Embedded Decision-Tree Sensor for Local Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sensor devices require external computational support for the training step of machine learning algorithms, leading to high computational resource consumption and inefficiency in anomaly detection.
Innovation Solution
A sensor device with a microcontroller unit and a processing circuit that includes a conversion stage, a data generation stage, and a decision stage, capable of performing machine learning tasks such as data vector generation and decision tree classification without external support.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning algorithms are implemented in sensor devices with microcontrollers, then anomaly detection capability is improved, but computational resource consumption increases
Solution Approach 1:
The patent segments the machine learning process into two distinct phases: a training phase that occurs externally during manufacturing, and an inference phase that occurs locally in the sensor device. This segmentation allows the sensor to have anomaly detection capability while minimizing ongoing computational resource consumption, as only the lightweight inference step runs on the microcontroller.
Solution Approach 2:
The training phase is performed in advance during the manufacturing process, before the sensor device is deployed. This preliminary action pre-computes the model parameters and thresholds, so that the deployed sensor only needs to perform simple inference operations using pre-computed data, significantly reducing operational computational requirements.
2Reliability
If the microcontroller continuously performs inference step, then anomaly detection accuracy is improved, but computational resource consumption increases
Solution Approach 1:
Instead of continuous inference processing, the system uses periodic sampling of sensor data at predetermined intervals. The microcontroller performs inference only at these discrete time points, comparing sampled data against pre-computed thresholds. This periodic approach maintains detection accuracy for time-varying anomalies while dramatically improving computational efficiency compared to continuous processing.
3Measurement precision
If training step is performed externally, then model accuracy is improved, but device complexity increases
Solution Approach 1:
The system is designed so that the sensor device serves itself during the inference phase, using pre-computed models to autonomously detect anomalies without requiring external computational support during operation. The external training phase is a one-time setup that enables the device's self-sufficiency during deployment, balancing model accuracy with operational simplicity.
4Adaptability or versatility
If sensor device operates autonomously without external computer support, then operational independence is improved, but computational resource requirements increase
Solution Approach 1:
The patent changes the computational parameters by using simplified inference algorithms with pre-computed thresholds and decision trees, rather than running full machine learning models. This parameter optimization allows the sensor to operate autonomously with minimal computational resources, achieving operational independence without excessive energy consumption.
Data Source
AI summary
Sensor device with a microcontroller unit and a sensor including a transducer, which is coupleable to a device and generates a signal indicative of a physical quantity, and a processing circuit including: a conversion stage which generates samples of the physical quantity; a data generation stage which generates data vectors as a function of the samples, each data vector being formed by programmable quantity values; and a decision stage. The microcontroller unit programs the decision stage so that it classifies the data vectors by executing a decision tree having a structure and thresholds. In a configuration mode, the microcontroller unit programs the data generation stage; in a calibration mode, the microcontroller unit acquires a corresponding set of data vectors, determines, for each programmable quantity, a corresponding range of admissible values and programs the thresholds as a function of the ranges of admissible values; in a detection mode, the decision stage classifies the data vectors, by executing the decision tree on the basis of the programmed thresholds.