The additives concentration attenuation in lubricating oil will weaken lubricity performance and protective ability for friction pairs. Monitoring lubricant additives attenuation plays an important role in normal operation of mechanical equipment. This paper proposes a multi-feature fusion hybrid deep learning model (MFHDL) to detect multiple additive content, thereby achieving the purpose of monitoring additive state. PAO40 was selected as base oil. MoDTC, ZDDP, and L135 were used as additives to prepare lubricating oil samples. The spectral data was obtained and three branch networks of MFHDL were constructed. In multi-scale dilated convolution network, more local features were obtained by superimposing multiple dilated convolution filters, expanding model perception range. In multi-head attention network, attention mechanism was utilized to capture the correlations among spectral data. In Bi-directional LSTM network, effective transmission of information was achieved by combining forward and backward LSTM. The output features of three networks were concatenated to obtain multi-feature fusion matrix. The contents of three additives were predicted. The results showed the predicted R 2 of MoDTC, ZDDP and L135 reached 0.940, 0.991 and 0.985, indicating the accuracy and robustness of MFHDL. The four-ball wear testing was conducted to verify monitoring ability of MFHDL. The results showed the relative error of average friction coefficient between aged oil sample and validation oil sample was less than 5.2%, the relative error of wear diameter was less than 4.5%, indicating MFHDL is suitable for monitoring additive concentration attenuation.