). The

embeddings generated by the quantization module

then serve as the targets for the model to predict

during pretraining (Baevski, 2020). We have

conducted our own experimentation on the base

model of Wav2Vec 2.0, which is the Wav2Vec model

to generate the embeddings via feature encoder. But

this model has some limitations. This model takes raw

signals as input. As we are working with ultrasound

signals, we have passed the signal directly to the

model. The Vector embeddings generated by this

model has the dimensions of 512*10. To create a

search index and perform similarity search to validate

the generated embeddings is quite challenging

because of this huge dimension. Even this dimension

is too big to perform vector search in renowned

Vector Databases such as Atlas MongoDB, Azure

Qdrant and Azure Cosmos DB.

Depending on the above-mentioned conclusion,

we have shifted our focus to frequency domain.

Spectrograms represent the echo wave characteristics

much better than the raw signals. Different

applications in the field of audio, music and speech

use pre-processed spectrograms and Mel-

spectrograms as the input data of neural networks

(Alnuaim, 2022). Spectrograms allow us to visualize

which frequencies are present for a specific material

and how they change. We have used the ResNet50

which is pretrained on ImageNet dataset and fine-

tuned on our own dataset. We have utilized a method

called transfer learning for efficiency to generate

vector embeddings. (Hossain, 2022) (Adebanjo,

2020).

3 METHODOLOGY

To produce vector embeddings from the reflected

echo signals we have followed certain steps. All the

experiments were conducted at the Computational

Intelligence Laboratory of Frankfurt University of

Applied Sciences.

3.1 Experimental Setup

To build our experimental setup, we have used one

ultrasonic sensor mounted on top of an embedded

system known as Red Pitaya. Figure 2, depicts the

visual representation of our setup and the way we

have mounted the ultrasonic system on the top of a

tripod. On the ground, yellow tapes are the markings

of the maximum reach of the signal in terms of angle

and space. The middle point is also marked at the

center with yellow tape. We have put all our materials

at the center to get the readings from the RedPitaya.

Our embedded system is connected to a laptop over

Ethernet cable, where monitoring of the signal

readings occurred.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

706

Figure 2: Experimental setup for data acquisition at

Computational Intelligence Lab (Frankfurt University of

Applied Sciences).

3.2 Data Acquisition

The system has sampled the analog signal with a

sampling rate of 1.95 MHz. Because of the extended

range of the system, it can analyse multiple materials

in a single shot. For experimental purpose, readings

of reflected echoes for 3 materials have been taken

such as plastic, wooden box, steel individually. For

each material 100 sampled time signals have been

taken for fine-tuning purpose. The distance of the

material from the sensor have been adjusted multiple

times as the strength of the signal gradually decreases

with the increasing distance. To avoid overfitting,

readings were taken from 35cm, 64cm and 150cm for

each of these materials.

3.3 Vector Embeddings Creation

Embeddings are the compact representation of given

input signal which contain the learned features. Our

captured Analog signals have converted to Analog-

to-Digital signals. After the preprocessing, we have

generated, spectrograms from these signals.

3.3.1 Labelling the Data

To label acquired data, human eye observations were

used. As we have acquired data for three materials,

we have labelled them as plastic, wood and steel.

3.3.2 Generating Spectrograms

To make the conversion from time domain to

frequency domain Fourier transformation is used

which reveal the frequency spectrum of the original

signal. We have ultrasound signals of different

amplitudes in this scenario. To measure the Power

Spectral Density (PSD) the sliding window

(Richardson, 2018) technique was used with the

window size of 256 with sliding step of 128 and

calculated the First Fourier Transform (FFT) of the

windowed portion. The mathematical representation

of the FFT on a set of N samples {𝑥



} as follows:

(Alnuaim, 2022)

𝑋



∑

𝑥







𝑒

(





)

𝑘= 0,1,..,𝑁−1 (1)

By squaring the FFT portions and divide them by

window size PSD was calculated. Then we have

generated spectrograms with these PSD values and

resized them into 256*256 pixels. These

spectrograms then converted to logarithmic scale for

better view. Figure 3, below shows the spectrogram

which has been generated by taking only the echo

portion of the whole signal reflected from plastic.

Figure 3: Spectrogram in Logarithmic scale.

3.3.3 ResNet50 Architecture with Custom

Layers

Residual Neural Network is one of Convolutional

Neural Networks (CNN). In CNN, a convolution

operation is conducted on the input data to learn the

features of the image. But this Deep CNN has some

drawbacks such as vanishing gradients, overfitting,

degradation and exploding gradients (LeCun, 2015).

