Benchmarking Transfer Learning Strategies in Time-Series Imaging: Recommendations for Analyzing Raw Sensor Data

With the growing availability and complexity of time-series sequences, scalable robust machine learning approaches are required that overcome sampling challenge quantitatively sufficient training data. Following research trend towards deep learning-based analysis encoded as images, this study proposes a imaging workflow overcomes limited sensor data across domains (i.e., medicine engineering). After systematically identifying three relevant dimensions affect performance visualized data, we performed benchmarking evaluation with total 24 unique convolutional neural network models. two-level transfer investigation, reveal fine-tuning mid-level features results in best classification performance. As result, present an optimized representation VGG16 network, which outperforms previous studies field. Our approach is accurate, robust, manifests internal external validity. By only using raw our model does not require manual feature engineering, being high practical relevance. post-hoc reveals allows automated extraction meaningful based on underlying also adds to explainable artificial intelligence. Furthermore, proposed reduces sequence length input while preserving all information. Especially hurdle long-term dependencies sequential related work’s limitation vanishing gradients problem contribute theory