With the proliferation of the Internet of Things (IoT) and the increasing integration of

smart technologies into everyday life, smart homes are becoming a critical part of modern digital
infrastructure. These homes are equipped with various interconnected devices such as motion
sensors, door sensors, temperature regulators, smart locks, and home appliances that constantly
generate large volumes of heterogeneous data. This data reflects the routine activities and
behaviors of the residents and is highly valuable for intelligent automation, energy efficiency,
and enhanced user convenience.

However, smart homes are also vulnerable to unexpected behaviors or unusual events,
such as intrusion attempts, equipment malfunctions, or sudden health-related emergencies of the
residents (e.g., falls or long periods of inactivity). These events, known as anomalies, may not
follow previously learned or expected patterns. Detecting such anomalous activities is essential
for ensuring the safety, privacy, and security of the residents. Given the diverse nature of smart
home data and the difficulty in obtaining comprehensive labeled datasets for all possible
activities, unsupervised learning techniques have emerged as a promising approach for
anomaly detection.

This paper explores how unsupervised machine learning algorithms can be effectively
applied to detect anomalous behavior in smart homes. Unlike supervised methods that require
large amounts of annotated data for training, unsupervised methods learn the underlying
structure of the data without requiring labels. This makes them suitable for domains like smart
homes, where collecting labeled anomaly data is not only challenging but also may raise privacy
concerns.

We focus on evaluating and comparing several prominent unsupervised algorithms, including:

 Isolation Forest (IF): An ensemble-based method that isolates outliers by randomly

selecting features and split values to build trees. Anomalies are expected to be isolated
quickly.
 DBSCAN (Density-Based Spatial Clustering of Applications with Noise): A
clustering algorithm that groups data based on density and identifies low-density points
as anomalies.
 Autoencoders: Neural network-based models trained to reconstruct input data; anomalies
produce higher reconstruction errors.
 One-Class SVM (Support Vector Machine): Learns a boundary around normal
instances and flags points lying outside this region as anomalies.

To evaluate these models, we use publicly available datasets, such as:

1. CASAS (Center for Advanced Studies in Adaptive Systems) Smart Home datasets –
collected from real smart homes with motion, temperature, and usage sensors.
2. ARAS Human Activity Dataset – multi-resident activity data from sensor-equipped
smart environments.
3. KDE Smart Home Dataset – containing user activity sequences from ambient sensors.

These datasets include a mixture of normal daily activities (such as cooking, sleeping, and
bathing) and some injected anomalies or irregularities for benchmarking.

The raw sensor data is often noisy, sparse, and asynchronous. Therefore, before applying the
algorithms, we develop a robust preprocessing pipeline. This pipeline involves:

 Data cleaning: Removing missing or duplicated sensor events.

  Time windowing: Aggregating events into fixed or sliding time windows (e.g., every 30
seconds or 5 minutes).
 Feature extraction: Generating numerical features such as sensor activation counts,
transitions, and temporal patterns.
 Dimensionality reduction: Applying PCA or t-SNE for visualization and compact
representation of high-dimensional sensor features.

Each algorithm is then trained and evaluated on the preprocessed datasets. For evaluation, we
use metrics such as:

 Precision: The proportion of detected anomalies that were truly anomalous.

 Recall: The proportion of true anomalies that were correctly identified.
 F1-score: The harmonic mean of precision and recall.
 ROC-AUC: For algorithms that provide anomaly scores.
 Computation time: To assess real-time applicability.

Findings:

 Isolation Forest demonstrated high efficiency and scalability, making it well-suited for
large-scale, real-time anomaly detection in streaming environments.
 DBSCAN effectively identified spatial clusters of normal activities, but was sensitive to
parameter tuning (e.g., epsilon and minimum samples).
 Autoencoders provided high recall rates and performed well on complex temporal
patterns, though training required GPU acceleration and careful hyperparameter tuning.
 One-Class SVM produced reliable results on smaller datasets with limited variability but
was computationally expensive and sensitive to kernel choice.

Additionally, we implemented a hybrid ensemble model that combines the outputs of multiple
unsupervised models using majority voting or weighted anomaly scores. This fusion approach
reduced false positives and improved detection robustness, especially in overlapping activity
contexts.

Challenges and Open Problems:

Despite promising results, several challenges remain:

 Concept drift: Over time, resident behavior patterns may evolve (e.g., due to illness or
new routines), making existing models obsolete. Adaptive learning strategies are needed
to handle such drift.
 Multimodal data fusion: Integrating audio, video, environmental, and wearable data can
improve accuracy but increases complexity and privacy concerns.
 Edge computing: Deploying models on edge devices within the smart home reduces
latency but limits computational power and storage.
 Data imbalance: Anomalies are rare by nature, making it difficult to evaluate
performance reliably.
 Privacy and security: Anomaly detection models must be designed with privacy-
preserving mechanisms (e.g., federated learning or differential privacy).

Future Directions:

To improve real-world applicability, future research could focus on:

  Developing online learning algorithms that adapt to new activity patterns over time.
 Exploring self-supervised learning approaches that use data augmentation or contrastive
learning to pretrain models.
 Designing lightweight models optimized for edge deployment with minimal power
consumption.
 Creating standardized benchmark datasets with real anomalies rather than synthetically
injected ones.
 Investigating explainable AI methods to provide interpretable reasons for why an
activity is classified as anomalous, which is crucial for gaining user trust.