Residual Neural Network can overcome these

challenges by adding the “Residual block” in the

network which is similar to “skip connection” and is

responsible to feed information deeper into the

network (Adebanjo, 2020).

In this paper, we have used the base ReNet50

architecture which has 50 layers for our training

purpose. In Figure 4, the curved lines define the

residual connections or the skip connections which is

responsible to feed the weights of previous layer into

the next deeper layer. This residual connection is

responsible to overcome the challenges like

exploding or vanishing gradients and degradation

Ultrasonic Large Scenario Model (ULSM): Vector Embedding System for Ultrasonic Echo Wave Characteristics

707

problems. ResNet50 have 5 convolutional blocks

known as Conv1, Conv2, Conv3, Conv4 and Conv5.

Each of these blocks has their own fixed size kernels

and convolution layers. In Figure 4, The first layer of

ResNet50 has 64 filters of size 7 × 7. The next layer

is a 3 × 3 sized maxpooling layer. There are three

identical grey color blocks, four identical orange

color block, four identical yellow color blocks and

lastly three identical brown color blocks depicted in

figure 4. The curved lines marked with blue color

represents the residual connection and the black color

represents the identity connection (Biswas, 2019).

Figure 4: ResNet50 Architecture with our custom Layers

(Biswas, 2019).

If the input and output dimensions of the

connecting blocks are same, then identity connections

are used. On the other hand, if the dimensions differ

then residual connections are used. We have not used

the Average pooling layer and the fully connected

1000 Dense layer of the traditional ResNet50. We

have added two custom layers to generate our desired

vector embeddings. We have added the Global

Average pooling layer at the end point of Conv5

block. This layer is responsible for reducing the

spatial dimensions which is basically height and

weight of the feature map of Conv5 block. Then we

have connected one dense layer to produce the 1024-

dimensional vector embeddings.

3.3.4 Pre-Training and Fine Tuning

We have used a compelling method known as Transfer

Learning to train our ResNet50. We have frozen the

Conv1, Conv2, and Conv3 blocks, which means we

have pre-trained them with the benchmark ImageNet

dataset using its weight (Adebanjo, 2020). Then for

fine tuning we have used our own 300 spectrograms (in

RGB scale) generated from ultrasound signals for

plastic, wood and metal because we have analysed that

the first three blocks extract general patterns and the

last two blocks Conv4 and Conv5 are responsible for

learning abstract pattern specific to ultrasound signals.

Our generated spectrograms have the size of 256*256

pixels and the ResNet50 can convert it to 224*224

pixels automatically.

3.3.5 Vector Embeddings

After the fine-tuning is done, we have generated 225

completely new spectrograms and then generated

vector embeddings with the size of (1, 1024)

dimensions.

3.3.6 Calculate Similarity Scores

For ensuring the correctness of our generated

embeddings, we have used these 225 newly generated

vector embeddings are our vector search space. This

search space is used to calculate the cosine similarity.

We have created three new spectrograms, one for each

material. After generating embeddings from these

three spectrograms, we have calculated the cosine

similarity scores within our vector space. Cosine

similarity is basically calculating the angle between

two vectors. Suppose we have two vectors called X and

Y. Then the cosine similarity between them can be

calculated by the following equation (Nguyen, 2010),

𝑐𝑜𝑠𝑖𝑛𝑒 𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 (𝐶𝑆) =





.





(2)

Here 𝐴



.𝐵 denotes the dot product between two

vectors. By using the dot product of the normalized

vectors, we have calculated the cosine similarity. As

we wanted to calculate pattern similarities between

spectrogram embeddings we have chosen this

method. If the value of 𝐶𝑆 is closer to 1 then we have

a perfect match and if the value is 0 then the vectors

are orthogonal means they are completely different

(Nguyen, 2010).

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

708

4 ANALOGY

In the current era of transformer models, word

embeddings are used for sentence structuring and

word prediction. A multidimensional array can be an

example of a word embedding which is simply

defining how semantically close the current word is

with other words of the sentence. A relation between

words in multi-dimension helps the model to even

define the type of the word (e.g. pronoun type or verb

type word). With these embeddings, a sentence is

structured and predicted in LLM models. In the case

of waves, these embeddings may get embedded in the

phoneme level of a word. Each of the words then gets

separated into fractions of phonemes. The whole

word embedding holds the audio batch size of the

word, audio channels, and phoneme characteristics.

