Benchmarking Deep Neural Network Architectures for Machining Tool Anomaly Detection

With the democratization of cyber-physical systems, edge computing, and large-scale data infrastructure, the volume of operational data available is continuously increasing. One of the significant challenges in current industrial research is defining a robust and scalable approach for machine health monitoring and anomaly detection. The methods that exist for such purposes rely extensively on feature engineering and are strongly dependent on the expertise of the operator, hence limiting their generalization. Deep learning techniques, on the other hand, are known to automate feature engineering and allow complex abstractions to be learned, making them particularly suitable for machine health monitoring. This paper presents a benchmarking of deep neural network architectures for the identification of machining tool anomalies on a lathe machine. The features are generated using indirect metrics such as sensor data and process variables from the machine controller, but without direct metrics like surface roughness or finished part quality. The ability of different architectures to identify incipient anomalies in tool quality is compared. A detailed treatment of various subsets of features is provided along with their relative importance to identify the minimum required parameters for accurately identifying the tool anomaly for each architecture considered. Finally, a recommendation is provided based on the results obtained on the type of architecture that is appropriate for the identification of machining tool anomalies. It is expected that the embeddings of the training data learned by the chosen network can be used for other learning tasks, such as transfer learning to another machine or anomaly type. The methodology described in this paper lends itself well to continuous monitoring through the use of scalable robust models and appropriate units of analysis for each model.