Conclusion:

Unsupervised anomaly detection in smart homes is a rapidly evolving field with significant
practical relevance. This study demonstrates that unsupervised learning techniques, particularly
Isolation Forests, Autoencoders, and clustering-based approaches like DBSCAN, are capable of
effectively identifying abnormal behaviors in complex sensor environments. By applying
advanced preprocessing, robust feature engineering, and model ensemble techniques, these
algorithms can enhance the reliability and safety of smart homes. Moreover, the integration of
such models into smart home systems can pave the way for proactive alerts, context-aware
automation, and personalized assistance for elderly or vulnerable populations. As smart home
adoption continues to rise, intelligent anomaly detection will remain a cornerstone of secure and
responsive home automation systems.
Narsalar Internetining (IoT) tarqalishi va aqlli texnologiyalarning kundalik hayotga tobora ko'proq
integratsiyalashuvi bilan aqlli uylar zamonaviy raqamli infratuzilmaning muhim qismiga aylanmoqda.
Ushbu uylar doimiy ravishda katta hajmdagi heterojen ma'lumotlarni ishlab chiqaradigan harakat
sensorlari, eshik sensorlari, harorat regulyatorlari, aqlli qulflar va maishiy texnika kabi bir-biriga
bog'langan turli xil qurilmalar bilan jihozlangan. Ushbu ma'lumotlar aholining muntazam faoliyati va
xatti-harakatlarini aks ettiradi va aqlli avtomatlashtirish, energiya samaradorligi va foydalanuvchilarga
qulaylik yaratish uchun juda muhimdir. Biroq, aqlli uylar kutilmagan xatti-harakatlarga yoki g'ayrioddiy
hodisalarga, masalan, bostirib kirishga urinishlar, uskunaning noto'g'ri ishlashi yoki aholining sog'lig'i
bilan bog'liq to'satdan favqulodda vaziyatlarga (masalan, yiqilish yoki uzoq vaqt harakatsizlik) ham
zaifdir. Anomaliyalar deb nomlanuvchi bu hodisalar ilgari o'rganilgan yoki kutilgan naqshlarga amal
qilmasligi mumkin. Bunday g'ayritabiiy harakatlarni aniqlash aholining xavfsizligi, shaxsiy hayoti va
xavfsizligini ta'minlash uchun juda muhimdir. Aqlli uy ma'lumotlarining xilma-xilligini va barcha mumkin
bo'lgan tadbirlar uchun keng qamrovli etiketli ma'lumotlar to'plamini olishda qiyinchiliklarni hisobga
olgan holda, nazoratsiz o'rganish texnikasi anomaliyani aniqlash uchun istiqbolli yondashuv sifatida
paydo bo'ldi. Ushbu maqola aqlli uylarda anomal xatti-harakatlarni aniqlash uchun nazoratsiz mashinani
o'rganish algoritmlarini qanday samarali qo'llash mumkinligini o'rganadi. O'qitish uchun katta
miqdordagi izohli ma'lumotlarni talab qiladigan nazorat qilinadigan usullardan farqli o'laroq, nazoratsiz
usullar yorliqlarni talab qilmasdan ma'lumotlarning asosiy tuzilishini o'rganadi. Bu ularni aqlli uylar kabi
domenlarga moslashtiradi, bu erda etiketli anomaliya ma'lumotlarini yig'ish nafaqat qiyin, balki maxfiylik
muammolarini ham keltirib chiqarishi mumkin. Biz bir nechta taniqli nazoratsiz algoritmlarni baholash va
taqqoslashga e'tibor qaratamiz, jumladan: * Izolyatsiya o'rmoni (IF): daraxtlarni qurish uchun
xususiyatlarni tasodifiy tanlash va qiymatlarni ajratish orqali chegaralarni ajratib turadigan ansamblga
asoslangan usul. Anomaliyalar tezda ajratilishi kutilmoqda. * Dbscan (shovqin bilan dasturlarning
zichlikka asoslangan fazoviy klasteri): ma'lumotlarni zichlikka qarab guruhlaydigan va past zichlikdagi
nuqtalarni anomaliya sifatida aniqlaydigan klasterlash algoritmi. * Autoencoders: kirish ma'lumotlarini
rekonstruksiya qilish uchun o'qitilgan neyron tarmoqqa asoslangan modellar; anomaliyalar yuqori
rekonstruksiya xatolarini keltirib chiqaradi. * Bir sinf SVM (qo'llab-quvvatlash vektor mashinasi):
anomaliya sifatida bu mintaqa tashqarisida yotgan normal hollarda va bayroqlar nuqtalari atrofida bir
chegara o'rganadi. Ushbu modellarni baholash uchun biz ommaviy ma'lumotlar to'plamidan
foydalanamiz, masalan: 1. CASAS (adaptiv tizimlarda ilg'or tadqiqotlar markazi) aqlli uy ma'lumotlar
to'plamlari – harakat, harorat va foydalanish sensorlari bilan haqiqiy aqlli uylardan to'plangan. 2. ARAS
inson faoliyati ma'lumotlar to'plami-sensor bilan jihozlangan aqlli muhitlardan ko'p rezidentli faoliyat
ma'lumotlari. 3. KDE aqlli uy ma'lumotlar to'plami-atrof-muhit sensorlaridan foydalanuvchi faoliyati
ketma-ketligi. Ushbu ma'lumotlar to'plamiga odatdagi kundalik faoliyat (masalan, pishirish, uxlash va
cho'milish) aralashmasi va benchmarking uchun ba'zi AOK qilingan anomaliyalar yoki nosimmetrikliklar
kiradi. Xom sensor ma'lumotlari ko'pincha shovqinli, siyrak va asenkron. Shuning uchun, algoritmlarni
qo'llashdan oldin, biz mustahkam ishlov berish quvurini ishlab chiqamiz. Ushbu quvur liniyasi
quyidagilarni o'z ichiga oladi: * Ma'lumotlarni tozalash: yo'qolgan yoki takrorlangan sensor hodisalarini
olib tashlash. * Vaqt oynasi: voqealarni sobit yoki toymasin vaqt oynalariga yig'ish (masalan, har 30
soniyada yoki 5 daqiqada). * Xususiyatlarni ajratib olish: sensorni faollashtirish soni, o'tish va
vaqtinchalik naqshlar kabi raqamli xususiyatlarni yaratish. * O'lchovni kamaytirish: vizualizatsiya va
yuqori o'lchovli sensor xususiyatlarini ixcham namoyish qilish uchun PCA yoki t-SNE-ni qo'llash. Keyin har
bir algoritm oldindan ishlov berilgan ma'lumotlar to'plamida o'qitiladi va baholanadi. Baholash uchun biz
quyidagi ko'rsatkichlardan foydalanamiz: * Aniqlik: haqiqatan ham anomal bo'lgan aniqlangan
anomaliyalarning ulushi. * Eslatib o'tamiz: to'g'ri aniqlangan haqiqiy anomaliyalarning nisbati. * F1-ball:
aniqlik va eslashning Harmonik o'rtacha qiymati. * ROC-AUC: anomaliya ballaringizni ta'minlaydigan
algoritmlar uchun. * Hisoblash vaqti: Real vaqtda qo'llanilishini baholash. Topilmalar: * Izolyatsiya
o'rmoni yuqori samaradorlik va miqyoslilikni namoyish etdi, bu esa oqim muhitida keng ko'lamli, Real
