🚦 Air Compressors Anomaly Detector

📑 Project Description

A project from previous work experiences.

Detect air compressors' potential malfunction via DBSCAN.

☁️ Clustering-Based Detector

📈 Available Features

Each air compressor only has minutely runtime series as the only feature, making it hard for supervised learning. Thus, clustering is chosen as the method.

🛠️ Selected Clustering Algorithm: DBSCAN

https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html

DBSCAN outputs a cluster label for each data point. Labels include 3 types:

1. -1 represents outliers.

2. 0 means majority cluster.

3. 1, 2, ..., k all refer to minority clusters.

Minority clusters imply abnormality, as do outliers which are even more severe.

⛓️ Detector Pipeline

1. Read each air compressor's hourly runtime series from database.
2. Calculate last 7 days' average hourly runtime series.
3. Do DBSCAN on the differential of hourly runtime - last 7 days' average hourly runtime.
4. Identify the hours when minority clusters and outliers happen. Among them, only consider positive differentials which imply runtime spikes.
5. From the hours in step 4, find out these hours' respective runtime level.
6. If runtime level from step 5 hits threshold, mark these hours as abnormalities, as runtime spikes combined with above-threshold runtime level = red flag 🚩.
7. Store these abnormal hours in database as potential malfunction alerts.