Sensor Anomaly Detection Using Cross-Sensor Regression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing predictive maintenance techniques are inadequate when historical sensor data is insufficient, as they rely on pattern identification from available data to detect impending failures, leading to inefficiencies in anomaly detection.

Innovation Solution

A system that calculates regressions from time-series data across multiple sensors to predict values and compare them with observed values, determining anomalies by aggregating root mean squared error, and issuing notifications when thresholds are exceeded, allowing for flexible integration with analytics workflows.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If historical sensor data is used for pattern identification to detect impending failures, then predictive maintenance capability is improved, but the method becomes unsuitable when sufficient training data is not available

Engineering Contradiction:
Improvepredictive maintenance capabilityVSAvoidapplicability with limited data
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary actions by using available sensor data to train prediction models before actual failure detection is needed. The models are pre-trained on whatever historical data exists, and these pre-trained models are then used to detect anomalies in real-time operation, allowing the system to function even with limited initial training data

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

Prediction models serve as intermediaries between the available sensor data and the failure detection task. These models act as mediators that can generalize from limited training examples to detect patterns indicating impending failures, bridging the gap between insufficient historical data and the need for reliable predictive maintenance

Inventive Principle:
Principle #24Intermediary (Mediator)

2Loss of time

If regression models are trained on limited historical data, then model training speed is improved, but prediction accuracy may deteriorate

Engineering Contradiction:
Improvemodel training timeVSAvoidprediction accuracy
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The system applies partial action by training models on whatever portion of historical data is available, rather than waiting for complete datasets. The models are trained on partial data and then continuously refined through iterative learning from new sensor data, achieving acceptable prediction accuracy without requiring extensive initial training data

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The prediction models are designed to be dynamic and adaptive, continuously updating their parameters as new sensor data becomes available. This allows the models to improve their prediction accuracy over time through online learning, starting from limited initial training data and progressively becoming more accurate as they accumulate experience from operational data

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3364262B1Sensor data anomaly detection
Publication Date: 2021.03.17 SAP SE
  • EP3364262B1 patent drawingFigure 1
  • EP3364262B1 patent drawingFigure 2
  • EP3364262B1 patent drawingFigure 3~4

AI summary

Some embodiments include reception of a time-series of a respective data value generated by each of a plurality of sensors, calculation of a regression associated with a first sensor of the plurality of sensors based on the received plurality of time-series, the regression being a function of the respective data values of the others of the plurality of data sources, reception of respective data values associated with a time from and generated by each the plurality of respective sensors, determination of a predicted value associated with the time for the first sensor based on the regression associated with the first sensor and on the respective data values associated with the time, comparison of the predicted value with the received value associated with the time and generated by the first sensor, and determination of a value indicating a likelihood of an anomaly based on the comparison.