vaqtda anomaliyani aniqlash uchun juda mos keladi. * DBSCAN normal faoliyatning fazoviy klasterlarini
samarali aniqladi, ammo parametrlarni sozlashga sezgir edi (masalan, epsilon va minimal namunalar). *
Autoencoders yuqori eslamoq sur'atlarini taqdim va ta'lim GPU tezlashtirish va ehtiyot hyperparameter
tuning zarur bo'lsa-da, murakkab vaqtinchalik naqsh yaxshi amalga. * Bir sinfli SVM cheklangan
o'zgaruvchanlikka ega bo'lgan kichik ma'lumotlar to'plamlarida ishonchli natijalar berdi, ammo hisoblash
uchun qimmat va yadro tanloviga sezgir edi. Bundan tashqari, biz ko'pchilik ovoz berish yoki vaznli
anomaliya ballari yordamida bir nechta nazoratsiz modellarning natijalarini birlashtirgan gibrid ansambl
modelini amalga oshirdik. Ushbu termoyadroviy yondashuv noto'g'ri pozitivlarni kamaytirdi va
aniqlashning mustahkamligini yaxshiladi, ayniqsa bir-birining ustiga chiqadigan faoliyat sharoitida.
Muammolar va ochiq muammolar: Istiqbolli natijalarga qaramay, bir nechta muammolar qolmoqda: *
Tushunchasi drift: vaqt o'tishi bilan, istiqomat xulq naqsh evolyutsiya mumkin (masalan, kasallik yoki
yangi muolajalarni tufayli), mavjud modellar eski qilish. Bunday driftni boshqarish uchun adaptiv ta'lim
strategiyalari zarur. * Multimodal data fusion: audio, video, ekologik va taqiladigan ma'lumotlarni
birlashtirish aniqlikni oshirishi mumkin, ammo murakkablik va maxfiylik muammolarini oshiradi. *
Chekka hisoblash: aqlli uy ichidagi chekka qurilmalarda modellarni joylashtirish kechikishni kamaytiradi,
ammo hisoblash quvvati va saqlashni cheklaydi. * Ma'lumotlar muvozanati: anomaliyalar tabiatan kam
uchraydi, bu esa ishlashni ishonchli baholashni qiyinlashtiradi. * Maxfiylik va xavfsizlik: Anomaliyani
aniqlash modellari maxfiylikni saqlash mexanizmlari bilan ishlab chiqilishi kerak (masalan, federatsiya
o'rganish yoki differentsial maxfiylik). Kelajak Yo'nalishlari: Haqiqiy dunyoda qo'llanilishini yaxshilash
uchun kelajakdagi tadqiqotlar quyidagilarga qaratilishi mumkin: * Vaqt o'tishi bilan yangi faoliyat
shakllariga moslashadigan onlayn ta'lim algoritmlarini ishlab chiqish. * Modellarni oldindan o'rgatish
uchun ma'lumotlarni ko'paytirish yoki kontrastli o'rganishdan foydalanadigan o'z-o'zini boshqaradigan
ta'lim yondashuvlarini o'rganish. * Minimal quvvat sarfi bilan chekka joylashtirish uchun
optimallashtirilgan engil modellarni loyihalash. * Sintetik AOK qilingan emas, balki haqiqiy anomaliyalar
bilan standartlashtirilgan benchmark ma'lumotlar to'plamlarini yaratish. * Faoliyatning anomal deb
tasniflanishining izohlanadigan sabablarini ta'minlash uchun tushuntiriladigan AI usullarini o'rganish, bu
foydalanuvchi ishonchini qozonish uchun juda muhimdir. Xulosa: Aqlli uylarda nazoratsiz anomaliyani
aniqlash muhim amaliy ahamiyatga ega bo'lgan tez rivojlanayotgan sohadir. Ushbu tadqiqot shuni
ko'rsatadiki, nazoratsiz o'rganish texnikasi, xususan izolyatsiya o'rmonlari, Avtoenkoderlar va dbscan
kabi klasterga asoslangan yondashuvlar murakkab sensor muhitida g'ayritabiiy xatti-harakatlarni
samarali aniqlashga qodir. Ilg'or ishlov berish, mustahkam xususiyat muhandisligi va namunaviy ansambl
texnikasini qo'llash orqali ushbu algoritmlar aqlli uylarning ishonchliligi va xavfsizligini oshirishi mumkin.
Bundan tashqari, bunday modellarning aqlli uy tizimlariga qo'shilishi proaktiv ogohlantirishlar,
kontekstdan xabardor avtomatlashtirish va keksa yoki zaif aholi uchun shaxsiy yordamga yo'l ochishi
mumkin. Aqlli uyni qabul qilish o'sishda davom etar ekan, aqlli anomaliyani aniqlash xavfsiz va sezgir uy
avtomatlashtirish tizimlarining asosi bo'lib qoladi.
 Aqlli uylarda mashinaviy o’qitishning nazoratsiz o’rganish usullaridan
foydalangan holda, anomal faolliklarni aniqlash.
Annotatsiya. O’tgan asr so’ngidan boshlab texnologiyalarga bo’lgan talab
yuqorila boshladi. Asrimiz boshlarida raqamli texnologiyalar rivoji ham yuqorilab
bormoqda. Jarayonlarni avtomatlashtirish, tizimlashtirish jadal tus olmoqda. Jadal
tus olishi natijasida uy qurulishi, yashash uchun sharoitlarda ham yengilliklar
keltirib chiqa boshladi. Turar joylar zamonaviy texnologiyalar asosida ta’mirlana
boshlanib, aqlli uylar terminlari yuzaga keldi. Aqlli uylar, bu xonadon
jihozlarining to’liq avtomatlashgan tizim asosida qurilgan uylar bo’lib, xonadon
elektr ta’minotiga bog’langan jihozlari to’liq xonadon sohibi tomonidan maxsus
platformalarda oson boshqarish imkonini beradi.
 Since the end of the last century, the demand for technology has been
increasing. At the beginning of our century, the development of digital
technologies has also been increasing. Automation and systematization of
processes are gaining momentum. As a result of this rapid development, housing
construction and living conditions have also become easier. Residential areas have
begun to be renovated based on modern technologies, and the term smart homes
has emerged. Smart homes are houses built on a fully automated system of
household appliances, which allow the owner of the house to easily control all the
appliances connected to the household electricity supply on special platforms.
These systems monitor user behavior in real time using various IoT devices and
provide various automated services based on them. Smart homes work in
integration with security, energy efficiency, comfort, and healthcare systems. At
the same time, the issue of early detection of abnormal movements or unexpected
events occurring in this environment is of great importance. Such events can be
associated with internal or external threats (for example, unauthorized access,
device malfunction, unexpected inactivity, health emergencies). Identifying such
situations - detecting anomalies - is one of the priorities in ensuring the security
and reliability of smart homes today.

