Search

This study adopted multiscale convolutional neural networks (MCNNs) to increase the volume of feature extraction, and a long short-term memory (LSTM) network was employed to rank time-series signals. Three simulated marine environments were examined, namely the Kinmen Sea, the southwestern sea, and the eastern sea areas of Taiwan. In addition, signal transmission and reception experiments were performed in a small fishpond and a swimming pool. To eliminate the intersymbol interference (ISI) induced by multipath interference, a virtual time reversal mirror (VTRM) was used to optimize the data collected from the simulated underwater environments. The results of this study reveal that after the data collected from the southwestern sea area were optimized and used for training, the resulting models could be used to demodulate the data collected from the Kinmen and eastern sea areas. In the simulations, when the signal-to-noise ratio (SNR) in the Kinmen Sea area was −14 and 14 dB, the bit error rate (BER) of the aforementioned model was 0.00145 and 0.00019, respectively. In the experiments, when the MCNN–LSTM model was trained using fishpond data under a transmit power of 0.003 and 0.01 W, the BER of the model was 0.000083 and 0.000025, respectively. When this model was trained using swimming pool data under a transmit power of 0.003 and 0.01 W, the BER was 0.000008 and 0.000004, respectively. The simulation and experimental results indicated that when model training and testing were performed using data collected from the same sea area, the MCNN–LSTM model was more accurate than the convolutional neural network (CNN)–LSTM model.

This study employed wearable inertial sensors integrated with an activity-recognition algorithm to recognize six types of daily activities performed by humans, namely walking, ascending stairs, descending stairs, sitting, standing, and lying. The sensor system consisted of a microcontroller, a three-axis accelerometer, and a three-axis gyro; the algorithm involved collecting and normalizing the activity signals. To simplify the calculation process and to maximize the recognition accuracy, the data were preprocessed through linear discriminant analysis; this reduced their dimensionality and captured their features, thereby reducing the feature space of the accelerometer and gyro signals; they were then verified through the use of six classification algorithms. The new contribution is that after feature extraction, data classification results indicated that an artificial neural network was the most stable and effective of the six algorithms. In the experiment, 20 participants equipped the wearable sensors on their waists to record the aforementioned six types of daily activities and to verify the effectiveness of the sensors. According to the cross-validation results, the combination of linear discriminant analysis and an artificial neural network was the most stable classification algorithm for data generalization; its activity-recognition accuracy was 87.37% on the training data and 80.96% on the test data.