Purpose <p>This study proposed a methodology for detecting ball screw preload loss in ball screw feed drive systems of machine tools by using vibration signals. The methodology aimed to classify different severities of preload loss with minimal uncertainty, thereby reducing the possibility of diagnostic errors. This study focused on improving detection effectiveness by improving signal processing, noise reduction, and feature extraction.</p> Methods <p>The methodology involved directly acquiring vibration amplitudes and phases of a rotating ball screw at the ball pass frequency and its multipliers (n × BPF) in the time domain by using a triaxial accelerometer mounted on the ball screw nut; numerical experiments were conducted using Vold-Kalman filtering order tracking (VKFOT). Holospectra, representing the trajectory of the rotating ball screw at the ball pass frequency and its multipliers (n × BPF), were constructed and then quantified into metrics or features, which indicate the severity of ball screw preload loss, by using a self-organizing map (SOM).</p> Results <p>This study investigated the effectiveness of VKFOT and the features extracted from the vibration holospectra for clustering ball screw preload loss severity. Moreover, this study examined the influence of noise on ball screw preload detection and the associated noise reduction methods.</p> Conclusions <p>This study presented simulations and experimental validations of the proposed methodology, demonstrating significant improvements in sensitivity and reliability.</p>