The article studies scientific and practical approaches to detecting

anomalous activities occurring in the smart home environment based on
unsupervised learning methods. These approaches, unlike traditional, i.e.
supervised learning, are trained without predefined labels. Because in practice, it is
impossible to identify and define all possible anomalous situations in advance.
Therefore, the article covers the stages of working with large-scale and high-
dimensional data collected from various sensors, their cleaning, feature extraction,
and anomaly detection using unsupervised algorithms.

The study investigated the application and effectiveness of the following

unsupervised learning algorithms:

Isolation Forest (IF) – a tree-based ensemble algorithm that quickly isolates

outliers.

DBSCAN (Density-Based Spatial Clustering of Applications with Noise) –

an approach designed to cluster based on density and identify outliers as
anomalies.
 Autoencoder – a deep learning-based neural network model that detects
anomalies based on the error in the reconstruction of incoming data.

One-Class SVM – trained only on “normal” cases and evaluates data that
deviates from this case as abnormal.

The research was carried out using open source databases such as CASAS,
ARAS, KDE Smart Home. The sensor data stream was pre-cleaned, time-spliced,
statistical features extracted, and made suitable for analysis using dimensionality
reduction techniques (PCA, t-SNE). The effectiveness of each algorithm was
evaluated based on standard criteria such as Precision, Recall, F1-score, ROC-
AUC.

The results show that Autoencoder has high accuracy and gives the best
results in detecting subtle anomalies. Isolation Forest is convenient for use in real-
time systems due to its speed and simple structure. Although DBSCAN was able to
distinguish clusters based on density well, it could not always give stable results
due to its sensitivity to parameters. One-Class SVM worked well on small amounts
of data, but the computational complexity was high on large datasets.
The paper also used an ensemble approach (ensemble learning) that
combines different models, balancing the advantages and disadvantages of each
model. This approach led to a reduction in false positives and false negatives in a
real environment.
The following issues were considered separately in the article:

Concept drift – changing user habits require retraining of the model.

Multi-data integration – the need to take into account the relationship

between audio, video, temperature, motion and other sensors.

