Polyn Debuts Power-saving AI Chip for Vibration-monitoring Sensors

April 24, 2023 by Jake Hertz

A new neuromorphic chip hopes to enable low-power predictive maintenance applications.

While predictive maintenance has become a critical technology for reducing downtime of vital production equipment, it’s not without its challenges and costs. Recently, Polyn Technology announced VibroSense, a new neuromorphic IC that aims to address these issues.

Last week, All About Circuits Director of Engineering, Dale Wison was able to meet with Polyn at their booth at Hannover Messe 2023, a major trade show for the industrial control and automation industry. Polyn used the event as its first public unveiling of VibroSense.


This YouTube video from Polyn explains how the VibroSense neuromorphic front-end chip works.


In this article, we will explore predictive maintenance and the challenges associated with it. We’ll also discuss Polyn’s new solution and share perspectives from Dale’s interview at Hannover Messe with Eugene Zetserov, VP of Marketing and Business Development at Polyn.


Predictive Maintenance

Predictive maintenance is a proactive approach to equipment maintenance that utilizes sensors and data analysis to predict and prevent equipment failures before they occur. By monitoring and analyzing data from various sensors, such as vibration sensors and acoustic sensors, predictive maintenance can detect changes in equipment behavior and alert maintenance teams to potential issues before they cause significant problems.


Predictive maintenance is a new paradigm in equipment maintenance.

Predictive maintenance is a new paradigm in equipment maintenance. Image used courtesy of Tibco


To implement predictive maintenance, sensors are installed on the equipment, vehicles, or other assets being monitored. In most cases, predictive maintenance is a cloud-based service, meaning that sensors continuously collect data and transmit it to a local server or cloud for storage and analysis. The data collected by these sensors is then analyzed using machine learning (ML) algorithms to identify patterns, anomalies, and potential equipment failures.

Vibration-based condition monitoring is one of the basic types of predictive maintenance. Here, vibration sensors are used as the main means to detect machine failure. Within this, different types of vibrations, such as displacement, velocity, and acceleration, can be measured using various measuring technologies, including piezoelectric sensors, microelectromechanical sensors, and others.


Challenges with Predictive Maintenance 

While predictive maintenance is an extremely powerful tool, the implementation of predictive maintenance can be costly. A major challenge with predictive maintenance is that it often relies on cloud computing. Because of this reliance on the cloud for ML analysis, it is required for all data produced by onboard sensors to be transmitted to the cloud for storage and processing.


Predictive maintenance often relies on cloud computation.

Predictive maintenance often relies on cloud computation. Image used courtesy of SenseGrow


Naturally, transmitting the large amount of data generated by the system sensors takes a significant toll on the system. Between the power expenditure and latency associated with data transmission to and from the cloud, predictive maintenance can ultimately result in increased operational expenses (OPEX). As a result, the cost of implementing predictive maintenance can make it prohibitive for some companies.


VibroSense Enables Always-on Sensors

Polyn positions VibroSense as an IC for enabling low-power vibration-based predictive maintenance solutions. VibroSense is a Neuromorphic Analog Signal Processing (NASP) device that relies on a Neuromorphic Front End (NFE) to offer always-on sensor-level solutions.

At a high level, VibroSense works to perform signal processing on the edge, whereby the chip can detect unique patterns from a vibration sensor's raw signal output in order to determine areas where there is a high likelihood of useful information.

With the embedded neuromorphic computing engines in VibroSense, the chip takes the vibration sensor’s output as its input, classifies the data looking for signals that might be relevant to predictive maintenance, and passes these along to the cloud.


VibroSense decreases the amount of data transmitted to the cloud.

VibroSense decreases the amount of data transmitted to the cloud. Image used courtesy of Polyn


Instead of a traditional always-on predictive maintenance approach, where all data is sent from the sensor to the cloud, a VibroSense solution only transmits relevant data to the cloud. In this way, VibroSense significantly limits the power consumption and bandwidth requirements for a predictive maintenance system.

According to Polyn, VibroSense can help reduce the amount of data required for transmission to the cloud by 1,000×. The chip itself only consumes 100 µW of power while operating as an always-on solution, enabling exceptionally low power systems overall.


Getting Deeper About NASP Technology

The Neuromorphic Analog Signal Processing (NASP) technology can be used on many types of 1-dimensional data, typically time series. Zetserov believes that their “thin-edge” analog AI processing is “changing the paradigm” for these types of applications. At Hannover Messe, Polyn was demonstrating VibroSense for data pre-processing of vibrations from rotating machinery and motors.

The analog AI system uses Ohm’s Law (V = I × R) to perform the multiplication function of current inputs by weighted resistive elements. The output from the parallel set of analog multiplications is summed to perform the accumulation. The result is a power efficient analog implementation of the familiar multiply-accumulate function that forms the basis of DSP and GPU processing.

All of the ML training is done in the digital domain using simulation tools. The simulation tools are tuned to mimic the effective quantization of the analog system based on process and individual device variation.

Their software then compiles the digital model to a metal mask layout for implementation in silicon. While he said that any analog CMOS process can be used, they are currently implementing the product in a Global Foundries 90 nm process and targeting 22 nm for future iterations.

According to Zetserov, all rotating machinery provides a similar signature, so they can apply a single implementation of the VibroSense to different motors. They can change the implementation of the neural net through a relatively low cost metal mask change to target other data pre-processing applications.

At Polyn’s booth at Hannover Messe 2023, Eugene Zetserov explains the VibroSense output data map.

At Polyn’s booth at Hannover Messe 2023, Eugene Zetserov explains the VibroSense output data map.


Zetserov was careful to point out this is a preprocessor to massively reduce the data and power. The high-bandwidth vibration sensor data is passed through the VibroSense to perform the AI inference. It does not perform classification.

The analog output is then passed through a digital-to-analog converter. The low-bandwidth digital output is then passed upstream through a digital serial peripheral interface.

A separate system must be used to classify the results as good or bad based upon testing of an individual rotating system. Zetserov went on to explain that each machine may have a different digital signature that is received from the VibroSense.


Still Early Days But a Bright Future Expected

The startup is still early in their business development. According to Zetserov, a portion of their current work is focused on educating three different segments of the industry about their technology:

  • Discrete sensor makers
  • Sensor node makers
  • System integrators

Polyn believes they have a superior solution, but that it will take some time to convince the industry that the future is analog, at least for these types of applications.