| CPC H04L 41/0631 (2013.01) [H04L 41/16 (2013.01)] | 4 Claims |
|
1. An abnormal data detection method for industrial Internet, comprising the following operations:
S1: acquiring real data distribution of a node and obtaining potential representation distribution based on the real data distribution; performing normal feature extraction on the potential representation distribution to obtain first normal data distribution; and comparing the real data distribution with the first normal data distribution to obtain a first anomaly score;
the operation of obtaining the potential representation distribution being implemented through a minimax adversarial encoder; and the minimax adversarial encoder being used for converting the real data distribution into the potential representation distribution;
the operation of obtaining the first normal data distribution being implemented through a minimax adversarial generator; and the minimax adversarial generator being used for converting the potential representation distribution into the first normal data distribution;
before the generator is subjected to minimax adversarial processing, further comprising performing evolutionary training on the generator;
the operation of the evolutionary training specifically comprising: step 1: converting normal data distribution in a node training set into potential representation distribution, and inputting the potential representation distribution into the generator to obtain a parent sample; step 2: performing variation on the parent sample to obtain a child sample; step 3: obtaining a quality evaluation score based on feature expression of the child sample, wherein the quality evaluation score is obtained based on a generated sample quality score and a generated sample diversity score, the generated sample quality score is an expected value of normal feature representation in the child sample, and the generated sample diversity score is obtained based on the normal feature representation in the normal data distribution and the normal feature representation in the child sample; step 4: if the quality evaluation score is less than a first predetermined value, eliminating the corresponding child sample; if the quality evaluation score is not less than the first predetermined value, performing step 5; and step 5: if the quality evaluation score is less than a second predetermined value, performing step 2 on the corresponding child sample as a new parent sample; and if the quality evaluation score is not less than the second predetermined value, finishing the evolutionary training;
S2: performing normal feature enhancement on the real data distribution of the node to obtain second normal data distribution; and comparing the real data distribution with the second normal data distribution to obtain a second anomaly score;
the operation of obtaining the second normal data distribution being implemented through a minimax adversarial discriminator; and the minimax adversarial discriminator being used for converting the real data distribution into the second normal data distribution;
S3: calculating a total anomaly score based on the first anomaly score and the second anomaly score;
S4: determining an anomaly risk level by comparing the total anomaly score with thresholds of different risk levels, wherein:
if the total anomaly score is less than a first threshold, the abnormal data risk level of the node being a first risk level;
if the total anomaly score is not less than the first threshold and not greater than a second threshold, the abnormal data risk level of the node being a second risk level;
if the total anomaly score is greater than the second threshold and less than a third threshold, the abnormal data risk level of the node being a third risk level; and
if the total anomaly score is not less than a third threshold, the abnormal data risk level of the node being a fourth risk level; and
S5: updating a communication permission of the node based on the determined anomaly risk level.
|