Compound Fault Detection of Rotating Machinery in Unobserved Conditions Using Missing Data

In recent years, there has been a rise in the popularity of using data-driven artificial intelligence models for detecting faults in rotating machinery. The challenge lies in creating a model that can be used even when sensor data is not available and the operating conditions differ from those observed during development. This article addresses the issue of potential failures in gear, bearing, and shaft components and suggests two strategies - adjusting entry and cost functions - to address these challenges in developing a one-dimensional convolutional neural network model. These strategies enable the model to extract features from the input signal with minimal dependency on operating conditions. By analyzing the 2009 PHM (Prognostics and Health Management Society) challenge competition dataset, the model achieved its highest accuracy by using the frequency spectrum of velocity and acceleration from vibrational signals. The model’s average accuracy for signals recorded by any arbitrary sensor is 79.6%, even if some operating speeds were not observed during training. Incorporating a suggested penalty function based on p-value into the cost function increased accuracy by up to 13.6%. Consequently, implementing the proposed strategies in similar cases is highly recommended, as demonstrated by successful application in two industrial cases.