ГІБРИДНІ ТЕХНОЛОГІЇ МАШИННОГО НАВЧАННЯ ДЛЯ ПРОГНОЗУВАННЯ КОМПЛЕКСНОЇ ДІЯЛЬНОСТІ ПРОМИСЛОВОГО ПЕРСОНАЛУ ЗА ДАНИМИ СМАРТ-ГОДИННИКІВ

Context. In today’s industrial development, significant attention is paid to systems for recognizing and predicting human activity in real time. Such technologies are key to the transition from the concept of Industry 4.0 to Industry 5.0, as they allow for improved interaction between man and machine, as well as to ensure a higher level of safety, adaptability and efficiency of production processes. These approaches are particularly relevant in the field of internal logistics, where cooperation with autonomous vehicles requires a high level of coordination and adaptability.Objective. To create a technological solution for the prompt detection and prediction of complex human activity in the internal logistics environment by using sensor data from smart watches. The main goal is to improve cooperation between employees and automated systems, increase occupational safety and efficiency of logistics processes.Method. A decentralized data collection system using smart watches has been developed. A mobile application in Kotlin was created to capture sensor readings during a series of logistics actions performed by five workers. To process incomplete or distorted data, anomaly detection algorithms were applied, including STD, logarithmic transformation of STD, DBSCAN, and IQR, as well as smoothing methods such as moving average, weighted moving average, exponential smoothing, local regression, and Savitsky-Goley filter. The processed data were used to train models, with the employment of such advanced techniques as transfer learning, continuous wavelet transform, and classifier stacking.Results. The pre-trained deep model with the DenseNet121 architecture was chosen as the base classifier, which showed an F1- metric of 91.01% in recognizing simple actions. Five neural network architectures (single- and multi-layer) with two data distribution strategies were tested to analyze complex activity. The highest accuracy – F1-metric 87.44% – was demonstrated by the convolutional neural network when using a joint approach to data distribution.Conclusions. The results of the study indicate the possibility of applying the proposed technology for real-time recognition of complex human activities in intra-logistics systems based on data from smart-watch sensors, which will improve human-machine interaction and increase the efficiency of industrial logistics processes