Sensor Drift Detection Using Differential Value Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing monitoring systems for infrastructure and manufacturing systems face challenges in efficiently detecting sensor drift due to environmental changes or instrument aging, leading to potential erroneous detections and difficulties in determining the optimal timing for updating machine learning models.
Innovation Solution
A data processing apparatus that calculates first and second differential values from measurement data of a sensor over different time periods, generates differential value distributions, and outputs these distributions for comparison, allowing users to detect drift in time derivatives and improve detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are continuously used for anomaly detection, then detection coverage is maintained, but model deterioration leads to increased erroneous detections
Solution Approach 1:
The system implements feedback by continuously monitoring sensor measurement values and comparing them against machine learning model predictions. When deviations exceed thresholds, the system triggers investigations and potential model updates, creating a closed-loop feedback mechanism that maintains detection accuracy while enabling timely model maintenance
Solution Approach 2:
The system performs preliminary actions by proactively monitoring model performance metrics and sensor drift patterns to determine optimal update timing before significant deterioration occurs. This allows the system to maintain high detection accuracy by updating models at appropriate intervals without waiting for failure modes to manifest
2Measurement precision
If sensor drift detection is performed on instantaneous values only, then detection simplicity is maintained, but time derivative drift cannot be detected leading to erroneous results
Solution Approach 1:
The system transitions from one-dimensional instantaneous value monitoring to two-dimensional analysis by incorporating time derivatives. It calculates both the sensor measurement value and its time derivative, then compares these against model predictions for the corresponding differential values, enabling detection of drift patterns that would be invisible in instantaneous values alone
Solution Approach 2:
The detection method is segmented into distinct components: instantaneous value comparison and time derivative comparison. This segmentation allows the system to independently evaluate different aspects of sensor behavior and model accuracy, improving overall detection precision while maintaining methodological clarity
3Reliability
If the number of sensors is increased to monitor a large system, then monitoring coverage is improved, but the number of people required to monitor all sensors increases
Solution Approach 1:
The machine learning model serves as an intermediary that processes sensor data automatically, comparing predicted values against actual measurements and identifying anomalies. This intermediary capability allows the system to maintain comprehensive monitoring coverage across numerous sensors while requiring minimal human intervention for data analysis and interpretation
Data Source
Figure 1
Figure 2
Figure 3~5
AI summary
According to one embodiment, a data processing apparatus (100) includes a first acquisition unit (103), a second acquisition unit (104), a first differential unit (105), a second differential unit (106), a first distribution unit (107), a second distribution unit (108), a first output unit (119) and a second output unit (120). The first differential unit (105) calculates, from the first measurement data, a first differential value set that is a set of first differential values in a time direction at a time included in the first period of the measurement values of the sensor of interest. The second differential unit (106) calculates, from the second measurement data, a second differential value set that is a set of second differential values in a time direction at a time included in the second period. A first differential value distribution and a second differential value distribution are generated.