Anomaly detection with dynamic density estimation

Patent metadata

Field	Value
Publication number	US-12499352-B2
Application number	US-202318156490-A
Country	US
Kind code	B2
Filing date	Jan 19, 2023
Priority date	Jan 19, 2023
Publication date	Dec 16, 2025
Grant date	Dec 16, 2025

Abstract

In detecting anomaly in samples, convolutional neural network (CNN) and machine-learning classifier modelled with support vectors are used. The CNN and classifier are initially trained with normal samples, and incrementally trained in retraining sessions each with self-generated anomalous samples identified in inference preceding a retraining session under consideration, thereby continually improving the anomaly-detection performance without a need to seek anomalous samples for initializing the CNN and classifier. The support vectors are selected as feature k-centers of output feature map of the CNN. Dynamic density estimation is used to determine which feature k-centers in existing support vector set are retainable in updating the support vector set in the retraining session. As such, not all feature k-centers need to be recomputed to give the support vectors in the updated support vector set. Computation effort in updating the support vector set is reduced in comparison to generating this set from scratch.

First claim

What is claimed is:

1 . A method for testing a plurality of manufactured articles to determine whether an individual manufactured article is a defective item, the method comprising:
generating a plurality of samples from the plurality of manufactured articles, wherein in the plurality of samples, a corresponding sample generated from the individual manufactured article is a one-dimensional or multi-dimensional signal obtained from sensing the individual manufactured article; and executing a computer-implemented process of detecting sample anomaly in the plurality of samples, wherein if the corresponding sample is determined to be anomalous, the individual manufactured article is determined to the defective item, and wherein the computer-implemented process comprises:
processing the plurality of samples for anomaly detection in an inference stage, inserting a cold-start stage preceding the inference stage, dividing the inference stage into a plurality of inference-stage sessions, and inserting a retraining session between any two successive inference-stage sessions;
in the inference stage, using a convolutional neural network (CNN) to extract features of an individual sample of the plurality of samples to thereby generate a feature map, and using a machine-learning classifier modelled with one or more support vector sets to process the feature map to determine if the individual sample is anomalous;
in the cold-start stage, initializing the one or more support vector sets according to an initial training set of feature maps generated from processing a set of normal samples with the CNN after the CNN is pretrained; and
in the retraining session, finetuning the CNN according to at least an interim set of self-generated anomalous samples identified during an inference-stage session immediately preceding the retraining session, and updating the one or more support vector sets according to an intermediate training set of feature maps generated from processing the set of normal samples with the CNN after the CNN is finetuned, thereby allowing an anomaly-detection performance to be continually improved due to introducing newly-identified self-generated anomalous samples in incrementally training the CNN and classifier while avoiding a need for seeking an initial training set of anomalous samples for initializing the CNN and classifier in the cold-start stage.

2 . The method of claim 1 , wherein the computer-implemented process further comprises:
in the cold-start stage, generating a first plurality of feature k-centers from extracted features in the initial training set of feature maps, and selecting respective support vectors in the initialized one or more support vector sets from the first plurality of feature k-centers; and in the retraining session:
determining whether an individual support vector in the one or more support vector sets is no longer a feature k-center according to the intermediate training set of feature maps, whereby respective support vectors in the one or more support vector sets are divided into retainable support vectors and discardable support vectors for updating the one or more support vector sets;
generating a second plurality of feature k-centers from extracted features in the intermediate training set of feature maps under a condition that the retainable support vectors are in the second plurality of feature k-centers, thereby reducing computation effort in comparison to generating the second plurality of feature k-centers from scratch; and
selecting respective support vectors in the updated one or more support vector sets from the second plurality of feature k-centers.

3 . The method of claim 2 , wherein each of the first and second pluralities of feature k-centers is obtained by using a greedy k-center algorithm.

4 . The method of claim 2 , wherein the one or more support vector sets consist of multiple support vector sets, an individual support vector set collecting respective support vectors located on a preselected region of the feature map, respective preselected regions for the multiple support vector sets being non-overlapping, and wherein the computer-implemented process further comprises:
in the inference stage, processing the feature map with each of the multiple support vector sets for determining any location on the individual sample where anomaly occurs.

5 . The method of claim 2 , wherein the one or more support vector sets consist of a single support vector set.

6 . The method of claim 2 , wherein:
the CNN includes an averaging pooling layer such that the feature map is reduced to a feature vector; and the one or more support vector sets consist of a single support vector set.

7 . The method of claim 1 , wherein the computer-implemented process further comprises:
in the cold-start stage, loading pre-stored CNN model parameters into the CNN for pretraining the CNN.

8 . The method of claim 1 , wherein in the retraining session, the finetuning of the CNN comprises:
augmenting the set of normal samples with the interim set of self-generated anomalous samples by copying selected anomalous portions of respective anomalous samples in the interim set of self-generated anomalous samples onto one or more normal samples in the set of normal samples, thereby creating a set of synthetic training samples for enriching training set variety in finetuning the CNN; and updating model parameters of the CNN according to the set of synthetic training samples by performing multiple iterations of optimizing the model parameters to minimize a loss function in each iteration and by using a center loss and a diversity loss alternately and recursively as the loss function over the iterations of model parameter optimization.

9 . The method of claim 1 , wherein the individual sample is an image, and the CNN is implemented with a two-dimensional convolution operation.

10 . The method of claim 1 , wherein the individual sample is an image, and wherein the finetuning of the CNN further comprises receiving the interim set of self-generated anomalous samples in which an individual anomalous sample is pixelwise labelled for increasing CNN model accuracy in retraining the CNN.

11 . The method of claim 1 , wherein the classifier adopts an L2 distance as a criterion in classification.

12 . The method of claim 1 , wherein the corresponding sample generated from the individual manufactured article is a two-dimensional (2D) image of the individual manufactured article.

13 . The method of claim 1 , wherein the corresponding sample generated from the individual manufactured article is a one-dimensional (1D) data stream given by a time series of sensor data for measuring an activity done by the individual manufactured article.

14 . The method of claim 13 , wherein the individual manufactured article is a car.

15 . The method of claim 1 , wherein the corresponding sample generated from the individual manufactured article is a video clip recording motion of the individual manufactured article during operation.

16 . The method of claim 15 , wherein the individual manufactured article is a robotic arm.

Assignees

• Hong Kong Applied Science & Tech Research Inst Co Ltd

Inventors

• Li Peiqin
• Shu Jiayu

Classifications

• G06T7/0004

  Industrial image inspection · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/30108

  Industrial image inspection · CPC title

• G06N3/08 Primary

  Learning methods · CPC title