OutlierNets: AI for On-Device Acoustic Anomaly Detection in Manufacturing
by Dr. Alexander Wong
By: Dr. Alexander Wong
In manufacturing facilities, human operators often can tell if machinery requires maintenance by the sounds it is making. If the pitch of a fan changes or something starts clanging, that’s a pretty clear indicator that something has changed and should be investigated immediately. Manufacturing operators require training and experience to listen for slight changes in sound that can result in an expensive shutdown or a potentially dangerous scenario if not addressed quickly.
Unfortunately, many of the currently available AI solutions require large amounts of prohibitively expensive computational resources to run for such a use case.
To help solve this problem our researchers at DarwinAI and our friends at the University of Waterloo have worked together to create OutlierNets, a collection of compact deep autoencoder network architectures that are compact enough to run on edge devices (like CPUs and microcontrollers) in real time while maintaining or exceeding the accuracy rates provided by much larger solutions containing millions of parameters.
Enter The OutlierNets Solution
With the new OutlierNets AI architectures available to run on edge devices in real time, manufacturing facilities can get all the benefits of immediately and accurately detecting audio anomalies in their equipment at an affordable setup cost, allowing them to detect minor issues before they become major issues. Earlier detection also allows for maintenance to be scheduled in an off-peak time, instead of requiring an immediate shutdown, or, worse, having the machinery break down, leading to expensive repairs in addition to an extended work stoppage.
How We Built OutlierNets
To create OutlierNets, the team started with the MIMII dataset, which contains audio recordings of fans, sliders, valves, pumps, etc. In this case the team focused on fans and sliders. The training set contained only normal sounds while the test set included an equal split of normal and abnormal sounds. Each audio clip was limited to one second in length. This helped provide a more uniform sample for the autoencoder to learn, it proved that OutlierNets only required one second to identify a sound, and significantly reduced lag due to a reduction in the size of the samples.
Once the samples were prepared, the team used Generative Synthesis to automatically explore and identify the best deep neural network architectures tailored specifically for the task of detecting acoustic anomalies on different machinery, thus finding the best trade-offs between size and accuracy based on the operational requirements specified by the team. The results are highly unique, customized deep neural network architectures that are highly accurate but still small enough to run on a variety of microcontrollers.
Phenomenal Results Achieved
Some of our new OutlierNet architectures operate with as few as 686 parameters and are as small as 2.7KB in size (orders of magnitude smaller than many of the available solutions), reducing CPU latency by as much as 21x.
We are extremely excited about our achievements to help manufacturing enterprises with acoustic anomaly detection and we look forward to seeing where our continued research in this space will take us.