Seksan Laitrakun

Seksan Laitrakun, Somrudee Deepaisarn, Sarun Gulyanon, Chayud Srisumarnk, Nattapol Chiewnawintawat, Angkoon Angkoonsawaengsuk, Pakorn Opaprakasit, Jirawan Jindakaew, and Narisara Jaikaew. Journal of information and communication convergence engineering 2023;21:208-15. https://doi.org/10.56977/jicce.2023.21.3.208

Human activity recognition (HAR) has been applied to several advanced applications, especially when individuals may need to be monitored closely. This work focuses on HAR using wearable sensors attached to various locations of the user body. The data from each sensor may provide unequally discriminative information and, then, an effective fusion method is needed. In order to address this issue, inspired by the squeeze-and-excitation (SE) mechanism, we propose the merging-squeeze-excitation (MSE) feature fusion which emphasizes informative feature maps and suppresses ambiguous feature maps during fusion. The MSE feature fusion consists of three steps: pre-merging, squeeze-and-excitation, and post-merging. Unlike the SE mechanism, the set of feature maps from each branch will be recalibrated by using the channel weights also computed from the pre-merged feature maps. The calibrated feature maps from all branches are merged to obtain a set of channel-weighted and merged feature maps which will be used in the classification process. Additionally, a set of MSE feature fusion extensions is presented. In these proposed methods, three deep-learning models (LeNet5, AlexNet, and VGG16) are used as feature extractors and four merging methods (addition, maximum, minimum, and average) are applied as merging operations. The performances of the proposed methods are evaluated by classifying popular public datasets.

As short-term load forecasting is essential for the day-to-day operation planning of power systems, we built an ensemble learning model to perform such forecasting for Thai data. The proposed model uses voting regression (VR), producing forecasts with weighted averages of forecasts from five individual models: three parametric multiple linear regressors and two non-parametric machine-learning models. The regressors are linear regression models with gradient-descent (LR), ordinary least-squares (OLS) estimators, and generalized least-squares auto-regression (GLSAR) models. In contrast, the machine-learning models are decision trees (DT) and random forests (RF). To select the best model variables and hyper-parameters, we used cross-validation (CV) performance instead of the test data performance, which yielded overly good test performance. We compared various validation schemes and found that the Blocked-CV scheme gives the validation error closest to the test error. Using Blocked-CV, the test results show that the VR model outperforms all its individual predictors.

Human activity recognition (HAR) is an emerging filed to classify daily activities and is applicable to many applications including healthcare. Several machine learning models and approaches have been investigated to improve the classification performances. In this paper, we investigate several fusion approaches of heterogeneous multichannel models where the channels use different kinds of deep learning (DL) algorithms to better the HAR performances. The DL algorithms exploited in those channels are convolutional neural networks (CNNs), long short-term memory networks (LSTMs), CNN-LSTMs, and convolutional LSTMs. Six different fusion approaches based on feature fusion and decision fusion are applied to combine the outputs of the channels. Their classification performances are evaluated and compared. The best fusion approaches achieve 98.96% and 96.35% of accuracy when using the PAMAP2 and DaLiAc datasets, respectively.

We consider the design of distributed detection algorithms for single-hop, single-channel wireless sensor networks in which sensor nodes send their local decisions to a fusion center (FC) by using a random access protocol. There is also limited time to collect local decisions before a final decision must be made. We thus propose and analyze a modified random access protocol in which the FC combines slotted ALOHA with a population-splitting algorithm called population-splitting-based random access (PSRA) and collision-aware distributed detection according to an estimate-then-fuse approach. Under the PSRA, only sensor nodes whose observations fall in a particular range of reliability will send their decisions in a specific frame by using slotted ALOHA. At the end of the collection time, the FC applies the collision-aware distributed detection to make a final decision. Here, the FC will first observe the state of each time slot—idle, successful, collision—in each frame, use this information to estimate the number of sensor nodes participating in each frame, and, then, compute a final decision using a population-based fusion rule. An approximation of the optimal transmission probability of the slotted ALOHA is determined to minimize the probability of error. Numerical results show that, unlike slotted-ALOHA-based data networks, the transmission probability maximizing the number of successful time slots does not optimize the performance of the proposed distributed detection. Instead, the proposed distributed detection performs best with a transmission probability that induces many collisions.

Human activity and fall recognition have been applied into many applications, especially, daily activity and healthcare systems. The recognition performance is up to the features used in the classification. Applying all features might degrade the performance since irrelevant features are used. In this paper, we study and compare four filter methods of feature selection: Chi square, mutual information, ANOVA, and Pearson’s coefficient. Six supervised classifiers are used to evaluate the classification performance according to the sets of features obtained from these methods. The dataset UniMiB SHAR, which consists of nine basic daily activities and eight types of falls, is investigated in the experiment. As a result, among all of the considered feature-selection methods and classifiers, the XGBoost incorporating with the mutual information method gives the best performance whose accuracy, precision, recall, and F1 scores are 0.9122, 0.8727, 0.8629, and 0.8640, respectively.

We consider the composite hypothesis testing problem of time-bandwidth-constrained distributed detection. In this scenario, the probability distribution of the observed signal when the event of interest is happening is unknown. In addition, local decisions are censored and only those uncensored local decisions will be sent to the fusion center over a shared and noisy collision channel. The fusion center also has a limited time duration to collect transmitted decisions and make a final decision. Two types of medium access control that the sensor nodes apply to send their decisions are investigated: time division multiple access and slotted-Aloha. Unlike using the time division multiple access protocol, the slotted-Aloha-based distributed detection will experience packet collisions. However, in this article, since only uncensored decisions are sent, packet collisions are informative. We derive fusion rules according to generalized likelihood ratio test, Rao test, and Wald test for both the time division multiple access–based distributed detection and the slotted-Aloha-based distributed detection. We see that the fusion rules for the slotted-Aloha-based distributed detection here also exploit packet collisions in the final decision-making. In addition, the asymptotic performances and energy consumption of both schemes are analyzed. Extensive simulation and numerical results are provided to compare the performances of these two schemes. We show that, for a given time delay, the slotted-Aloha-based distributed detection can outperform the time division multiple access–based distributed detection by increasing the number of sensor nodes which results in higher energy consumption.

Thailand Provincial Electricity Authority (PEA) has 19.36 million residential customers and approximately 73% of them have late or unpaid electricity payments. This causes several problems including reduction of PEA’s income. Therefore, we aim to classify and find data mining algorithms to predict classes of electricity payment behaviors of residential customers in order to improve customer service and to reduce the number of late or unpaid payment of electrical bills. In this work, the data set consists of payment behaviors and demographic characteristics of customers, which are gathered from 1,079 customers. The prediction process is divided into two main parts: clustering classes of customers by using k-means algorithm, and using data mining algorithms for customer model prediction. We performed experiments using five data mining algorithms: logistic regression, decision tree, random fores, support vector machine (SVM) and XGBoost. Our results show that the random forest algorithm is a good choice for this predictive classes of customers to develop appropriate strategies to increase satisfaction of customers and revenue of PEA.