TY - GEN A1 - Haddadi Esfahani, Ali A1 - Maye, Oliver A1 - Frohberg, Max A1 - Speh, Maria A1 - Jöbges, Michael A1 - Langendörfer, Peter T1 - Machine Learning based Real Time Detection of Freezing of Gait of Parkinson Patients Running on a Body Worn Device T2 - IEEE/ACM international conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE 2023), 181 (2023) N2 - For those who have Parkinson's disease, one of the most incapacitating symptoms is Freezing of Gait (FOG). Gait impairment and disruptions limit everyday activities and reduce quality of daily life along with the increase in the risk of falling [1]. Thanks to recent advancement in embedded electronics and sensors as well as their adaptation in the wearable device market, low power devices are becoming more and more capable running neural networks. This enables researchers to implement complex models on wearable devices that capture and analyze sensor data to detect FOGin real-time. KW - Parkinson's disease KW - Wearable computers KW - Neural networks KW - Machine learning KW - Real-time systems Y1 - 2023 SN - 979-8-4007-0102-3 U6 - https://doi.org/10.1145/3580252.3589423 SN - 2832-2975 SP - 181 EP - 182 ER - TY - GEN A1 - Assafo, Maryam A1 - Langendoerfer, Peter T1 - Tool remaining useful life prediction using feature extraction and machine learning-based sensor fusion T2 - Results in engineering N2 - Tool remaining useful life prediction (RUL) is a critical task for predictive maintenance in manufacturing. Common limitations of existing data-driven solutions include: 1) Dependence on tool wear labels which are intricate to obtain on shop floors. 2) High resource requirements, affecting applicability on resource-constrained Internet-of-things devices. 3) Heavy feature engineering. To address these limitations, we present a methodology aiming at accurately predicting RUL without using wear labels, while ensuring implementation efficiency and minimal feature engineering. It involves extracting time-domain features and multiscale features using maximal overlap discrete wavelet transform (MODWT) from three cutting-force sensor signals. Without undergoing any feature selection or dimensionality reduction, the features are fed to machine learning (ML) regression models where they are fused into an RUL decision. For this purpose, one-to-one and sequence-to-sequence regression using random forest (RF) and different long short-term memory (LSTM) networks were used, respectively. The 2010 PHM Data Challenge milling dataset was used for validation. The results highlighted the significant role of sensor fusion in reducing prediction errors and increasing the performance consistency over three test cutters, compared to single sensors. Global interpretations were provided using RF-based feature importance analysis. Our methodology was compared with six existing state-of-the-art works, including different end-to-end deep learning (DL) models using raw data as input, and works coupling heavy feature engineering with DL. The results showed that our methodology consistently outperformed all the comparative methods over the test cutters, despite using fewer sensors, which further proves its competitiveness and suitability in resource- and sensor-constrained environments. KW - Cutting tool KW - Data-driven models KW - Feature extraction KW - Machine learning KW - Predictive maintenance KW - Remaining useful life prediction KW - Sensor fusion Y1 - 2025 U6 - https://doi.org/10.1016/j.rineng.2025.107297 SN - 2590-1230 VL - 28 SP - 1 EP - 14 PB - Elsevier BV CY - Amsterdam ER -