Preventing Mechanical Failures: Industrial Audio-Based Anomaly Detection Systems

Problem & Motivation

Factory technicians often experience mechanical failures in their maintenance workflows. These failures can result in long and costly downtimes as well as tragic accidents. A maintenance workflow consists of a predefined, repeatable outline that entails systematic procedures for managing available resources, people, and technology for the work that needs to be accomplished in the factory¹. Due to these failures large facilities lose an average of 27 hours a month to equipment failures, at a cost of $532,000 for each hour of unplanned downtime². Additionally, these industrial mechanical failures also cause tragic on-the-job accidents and injuries. According to the U.S. Occupational Safety and Health Administration (OSHA), heavy equipment and heavy machinery accidents are among the leading causes of fatal and serious accidents that take place on construction sites³. Therefore, the question we want to answer is: how can we leverage machine-generated audio data to help prevent mechanical failures to minimize workflow disruptions and enhance workplace safety?

Data Source & Data Science Approach

We utilize the sound dataset for malfunctioning industrial machine investigation and inspection (MIMII)⁴. It contains the sounds generated from four types of industrial machines, i.e. valves, pumps, fans, and slide rails. Each type of machine includes 4 individual product models, and the data for each model contains normal sounds and anomalous sounds which are all 10-seconds. We implement the YOLOv8 (You Only Look Once) model which is a state-of-the-art (SOTA) computer vision model which performs remarkably well for image classification tasks. Our approach consists of taking the labeled audio files between normal and abnormal sounds to generate images corresponding to the audio file’s mel spectrogram representations. We use these images to train four different types of models for fans, pumps, guard rails, and valves.

Evaluation

The YOLOv8 model is highly performant and efficient. The performance metrics on F1 scores, precision, recall, and accuracy are above 99.00% on each of the test sets for all industrial machine types. The inference times for each prediction take less than 0.50 seconds across all types of audio files and devices. Other deep learning approaches such as Convolutional Neural Networks and DenseNet for image classification did not achieve performance metrics above 90%. Our YOLOv8 (You Only Look Once) model is a state-of-the-art model for image classification and object detection which was initially pre-trained on millions of images from the ImageNet dataset. It processes the mel spectrogram generated from our 10-second audio segments and classifies them based on learned patterns and features. By leveraging YOLOv8, our model can accurately identify and classify different types of audio events, providing reliable real-time analysis and detection capabilities.

Key Learnings & Impact

Using the MIMII dataset we have demonstrated that we can provide a successful machine learning solution for factory maintenance technicians to minimize costly machinery downtimes while increasing workplace safety. According to the U.S. Occupational Safety and Health Administration (OSHA), heavy equipment and heavy machinery accidents are among the leading causes of fatal and serious accidents that take place on construction sites. In the industrial sector, the world’s largest manufacturers lose approximately $1 trillion per year due to machine failures according to the International Society of Automation⁵. Given the results from our work, we believe that we can help minimize these failures and accidents by providing this near real-time service to factory maintenance technicians.

Acknowledgements

We would like to thank our instructors, Fred Nugen and Danielle Cummings, for their valuable feedback. In addition, we are grateful to Jason Rudianto for ensuring our work adheres to the ethical standards.

References

1. Watsoncreative. (2024, July 1). Understand maintenance management & its workflow. NEXGEN. https://www.nexgenam.com/blog/what-is-maintenance-management-workflow-and-why-you-need-to-follow-it/

2. McCallion, R. (2021, July 28). What’s the true cost of machine downtime? MarketScale. https://marketscale.com/industries/engineering-and-construction/whats-the-true-cost-of-machine-downtime/ 

3. Workers’ compensation costs. Injury Facts. (2023, December 1). https://injuryfacts.nsc.org/work/costs/workers-compensation-costs/

4. H. Purohit, R. Tanabe, K. Ichige, T. Endo, Y. Nikaido, K. Suefusa, and Y. Kawaguchi, “Mimii dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,” 2019.

5. Aliano, S. (2021, September 21). World’s largest manufacturers lose $1 trillion/year to machine failure. World’s Largest Manufacturers Lose $1 Trillion/Year to Machine Failure. https://blog.isa.org/worlds-largest-manufacturers-lose-1-trillion/year-to-machine-failure