Time-Series Cluster Classification for New Machine Behavior Labels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing AI-based monitoring mechanisms for machines, particularly for electric submersible pumps, face challenges in formulating and maintaining rules, lack transparency in machine learning outputs, require large amounts of manually prepared training data, and struggle to address new behavior types immediately.

Innovation Solution

A device and method that segments time series data into clusters based on probabilistic models, utilizes existing label information for classification, and allows user input to generate new label information, resolving conflicts through probabilistic models and user interaction.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning is used for classification, then classification accuracy is improved, but the system becomes a black-box approach offering no explanations and requiring large amounts of manually prepared training data

Engineering Contradiction:
Improveclassification accuracyVSAvoidtraining data preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary clustering of time series data into behavior patterns before classification. This preliminary segmentation organizes the data structure, allowing the classification stage to work with pre-grouped patterns rather than raw data, reducing the need for extensive manual training data preparation while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The time series data is segmented into clusters representing different behavior patterns. This segmentation divides the complex classification problem into smaller, more manageable pattern recognition tasks, reducing the computational complexity and training data requirements while preserving classification accuracy.

Inventive Principle:
Principle #1Segmentation

2Ease of operation

If pre-defined rules are used for classification, then transparency and explainability are improved, but the system complexity increases and requires domain expert formulation

Engineering Contradiction:
ImprovetransparencyVSAvoidrule formulation complexity
Core Design Contradiction:
Ease of operationVSDevice complexity

Solution Approach 1:

The system automatically learns behavior patterns from the data through clustering and classification stages, eliminating the need for manual rule formulation by domain experts. The system serves itself by automatically discovering patterns and creating classification rules from the clustered data, maintaining transparency while reducing complexity.

Inventive Principle:
Principle #25Self-service

3Adaptability or versatility

If machine learning is used, then new behavior types can be learned, but the system cannot address new behavior types immediately until the machine learning process is completed

Engineering Contradiction:
Improvenew behavior type detectionVSAvoidtime to address new behavior
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system dynamically adapts to new behavior types through an iterative process. When new time series data is received, it is clustered into existing or new behavior patterns, and the classification model is continuously refined. This dynamic approach allows the system to immediately address new behavior types by clustering them and progressively improving classification accuracy over time.

Inventive Principle:
Principle #15Dynamics

4Measurement precision

If manual labeling of time segments is required, then classification accuracy is improved, but the productivity decreases due to tedious data preparation

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata preparation efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

Instead of requiring manual labeling of all time segments, the system applies partial action by automatically clustering the majority of data and only requiring minimal user input for ambiguous cases. This approach maintains high classification accuracy while dramatically improving productivity by eliminating the need for exhaustive manual labeling.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3918537B1Cluster based classification for time series data
Publication Date: 2026.04.08 SIEMENS AG
  • EP3918537B1 patent drawingFigure 1
  • EP3918537B1 patent drawingFigure 2
  • EP3918537B1 patent drawingFigure 3

AI summary

For analyzing time series data monitored on a machine, a device (200) segments the time series data into multiple time segments (230). Further, the device (200) determines a cluster of time segments estimated to have the same dynamics of the time series data. The device (200) then classifies the cluster based on label information associated with at least one of the time segments. If none of the time segments of the cluster has associated label information, the device (200) presents at least a part of the time series data of the cluster to a user (10). Based on a user input received in response to said presenting the time series data, the device (200) classifies the cluster and generates label information associated with the time segments of the cluster. The generated label information indicates a result of classifying the cluster.