Reliable condition assessment of complex technical systems based on vibroacoustic signals is challenging due to strong non-stationarity, high noise levels, and overlapping spectral components. This paper proposes an integrated signal processing and machine learning framework for automated technical condition prediction. The approach is based on adaptive time-frequency decomposition using the Hilbert-Huang methodology and Variational Mode Decomposition (VMD), which enables the extraction of physically meaningful intrinsic mode functions while suppressing noise-induced distortions. The decomposed components are subsequently used as informative inputs to neural network classifiers. Special attention is given to a hybrid architecture combining VMD-based preprocessing with a Kolmogorov-Arnold Network and a Long Short-Term Memory model (VMD-KAN-LSTM), designed to capture both nonlinear feature interactions and temporal dependencies. The proposed methodology is evaluated using monitoring data obtained from an SKA60/200 reciprocating compressor under normal operation and three characteristic fault conditions. Comparative experiments involving LSTM, deep autoencoder, KAN, and their VMD-enhanced counterparts are conducted using confusion-matrix-based performance analysis. The results demonstrate that VMD significantly improves class separability in noisy operating conditions, while the proposed VMD-KAN-LSTM architecture achieves the highest diagnostic accuracy and robustness among the considered models. The findings confirm the effectiveness of decomposition-informed learning for condition monitoring of complex technical systems.