Recent advancements in the domain of human activity recognition (HAR) are increasingly aimed at developing methodologies, approaches, and models for real-time, multi-step activity recognition and analysis. This work presents a smartwatch-based approach for complex, real-time HAR that is applicable but not limited to internal logistics systems that use autonomous guided vehicles. A distributed smartwatch-based data collection system was developed, and a dataset was gathered and published, containing readings from human staff representatives executing activity sequences representing typical internal logistics tasks. A HAR-specific, pre-trained DenseNet121 was used for basic activity classification, achieving an F1-score of 91.01%. For multi-step activity classification, we compared models based on CNN, LSTM, BiLSTM, GRU, and BiGRU as meta-classifiers, employing different dataset-splitting strategies and models’ configurations. The best-performing CNN-based model achieved an F1-score of 87.44% using the shared dataset utilization approach. Despite the challenges faced, the adaptability of the proposed approach suggests that it can be integrated into an intelligent enterprise management system to provide a robust and flexible HAR framework that enhances production efficiency.

Modern methods for solving the AGV battery cell voltage prediction problem include a symbiosis of probabilistic and machine learning methods. In this study, we propose to use two RNN-based approaches with preliminary data analysis and processing. Five filters were selected for smoothing: Moving average; Weighted moving average; Exponential smoothing; Smoothing with local regression; Lowess smoothing. RNN forecasting was performed on the basis of a centered stationary signal obtained by subtracting the value obtained with the five filters from the original signal.The forecast accuracy was evaluated using the following errors: MAPE, RMSE, MSPE, RMSPE, MFE, MBE, SEBIAS, ME. Adequacy of the proposed neural network model based on the determination coefficient R2. The coefficient of determination for all the centered signals obtained using the above five methods yielded adequate predictive models. Exponential smoothing (MAPE=0.002; RMSE=141.481; R2=0.94) and LS (MAPE=0.002; RMSE=124.239; R2:0.96) proved to be the most effective methods for short-term forecasts. Moving average (MAPE=0.00207; RMSE=146.589; R2=0.94) is recommended for long-term forecasts with additional smoothing of the forecasted values because of the noise in them.

We reviewed the existing and currently used approach in processing the residual charge of an AGV battery. The method of setting up the experiment for collecting the historical data for an AGV Formica 1 of the AIUT company was proposed and implemented. The collected properties of the time series were analyzed and the algorithm for the necessary data pre-processing was selected. This algorithm includes padding any the suppression spontaneous peaks, the recovery of any lost data and data normalization.The collected data for the AGVs were also analyzed using the correlation analysis methods (Pearson, Spearman and Kendall correlations). These determined the parameters on which the AGV battery discharge depends. A battery discharge prediction approach that is based on the quasi-stochastic signal's probabilistic characteristics is suggested.A Multiparameter ANN model using a time window was developed. The dependence of the forecast error on the length of the time window was also investigated. The optimal parameters of the ANN were selected experimentally. The mean absolute percentage error for the AGV short-term forecast of a battery discharging was less than 1%. For the other parameters on which it depends, the AGV battery discharging was less than 9%. All of the studies were conducted within the framework of the "Automated Guided Vehicles integrated with Collaborative Robots for Smart Industry Perspective" project.