Privacy and privacy – the need to protect data when working with user data.
 Resource constraints – the need for lightweight versions of the model to
work on edge devices.

At the end of the study, the following future directions were indicated:

Adaptive, online learning systems – models that adapt to changes in user

habits.

Federated learning – the ability to update the model on the user’s device
without sending data to a central server.
Explainable AI – systems that can explain why a given situation was
considered an anomaly.

Context-aware learning is the training of models that take into account time,
weather, calendar, and other external factors.

In conclusion, unsupervised learning methods can effectively detect

anomalous activities occurring in the smart home environment. This has a
significant positive impact on ensuring the safety, comfort, and health of users.
The methods described in this article can be adapted to real-life smart home
systems and widely implemented in practice.
The need for technology has been growing since the turn of the century. The growth of digital
technology has also been accelerating at the start of this century. Process automation and
systematisation are becoming more popular. This quick development has also made it easier to
build homes and improve living circumstances. The phrase "smart homes" has evolved as a
result of residential spaces starting to be updated using contemporary technologies. Smart homes
are constructed with a completely automated system of appliances that enables the homeowner to
conveniently operate any device linked to the home's electrical supply from specialised
platforms. Using a variety of IoT devices, these systems track user behaviour in real time and
offer a range of automated services. Security, energy efficiency, comfort, and healthcare systems
are all integrated into smart houses. At the same time, it is crucial to address the problem of early
detection of anomalous movements or unforeseen occurrences that take place in this setting.
These occurrences may be linked to either external or internal risks (e.g., unexpected inaction,
gadget breakdown, unauthorised access, health emergencies). One of the top goals for
guaranteeing the security and dependability of smart homes nowadays is recognising such
circumstances, or detecting anomalies.
Based on unsupervised learning techniques, the essay examines both scientific and practical
strategies for identifying unusual activity taking place in the smart home environment. These
methods are learnt without specified labels, in contrast to standard methods, such as supervised
learning. Because it is impossible to foresee and characterise every potential aberrant scenario in
practice. As a result, the paper discusses the steps involved in working with high-dimensional,
large-scale data that has been gathered from several sensors, including cleaning, feature
extraction, and anomaly detection using unsupervised algorithms.
The following unsupervised learning algorithms were examined in the study for their use and
efficacy:

A tree-based ensemble approach called Isolation Forest (IF) rapidly separates outliers.

Density-Based Spatial Clustering of Applications with Noise, or DBSCAN, is a method for

grouping applications according to density and identifying abnormalities in outliers.

An autoencoder is a neural network model that uses deep learning to identify abnormalities by
calculating the error in the reconstruction of incoming data.

Only "normal" cases are used to train One-Class SVM, which classifies data that differs from
these cases as abnormal.
Open source databases including CASAS, ARAS, and KDE Smart Home were used in the study.
Dimensionality reduction techniques (PCA, t-SNE) were used to pre-clean, time-splice, extract
statistical features, and prepare the sensor data stream for analysis. Standard metrics including
precision, recall, F1-score, and ROC-AUC were used to assess each algorithm's efficacy.
The findings demonstrate that Autoencoder is highly accurate and provides the best results when
it comes to identifying minute irregularities. Isolation Forest's quickness and straightforward
design make it ideal for usage in real-time systems. Due to its sensitivity to settings, DBSCAN
was not always able to produce steady findings, despite its ability to differentiate clusters based
on density. On small datasets, One-Class SVM performed well, but on large datasets, its
computational cost increased.
In order to balance the benefits and drawbacks of each model, the study also employed an
ensemble technique, or ensemble learning. In a real-world setting, our method reduced the
number of false positives and false negatives.
The essay examined each of the following concerns separately:

Concept drift: The model needs to be retrained in light of shifting user behaviours.
The necessity of considering the connections between motion, temperature, video, audio, and
other sensors is known as multi-data integration.

When working with user data, privacy and privacy are essential.

Resource limitations: in order for the model to function on edge devices, it must be lightweight.

The study concluded with recommendations for the following future directions:

Online learning systems that adjust to shifting user habits are known as adaptive models.
With federated learning, the user's device can update the model without transmitting data to a
central server.
Systems that can explain why a particular circumstance was deemed unusual are known as
explainable AI.

