Time-Series Anomaly Detection with Peak-Based Data Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods for time-series data in manufacturing processes are sensitive to preprocessing methods, leading to inconsistent performance of artificial neural network models, and there is a need for effective data preprocessing and anomaly detection techniques.
Innovation Solution
The method involves acquiring time-series data, dividing it into sub-time-series data based on time intervals or histogram data, adjusting variable values using scaling functions, and using a neural network-based detection model to determine anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If time-series data with various peak values is input directly to the neural network model without preprocessing, then the model training may succeed or fail depending on the data distribution, but the anomaly detection performance becomes inconsistent and unreliable
Solution Approach 1:
The patent divides time-series data into multiple sub-time-series data segments based on peak values. Each segment represents a distinct operational mode or product type. By segmenting the data, the neural network can be trained on more homogeneous subsets, leading to consistent anomaly detection performance across different operating conditions without requiring complex global preprocessing.
Solution Approach 2:
The patent transforms the original time-series data by changing its parameter representation - specifically by creating segmented representations based on peak value thresholds. This parameter transformation converts a single complex time-series into multiple simpler sub-series, each with more consistent statistical properties, thereby improving model reliability.
2Measurement precision
If the neural network model is trained on multi-modal time-series data with various peak values, then the model must handle complex data distributions, but the training success becomes dependent on preprocessing methods rather than the model itself
Solution Approach 1:
By segmenting multi-modal time-series data into sub-time-series based on peak values, the patent creates homogeneous training subsets. This segmentation allows the neural network to learn anomaly patterns from consistent data distributions, improving measurement precision for anomaly detection while simplifying the training process by eliminating the need to handle complex multi-modal distributions as a single dataset.
Solution Approach 2:
The patent performs preliminary segmentation of time-series data before feeding it to the neural network. This preliminary action of dividing data into mode-specific segments prepares the data in advance, ensuring that the model receives consistent input patterns during training, thereby improving both anomaly recognition accuracy and training ease.
3Reliability
If a single neural network model is used to detect anomalies across all operational modes, then the model structure remains simple, but the detection accuracy varies depending on the operational mode and data distribution
Solution Approach 1:
The patent segments time-series data into multiple sub-time-series based on peak values that correspond to different operational modes or product types. This segmentation enables a single neural network model to process different operational modes through distinct data segments, improving detection reliability across products while maintaining model simplicity. The model adapts to different modes by receiving mode-specific segmented data rather than requiring multiple specialized models.
Data Source
AI summary
An anomaly data detecting method performed by a computing device having at least one processor includes acquiring first time-series data, dividing the first time-series data into a plurality of sub time-series data, adjusting scales of variable values included in at least one sub time-series data among the plurality of sub time-series data and determining whether the first time-series data is abnormal by inputting scaled first time-series data to a neural network based detection model.