Resumen de Automatic feature extraction for time series analysis using deep and machine learning
Time series analysis is crucial in understanding and extracting valuable insights from temporal data, capturing the inherent patterns, trends, and dependencies that evolve over time. This study concisely overviews the state of the art key components and methodologies involved in time series analysis and classification for both industrial and Electrocardiogram (ECG) signals using deep learning approaches like CNNs, LSTMs and transformers etc.
The analysis begins with a use case study of industrial collaboration with the lens manu- facturing industry where the process models called CRISP-DM and DMME for industrial data science are implemented to an industrial production dataset. The collaboration im- proved industrial production by decreasing the downtime for cleaning the machinery. The study also delves into the shortcomings of the process models and the actual challenges faced during the implementation of data science in real industrial projects. Furthermore, the study explores fundamental concepts such as the methodology of analyzing ECG in particular and time series in general. To lay the groundwork for subsequent advanced techniques of classification by deep learning, the basic construct of the human heart and ECG signals is also presented in detail.
The next vital TSA subject probed was fall detection using ECG signals. The paper introduces a novel approach using electrocardiogram (ECG) signals for fall detection and activity classification. An algorithm employing pre-trained convolutional neural networks (AlexNet and GoogLeNet) as classifiers is proposed, achieving a significant val- idation accuracy of 98.08% for distinguishing between fall and no-fall scenarios in the first model. The signals are pre-processed to reduce noise, and frequency-time representations (scalograms), which are obtained through continuous wavelet transform, serve as feature extractors. The trained model accurately distinguishes ECGs with fall activity from those without at an accuracy of 98.02%. The robustness of the algorithm is verified by augmenting the experimental dataset with publicly available datasets, achieving a clas- sification accuracy of 98.44% in the second model, which classifies fall, daily activities, and no activities. The models, developed through transfer learning from real images to medical images, offer a lightweight solution compared to traditional deep learning approaches, avoiding redundant computational efforts.
In recent studies on electrocardiogram (ECG) signal classification using deep learning (DL), the focus has been on complex DL methods like transfer learning or feature ex- traction based on domain knowledge. As the next steps, this study challenges the common assumption that deeper and more complex DL models lead to better learn- ing. Instead, the authors propose two novel DL models: a CNN-LSTM hybrid and an attention/transformer-based model with wavelet transform for dimensional embedding.
These models extract features from ECG signals in the initial layers, demonstrating per- formance on par with or surpassing many contemporary deep neural networks. Validation using three publicly available datasets shows benchmark accuracy, reaching 99.92% for fall detection and 99.93% for PTB database classification of myocardial infarction versus normal heartbeat.
While transformer models have demonstrated superior performance in natural language processing, their careful adoption is essential for achieving comparable results in the realm of time series classification. This study, to the best of our knowledge, for the first time, explores the impact of various dimensional embedding techniques in time series classifications. The exploration includes the use of wavelet transformation, discrete and continuous wavelets, scattering, and feature maps from convolutional neural networks for performance comparison. Several ECG datasets from UCR dataset and Physionet are employed for both multi-class and binary classification. The experimental results con- sistently reveal that incorporating relevant feature extraction techniques as dimensional embedding outperforms a plain transformer approach.
In conclusion, this dissertation offers a comprehensive overview of the multifaceted realm of automatic feature extraction for time series classification, serving as a guide for re- searchers, practitioners, and enthusiasts seeking to work with temporal data and harness its capability for informed decision-making.