As Ultrasonic Sounds are also waves with higher

frequency than normal human audible range, the

behaviour of Ultrasonic echo is also like sound waves

of words created while reading a book. The analogy

of each word is to make a whole sentence that is

meaningful. But in the case of a Scenario analysis

Ultrasonic sound must produce which type of

material (e.g. Solid Steel, Soft Foam, Solid Wooden

structure with a hollow in the middle) is present and

in which direction with the information of the current

distance. If for example a sentence “I have seen this

before” produces an embeddings representation in

binary dimension analysis like yellow-coloured

boxes in Figure 3, each of the materials in front of the

Ultrasonic Sensor in office space would create the

embeddings representation given in green-coloured

boxes based on the material of that object. Ultrasonic

Sensors are well known for the capability of distance

measurement, which will interpret the scenario for

example “Plastic at 20cm distance, Wood in 1 meter

distance, Steel cube in 5 meters distance

corresponding a Concrete wall behind the steel frame

at 6 meters from the Ultrasonic sensor”.

5 FUTURE WORK

In this paper, the embeddings of ultrasonic sound

echoes using spectrograms were processed for a far-

reaching achievement of Scenario analysis for trained

material in a live location. The forward plan

corresponds to the further training of the Neural

Network dedicatedly working for processing

Ultrasonic sound with larger dataset. This includes

comparing the embedding creation from Facebook’s

“fairseq” library with own built embedding creation

library (Baevski, 2020). With this achievement, the

process of using the embeddings for material

classification requires an LSM model combined with

LLM which will be capable of interpreting live

scenarios with material positioning and shaping

information. The building phase of the transformer

model for Large Scenario Model (LSM) analysis is

ongoing and will be included with the result in future

works. Required feature engineering and Data

preprocessing will be included in the journey of

getting the highest accuracy of material classification

with positional changing information to build the

scenario into an HCI vision model. Building database

with Log-spectrogram for aiding embedding creation

using CNN feature extraction is also in the future

work plan.

Figure 5: Analogy of Ultrasonic Signal Embeddings with respect to Audio Signal Embeddings.

Ultrasonic Large Scenario Model (ULSM): Vector Embedding System for Ultrasonic Echo Wave Characteristics

709

6 RESULT ANALY SIS

we have fine-tuned our model with 300 labelled

spectrograms in the training phase and for validation

75 new spectrograms were used. Figure 6, depicts the

convergence of our network in training and validation

phases.

Figure 6: Accuracy Distribution in training process.

For our search space we have generated 225

vector embeddings. Each of these embeddings has the

size of (1, 1024). To visualize such a huge dimension,

we have used the t-Distributed Stochastic Neighbour

Embedding (t-SNE) to reduce the high-dimensional

data in a lower dimensional space (Arora, 2018, July).

From figure 7, it is clearly noticeable that the

generated embeddings formed similar cluster or

patterns. Vectors generated for same material are

closed to each other.

Figure 7: Scatter plot of the vector embeddings.

To verify our experimentation, we have taken

three new samples (one from each material) and

generated vector embeddings. We have created our

search space to calculate the cosine similarity among

these three vectors and the 225 vectors. From Table

1, it can be demonstrated that the top five similarity

scores are closer to 1 which means they are close to

each other, they have similar patterns and they are in

a same cluster.

Table 1: Top five similarities for each material.

Materials Top Five Cosine Similarity Score

from the Search S

ace

Book

(

'Book', 0.9986334

)

('Book', 0.998229)

('Book', 0.9974883)

(

'Book', 0.9972976

)

(

'Book', 0.99661756

)

Plastic

(

'Plastic', 0.99962866

)

('Plastic', 0.9996282)

('Plastic', 0.99961853)

('Plastic', 0.9995949)

(

'Plastic', 0.9995947

)

Steel

(

'Steel', 0.9998824

)

(

'Steel', 0.99977136

)

('Steel', 0.9997442)

('Steel', 0.99968445)

(

'Steel', 0.99959064

)

7 APPLICATION

With these embeddings currently, transformer model

creation and training-testing work is going on. This

will enable the HCI system to build an image of the

surrounding scenario of a vision-impaired person to

assist with sound or neural image transferring using

cheap ultrasonic sensors.

The material classification from the echo signal

will enable a couple of new options for the auto

industry to build vehicles with more autonomous

functions cheaper to build with control of scenario

analysis. Doppler effect analysis for echo

spectrogram can be used for rapid position-changing

effect analysis and emergency analysis on roads (Raj,

2012). It can open the option of cheaper sensors for

accident prediction.

