Workflow Development of AI Based Spectrogram Analysis with Real-Time Out of Distribution Detection

The aim of this paper is to investigate possible workflows for OOD pattern recognition in AI-based spectrogram analysis, applied in industrial manufacturing environment. First, we attempt to identify and articulate the challenges associated with OOD recognition in the context of spectrogram analysis, where the acoustic sources are subtle and often complex signals. These deserve particular attention, since the effectivity of OOD detection algorithms are acceptable in case of significant deviations, however, it is questionable for fine anomalies. In addition, it is also discussed here, how OOD records can affect the accuracy and reliability of AI models in terms of equipment failure identification and process inefficiencies. Last, methodes are proposed for OOD-pattern recognition. The integrability of these methods into existing manufacturing workflows in terms of practicality, adaptability and effectiveness are also investigated.