THE OPTIMAL STATE ESTIMATION FUSION ALGORITHM IN DISCRETE-CONTINUOUS AUV SYSTEMS

Recently, neural network models have become one of the most promising directions in the field of automaticfeature extraction from digital signals. Traditional approaches, such as statistical, time-domain,frequency-domain, and time-frequency analysis, require significant expert knowledge and often prove insufficientlyeffective when dealing with non-stationary and complex signals, such as biomedical signals (ECG,EEG, EMG) or industrial signals (e.g., currentgrams). These methods have several limitations when it comesto analyzing multichannel data with varying frequency structures or when signal labeling is too laborintensiveor expensive. Modern neural network architectures, such as transformers, have demonstrated highefficiency in automatic feature extraction from complex data. Transformers have outperformed traditionalconvolutional and recurrent neural networks in many key metrics, particularly in tasks involving time seriesforecasting, multimodal data classification, and feature extraction from sequences. Their ability to modelcomplex temporal dependencies and nonlinear relationships in data makes them ideal for tasks such as noisefiltering and multimodal signal processing. This paper proposes a method for feature extraction from digitalsignals based on a modified transformer architecture that incorporates a nonlinear layer after the selfinspectionmodule. This approach improved the ability of the model to detect complex and nonlinear dependenciesin the data, which is particularly important when dealing with biomedical and signals obtained fromindustrial systems. A description of the architecture and the experiments performed are presented, demonstratingthe high performance of the model in solving signal classification, prediction and filtering problems.It is expected that the model can be applied to a wide range of applications including disease and faultdiagnosis, signal parameter prediction and system modelling.