Searching for Under-Explored Signals in Apollo Seismic Data by Deep Learning and Template Matching

Apollo lunar seismic data are essential for understanding the Moon’s internal structure and geological history. Despite being collected over five decades ago, the Apollo dataset remains the only available source of lunar seismic data, continuing to provide valuable insights into the interior of the Moon and its seismic activity. Recent advances in artificial intelligence, particularly deep learning techniques, have significantly enhanced planetary seismology by providing novel and powerful methods for analyzing previously under-explored, or even unrecognized seismic signal types. In this study, we apply deep learning for unsupervised clustering of lunar seismograms, revealing a new kind of long-period seismic signal that persisted every lunar night from 1969 to 1976. Through a detailed analysis of its timing, frequency, polarization, and temporal distribution, we concluded that this signal is likely induced by the cyclic heater, rather than being an artifact of voltage changes or other artificial sources. In addition to this newly identified signal, the unsupervised clustering algorithm also revealed a class of step/spike signals in acceleration (ACC-Step/Spike) similar to calibration signals. We built a comprehensive search of these signals using template matching, and then analyzed their features. These signals are particularly prevalent during lunar sunrise, sunset, and noon, and their amplitude range varies with temperature as well. Unlike the calibration signals with linear polarization, these ACC-Step/Spike signals exhibit elliptical polarization. Their incidence angles occasionally show noticeable variation during sunrise and sunset. Their characteristics in terms of azimuth and incidence angles also exhibit significant differences between the vertical and horizontal components. For example, in the horizontal component, the azimuth distribution is relatively uniform, and the incidence angle is nearly vertical. In contrast, in the vertical component, the azimuth distribution is sometimes more stable, and the incidence angle distribution is more uniform. Furthermore, our clustering results uncovered short-period abnormal signals near lunar noon and those caused by instrument malfunctions. Our research introduces a novel method for discovering new types of planetary seismic signals and enhances our understanding of Apollo seismic data. The discovery of long-period signals and the ACC-Step/Spike catalogs provide valuable references for future lunar seismic observations and data interpretation, thereby benefiting the analysis of lunar seismic signals.