Time-Series Sensor Segmentation for AI Anomaly Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based anomaly detection systems for technical systems require manually labeled training data, which is tedious and often inaccurate, leading to inefficient and unreliable training processes.
Innovation Solution
An assistance device automatically generates training data by dividing historical sensor data into segments with similar data distributions, determining neighborhood patterns, and selecting subsequences based on the most frequent pattern to represent normal operation, thereby ensuring accurate representation of normal behavior for AI training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual labelling of training data is performed, then training data can be provided for AI model training, but the process is tedious and time-consuming
Solution Approach 1:
The system automatically identifies and segments normal operation data without human intervention. The processor autonomously analyzes sensor data, determines normal operation periods, and generates training data sequences, making the system self-sufficient in data preparation tasks.
Solution Approach 2:
Manual labelling operations are replaced by automated computational processes. The processor uses algorithms to analyze sensor data patterns, identify normal operation segments, and generate training data, substituting human manual work with automated mechanical computation.
2Manufacturing precision
If manual setting of time window width is performed, then training data segments can be created, but the setting is often inaccurate and too wide
Solution Approach 1:
The time window width is not fixed but dynamically adapted based on the actual duration of normal operation periods detected in the sensor data. The system adjusts segment lengths to match the genuine temporal characteristics of normal operations, making the segmentation flexible and data-driven rather than static and predetermined.
Solution Approach 2:
The system performs preliminary analysis of the sensor data to identify normal operation periods before creating training segments. By first determining which periods represent normal operation through pattern recognition, the system establishes accurate temporal boundaries before segmenting the data into training units.
3Productivity
If shuffling methods are used to divide time series data, then training data can be generated quickly, but the method is susceptible to wrong or unsuitable data selection
Solution Approach 1:
Before shuffling or randomizing data for training, the system performs preliminary identification of normal operation periods by analyzing sensor data patterns. This preliminary step ensures that only data from genuinely normal operations is selected and shuffled, preventing inappropriate data from being included in the training set.
Solution Approach 2:
The system replaces manual data selection and validation processes with automated computational analysis. The processor uses algorithms to objectively determine normal operation periods based on sensor data characteristics, eliminating human subjectivity and potential errors in data selection while maintaining high speed.
Data Source
AI summary
Assistance device for automatically generating training data of a time series of sensor data, further on called temporal sensor data, applied to train an Artificial Intelligence system used for detecting anomalous behavior of a technical system, including a processor configured to perform—obtaining historical temporal sensor data, dividing the historical temporal sensor data into a temporal sequence of segments and assigning one segment type out of several different segment types to each segment, iteratively for each segment, determining a neighborhood pattern of segment types, determining the most frequently occurring neighborhood pattern from all determined neighborhood patterns as reference pattern for normal operation of the technical system, —selecting a subsequence of segments out of the historical temporal sensor data, which is ordered according to the reference pattern, and—outputting the subsequence of segments for applying as training data.