Loading...
Hierarchical Activity Recognition for PD Patients by Means of an IMU-Based Wearable System
Mohammadzadeh Ghahfarokhi, Mohammad | 2024
3
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 56953 (08)
- University: Sharif University of Technology
- Department: Mechanical Engineering
- Advisor(s): Behzadipour, Saeed
- Abstract:
- Parkinson's disease is a progressive and severe neurodegenerative disorder in the central nervous system that is common in adults and, to some extent, in young people. One way to control and improve this disease is through rehabilitation, specifically in the form of rehabilitation activities. Due to the difficulty of travelling to and from the clinic, systems have been developed today that, using inertial sensors, provide the possibility of remotely monitoring this type of treatment up to 95%. In this regard, a system named SEPANTA has been developed at the Djavad Mowafaghian Research Center to diagnose and monitor the rehabilitation activities performed by patients with Parkinson's disease. This system includes a smartphone and four inertial sensors connected to the wrists and thighs of Parkinson's patients. The inertial sensors in this system obtain movement information, including angular velocities and acceleration of their attachment points to the body, and send them to a smartphone to detect the type of activity performed by the patient. The classifier running on the smartphone is tasked with using the signals sent from the inertial sensors to detect in real time the type of rehabilitation activities performed by the patient. Recent studies conducted on the SEPANTA system have identified shortcomings in its performance in the real-time detection of rehabilitation activities, indicating a need for improvement. The purpose of this project is to enhance the capability of the SEPANTA system in the real-time diagnosis of a variety of rehabilitation activities suggested for patients with Parkinson's disease. To achieve this goal, a hierarchical classification method has been employed. This classification method, unlike conventional approaches, detects activity types in multiple stages. The therapeutic activities outlined in this research can be categorized based on their similarity to one another. Using this method, it becomes feasible to tackle the primary challenge in conventional activity classification methods, which is ignoring the similarity across different activities during classification. For this purpose, in this research, a hierarchical classification was implemented, which doesn't limit the detecting process to a single classification. Instead, based on the similarity of therapeutic movements, activities are placed in different groups. Specific classifications designed for each activity group are then used to identify the type of activities performed. The proposed classification method is not restricted to a single category; it includes steps such as data generation, detection of therapeutic movements from unknown movements, adaptive windowing, design and training of hierarchical classification, and performance evaluation of the activity recognition system. Various data pre-processing methods have been employed for each section of the classifier system. Among the pre-processing steps performed are filtering and removing faulty parts of data, interpolating data, and grouping activities. Data generation is accomplished using a generative adversarial network, recognition of therapeutic motions from unknown motions by an Autoencoder model, and adaptive windowing by a Q deep learning model. The designed hierarchical classifier consists of a two-layer attention-based classifier that utilizes pre-processed data for training. At the end of the current research, the effect of each classifier’s part was investigated, and the results obtained from the proposed model were compared with the results of other existing research. The comparison results demonstrated that the proposed classification method is superior in performance, achieving a recognition accuracy of 93% in an average time of 3.48 seconds
- Keywords:
- Parkinson Disease ; Hierarchical Classification ; Data Generation ; Generative Adversarial Networks ; Anomaly Detection ; Autoencoder ; Adaptive Windowing ; Wearable Activity Detection System ; Attention Networks
-
محتواي کتاب
- view
- فصل 1 مقدمه
- فصل 2 مفاهیم اولیه
- 2 - 1 - مقدمه
- 2 - 2 - بیماری پارکینسون
- 2 - 1 - حرکات درمانی در بیماری پارکینسون
- 2 - 2 - سامانه پوشیدنی تشخیص فعالیت
- 2 - 3 - انواع پیشپردازش دادهها
- 2 - 4 - ساختارهای مورد استفاده برای تشخیص فعالیت
- 2 - 4 - 1 - شبکههای پرسپترون چندلایه
- 2 - 4 - 2 - شبکه عصبی پیچشی (CNN)
- 2 - 4 - 3 - شبکههای بازگشتی (RNN)
- 2 - 4 - 4 - حافظه طولانی کوتاه-مدت (LSTM)
- 2 - 4 - 5 - واحد برگشتی دروازهای (GRU)
- 2 - 4 - 6 - مکانیزمهای خود توجه (self-attention mechanism)
- 2 - 4 - 7 - لایهی تماماً متصل57F
- 2 - 4 - 8 - لایه softmax
- 2 - 5 - شبکه مولد متخاصم
- 2 - 6 - شبکه خودرمزگذار
- 2 - 7 - یادگیری تقویتی
- فصل 3 مرور ادبیات
- فصل 4 تولید داده
- 4 - 1 - مقدمه
- 4 - 2 - پیشپردازش
- 4 - 3 - ساختار مدل تولید داده
- 4 - 4 - ساختمان شبکه مولد متخاصم
- 4 - 5 - فرایند آموزش شبکه مولد متخاصم
- 4 - 6 - تولید دادههای فعالیتهای پیچیده
- 4 - 7 - ارزیابی دادههای تولیدی
- 4 - 8 - تولید داده با استفاده از روشهای داده افزایی
- 4 - 9 - عملکرد مدل مولد متخاصم پیشنهادی در برابر سایر روشهای تولید داده
- 4 - 10 - جمعبندی
- فصل 5 توسعه دستهبند سلسله مراتبی
- 5 - 1 - مقدمه
- 5 - 2 - دستهبندی سلسله مراتبی
- 5 - 3 - گروهبندی حرکات درمانی
- 5 - 4 - پیشپردازش
- 5 - 5 - ساختار مدل تشخیص فعالیت سلسله مراتبی
- 5 - 6 - نتایج دستهبند سلسله مراتبی پیشنهادی در تشخیص حرکات درمانی
- 5 - 7 - ماژولهای کمکی دستهبند
- 5 - 8 - مقایسه نتایج حاصل از دستهبند پیشنهادی و سایر دستهبندهای سامانه تشخیص فعالیت سپنتا
- فصل 6 بحث و نتیجهگیری
- فصل 7 مراجع