Unsupervised Anomaly Detection for Time Series Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current anomaly detection methods in computer systems, especially in unsupervised scenarios, face challenges in identifying abnormal data points in real-time within evolving time series without labeled data, leading to inefficiencies and potential false positives or negatives.

Innovation Solution

A system and method utilizing a combination of statistical and machine learning techniques, including the Z-test and Gaussian distribution analysis, to automatically detect anomalies in time series data without prior labeling, adapting to dynamic changes and minimizing false positives and negatives by calibrating algorithms with a small number of data points.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If unsupervised anomaly detection is used without labeled data, then data processing efficiency is improved and human intervention is reduced, but detection accuracy deteriorates leading to false positives and negatives

Engineering Contradiction:
Improvedata processing efficiencyVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary calibration using a small number of labeled data points to establish baseline statistical parameters (mean, standard deviation) before processing the main unlabeled time series data. This preliminary action enables the system to achieve both automated processing efficiency and acceptable detection accuracy without requiring extensive labeled data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system incorporates feedback mechanisms where detected anomalies are reviewed and used to refine the calibration model. The calibration parameters are continuously updated based on feedback from anomaly reviews, allowing the system to improve detection accuracy over time while maintaining automated processing capabilities.

Inventive Principle:
Principle #23Feedback

2Loss of time

If real-time anomaly detection is implemented in evolving time series, then response time is improved, but system complexity increases due to non-stationary data characteristics

Engineering Contradiction:
Improvemean time to detectVSAvoidsystem complexity
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The system dynamically adapts its calibration parameters in real-time as the time series evolves. Instead of using a fixed calibration model, the system continuously updates statistical parameters based on incoming data, allowing it to maintain detection accuracy in non-stationary environments while managing complexity through adaptive rather than static approaches.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes its operational parameters (calibration mean, standard deviation, detection thresholds) based on the evolving characteristics of the time series data. This parameter adaptation enables real-time detection in changing conditions without requiring a completely complex redesign of the detection algorithm.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If statistical methods like Z-test and Gaussian distribution analysis are used, then detection reliability is improved through mathematical rigor, but computational resources are consumed

Engineering Contradiction:
Improvedetection reliabilityVSAvoidcomputational resources
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system applies statistical methods selectively rather than universally. It uses rigorous Z-test and Gaussian distribution analysis only for the calibration phase and for points that require detailed analysis, while using simpler heuristic methods for routine anomaly detection. This partial application of statistical methods maintains reliability where needed while reducing overall computational resource consumption.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3671466B1Unsupervised anomaly detection for arbitrary time series
Publication Date: 2022.01.05 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3671466B1 patent drawingFigure 1a
  • EP3671466B1 patent drawingFigure 1b
  • EP3671466B1 patent drawingFigure 2a

AI summary

Examining time series sequences representing performance counters from executing programs can provide significant clues about potential malfunctions, busy periods in terms of traffic on networks, intensive processing cycles and so on. An unsupervised anomaly detector can detect anomalies for any time series. A combination of known techniques from statistics, signal processing and machine learning can be used to identify outliers on unsupervised data, and to capture anomalies like edge detection, spike detection, and pattern error anomalies. Boolean and probabilistic results concerning whether an anomaly was detected can be provided.