On-Device Learning Sensor Unit for Machine Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machines lack advanced sensor systems, making them prone to breakdowns due to undetected deterioration, which is costly to address with full replacements.
Innovation Solution
A low-power sensor unit with an onboard processing unit that learns standard operating modes and detects anomalies using Mahalanobis distance calculations, outputting warnings for timely maintenance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machines are equipped with advanced sensor systems to detect operational anomalies, then reliability and preventive maintenance capability are improved, but device complexity and cost increase
Solution Approach 1:
The sensor unit performs self-learning during a training phase where it automatically captures sensor data during normal machine operations and generates classification vectors and Mahalanobis distance thresholds without human intervention. This self-service capability eliminates the need for complex manual programming while enabling the system to detect anomalies and predict failures, thereby improving reliability without proportionally increasing device complexity
Solution Approach 2:
The system performs preliminary learning during a training phase before actual anomaly detection begins. During this phase, the sensor unit collects data from normal operations and pre-computes classification vectors and distance thresholds. This preliminary action prepares the system in advance, enabling it to quickly and accurately detect deviations from normal operation without requiring complex real-time analysis, thus improving reliability while keeping the operational complexity low
2Measurement precision
If sensor data is continuously monitored and analyzed in real-time, then anomaly detection precision is improved, but energy consumption increases
Solution Approach 1:
The system pre-computes classification vectors from training data and calculates Mahalanobis distance thresholds during a training phase before actual monitoring begins. By performing these computationally intensive operations in advance, the system avoids continuous complex calculations during real-time operation. During monitoring, only simple distance comparisons are needed, which maintains high detection precision while significantly reducing energy consumption
Solution Approach 2:
The patent replaces continuous complex mechanical/computational analysis with a statistical mathematics approach using Mahalanobis distance. By substituting the need for continuous sophisticated signal processing with pre-computed statistical thresholds and simple distance calculations, the system achieves high anomaly detection precision with minimal ongoing computational energy expenditure
Data Source
Figure 1
Figure 2A
Figure 2B~2E
AI summary
A sensor unit (102) is coupled to a machine (100) and configured to detect anomalous behavior of the machine (100). The sensor unit (102) includes a low power microcontroller that learns to recognize a plurality of operations of the machine (100). The sensor unit (102) generates mean vector (M) and inverse of a Cholesky decomposition matrix for each operation. During a detection mode the sensor unit (102) computes a Mahalanobis distance for each feature vector (F), mean vector (M) and first matrix. The sensor unit (102) detects anomalous behavior or classifies the operation of the machine (100) based on the Mahalanobis distances.