Keywords: Timeseries, Clustering
In refrigeration systems, compressor cycles create a recurring pattern in temperature recordings, causing the temperature to oscillate between high and low ranges. Any deviation from this temperature cycle may indicate potential issues with the system. This article presents an anomaly detection algorithm designed to monitor changes in the temperature cycles and generate alerts when abnormalities occur.
Wavelet decomposition is used to transform the temperature waveform into the frequency domain. This technique incorporates historical information into each sample. A typical period is selected for training the Principal Component Analysis (PCA) model.
PCA compresses and reconstructs the frequency waveform obtained from the wavelet decomposition. Since the PCA model is trained on a normal operation period, it can efficiently compress and reconstruct the waveform with minimal loss. The more abnormal an observation is, the greater the loss incurred during the compression and reconstruction process.
The Mean Squared Error (MSE) quantifies the loss resulting from the PCA compression and reconstruction. Higher MSE values indicate a greater deviation from normal operation.
To determine the alarm threshold, a logistic regression model is trained using previous examples of abnormal refrigeration cycles. This threshold helps to differentiate between normal and abnormal operation and trigger alerts when necessary