Predicting tribological behaviour in multi-component wear-resistant alloys is challenging due to their complex microstructures and nonlinear phenomena in the contact zone. The Fe–Mn–C–B–Si–Ni–Cr coating combines hard carbide and boride phases with a ductile austenitic matrix. A custom recurrent neural network was developed to predict its friction behaviour. Tribological tests over a sliding distance of 2000 m on a high-temperature ball-on-disc tribometer (THT 1000) gave a mean coefficient of friction of 0.739 ± 0.105 at ambient temperature and 0.600 ± 0.077 at 300 °C. Because pointwise analysis could not capture the non-stationary nature of the friction signal, the recorded signals, totalling about 390,000 samples, were processed using a moving-window technique to extract statistical descriptors. A custom two-branch recurrent network, TriboRNN, was developed. The recurrent branch comprises a bidirectional gated recurrent unit followed by a long short-term memory layer with additive attention. It processes the extracted descriptors, whereas a dense branch encodes alloy characteristics and tribological test parameters. Averaged over five random initialisations, the model achieved a pooled coefficient of determination of 0.9113 ± 0.0005 and a root-mean-square error of 0.0312 ± 0.0001 on the withheld test runs. It matched the accuracy of a stacked long short-term memory baseline (0.9111 ± 0.0050) but proved substantially more reproducible, with a seed-to-seed standard deviation approximately one order of magnitude lower. The proposed method provides an effective framework for friction prediction under the limited-data conditions typical of tribological research.