Time-Series State Classification for Variable Anomaly Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of state prediction in monitoring targets, such as manufacturing plants, is compromised due to high variability in labels assigned to anomalous measurement data, leading to discrepancies between predicted and actual states.
Innovation Solution
A time-series data processing method that generates state information using a learned generator for both original and divided time-series data, allowing for classification and improved prediction accuracy by refining labels based on similarity analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If labels are given by persons to anomalous measurement data, then the degree of freedom of label content is high, but the accuracy of state prediction lowers due to label variability
Solution Approach 1:
The patent transforms the label assignment process from manual human labeling to automated machine learning-based labeling. The system changes the parameter of label generation method, using a trained model to predict state information and generate standardized labels, thereby reducing label variability while maintaining adaptability through the model's learning capability
Solution Approach 2:
The patent replaces the mechanical human labeling process with an automated computational system. Instead of persons manually assigning labels to measurement data, the system uses machine learning models to automatically generate state information and labels, eliminating human subjectivity and improving prediction accuracy
2Device complexity
If time-series data is analyzed with fixed time width, then processing is simple, but prediction accuracy lowers when actual anomaly duration varies
Solution Approach 1:
The patent introduces dynamic time width adjustment based on the learned state information. Instead of using a fixed time width for all measurements, the system adapts the time width according to the predicted state and characteristics of the measurement data, allowing the processing parameters to change dynamically to match actual anomaly patterns
Solution Approach 2:
The patent segments the time-series data into multiple intervals with different time widths based on the predicted state. By dividing the data processing into segments with optimized time widths for different states, the system achieves both processing efficiency and high prediction accuracy for anomalies of varying durations
Data Source
AI summary
A time-series data processing apparatus according to the present invention includes: a generating unit configured to generate a generator having learned so as to generate state information representing a state of time-series data having a predetermined time width in accordance with a label given to the time-series data; a state information generating unit configured to generate, by using the generator, state information representing a state of division time-series data obtained by dividing the time-series data by a shorter time width than the predetermined time width; and a classifying unit configured to classify a plurality of division time-series data based on state information of the plurality of division time-series data.