Supplier eBooks

ST/Industrial Sensing Solutions 14 Figure 7: A 5-level decision tree. Figure 8: Machine Learning Processing Strategy. Machine Learning Core Before describing the implementation of the machine learning capability in the latest sensors, let us review the definition of decision tree technique, which is one of the core elements of the machine learning processing in these sensors. A decision tree is one of the methods for super- vised learning and used to predict the value of a target variable based on a sequence of Yes/No questions ("deci- sions") about one or more variables. Illustrated below is a 5-level decision tree (Figure 7). The nodes represent the questions, and the left/right branches represent the yes/no answers. The collected data points are used to estimate the value of the target. The closer the expected values is at the leaves to the actual data, the more accurate the predictive model will be. It is essential to distinguish between a classification tree, which is when the target variable is categorical, and a regression tree when the target variable is continuous. The implementation of machine learning processing in the latest sensors offers ease of use and a high level of flexibili- ty using the decision tree technique. The machine learning processing can be broken down in three steps as illustrat- ed in Figure 8 below to explain how it works. Each appli- cation requires a different set of sensor data. The first step would be to define the sensor type that the target applica- tion will use. The next step is to determine the computation block in which the features/parameters gets calculated. The final step would be to implement the trained decision tree model and to configure the output of the model. As far as the sensor inputs for the target applications are concerned, one can use a three-axis accelerometer, a three-axis gyroscope, and an external sensor, e.g., a three-axis magnetometer. The magnitude and square magnitude from the three sensors gets computed. These values get employed as inputs for the embedded filters to create additional inputs. The choice of filtering is vast and it has gotten designed with smart memory allocation in mind. Each filter may get applied to all three components (sensor data, magnitude, or square magni- tude). The beauty of this solution is that the de- cision trees may get configured with a high degree of flexibility (Figure 9). Each node is composed of an input provided by the computational block, a threshold, a connection for the true path, and another connection for the false one. Every decision tree once executed will output a temporary result. Sensor-Hub Some of the latest MEMS sensors, in partic- ular, the Inertial Measurement Units (IMUs), are equipped with a sensor hub function- ality. That means that the sensor may get used as a hub to which external sensors can be connected. There is to external sensors a dedicated connection mode available, which allows the implementation of the sensor hub functionality. In sensors with embedded hub functionality, there is an I2C master available to which up to four external sensors can be connected when the sensor gets operated in the "sensor hub" mode. Shown below is a block diagram of a sensor hub embedded in the latest IMUs offered by ST (Fig- ure 10). The main advantages of the sensor hub function- ality include data coherency/data synchronization, more natural placement and routing, and lower system power consumption. FIFO The MEMS sensors have an integrated FIFO (First-In, First- Out) buffer allowing the user to store data to limit interven- tion by the host processor. It helps reduce overall system power consumption because this decreases the host processor interaction with the sensor. Other advantages include the possibility of saving the history of an event

• Cover