Training models that consider time, weather, calendar, and other external aspects is known as
context-aware learning.
To sum up, unsupervised learning techniques are capable of efficiently identifying unusual
activity taking place in the context of smart homes. The safety, comfort, and health of users are
significantly improved by this. The techniques discussed in this article can be broadly applied in
reality and modified for use with actual smart home systems.
 Texnologiyaga bo'lgan ehtiyoj asrning boshidan beri o'sib bormoqda.
Raqamli texnologiyalarning o'sishi ham shu asrning boshida tezlashdi. Jarayonlarni
avtomatlashtirish va tizimlashtirish tobora ommalashib bormoqda. Ushbu tez
rivojlanish uy-joy qurish va yashash sharoitlarini yaxshilashni ham osonlashtirdi.
"Aqlli uylar" iborasi zamonaviy texnologiyalardan foydalangan holda turar-joy
binolari yangilana boshlagani natijasida rivojlandi. Aqlli uylar to'liq
avtomatlashtirilgan maishiy texnika tizimi bilan qurilgan bo'lib, bu uy egasiga
ixtisoslashgan platformalardan uyning elektr ta'minotiga ulangan har qanday
qurilmani qulay boshqarish imkonini beradi. Turli xil IoT qurilmalaridan
foydalangan holda, ushbu tizimlar real vaqtda foydalanuvchi xatti-harakatlarini
kuzatib boradi va bir qator avtomatlashtirilgan xizmatlarni taklif qiladi. Xavfsizlik,
energiya samaradorligi, qulaylik va sog'liqni saqlash tizimlari aqlli uylarga
birlashtirilgan. Shu bilan birga, ushbu sharoitda sodir bo'ladigan anomal harakatlar
yoki kutilmagan hodisalarni erta aniqlash muammosini hal qilish juda muhimdir.
Ushbu hodisalar tashqi yoki ichki xavflar bilan bog'liq bo'lishi mumkin (masalan,
kutilmagan harakatsizlik, gadjetning buzilishi, ruxsatsiz kirish, sog'liq uchun
favqulodda vaziyatlar). Bugungi kunda aqlli uylarning xavfsizligi va
ishonchliligini kafolatlashning asosiy maqsadlaridan biri bunday holatlarni tan
olish yoki anomaliyalarni aniqlashdir.
Nazoratsiz o'qitish usullariga asoslanib, insho aqlli uy sharoitida sodir
bo'layotgan noodatiy faoliyatni aniqlashning ilmiy va amaliy strategiyalarini ko'rib
chiqadi. Ushbu usullar nazorat ostida o'rganish kabi standart usullardan farqli
o'laroq, aniq belgilarsiz o'rganiladi. Chunki amalda har bir potentsial aberrant
stsenariyni oldindan ko'rish va tavsiflash mumkin emas. Natijada, maqolada
nazoratsiz algoritmlar yordamida tozalash, xususiyatlarni ajratib olish va
anomaliyalarni aniqlash kabi bir nechta sensorlardan to'plangan yuqori o'lchamli,
keng ko'lamli ma'lumotlar bilan ishlash bosqichlari muhokama qilinadi.
Tadqiqotda quyidagi nazoratsiz o'rganish algoritmlari ulardan foydalanish va
samaradorligi uchun tekshirildi:
Izolyatsiya o'rmoni (IF) deb ataladigan daraxtga asoslangan ansambl
yondashuvi tez sur'atlar bilan ajralib turadi.
Shovqinli ilovalarni zichlikka asoslangan fazoviy klasterlash yoki DBSCAN
- bu ilovalarni zichlikka ko'ra guruhlash va chetdagi anormalliklarni aniqlash usuli.
Avtokoder - kiruvchi ma'lumotlarni qayta tiklashdagi xatoni hisoblash orqali
anormalliklarni aniqlash uchun chuqur o'rganishdan foydalanadigan neyron tarmoq
modeli.
One-Class SVMni o'rgatish uchun faqat "oddiy" holatlardan foydalaniladi,
bu holatlardan farq qiluvchi ma'lumotlarni g'ayritabiiy deb tasniflaydi.
Tadqiqotda CASAS, ARAS va KDE Smart Home kabi ochiq manba
maʼlumotlar bazalaridan foydalanilgan. O'lchovni qisqartirish usullari (PCA, t-
SNE) oldindan tozalash, vaqtni ajratish, statistik xususiyatlarni ajratib olish va
sensor ma'lumotlar oqimini tahlil qilish uchun tayyorlash uchun ishlatilgan. Har bir
algoritm samaradorligini baholash uchun standart ko'rsatkichlar, jumladan, aniqlik,
eslab qolish, F1-ball va ROC-AUC ishlatilgan.
Topilmalar shuni ko'rsatadiki, Autoencoder juda aniq va daqiqali
tartibsizliklarni aniqlashda eng yaxshi natijalarni beradi. Izolyatsiya o'rmonining
tezkorligi va sodda dizayni uni real vaqtda tizimlarda foydalanish uchun ideal
qiladi. Sozlamalarga nisbatan sezgirligi tufayli, DBSCAN zichlikka qarab
klasterlarni farqlash qobiliyatiga qaramay, har doim ham barqaror topilmalar
chiqara olmadi. Kichik ma'lumotlar to'plamlarida One-Class SVM yaxshi ishladi,
lekin katta ma'lumotlar to'plamlarida uning hisoblash narxi oshdi.
Har bir modelning afzalliklari va kamchiliklarini muvozanatlash uchun
tadqiqotda ansambl texnikasi yoki ansambl o'rganish ham qo'llanilgan. Haqiqiy
dunyo sharoitida bizning usulimiz noto'g'ri ijobiy va noto'g'ri salbiy holatlar sonini
kamaytirdi.
Insho quyidagi tashvishlarning har birini alohida ko'rib chiqdi:
Kontseptsiyaning o'zgarishi: Modelni foydalanuvchi xatti-harakatlarining
o'zgarishi nuqtai nazaridan qayta tayyorlash kerak.
Harakat, harorat, video, audio va boshqa sensorlar o'rtasidagi aloqalarni
hisobga olish zarurati ko'p ma'lumotlar integratsiyasi deb nomlanadi.
Foydalanuvchi ma'lumotlari bilan ishlashda maxfiylik va maxfiylik muhim
ahamiyatga ega.
Resurs cheklovlari: model chekka qurilmalarda ishlashi uchun u engil
bo'lishi kerak.
Tadqiqot quyidagi kelajakdagi yo'nalishlar bo'yicha tavsiyalar bilan
yakunlandi:
Foydalanuvchilarning o'zgaruvchan odatlariga moslashadigan onlayn ta'lim
tizimlari adaptiv modellar deb nomlanadi.
Federativ ta'lim bilan foydalanuvchi qurilmasi ma'lumotlarni markaziy
serverga uzatmasdan modelni yangilashi mumkin.
Muayyan vaziyat nima uchun g'ayrioddiy deb topilganligini tushuntira
oladigan tizimlar tushuntiriladigan AI deb nomlanadi.
Vaqt, ob-havo, kalendar va boshqa tashqi jihatlarni hisobga oladigan o'quv
modellari kontekstdan xabardor o'rganish deb nomlanadi.
Xulosa qilib aytganda, nazoratsiz o'qitish usullari aqlli uylar kontekstida
sodir bo'layotgan noodatiy faoliyatni samarali aniqlashga qodir. Bu
foydalanuvchilarning xavfsizligi, qulayligi va salomatligi sezilarli darajada
yaxshilanadi. Ushbu maqolada muhokama qilingan usullar haqiqatda keng
qo'llanilishi va haqiqiy aqlli uy tizimlarida foydalanish uchun o'zgartirilishi
mumkin.
