👩🏾 👩🏼‍💼 🌡️ “Listen to find a breakdown”: audio recordings of failed industrial machines published 👨‍👧‍👧 🦓 🐩

Engineers from Hitachi have produced thousands of audio recordings depicting the operation of industrial equipment in combat conditions. With the help of such a library, machine algorithms will be able to detect abnormal operation of systems and predict potential breakdowns.

We tell what went into the data set and who is still working on similar projects.

^{Photo abi ismail / Unsplash}

Why is it necessary

Doctors use stethoscopes to diagnose certain diseases. But the accuracy with which pathology can be detected in this way is 20–40%. Therefore, today electronic stethoscopes with artificial intelligence systems come to the rescue. For example, such a device is made by specialists from the Northwest Memorial Hospital of Chicago. Smart algorithms help increase the accuracy of diagnosis up to 97%.

A similar approach is gaining momentum in the field of production - machine learning models reveal malfunctions of industrial machines: unnatural noises in the operation of transmissions, pumps, fans or valves. To improve the accuracy and quality of machine learning, Hitachi prepared and made publicly available a set of audio recordings of industrial machines. The work has been done in real factory conditions.

According to the authors, this is the first such data set in the world . It is called MIMII Dataset - Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection.

What's inside

The compilation contains the sounds of valves, pumps, fans and rail guides. All of them are from different manufacturers. The archive contains more than 26 thousand ten-second samples with the sounds of machines operating under normal conditions. Also, there are 6 thousand audio files with recordings of the work of faulty devices and machines that are in an “imperfect state”.

Among these conditions: damage to the guides and the lack of lubrication on them, imbalance during rotation of the blades and their jamming, oil leakage and power surges.

^{Photo Sergei Akulich / Unsplash}

The duration of the audio samples is 10 seconds, all of them are recorded in WAV format with a sampling frequency of 16 kHz. Recording was carried out at once at several production factories. Eight microphones were used, assembled in a circular array and installed at a distance of 10-50 cm from machine tools and equipment (the diagram can be found in whitepaper on page 3 ).

The kit is licensed under CC BY-SA 4.0 . The archives are on the Zenodo site , but their total weight exceeds 150 GB . You can listen to several audio recordings here .

Author plans

For each model of an industrial machine - since they all have their own acoustic characteristics - engineers developed anomaly detectors. Test launches showed that trained intelligent systems better detect failures in fans and rail guides. But the detectors experienced difficulty analyzing the operation of the valves.

Engineers explained this by the fact that the sound of opening and closing the valve is short and rarely occurs. Thus, it is more difficult to identify malfunctions than in the case of static and continuous sounds of other mechanisms. The development of effective algorithms to detect anomalies in the operation of valves, Hitachi specialists will deal with in the future.

Similar projects

In the first subsection, we noted that artificial intelligence systems penetrate the healthcare sector. Therefore, a large number of data sets for training algorithms that diagnose diseases of internal organs also appear in this area. For example, an engineer from Stanford University has publicly published a classification of heartbeat abnormalities.

It is used to develop smart stethoscopes - the data set has been downloaded more than 7 thousand times by experts from India, the USA, Canada, France, Germany and other countries.

^{Photo Marcelo Leal / Unsplash}

Another example is from the world of cars. The Israeli company 3dsignals is developing an intelligent diagnostic system. Thanks to her, car owners will be able to receive timely information about malfunctions. The authors claim that the system is able to predict the time interval during which a breakdown will occur.

The accuracy of such diagnostics, according to the test results, is 98%. Unfortunately, the data set on which the neural network was trained is not disclosed in 3dsignals. The company's solution is also suitable for monitoring large industrial installations. For example, it is already being used at Enel Green Power Corporation to evaluate the condition of energy turbines.

Additional reading from our blog: