Reciprocating mechanical vibration signals are characterised by multi-source impact coupling and varied time-domain intervals. Different sub-impacts contain rich information closely related to the operational states during different strokes of the working cycle. Therefore, accurate and efficient decomposition of different sub-impacts in the time domain is an important aspect of state monitoring and fault diagnosis in reciprocating machinery. Existing time-domain decomposition methods that primarily focus on the concentration of signal energy lack robustness against noise. Furthermore, these methods suffer from issues, such as excessive reliance on prior parameters, low computational efficiency, and signal distortion. In this study, impact time domain decomposition (ITDD), a novel adaptive decomposition method is introduced. ITDD uses the envelope energy gradient (EEG) characteristics of a signal to extract sub-impacts. An impact-adaptive localisation strategy is designed based on the EEG characteristics of the noisy signals. A dimensionless evaluation metric called the impact distortion index is then constructed to optimise the smoothed EEG sequence. Finally, a residual amplitude estimation method for adaptively setting the impact identification thresholds is proposed. The effectiveness and superiority of the ITDD over existing time-domain decomposition algorithms are demonstrated using simulations and actual analyses of reciprocating machinery signals. These results indicate that the ITDD can precisely and efficiently decompose multi-source impact signals, and it significantly enhances decomposition precision and speed compared to existing methods, with a decomposition time < 1 s, thus demonstrating substantial potential for engineering applications.