Recent attention to privacy issues demands a communication-safe method for training human activity recognition (HAR) models on client activity data. Federated learning (FL) has become a compelling technique to facilitate model training between the server and clients while preserving data privacy. However, classical FL methods often assume independent and identically distributed (IID) data among clients. This assumption does not hold true in practical scenarios. Human activity in real-world scenarios varies, resulting in skewness where identical activities are executed uniquely across clients. This leads to local model objectives drifting away from the global model objective, thereby impeding overall convergence. To address this challenge, we propose FedCoad, a novel federated model leveraging contrastive learning with adaptive control variates to handle the skewness among HAR clients. Model contrastive learning minimizes the gap in representation between global and local models to help global model convergence. During local model updates, the adaptive control variates penalize the local model updates with respect to the model weight and the rate of change from the control variates update. Our experiments show that FedCoad outperforms state-of-the-art FL algorithms on HAR benchmark datasets.

Sensor-based Human Activity Recognition (HAR) is increasingly utilized to automatically detect daily human activities, stimulated by the widespread adoption of wearable devices. To protect user privacy, the Federated Learning (FL) framework is often applied in sensor-based HAR, ensuring that raw data remains within the confines of the wearable device. This data isolation typically results in unlabeled raw data, as labeling is costly, time-consuming, and would require sending data to external experts. Moreover, implementing sensor-based HAR in real-world scenarios faces challenges such as computational constraints on wearable devices and the non-IID nature of FL data. In response, we propose a novel framework, LiCAFeL-STC, designed to train sensor-based HAR under conditions where only unlabeled data is available on wearable devices. These devices are limited in computational resources, and the data exhibits high heterogeneity, simulating the non-IID nature of FL data. Our framework employs signal transformation classification as an auxiliary self-supervised learning (SSL) technique to leverage large amounts of unlabeled data on wearable devices and incorporates a clustering mechanism to group similar devices, mitigating the non-IID problem. Our findings demonstrate that LiCAFeL-STC can outperform both the conventional method and baseline frameworks under similar experimental settings.

Detecting an anomaly is an essential task in the manufacturing operation. Due to the vast adaptation of machines for industry, AI has become an indispensable part of detecting anomalous instances. However, data scarcity and cost allocation pose a significant challenge for individual companies to train a model alone. Therefore, in Industries 5.0, companies need to collaborate to achieve the common goal. On the other hand, they also need to disclose sensitive information according to General Data Protection Regulation (GDPR). This work presents a secure collaborative framework with Federated Learning to enable the development of an anomaly detection model among multiple machines or clients in different companies. The proposed framework performs tasks such as managing secure connections among clients, transforming client data to a processable format, and conducting model training between clients simultaneously.

Due to the rapid development of sensor technology, wearable sensors have been widely applied in various real-life human applications and improved the mass adoption of smart environments. This recent technology offers a pioneering opportunity to recognize human daily behavior patterns from a large amount of collected data. In this work, we address the challenge of applying process mining to discover human daily behavior patterns from sensors. Sensor data could be seen as the execution of a process representing user daily activity. However, it requires additional tasks for investigating the perspective of sensor data that comes from diverse environmental settings. Therefore, we propose an analytical framework to discover human daily behavior patterns from various sensors in a smart environment. In order to evaluate the proposed framework, a real-world dataset in a smart environment is used. From the conducted experiment, this framework could be used to transform general human activity data and takes into account the process mining to discover human daily behavior in accordance with a multi-case perspective process model (i.e., user-based, time-based, and sensor flow pattern).

The concept of connected things by involving emotional aspects has been raised as a new research issue which is known as "emotional IoT". The deeper interaction between object and human shows an importance to develop a system with either cognitive or affective capabilities such as emotion. While the existing works on real time emotion recognition mostly rely on facial data, there are a few works dealing with real time emotion recognition based on physiological data using pervasive devices. In this work, we propose a framework to recognize emotion based on human physiological signals using the pervasive wearable device. This framework opposed most of the works which employed sensors which are expensive and complex in arrangement. The challenge on using pervasive devices is the low accuracy due to the low sampling rate. The approach is implemented in an end-to-end soft real time emotion recognition system using smartphone and smartwatch devices. The performance of our system was evaluated under a common environment and proved the system applicability throughout everyday life.

Rotation invariant texture descriptor plays an important role in texture-based object classification. However the classification accuracy may decrease due to the inconsistent performance of texture descriptor with respect to various rotated angles. In this paper we propose a consistent rotation invariant texture descriptor named Sorted Neighborhood Differences (SND). SND is derived from the integration of sorted neigh- borhood and binary patterns. Experimental results show that overall texture classification accuracy of SND with respect to different rotations using OUTEX TC 0010 texture database is 91.81% whereas those of LBPriu and LBP-HF are 86.42% and 88.28%, respectively. The texture and coin classification accuracies of SND are also consistent in various rotation angles and illumination levels.

Time series data, which are a series of one dimensional real numbers, have been studied in various database applications. We extend the traditional similarity search methods on time series data to support a multidimensional data sequence, such as a video stream. We investigate the problem of retrieving similar multidimensional data sequences from a large database. To prune irrelevant sequences in a database, we introduce correct and efficient similarity functions. Both data sequences and query sequences are partitioned into subsequences, and each of them is represented by a Minimum Bounding Rectangle (MBR). The query processing is based upon these MBRs, instead of scanning data elements of entire sequences. Our method is designed: (1) to select candidate sequences in a database, and (2) to find the subsequences of a selected sequence, each of which falls under the given threshold. The latter is of special importance in the case of retrieving subsequences from large and complex sequences such as video. By using it, we do not need to browse the whole of the selected video stream, but just browse the sub-streams to find a scene we want. We have performed an extensive experiment on synthetic, as well as real data sequences (a collection of TV news, dramas, and documentary videos) to evaluate our proposed method. The experiment demonstrates that 73-94 percent of irrelevant sequences are pruned using the proposed method, resulting in 16-28 times faster response time compared with that of the sequential search.