This paper is dedicated to validating the scalability and generalization of our previously proposed "machinery failure predictive scheme". Our aim is to have a generic core technology to provide a solution applicable in industry that is low-cost and low-intrusive.
Background: In our previous works, we proposed an unsupervised predictive scheme combining the use of full spectrum of vibration / audio data and data visualization techniques. We then proposed a real time data tracker (RTDT) and we applied our proposal on vibration data of bearings. In this paper, we are applying our predictive scheme on a facility (composite system) rather than a specific mechanical component (singular system). We chose to apply our proposal on the MIMII dataset as it was used in task 2 of the DCASE 2020 challenge for the detection of anomalous sounds given normal data only.
Methodology: We adopted two approaches: (1) the same scheme used in our application on bearing vibration data and (2) with a slightly modified approach where we apply a high pass filter (HPF) on the audio data to reduce the effect of the background noise. To effectively evaluate the accuracy of our scheme in detecting and recognizing anomalous sounds, we are comparing our results to the performance of the baseline system proposed by the organizers of the challenge as well as the results from the 40 participating teams. For the evaluation, we used the same metrics used in the challenge: the area under the receiver operation characteristic (ROC) curve (AUC) and the partial AUC (pAUC).
Results: We obtained satisfactory values of AUC and pAUC compared to the related works. We also outperformed the baseline system in 13 out of 16 machines in terms of AUC and 15 out of 16 machines in terms of pAUC.
Merits: Compared to the current related works, our "machinery failure predictive scheme" is featured by 0-training, and no complex preprocessing or de-noising techniques. Furthermore, our solution based on our scheme is provided as a white box, users will have the following merits: (1) our solution can be in operation right after a normal data set is obtained usually in a few days and (2) our solution can be built into conventional operation and maintenance systems without advanced background in artificial intelligence or data science.