The final Machine Learning Model will open a

new research option for Rovers. Mars or Moon

Rovers with LSM can define a newly found material

on an unknown surface depending on the atmospheric

effect of the planet on sound waves. By autonomous

calculation of deviation of Ultrasonic sound echo

from the earth's surface and unknown planet surface,

calibration of the robot for the atmosphere can be

more accurate. This will enable also material

classification of unknown objects and similarity

analysis with known earth objects.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

710

REFERENCES

Adebanjo, Q. A.-H. (2020). Breast Cancer Diagnosis in

Histopathological Images Using ResNet-50

Convolutional Neural Network. 2020 IEEE

International IOT, Electronics and Mechatronics

Conference (IEMTRONICS) (pp. 1-7). Vancouver, BC,

Canada: IEEE. doi:10.1109/IEMTRONICS51293.20

20.9216455

Agarwal, S. M. (2021). A comparison between VGG16,

VGG19 and ResNet50 architecture frameworks for

Image Classification. 2021 International Conference

on Disruptive Technologies for Multi-Disciplinary

Research and Applications (CENTCON), (pp. 96-99).

Bengaluru, India. doi:10.1109/CENTCON52345.202

1.9687944

Alnuaim, A. A. (2022). Speaker gender recognition based

on deep neural networks and ResNet50. Wireless

Communications and Mobile Computing, 1(4444388).

Baevski, A. Z. (2020). wav2vec 2.0: A framework for self-

supervised learning of speech representations.

Advances in neural information processing systems, (p.

33).

Biswas, Z. M. (2019). Transfer Learning Based Plant

Diseases Detection Using ResNet50. 4th International

Conference on Electrical Information and

Communication Technology (EICT), (pp. 1-6). Khulna,

Bangladesh. doi:10.1109/EICT48899.2019.9068805

Ghosh, A. C. (2023). UltraSense: A non-intrusive approach

for human activity identification using heterogeneous

ultrasonic sensor grid for smart home environment.

Journal of Ambient Intelligence and Humanized

Computing, 1-22.

Gumbel, E. J. (1954). Statistical theory of extreme values

and some practical applications: a series of lectures

(Vol. 33). US Government Printing Office.

Hossain, M. B. (2022). Transfer learning with fine-tuned

deep CNN ResNet50 model for classifying COVID-19

from chest X-ray images. Informatics in Medicine

Unlocked, 30(100916).

Jang, E. G. (2016). Categorical reparameterization with

gumbel-softmax. arXiv preprint arXiv:1611.01144.

LeCun, Y. B. (2015). Deep learning. nature, 521(7553),

436-444.

Ma, L. &. (2015). Using Word2Vec to process big text data.

2015 IEEE International Conference on Big Data (Big

Data) (pp. 2895-2897). IEEE.

Matz, V. K. (2006). Classification of ultrasonic signals.

International Journal of Materials and Product

Technology, 27(3-4), 145-155.

Nalmpantis, C. &. (2019). Signal2vec: Time series

embedding representation. International conference on

engineering applications of neural networks (pp. 80-

90). Cham: Springer International Publishing.

Nayan, N. M. (2016). An ultrasonic sensing system for

assisting visually impaired person. Jurnal Teknologi,

78(7-4).

Nguyen, H. V. (2010). Cosine similarity metric learning for

face verification. In Asian conference on computer

vision (pp. 709-720). Berlin, Heidelberg: Springer

Berlin Heidelberg.

Pagodinas, D. (2002). Ultrasonic signal processing methods

for detection of defects in composite materials.

Ultragarsas/Ultrasound, 45(4), 47-54.

Raj, B. K. (2012). Ultrasonic doppler sensing in hci.

IEEE

Pervasive Computing, 11(2), 24-29.

Richardson, L. F. (2018). The Sliding Window Discrete

Fourier Transform. arXiv preprint arXiv:1807.07797.

Sanders, D. H. (2021). Intelligent control and HCI for a

powered wheelchair using a simple expert system and

ultrasonic sensors. Intelligent Systems and

Applications: Proceedings of the 2020 Intelligent

Systems Conference (IntelliSys). 3, pp. 571-583.

Springer International Publishing.

Taiju Shibata, M. I. (2001). Ultrasonic signal characteristics

by pulse-echo technique and mechanical strength of

graphite materials with porous structure. Nuclear

Engineering and Design, 203(2–3), 133-141.

doi:https://doi.org/10.1016/S0029-5493(00)00305-8

Vaswani, A. S. (2017). Attention is all you need. 31st

Conference on Neural Information Processing Systems

(p. 30). Long Beach, CA, USA: Advances in neural

information processing systems.

Yuan, Y. X. (2019). Wave2vec: Deep representation

learning for clinical temporal data. Neurocomputing,

324, 31-42.

Zhou, Y. &. (2022). Intelligent analysis system for signal

processing tasks based on LSTM recurrent neural

network algorithm. Neural Computing and

Applications, 34(15).

Ultrasonic Large Scenario Model (ULSM): Vector Embedding System for Ultrasonic Echo Wave Characteristics

711