Issue link: https://resources.mouser.com/i/1481902
Intelligence at the Edge 30 Additional Data Sensor Data Storage Data Preprocessing Machine Learning Techniques • Noise removal • Oset removal • Outlier detection Final Data Set Microphone Predictive Model Mag, E-compass Pressure, Humidity, Temperature 6-axis IMU Intelligent Sensors with Finite State Machines & Machine Learning Core The goal of ML is to mine rules out of the data in real- time or as the data is pulled from storage. ML is useful when rules are not available beforehand as in traditional computer programing. Figure 2 illustrates a high-level general ML method based on sensor data that are delivered by sensors in real-time and the historical data that is available in data storage. As the figure indicates, the ML techniques require that clean and reliable data allows the system to extract reliable models that can be then used to predict future events. The quality of the predictive models is as good as the data fed into ML algorithms. A standard ML solution at the edge consists of a set of sensors and an external general-purpose processing unit. In this solution, the ML models run on the processing unit outside the sensor chip. The main advantages of such a solution include configurability and the number of algorithms that can be executed. However, this solution Figure 3: High-level block diagram of a) a standard solution b) an intelligent sensor module with embedded FSM and MLC. (Source: STMicroelectronics) comes with a few drawbacks. For example, with dedicated individual components, this solution is less suitable for a battery-operated industrial node subsystem that has size and power constraints. Figure 3a depicts a high-level block diagram of a standard ML solution. Figure 3b shows the block diagram of an intelligent sensor with embedded FSMs and MLC. The main advantage of this solution is the significant reduction of power consumption while running the algorithms. Common application cases show 20 to 100 times current consumption reduction compared to a standard solution (Figure 3a). The other advantage of an intelligent sensor module is the configurable embedded electronic blocks that can ensure fast and effective implementation of target algorithms. It is important to note that there are some differences between the embedded FSM and the MLC. The FSM is designed for deductive learning while the MLC performs inductive learning. That means an FSM can draw conclusions that lead to observations. The MLC, on the other hand, makes observations that lead to conclusions. Figure 2: General ML method based on real time sensor data and historical data. (Source: STMicroelectronics) Figure 3 Sensor Unit Sensor 1: : Sensor n Intelligent Sensor Module: Sensor 1... Sensor n + Finite State Machines + Machine Learning Core General Purpose Microcontroller Algorithms: Algo 1: : Algo m I2C/SP1 UART,... Gateway, ... Gateway, ... I2C/SP1 Figure 3 Sensor Unit Sensor 1: : Sensor n Intelligent Sensor Module: Sensor 1... Sensor n + Finite State Machines + Machine Learning Core General Purpose Microcontroller Algorithms: Algo 1: : Algo m b. Intelligent Sensor with embedded FSM and MLC I2C/SP1 UART,... Gateway, ... I2C/SP1 (Source: STMicroelectronics)