Recent advances in sensing technology, condition-based maintenance (CBM) is becoming popular for machinery equipment. CBM uses sensors to detect anomalies and performs maintenance when anomalies are detected. CBM is expected to reduce maintenance costs and time because of detect failures before they occur.
If the changes have tendency under normal operating, such as current and temperature, the anomaly can be detected through comparing with a present threshold value. However, when the observed object is vibration of rotating machinery such as motors, it is difficult to detect anomaly using only a simple threshold value because it is necessary to find anomaly features in the vibration waveform.
Therefore, we developed an anomaly detection method for vibration waveforms obtained from an acceleration sensor attached to a rotating machine. Then, we constructed a condition monitoring system using an edge computing in which these computational processes are performed by a single board computer (SBC) called Raspberry Pi installed on site, and the judgment results are sent to the cloud.
The anomaly detection method uses waveform processing to extract features, which are then input into an AI model for evaluation. Waveform processing as a pre-processing step for AI makes it possible to focus on small changes that are hidden by larger fluctuations. The features obtained by waveform processing are then classified into pre-assumed abnormality levels using three different AI models.
Specifically, an envelope is calculated from the vibration waveform obtained from the acceleration sensor, and the Welch's power spectrum density of the envelope used as the features.
Then, anomaly detection was attempted by combining methods such as Auto Encoder (AE), Support Vector Machine (SVM), and Mahalanobis Distance (MD) from the obtained features.
Condition monitoring systems can also be applied in the field of civil engineering where rotating machinery is used.