It seems like new applications for wireless sensors are appearing nearly every day. Two of the biggest growth areas are wake-word-driven systems like Alexa or Google and remote vibration monitoring of industrial machinery. Vibration monitoring provides useful information about the machinery, either in real time — condition monitoring; or to predict defects 60 or 90 days in advance — predictive maintenance.
Wireless always-on listening, whether wake-word-driven systems or vibration monitoring, is an exciting use of technology. However, as with any innovation, it is accompanied by new challenges. Battery life is major: If the factory maintenance staff has to change batteries too frequently, there will be a tendency to let some sensors drop out while they attend to more important tasks. This cancels the usefulness of the system. Similarly, users of wake word systems will also tire of too-frequent battery replacements. So, it’s critical that wireless sensors use as little power as possible.
A second problem is that, especially with industrial sensing, extracting the relevant data points that are buried in a lot of data that’s irrelevant, consumes system resources. With more irrelevant data, you need a larger processor and more — wasted — storage. This has led to the idea of moving a lot of data to the cloud and leveraging its powerful processing capabilities to extract information. Performance is better, with lower latency, but because there’s a tendency to want more and more information, you add more and more sensors. That leads to more data going to the cloud and clogging the pipes.
Preprocessing at the sensor (the edge) so you don’t have to transmit every byte of data to the cloud is a good way to reduce the burden on it. However, that means there’s an Analog to Digital Converter (ADC) and a Digital Signal Processor (DSP) or microcontroller residing at each sensor. The sensor data is collected in the analog domain and the processing happens in the digital domain, which is a waste of energy — you’re moving, digitizing, and processing data that’s irrelevant most of the time.
An Innovative Energy-Saving Architecture
According to Tom Doyle, CEO and founder of Aspinity (Pittsburgh, PA), they have come up with a method for reducing the amount of data, and therefore the power usage, at the edge: embed an analog preprocessor at the sensor. Their RAMP (Reconfigurable Analog Modular Processor) chip analyzes raw, unstructured analog data to determine what is relevant before digitizing and processing it in the digital domain. Essentially, RAMP keeps the system asleep unless it detects relevant data, which it sends to the ADC/DSP.
The key is using a neuromorphic processor to do the analog processing. It has the same classification capabilities you might find in a DSP or in a neural network in the digital world. Since it has a trainable and programmable core, it can be used for many different applications. The chip can be preprogrammed to process speech or other acoustic events like alarms or the vibration data from an accelerometer mounted to a rotating machine. Or a user can train it for a proprietary function by using an available Software Development Kit (SDK).
An example of the savings can be understood by considering wake word functions.
Those devices are always on, always processing, although general voice only occurs about 10 – 20% of the time — and voice commands occur far less frequently than that. So, 80 – 90% of the time, they’re processing and digitizing things like dog barks and other sounds as if they’re going to have the wake word in them, but they’re not. An analog processor listens to 100% of the sound but wakes up the digital processor only when it detects a signal of interest, such as speech.
An always-on analog processor, the RAMP will typically use 10s of microamps, while a typical Microcontroller Unit (MCU) requires 10s of milliamps when it’s running. So, according to Doyle, the analog-first architecture gives you a system power reduction of up to 10 times. On top of that, the amount of data that has to be processed in the cloud is reduced by up to 100 times.
Wake Word Processing
In an analog-first system, sensor interfacing is the initial step, followed by feature extraction. The data is then sent to a neural network that’s all analog. When speech is detected, a wake-up trigger is sent to the digital processor. A lot of the steps are the same as in the digital world, but they’re being done earlier in the chain and more efficiently. The analog processor doesn’t do everything, but it can be programmed to detect whether an alarm went off or someone broke a window or if speech was really present.
According to Doyle, “Really, Aspinity stands by itself in working with a wake word engine. We’re very accurate in detecting voice. We gate the entire system based on voice, but we also capture pre-roll, so the wake word accuracy is very high.” The wake word engines need to know up to 500 ms of sound before they can make a decision — a process called pre-roll. The RAMP chip compresses the data, so while capturing all of the pre-roll data and overwriting it, it’s captured at a fraction of the real-time size, but it can be stitched together and rebuilt to retain the wake word engine’s accuracy.
Vibration Monitoring for Predictive Maintenance
Vibration, or even acoustic, monitoring of moving machine parts can be used to infer a machine’s health. While this is an excellent method for monitoring machinery, it presents a lot of challenges. First, there is a massive amount of sensor data that has to be collected and analyzed. Optimally, the more sensors there are, the better the analysis, but the problem is compounded because the vibrational spectrum contains thousands of data points.
Traditionally, thousands of points are gathered and connected by cable to a gateway, then to the factory network where it is pushed to a maintenance worker or the cloud for analysis and decision-making. This is done by using a Fast Fourier Transform (FFT) to examine spectral peaks for frequency-amplitude pairs. Once again, if the edge processing is digital, it must be constantly digitizing these thousands of data points — and that’s a waste of resources. Analog processing can extract the significant data and only send forward a much-reduced set of data, as needed, based on the spectral points that are of particular interest. The system can be kept asleep and only awakened when there are changes in the amplitudes of the peaks.
Each machine is a little different. With rotating machinery, what typically goes bad first are the bearings, each of which has particular fault frequencies that provide a signature for normal operation. In order to identify the normal and abnormal patterns, the processor should be able to track the fundamental frequency and two to three harmonics.
“For the RAMP, we have an SDK development environment, so you don’t really have to be an analog expert. You can build your algorithm, then download it wirelessly to the microcontroller through a Serial Peripheral Interface (SPI) even if the chip is already deployed at a machine,” said Doyle.
Putting Analog to Work for Power-Efficient Wireless Sensing
As the quest for acquiring information from sensors in everything from voice activated devices to machine vibration data for predictive maintenance expands, there are clear advantages to transmitting data wirelessly. The expanding amount of data, however, places a burden on the system and batteries that transmit and process it. This burden can be reduced, however, by doing preprocessing in edge processors embedded in the sensors themselves. But wireless sensors require batteries to run their transmitters, so adding power-hungry digital processing can overly reduce their operating life. The solution — do preprocessing in the low-power analog domain and only send significant data to the digital circuits.
This article was written by Ed Brown, Editor of Sensor Technology. For more information, visit here .