Sensor Anomaly Detection Using Event Sequence Likelihood
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Identifying malfunctioning or inoperable sensors in a computationally efficient manner is challenging, especially in environments where sensors can fail or record spurious readings, which interferes with accurate monitoring.
Innovation Solution
A method involving a trained statistical model, such as a Markov model, is used to analyze sensor data to determine the probability and frequency of event sequences, combining these to calculate the likelihood of anomalies, allowing differentiation between one-time failures and persistent faults.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional sensor monitoring methods are used, then sensor failures can be detected, but the computational complexity increases and false positives occur
Solution Approach 1:
The sensor data stream is segmented into discrete event sequences that can be independently analyzed. Each sequence represents a specific pattern of sensor activations, allowing the system to evaluate individual event patterns rather than processing entire data streams, thereby reducing computational complexity while maintaining detection accuracy.
Solution Approach 2:
Event sequences are pre-defined based on expected normal operations. By establishing what constitutes normal sensor behavior patterns in advance, the system can quickly compare actual sensor data against these predefined sequences without performing complex real-time analysis, reducing computational load while improving reliability.
2Reliability
If frequent sensor monitoring is performed, then persistent faults can be identified, but one-time failures generate false positives
Solution Approach 1:
The system applies different evaluation criteria to different event sequences based on their characteristics. Frequently occurring sequences are evaluated for persistence patterns, while rare sequences are evaluated differently, allowing the system to distinguish between one-time anomalies and persistent faults without generating false positives.
Solution Approach 2:
The system uses feedback from the frequency and pattern of event sequence occurrences to adjust its anomaly detection behavior. By monitoring how often specific sequences occur and comparing them against expected frequencies, the system can identify persistent faults while filtering out one-time failures that deviate from normal patterns.
3Measurement precision
If complex statistical models are used for anomaly detection, then detection accuracy improves, but computational efficiency decreases
Solution Approach 1:
The system uses lightweight statistical models that can be quickly computed and discarded for each event sequence evaluation. Rather than maintaining complex continuous models, the system applies simple probability calculations to predefined sequences, achieving sufficient detection accuracy with minimal computational overhead.
Solution Approach 2:
The system changes the parameters of statistical analysis from continuous complex distributions to discrete event sequence frequencies. By transforming the problem from analyzing continuous sensor readings to counting and comparing discrete event patterns, the system achieves high detection accuracy using simple frequency-based statistics rather than complex continuous models.
Data Source
AI summary
A method of identifying anomalous data obtained by at least one sensor of a plurality of sensors located within an environment. The method includes identifying, based on sensor data obtained from the plurality of sensors, at least one instance of a sequence of events that occurred within the environment. A probability of the sequence of events occurring within the environment under non-anomalous conditions is obtained. A frequency characteristic dependent on a frequency at which the sequence of events occurred within the environment is determined. A likelihood of the sequence of events occurring within the environment at the frequency is determined, based on a combination of the probability and the frequency characteristic. It is identified, based on the likelihood, that at least a portion of the sensor data is anomalous.