Accurate prediction of a user's movement path has various advantages for many applications, such as optimising a nurse's trajectory in a hospital and assisting elderly or disabled people and making them feel secure and protected in the places where they live. Recently, researchers have suggested techniques based on machine learning and deep learning in this field. However, these approaches have drawbacks such as their low accuracy in classifying the extracted features into associated movement paths, high sensitivity to noisy data, and ignoring time dependencies within raw data. In this work, a threephase stacked method named CNN-LSTM-FSC is proposed, which uses the Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Fuzzy Soft-max Classifier (FSC) to overcome the mentioned constraints. In the first phase, the CNN structure extracts time dependencies within raw data using the stacked convolutional and pooling layer. In the second phase, the long-term time dependency of the user's movement path is learnt using the LSTM layers, and the user path is determined using a new innovated fuzzy soft-max classifier. Finally, in the post-processing phase, by performing a majority voting technique on the k-adjacent sample predictions of the classifier, the authors have tried to reduce the effects of noise in identifying the user's movement path. Experiments were conducted on the MovementAAL_RSS dataset. The proposed method has successfully reached 93.86%, 93.71%, and 93.26% accuracy rate on the Move-mentAAL_RSS datasets, with 0%, 5%, and 10% Gaussian noise, respectively, and demonstrates superior results in comparison to the previous literature research. K E Y W O R D Sconvolutional neural network, fuzzy soft-max classifier, long short-term memory, user's movement path | INTRODUCTIONDifferent user movement prediction methods attempt to derive patterns within input data that predict future events [1][2][3]. These patterns can be used effectively to optimise resources and provide intelligent services. Today, motion prediction methods are widely used in a variety of fields, including wireless cellular networks, Location-based Services (LBS) (which include navigation services, social networking services, monitoring and tracking objects in warehouses) [4,5], Ambient Assisted Living (AAL) [6,13] and various applications (including utilities-based applications, health care, safety systems, and alert applications). Two key components in movement prediction systems include hardware environment with its indoor positioning technique and its prediction or classification technique. Hardware environments with indoor positioning can be made in different ways, such as Wireless Sensor Networks (WSN) [5] or Ultra-Wideband (UWB) [7]. A WSN-based system consists of small sensors armed with environmental power sources, sensing devices, radio, and processor units. User's movement is predicted by utilising Received Signal Strength (RSS) information [8] between some WSN and a user of wearable sensors. Then data-drivenThis is an ope...

Human activity recognition has been a popular research topic in recent years. The rapid development of deep learning techniques has greatly helped researchers to achieve success in this field. During the training process with deep learning techniques, features and time dependencies between them are well learned. However, researchers generally ignore the distribution of extracted features in the coordinate space despite their significant effect on classification and network convergence status. The present article utilizes a simple but effective centralized coordinate learning method that dispersedly spans extracted features across the coordinate space. This method causes the angle between the features of different classes to increase significantly. A hybrid loss function is also suggested to enhance the discriminative power of learned features. Some experiments were carried out on the OPPORTUNITY and the PAMAP2 datasets. The results showed that the proposed method outperformed the recently proposed deep learning methods, including the Deep ConvLSTM, CNN‐LSTM‐ELM, and Hybrid methods. This high efficiency was due to the identification of discriminative features.