Anomaly Score Threshold Estimation from Negative Sample Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unsupervised defect detection methods lack an effective way to estimate a threshold for distinguishing between positive and negative samples, leading to false negatives due to sampling bias and requiring additional computing power and time for artificial defect generation.
Innovation Solution
A method that utilizes negative sample data to estimate a threshold by calculating anomaly scores, constructing a probability distribution model, and determining a confident interval to set a threshold based on the upper limit value, eliminating the need for positive samples and reducing computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user provides positive samples to determine threshold, then threshold can be determined, but sampling bias causes false negative determinations and threshold lacks statistical meaning
Solution Approach 1:
The patent inverts the conventional approach by determining the threshold based on negative samples (normal data) rather than positive samples (defective data). Specifically, it uses the maximum anomaly score from negative samples as the threshold, ensuring that normal samples will not be misclassified as defective, thereby eliminating false negatives while maintaining statistical rigor.
Solution Approach 2:
The method enables the system to automatically determine the threshold using only the available negative samples during the training phase, without requiring external user input of positive samples. The threshold is self-determined through the anomaly scoring mechanism applied to negative samples, making the process autonomous and statistically sound.
2Measurement precision
If artificial defects are generated to obtain probability distribution, then positive sample distribution can be obtained, but it consumes extra computing power and time
Solution Approach 1:
The patent extracts the threshold determination task from the conventional approach that requires positive samples and artificial defect generation. By taking out the necessity of positive samples and using only negative samples with the anomaly scoring mechanism, it eliminates the computationally expensive processes of artificial defect generation and multiple sampling iterations while maintaining statistical validity.
Solution Approach 2:
The threshold is predetermined during the training phase using negative samples, so that during the inference phase, no additional computational resources are required for threshold determination. This preliminary action eliminates the need for runtime threshold adjustment and artificial defect generation, significantly improving computational efficiency.
3Adaptability or versatility
If threshold is adjusted based on online samples, then adaptation to real data is achieved, but additional time and computational resources are required
Solution Approach 1:
The threshold is predetermined during the training phase using negative samples and the anomaly scoring mechanism. This preliminary determination ensures the threshold is adapted to the specific data distribution without requiring online adjustment, saving time and computational resources during deployment while maintaining adaptability to the specific application domain.
Data Source
AI summary
A threshold estimation method and an electronic device are provided. The electronic device includes a processing device and a storage device. Negative sample data is stored in the storage device. The processing device is electrically connected to the storage device and performs threshold estimation. The threshold estimation method includes: reading negative sample data, and extracting feature values; compressing the feature values into training data; calculating and recording anomaly scores of the training data to obtain an anomaly score distribution chart, the anomaly score distribution chart including a local extreme value and a global extreme value; estimating a probability distribution model corresponding to the anomaly score distribution chart based on the local extreme value and the global extreme value; calculating a confident interval based on the probability distribution model; and obtaining a threshold with meaning of negative sample probability distribution based on an upper limit value of the confident interval.