Paul Dvorak / Editor
A condition-monitoring firm from Israel has devised a system that listens for ultrasonic noise from rotating machinery, such as that in a wind turbine nacelle. It processes the signals using self-teaching algorithms and notifies maintenance when something needs attention. Amnon Shenfeld, CEO of 3DSignals, says the capability is possible, thanks to advances in signal processing, machine learning, and artificial intelligence (AI).
Deep learning is a recent field in AI. “There is a famous family of algorithms in computer science that deals with a teaching algorithm that recognizes phenomena. It has been used most recently around computer vision to recognize faces better than people can,” says Shenfeld. The system needs initial data sets for training, but once you teach the algorithms a few general characteristics or features, they have the capability to recognize other conditions without further training.
His team decided to take the family of algorithms and apply the same level of understanding to the acoustic realm or soundscape because there are specific and well-known behaviors in the ultrasonic domain for bearings.
Video involves moving pixels while sound involves moving frequencies in time. “So mathematically speaking, you can apply similar algorithms to a soundscape of an environment,” says Shenfeld. “An application could be hearing a motor, gearbox, or generator and decipher its rotational speed, and whether or not is it under stress. These are the applications that 3DSignals deals with today.”
Where used
He says his systems are used in hydro-electric facilities where in one case, the system recognized a valve failure that vibration and temperature sensors did not. In hydro plants, results were conclusive. “Acoustics picked up events earlier than vibration sensors, and there are many sound sources in a hydro plant,” says Shenfeld.
Ultrasonics can detect flaws earlier than vibration sensors because of their sensitivity. Vibration sensors must be close to the area of interest. “Most conventional vibration sensors are not sensitive enough to pick up flaws, but ultrasonic signals are the first signs of impending failures. So by the time a vibration sensor picks up a problem it may be too late. In some cases our system was the first and only one to report a failure,” he says.
The existing sensor would have to mount about 10 ft from the equipment of interest, but newer sensors, available later in the year, could be positioned up to 30 ft. from equipment. “Sensors are our designs, our integration. Because the system can learn to detect rpm, an encoder for that function is not necessary. The rpm detection is patented.”
Learning
Shenfeld says an algorithm can detect rotating systems out of the box, by their physical parameters. “If we know the bearing size, its acceptable speed, and its number of elements, the system can recognize its acoustic signature. We know right away, how far from the ideal the signature is and, therefore, how healthy the system is,” he says.

Easily read analyses of machine events can be sent to smartphones and tablets for prompt attention by operations and maintenance personnel.
Consider a system that was not previously monitored, a gearbox or wind turbine of a particular model from a specific manufacturer. “We first have reports of anomalies where the acoustic pattern out of the equipment differs from the expected acoustic signature. The system raises a flag with an alert initiated by some sort of acoustic anomaly. Then we alert a maintenance expert or engineer and let them listen to the gearbox before they leave their office.” Shenfeld says it is possible to collect feedback from experienced maintenance people as to whether the anomaly sounds critical or not, and what type of failure it might indicate.
An alert would also say – with a particular percent of confidence – that the problem is related to, for example, a lack of lubricant or misaligned shafts.
Human input trains the algorithm so next time the same signature happens, or any anomaly, it is compared by the deep-learning algorithm. Wind techs have mentioned that even at ground level, they can tell by the sound the turbine is making, that something is amiss in the nacelle.
“Lucky for us, mechanical engineers get intimate with their machines and over time they learn to recognize failure modes by the way things sound when not running right. In many cases, detecting and recording an anomaly to play back for a maintenance person takes just minutes. The expert decides what happens next.
“The learning process depends on getting enough acoustic signature failures from different types of equipment. In most cases, an anomaly will tell an expert how soon they must get to the machine. “We start giving value on day one, and start getting insight to the types of failures, if we have enough previous acoustic data.”
How many sensors for a turbine?
Shenfeld says two or three sensors would be sufficient for a turbine nacelle − one sensor with the gearbox and second near the generator. “I am confident that after one emergency stop, the signature coming out of the bearings would indicate a type of damage. So a good test would be to collect a sound sample at the start.”
Shenfeld thinks the timing is right for the introduction of this technology because experienced people are retiring and taking their experience and the close relationship with their machine with them. Although the system has not yet been applied to wind turbines, all applications so far have been with rotating equipment. So he is looking in the U.S. for wind farms that want to benefit from this technology.
Filed Under: News, O&M