 Texnologiyaga bo'lgan ehtiyoj asrning boshidan beri o'sib bormoqda.
Raqamli texnologiyalarning o'sishi ham shu asrning boshida tezlashdi. Jarayonlarni
avtomatlashtirish va tizimlashtirish tobora ommalashib bormoqda. Ushbu tez
rivojlanish uy-joy qurish va yashash sharoitlarini yaxshilashni ham osonlashtirdi.
"Aqlli uylar" iborasi zamonaviy texnologiyalardan foydalangan holda turar-joy
binolari yangilana boshlagani natijasida rivojlandi. Aqlli uylar to'liq
avtomatlashtirilgan maishiy texnika tizimi bilan qurilgan bo'lib, bu uy egasiga
ixtisoslashgan platformalardan uyning elektr ta'minotiga ulangan har qanday
qurilmani qulay boshqarish imkonini beradi. Turli xil IoT qurilmalaridan
foydalangan holda, ushbu tizimlar real vaqtda foydalanuvchi xatti-harakatlarini
kuzatib boradi va bir qator avtomatlashtirilgan xizmatlarni taklif qiladi. Xavfsizlik,
energiya samaradorligi, qulaylik va sog'liqni saqlash tizimlari aqlli uylarga
birlashtirilgan. Shu bilan birga, ushbu sharoitda sodir bo'ladigan anomal harakatlar
yoki kutilmagan hodisalarni erta aniqlash muammosini hal qilish juda muhimdir.
Ushbu hodisalar tashqi yoki ichki xavflar bilan bog'liq bo'lishi mumkin. Masalan,
kutilmagan harakatsizlik, gadjetlarning ishdan chiqishi, tizimga, qurilmaga
ruxsatsiz kirish, sog'liq uchun favqulodda vaziyatlar. Bugungi kunda aqlli
uylarning xavfsizligi va ishonchliligini kafolatlashning asosiy maqsadlaridan biri
bunday holatlarni tanish yoki anomaliyalarni aniqlashdir.
Nazoratsiz o'qitish usullariga asoslanib, insho aqlli uy sharoitida sodir
bo'layotgan noodatiy faoliyatni aniqlashning ilmiy va amaliy strategiyalarini ko'rib
chiqadi. Ushbu usullar nazorat ostida o'rganish kabi standart usullardan farqli
o'laroq, aniq belgilarsiz o'rganiladi. Chunki amalda har bir potentsial aberrant
stsenariyni oldindan ko'rish va tavsiflash mumkin emas. Natijada, maqolada
nazoratsiz algoritmlar yordamida tozalash, xususiyatlarni ajratib olish va
anomaliyalarni aniqlash kabi bir nechta sensorlardan to'plangan yuqori o'lchamli,
keng ko'lamli ma'lumotlar bilan ishlash bosqichlari muhokama qilinadi.
Tadqiqotda quyidagi nazoratsiz o'rganish algoritmlari ulardan foydalanish va
samaradorligi uchun tekshirildi:
Izolyatsiya o'rmoni (IF) deb ataladigan daraxtga asoslangan ansambl
yondashuvi tez sur'atlar bilan ajralib turadi.
Shovqinli ilovalarni zichlikka asoslangan fazoviy klasterlash yoki DBSCAN
- bu ilovalarni zichlikka ko'ra guruhlash va chetdagi anormalliklarni aniqlash usuli.
Avtokoder - kiruvchi ma'lumotlarni qayta tiklashdagi xatoni hisoblash orqali
anormalliklarni aniqlash uchun chuqur o'rganishdan foydalanadigan neyron tarmoq
modeli.
One-Class SVMni o'rgatish uchun faqat "oddiy" holatlardan foydalaniladi,
bu holatlardan farq qiluvchi ma'lumotlarni g'ayritabiiy deb tasniflaydi.
Tadqiqotda CASAS, ARAS va KDE Smart Home kabi ochiq manba
maʼlumotlar bazalaridan foydalanilgan. O'lchovni qisqartirish usullari (PCA, t-
SNE) oldindan tozalash, vaqtni ajratish, statistik xususiyatlarni ajratib olish va
sensor ma'lumotlar oqimini tahlil qilish uchun tayyorlash uchun ishlatilgan. Har bir
algoritm samaradorligini baholash uchun standart ko'rsatkichlar, jumladan, aniqlik,
eslab qolish, F1-ball va ROC-AUC ishlatilgan.
Topilmalar shuni ko'rsatadiki, Autoencoder juda aniq va daqiqali
tartibsizliklarni aniqlashda eng yaxshi natijalarni beradi. Izolyatsiya o'rmonining
tezkorligi va sodda dizayni uni real vaqtda tizimlarda foydalanish uchun ideal
qiladi. Sozlamalarga nisbatan sezgirligi tufayli, DBSCAN zichlikka qarab
klasterlarni farqlash qobiliyatiga qaramay, har doim ham barqaror topilmalar
chiqara olmadi. Kichik ma'lumotlar to'plamlarida One-Class SVM yaxshi ishladi,
lekin katta ma'lumotlar to'plamlarida uning hisoblash narxi oshdi.
Har bir modelning afzalliklari va kamchiliklarini muvozanatlash uchun
tadqiqotda ansambl texnikasi yoki ansambl o'rganish ham qo'llanilgan. Haqiqiy
dunyo sharoitida bizning usulimiz noto'g'ri ijobiy va noto'g'ri salbiy holatlar sonini
kamaytirdi.
Insho quyidagi tashvishlarning har birini alohida ko'rib chiqdi:
Kontseptsiyaning o'zgarishi: Modelni foydalanuvchi xatti-harakatlarining
o'zgarishi nuqtai nazaridan qayta tayyorlash kerak.
Harakat, harorat, video, audio va boshqa sensorlar o'rtasidagi aloqalarni
hisobga olish zarurati ko'p ma'lumotlar integratsiyasi deb nomlanadi.
Foydalanuvchi ma'lumotlari bilan ishlashda maxfiylik va maxfiylik muhim
ahamiyatga ega.
Resurs cheklovlari: model chekka qurilmalarda ishlashi uchun u engil
bo'lishi kerak.
Tadqiqot quyidagi kelajakdagi yo'nalishlar bo'yicha tavsiyalar bilan
yakunlandi:
Foydalanuvchilarning o'zgaruvchan odatlariga moslashadigan onlayn ta'lim
tizimlari adaptiv modellar deb nomlanadi.
Federativ ta'lim bilan foydalanuvchi qurilmasi ma'lumotlarni markaziy
serverga uzatmasdan modelni yangilashi mumkin.
Muayyan vaziyat nima uchun g'ayrioddiy deb topilganligini tushuntira
oladigan tizimlar tushuntiriladigan AI deb nomlanadi.
Vaqt, ob-havo, kalendar va boshqa tashqi jihatlarni hisobga oladigan o'quv
modellari kontekstdan xabardor o'rganish deb nomlanadi.
Xulosa qilib aytganda, nazoratsiz o'qitish usullari aqlli uylar kontekstida sodir
bo'layotgan noodatiy faoliyatni samarali aniqlashga qodir. Bu
foydalanuvchilarning xavfsizligi, qulayligi va salomatligi sezilarli darajada
yaxshilanadi. Ushbu maqolada muhokama qilingan usullar haqiqatda keng
qo'llanilishi va haqiqiy aqlli uy tizimlarida foydalanish uchun o'zgartirilishi
mumkin.
Modern technological advancements are radically altering every facet of social life. Specifically,
developments in automation, big data, cloud computing, artificial intelligence, and the Internet of
Things (IoT) are becoming more and more integrated into daily life. During this process, smart
home technologies are gaining prominence as cutting-edge solutions that help people live more
comfortably, stay secure, and use resources more wisely.

