Der Blätterkatalog benötigt Javascript.
Bitte aktivieren Sie Javascript in Ihren Browser-Einstellungen.
The Blätterkatalog requires Javascript.
Please activate Javascript in your browser settings.
16 Elektronik International 2020 Microcontroller AI and Machine Learning rely heavily on data this is generally accepted It is also generally believed that the processing power needed to handle that data is huge and here opinions start to diverge Yes it is indisputable that creating an AI model requires a large amount of data and an even larger amount of data processing but once the model is created it can be deployed in endpoints with much more modest resources This is the essential difference between training and inferencing The former requires large amounts of processing power often in the form of GPUs or HPCs running in racks found in data centers but the latter is effectively code that can be deployed on almost any platform It is also important to understand how AI and Machine Learning differ ML is a subset of AI and is most often deployed as a trained model that can infer from data provided what action to take In most cases and particularly in AIoT endpoints there will be little or no reinforcement learning at this point the inferencing will not get more or less accurate over time unless positive feedback is applied which would subsequently change the model This would typically require further training which is not the scenario most associated with putting ML into small lowpower endpoints such as smart sensors and smart actuators By making this declaration it becomes clearer how ML can be deployed in low power microcontrollers in small endpoints and the benefits this can bring It also highlights why those companies keen to deploy ML in endpoints will be more reliant on embedded technology providers to help them achieve that This is a real opportunity for growth in the embedded sector presented by creating an environment that supports ML in endpoints Real-World examples of ML in endpoints Initial examples of how ML is making endpoints smarter and having a measurable impact on profitability and safety involve using AI to move from periodic maintenance to predictive maintenance For example EDGXL is an ultralowpower inferencing technology developed by INFXL for deployment in resourcelimited endpoints The company took runtofailure data from 21 sensors in a turbo propeller aircraft to train its EDGXL module Based on a dataset of some 34 000 histories gathered using sensors including temperature pressure RPM fuel flow and fuelair ratio it was able to train the AI to predict a failure with 95% accuracy Despite this the model runs on a Cortex-M0 MCU requires just 17kbyte of memory and consumes μWs of power AquaSeca has developed an Arm Cortex-M4 based vibration sensor that attaches to a water pipe to form a simple and lowcost method for detecting the relative flow of water Any changes to the vibration signature would indicate cause for alarm such as the faster flow caused by a leak or the impeded flow caused by a blockage The telltale vibrations that result from these kinds of faults can be detected by the sensor and the causes inferred using ML The AIoT system can then alert the owner before the fault escalates with a realworld impact of detecting a boiler leak in a housing complex of over 5 000 units and preventing legionella in stagnating water in hospitals and care homes Today the AquaSeca sensor sends its data to the cloud but they plan to put all of the ML inference inside the sensor itself These and other examples like them illustrate how few resources ML really needs in an endpoint to deliver highly accurate predictive maintenance but it is also being used to predict other things For example a smart sensor equipped with ML inferencing is being used to detect when a vulnerable person may have fallen or has been inactive for too long What constitutes a fall or long periods of inactivity is of course subjective but through AI that subjectivity can be largely removed based on trend data used to train a model to recognise the pertinent signs Acommon theme with these examples is how movement in one form or another is being used to create datasets Rather than absolute motion the commonality here is relative movement and often something as simple as vibration This is a form of movement that modern sensors are extremely adept at detecting and one of the most versatile forms of data gathering vibrations convey more than just movement because the signatures can tell us much more Deploying ML in endpoint devices Conventional embedded software would normally be hand coded in a highlevel language such as Cwith a clear idea of the target architecture and its features For example while many MCUs are based on an Arm Cortex-Mprocessor different manufacturers implement vastly different peripherals memory structures and other system resources These variations would need to be understood by the developer and software development tools Conversely an ML inferencing model starts life in a framework created specifically to support AI Porting this model to an embedded platform must take into account the limited resources available which is becoming less of a challenge thanks to the work being done to develop a strong tool flow that supports ML in MCUs Figure 3 Inferencing in security cameras will enable faster actions by detecting the presence of an unknown person identifying a package and inferring whether the person is delivering the package or potentially stealing it photo Arm