Anomaly detection of 2.4l diesel engine using one-class svm with variational autoencoder

Despite intensive research, it is still at its early stage to prevent breakdown of construction machine, leading to a need to develop an autonomous and robust solution that minimizes equipment downtime and ensures the rigidity of equipment through predictive diagnostics. In particular, engine failure is critical because it causes the entire system to stop, highlighting the importance of determining and predicting the symptoms before failure occurs. However, at present, specific indicators based on domain knowledge should be set in order to judge a failure. This paper proposes an anomaly detection model for a 2.4L diesel engine, and verifies the model for two main faults. The proposed method based on deep learning extracts 130 feature parameters with autoencoder and distinguishes between normal and abnormal states by one-class support vector machine (OCSVM). An autoencoder can automatically extract useful features from multiple sensors on an excavator engine, and a variational autoencoder (VAE) extracts latent values from its input variables to generate new information. In this paper, a VAE is applied to extracting feature vector from the vibration signal for robust modeling, and OCSVM detects abnormal state and distinguishes between the two different faults and unknown factors. The experimental results show the accuracy of about 73%, and the false alarm related to the reliability of the model can be minimized to about 17%. Finally, to resolve the issues of reliability and interpretability of the model based on the deep learning, the Local Interpretable Model-agnostic Explanations (LIME) analysis is applied to listing the sensor data that affect the determination of the abnormal state. We intend to improve the accuracy of the model by adding expert knowledge to the data-driven model, because experts can easily make professional judgments about abnormal conditions and build a model with a continuously increasing sets of known data about faults and symptoms.