Unsupervised Anomaly Detection of Reciprocating Compressors using Auto-Encoder

This paper presents an approach for diagnosing faults in reciprocating compressors used in commercial refrigerators. With the increasing importance of fault detection in electrical drives, machine learning techniques, particularly unsupervised learning, have gained popularity due to the difficulty in obtaining faulty data. This study employs an auto-encoder (AE) based anomaly detection model trained exclusively on normal data. The AE compresses the input data into a latent space and reconstructs it, enabling the detection of anomalies when the reconstruction error exceeds a certain threshold. Key input features are selected and extracted without the need for additional sensors, utilizing harmonic information of the compressor’s torque derived through a heterodyning method. The effectiveness of the proposed model is verified on a compressor system using an STM32G431RBT6 digital signal processor (DSP). Experimental results demonstrate 100% accuracy in identifying various mechanical faults, including demagnetized compressors and those operating without oil or refrigerant.