Machine learning (ML) for sensors and signal data is becoming easier than ever. Hardware is becoming smaller and sensors are getting cheaper, making IoT devices widely available for a variety of applications ranging from predictive maintenance to user behavior monitoring. Vishal Goyal, senior technical marketing manager – Analog and MEMS Group, RF, Sensors and Analog Custom Products, Asean-ANZ and India of STMicroelectronics in an interaction with Baishakhi Dutta of the Electronicsforu.com Network shares how they have harnessed the power of ML to enhance modern-day sensors. Excerpt follows…
Q) What is the latest offering that STMicroelectronics has for the industry?
As our latest offering, we have integrated ML technology into our advanced inertial sensors to improve activity-tracking performance and battery life in mobiles and wearables.
ST has launched the LSM6DSOX iNEMO sensor which contains an ML core to classify motion data based on known patterns. Relieving this first stage of activity tracking from the main processor saves energy and accelerates motion-based apps such as fitness logging, wellness monitoring, personal navigation, and fall detection. The sensor can be easily integrated with the Android and iOS platform.
Q) How much current does the sensor consume for improving the battery life?
The LSM6DSOX contains a 3D MEMS accelerometer and 3D MEMS gyroscope and tracks complex movements using the ML core at low current consumption of just 0.55mA to minimize load on the battery.
Q) Is the new sensor ready to hit the market?
The LSM6DSOX is in full production and available now, priced from $2.50 for orders of 1000 pieces.
Q) How will machine learning help lower the power consumption of a device?
Imagine a scenario where you have billions of sensor nodes, and you are taking all sensor data to the cloud, and then you are doing the analytics and AI, ML on the cloud. Since sensors are generating a lot of data, that will choke the complete network. That will, in turn, lead to an increase in the cost of data, an increase in the cost of operation, and an increase in power consumption.
Through edge computing, instead of taking the decision at the cloud, the decision is taken at the node itself. So, with this, the amount of data which is going to the cloud reduces and thus the overall network bandwidth reduces, leading to a decrease in the overall cost of operation.
Usually, machine learning takes place inside the microcontroller. So, to reduce power consumption, we have put ML inside the sensor itself. The whole of the power consumed by the microcontroller has become zero, and the intelligence has been transferred into the sensor. So, just by changing the architecture, the low-level power consumption is achieved. And most importantly, this can be considered as a solution for data flooding.
Q) Any space challenge that is likely to take place with the implementation of ML?
There is no space challenge. In fact, it will reduce the space as the sensor size will continue to remain the same.
Q) Are the sensors calibrated and trained in the factory?
Yes, the sensors are calibrated in terms of their performance. This is done at the factory, while the algorithm is decided by the designer.
Q) In the process of implementing AI or ML, how crucial is open source for you right now?
Not too much open source thing is being used, but when it comes to edge computing, we are putting machine learning core inside the microcontroller. Generally, while doing a complex algorithm, people do use a lot of open-source models.
The open-source model is accurate, and it is a great thing to use. But I suggest not to blindly believe it because as that can be an individual perspective. So you should always validate whether using a particular tool is suiting your product or not.
Q) What latest design strategies are you following for all your sensors?
If we make a sensor which has extremely low power, but which will require a microcontroller to do a lot of tasks, then the overall system efficiency will go down. So, we are aiming to reduce the overall system power, which means the task of the microcontroller is offloaded to the sensor and reduction in the power in the microcontroller would prove to be far more beneficial.
Consider a sensor is detecting motion. One way is that the sensor is detecting the information and sending it to the processor continuously. It means your processor is on all the time and sending data that consumes more and more power. Thus, system power increases. But if we have some memory inside, then the sensor is storing the data internally the sensor is loading the entire TV and sending that data to the processor only for a fixed duration, maybe once in an hour or once in a minute. So, the overall time of your microcontroller in which it is on is dramatically reduced, which will help to achieve overall better power consumption. An application processor or the microcontroller interfaced with a sensor.
Q) What is unique about your production process?
We have two different manufacturing processes: Back-end and Front-end. Back-end refers to the outer black covering of a chip and front-end refers to silicon semiconductor present inside the back-end. To ensure no compromise service to our customers, we manufacture each component in two different places. The idea is to provide undisrupted service to our customers even at the times of natural calamity or any unwanted events which are likely to affect our manufacturing base.
Q) How have your sales been in terms of sensors?
For the sensors alone, we are serving more than 100,000 customers worldwide. We have leadership in waterproof and harsh environment pressure sensors. This type of pressure sensor is used to detect altitude and can even work in water that has chlorine or washing powder.
ST also has a sensor one in every three car navigators. At around 2000, ST did the first MEMS testing for inkjet printers. At around 2005, we entered included accelerometers and expanded our portfolio to magnetometer, humidity sensors, piezoelectric sensors, and all.
At around 2012, ST entered into the automotive domain and within five years, became the market leader in automotive and navigator applications.
At around 2014, ST entered into the industrial domain.
Q) What is ST’s current workforce and product portfolio in the MEMS and sensor division?
There are 7,400 people working in this division. Our number of products and software libraries are more than 550.
Q) How has ST performed in the last fiscal year?
Our revenue last year was US$ 9.66 billion.