Smart homes are systems that use a variety of sensors, Internet of Things devices, cameras, and
software to automatically monitor a user's everyday activities and offer relevant services. These
systems capture a wide range of data in real time, including movement, temperature, light,
location, human behaviour, opening doors and windows, and turning on gadgets. It will be
feasible to autonomously regulate the home atmosphere, improve security, lower energy usage,
and even keep an eye on health thanks to this data.

However, it is essential to differentiate between typical and aberrant activity for smart homes to
function properly. Because any departure from the user's typical way of life may indicate a
threat, technical malfunction, medical issue, or emergency. For instance, the system can detect
an irregularity and issue a warning if a person consistently gets up at 7:00 and goes into the
kitchen at 7:15, and then these routine behaviours are abruptly absent. Those who live alone, are
elderly, or have impairments should pay particular attention to this.

Because each user's "normal" can vary, identifying anomalies can be challenging. These systems
must therefore be adaptable, customised, and capable of self-learning. Machine learning
techniques are needed for this. In this context, unsupervised learning algorithms in particular
have excellent prospects. because it is typically impossible to predict whether a user's behaviour
will be "normal" or "anomaly." As a result, the model needs to be able to autonomously spot
hidden patterns in the data in real time, construct a model using common patterns, and show any
circumstance that does not fit those patterns as abnormal.

In addition, the data used in smart home systems is often time-bound, sequential (sequential), which
requires more complex approaches than classical anomaly detection methods. It is in such cases that it
is important to test the capabilities of unsupervised machine learning methods such as Autoencoder,
Isolation Forest, DBSCAN, k-Means.

At the same time, the topic of this research takes the interaction between people and technology to a
new level. Smart homes are not simple automated devices, but intelligent systems that learn human
behavior, adapt to it, and guarantee security. The main pillar of these systems is data analysis and
effective anomaly detection.

This research addresses this complex but important issue - the detection of anomalous activities
using unsupervised machine learning approaches in the context of smart homes.
Aqlli uylarda bajarilayotgan ishlar bo’yicha