 Hello, everyone. This is Ernesto Manuel Cantone, part of the Mems and Sensors Product Marketing team here in US. Today with us, I have Tina Shiwan, Senior Product Marketing Manager at Kekso, here to talk about Kekso AutoML on Steam and Electronics Development Kits. Hi, this is Tina from Kekso. We are the first company to automate end-to-end machine learning for the Edge. Industry-wide, we're seeing a prominent trend of intelligence moving from the cloud to the Edge. This means that the smart decisions are being made right on the device without having to go through the cloud. There are many benefits with doing this, including that it reduces latency, bandwidth, and power usage by eliminating data transfers. It removes dependency on the cloud for higher availability and flexibility, and addresses privacy and security concerns since data never has to leave the device. Data from IDC shows that there will be 41.6 billion connected IoT devices generating almost 80 zettabytes of data by 2025. Gartner says that 75 percent of enterprise-generated data will be created and processed at the Edge by the same time. In a typical sensor node application, the main processing unit collects sensor row data and then runs all the necessary algorithms. This is a traditional design approach, but not necessarily an optimized implementation when it comes to system power consumption. To allow a more efficient working principle, the machine learning core functionality in the sensor ASICs allows running a programmable machine learning classification algorithm directly. A system level displays a key role by allowing the host microcontroller to focus only on high-level processing triggered by the machine learning model implemented inside the sensor. The machine learning core brings unprecedented value to our mem sensors and enable a power and cost-optimized solution design. It's currently available in these products, recognizable by the X at the end of the part number. LSM60SOX, consumer-grade IMU, focused on low-power applications such as wearable asset trackers and IoT nodes. In this part number, there are 256 nodes available to be utilized in eight independent decision tree. LSM60SR, IMU with higher gyroscope full-scale and higher stability over time and temperature, focused on virtual and augmented reality. Navigation systems and high-performance sensor fusion-based applications with 512 nodes available in the machine learning core. Last, ISM320DHCX, industrial-grade IMU, featuring extended temperature range up to 105 Celsius and 10 years of public longevity commitment. Let me now introduce a couple of development kits that can be used to work on machine learning both on the OSTMCU and the Machine Learning Core, both enabled in Kexo AutoML. LSM60SOX can be found in the SensorTile.box, the easiest development kit to get familiar with ST sensors. This supports a sample of all the latest ST sensing technology, and it's the ideal starting point to build a Bluetooth low-energy battery-powered IoT design. Industrial-grade ISM320DHCX is part instead of ST-WIN, SensorTile Wireless Industrial Node, an easy-to-use development kit powered by our STM3214 Plus MCU, that includes a broad combination of industrial sensors embedded Bluetooth low-energy connectivity for sensor streaming and additional connectivity modules like Wi-Fi expansion for cloud enablement. Machine Learning Models for IMU with Machine Learning Core can be designed using ST-Nucleo. UNICO is a graphical user interface tool made available by STIMAC Electronics that allows to capture and label data, build, embed, and run decision trees within the ST sensors. Models running within the sensors can classify a variety of different classes defined by the user and wake up a platform that can run more complex machine learning on the main processing unit. Kixxel has been working closely with ST to support building machine learning models automatically. Currently, both the SensorTile.box and the SC-WIN are supported on Kixxel AutoML. Kixxel AutoML supports a wide range of deep learning and non-deep learning algorithms, including GBM, X-Tree Boost, Random Forest, Logistic Regression, Decision Tree, SVM, CNN, RNN, CRNN, ANN, Local Outlier Factor, and Isolation Forest. We are bringing value to the industrial predictive maintenance, smart home, wearables, automotive, mobile, and IoT markets. In our demo, we will showcase our easy-to-use interface for collecting and analyzing sensor data to build machine learning models. We'll be using the ST-WIN platform to detect two different machine anomalies using multi-class classification. In this video, we will be using Kixxel AutoML to create multi-class machine learning models for anomaly detection. We will begin by creating a multi-class classification project and select ST-WIN platform as the target hardware. First, we navigate to the Data Collection tab. Then we name our environment and select the sensors we want to use and their configurations. We will configure the accelerometer and gyroscope sensors to 6,667 Hertz with an FSR of 2G for the accelerometer and 125 degrees per second for the gyroscope. The Data Collection library will be automatically built and flashed to the device so that we can begin to collect data. We have set the simulator to approximately 1,500 rotations per minute and we'll collect data for all classes at this configuration. We start with collecting 200 seconds of data with the normal rotor. Let's label it regular. We wait for the fault simulator to reach 1,500 RPM before initiating data collection. Here we are showing a sample visualization of the data which can be accessed from the training tab after collection is completed. We will now add a screw to the edge of the normal rotor to create a major imbalance and collect 200 seconds of data for this imbalanced class. As you may know, 200 seconds is a very small amount for machine learning, but even with the small sample size we can achieve high accuracy with Kikso AutoML. With more data, the classification results can be even more robust. You can check whether or not additional data can help by clicking the Details icon in the Models tab to look at the learning curve after the model is built. To collect data for the eccentric class, the rotor will be replaced with an eccentric rotor on the fault simulator. We will collect 200 seconds of data for the eccentric class. Finally, we will collect an additional 200 seconds of data with the fault simulator off. After having completed data collection for the four classes, off, eccentric, imbalanced, and regular, we will proceed to train machine learning models. We skip grouping, select automatic sensor and feature group selection, then define the instance length and the classification interval. This can also be determined automatically by Kikso AutoML. We will have Kikso AutoML perform hyperparameter tuning and generate learning curves. Here, you can see the available machine learning algorithms. We will choose Logistic Regression, Random Forest, and XGBoost models for this demo. After training is complete, we can look at all of the model's metrics, including cross-validation accuracy, latency, size, and additional details in order to select the best model for our use case. With just one click, we can flash the Random Forest model to the STWIN for live testing. We can select how to connect our hardware to send back the classification results. Here, we will use USB. As we start the fault simulator, we can see the probability scores begin to change between the class labels. All of the inference is happening right on the STWIN locally. The laptop is only used for display. Now, we will add a screw to the edge of the rotor and create a major imbalance. Once the simulator reaches 1500 RPM, the speed at which we recorded the original data, we see that the machine learning model detects the imbalanced rotor. We will now replace the normal rotor with an eccentric rotor on the fault simulator. Finally, at 1500 RPM, the model is able to distinguish the eccentric rotor anomaly from the imbalanced rotor anomaly. I just want to highlight a couple of things from the video. On the left here, you can see all the sensors available on the STWIN platform. If a different hardware is selected, for example, the sensor tile.box, these will change accordingly. On the bottom, we see isolation forest, local outlier factor, and one class SVM. These are the three algorithms supported for single class classification. On the right, we see 10 different machine learning algorithms for multi-class classification, which was what we showed in a demo video earlier. And we're always working on adding more. Hope you've enjoyed our demo. For a limited time, you can experience our powerful tool at cakesill.com with ST platforms for free. Thank you, Tina. That was very interesting and informative. If you have any questions, we are available live during the event via chat, and for further information, please visit cakesill.com and st.com. Thank you.