An Industrial Internet Platform for Real-Time Fault Diagnosis using Statistical Methods and Neural Networks

摘要

Being able to detect, isolate and diagnose a fault is an important aspect of supervision systems in industrial environments, since enables advance asset management, in particular predictive maintenance, which greatly increases efficiency and productivity. In this work an industrial Internet platform, with real-time fault detection and diagnosis capabilities is designed and implemented in a pilot scale industrial motor. Fault detection and isolation is based on a statistical data-driven method using principal component analysis (PCA) and reconstruction based contribution (RBC). If the detection and isolation analysis indicates that one of the vibration measurements is responsible for the fault, a convolutional neural network (CNN) is used to identify the fault type from the raw vibration signals. The implemented platform was evaluated in its three functionalities: fault detection, fault isolation and fault identification in vibration signals. In the fault detection task, over 99% of detection rate was reached. On the other side, in the fault isolation and fault identification in the raw vibration signals tasks, after generating a series of artificial faults, in all cases the percentages of accuracy obtained is over 90%. In spite the fact that platform is currently implemented in the laboratory, it can be easily scaled to a full-size industrial process and due to its flexibility, can be adapted to work with different machines with minor modifications.