An AI-driven strategy for non-invasive fault detection: techniques, applications, and innovations

Abstract

This paper will provide a non-invasive fault detection solution with Artificial Intelligence (AI) techniques. The system uses already available information to collect visual, audio, and vibration data for diagnostics. Because the early signs of fault are numerous, subtle, complex, and difficult to classify and detect with mathematics and signal processing, the collected diagnostic data will be processed and analyzed for unique patterns and features to be used as input to AI tools. These features will be used to train the AI tools (Alexnet, Googlenet, Hybridnet, Yamnet, LSTM, Fuzzy Logic). The subtle characteristics are learned through training; when completed, the trained AI tools can detect them in real time. Each collected data will be classified as a fault with a soft value between [0-1]. Some faults are better detected by visual images than by vibration or audio, while others are better with audio and/or by vibration because the fault is embedded deep inside a machine. For example, an Instrumental Panel (IP) showing engine rpm, and a speedometer complemented by engine sound and wheel vibrations can reveal non-obvious anomalies that might not be shown solely on the IP. Sound and vibration would be able to provide the early telltale signs of anomalies inside the engine and wheels. These theories are experimented with and validated on an actual vehicle and will show that a non-invasive fault detection solution is a viable solution to early fault detection