Skip to content

17 May, 2022

Predictive Maintenance with Acoustic Sensors

Engineer use augmented reality software in smart factory production line with automated application . Futuristic machinery in working in concept of Industry 4.0 or 4th industrial revolution.

Article By

Martin Hangaard Hansen, Senior Data Scientist, Flowtale


When expensive mechanical machinery such as manufacturing machines or large engines break down, the costs can be enormous. These cost are due to:

  • Downtime
  • Manual labor
  • New parts and materials
  • Disposal or recycling

Downtime is often the most expensive. Imagine the cost of a 400 meter long container ship, stuck in a port due to an unscheduled problem with the engine. While it is stuck, some tens of thousands of tons of goods are sitting idle on board, while space in the port is also blocked, potentially delaying other ships and land based transports. Such events can lead to cascades of expensive delays.

To minimize the risk of downtime, it pays to do excessive maintenance, inspections and lubrication of equipment. Such cases also include machines for manufacturing. Even though they are not as big as a container ship, they are expensive and require the attention of specialized technicians.

Automation

Is the machine running correctly? Is it always lubricated sufficiently? Some problems may also be prevented if automatic systems are in place to warn and even make adjustments, if a sensor detects a problem. It may be possible to reduce scheduled maintenance, if the automation is so reliable that it can order the maintenance only when it is optimal.

To this end, various sensors are installed on the machines. Temperature sensors are slow to respond to changes and they only detect changes in temperature very locally where they are placed. That provides quite limited information about any potential problems in the operation of the machine. Fluid-level sensors can detect the presence or absence of almost any liquid, but are very sensitive for direct sunlight, dust and temperature, so they are not always applicable and they still provide limited information. Analyses of the oil or exhaust by samples or by sensors may provide information about problems, but by the time the chemical composition of oil or exhaust has changed noticably, a serious problem may have already occurred.

Combinations of multiple sensor types can improve predictive abilities, but this requires significant software and data engineering effort, and results will always depend on the sensors’ ability to pick up information about the relevant phenomena.

Emerging Technology

Modern vibro-acoustic sensors can be attached to the outer metal parts of machines and sensitively detect every tiny movement and vibration. They do this automatically and at a wide frequency range from Hz to Mega-Hertz, which ranges far beyond what the human ear can detect. The sound of anything that slides against the metal or any tiny crack that forms inside the metal will be carried throughout the metal part. Even the characteristics of a liquid or gas flowing though the part may in some cases be detected. A tremendous amount of information is therefore available to acoustic sensors. That also poses a challenge: To separate the important sounds from the unimportant ones, or in other words to separate the signal from the noise.

How do we even know what a problem sounds like? The sounds of a dangerous event such as elevated friction due to insufficient lubrication may be quite rare, because these events are not supposed to happen. To collect data and learn about “warning sounds”, it is beneficial to to experiment and provoke the problems, but in a machine that is in operation, the options to experiment may be quite limited. This poses a serious challenge to validate the automation solution, since real validation can only be proved when obtaining data in the production environment.

A predictive maintenance strategy depends on the ability of automation to detect anomalies and take action.

Every machine also has a different sound. Even identical machines are slightly different, so how can we be sure to detect problems and avoid false alarms in every case? Anything can be automated. Even tuning algorithms to fit the acoustic data from a new machine.

Cases

Great progress is being made on these long standing technical challenges, due to improvements in sensors, in software and in data science.

Flowtale recently proved the concept for a client with industrial machines for manufacturing. We delivered a proof-of-concept, which showed that the sensor could detect, at very high accuracy, whether or not the machine had sufficient lubrication. We did this with a combination of the best sensors on the market, as well as knowledge about sound signal processing and data science techniques including artificial neural networks. The proof-of-concept is currently being further developed into an application that can minimize both downtime and technician time.

We are currently pursuing more cases to leverage this technology for optimizing the worlds machinery. Do not hesitate to reach out to us, if you have a similar challenge.