Token-based multi-class anomaly detection method

Abstract

The invention discloses a Token-based multi-class anomaly detection method. The method comprises the following steps of 1) obtaining a data source, 2) judging whether the data source is a continuous number type data source or a continuous character type data source, 3) if the data source is judged to be the continuous number type data source, dividing the continuous number type data source into ngroups to obtain token mapping values, and if the data source is judged to be the continuous character type data source, directly obtaining a token mapping value, 4) constructing a deep learning network model, 5) accessing the data features of the multiple groups of tokens into the deep learning network model for training, and obtaining a prediction value of the target by an output layer through an activation function, 6) comparing the predicted value with the actual label to construct a ternary cross entropy loss function, and optimizing an output layer by using the ternary cross entropy lossfunction, and 7) using the deep learning network model to output an approximate value for judgment to obtain the result. According to the method, the complexity of the model is reduced by reducing the complexity of the decision-making space, the requirement for the number of training samples is reduced, and the amount of calculation parameters is small.

Description

technical field [0001] The invention belongs to the field of predictive data digitization, and in particular relates to a Token-based multi-category anomaly detection method. Background technique [0002] In existing applications such as machine abnormality monitoring and predictive maintenance, compared with public data such as image digital recognition and image data, the digitization process of these complex operation data requires high precision and takes a long time. It is still a huge challenge to build these learning models with only a small amount of high-quality data and obtain high recognition accuracy. [0003] In addition, the selection of the training time window in repetitive data has a great influence on the final result. In order to weaken or eliminate such influence, we introduce the concept of Token in the NLP language, discretize the continuous input data, and The overall model has a certain fault tolerance to the input data, and reduces the quality requi...

AI Technical Summary

Problems solved by technology

[0004] The existing binary cross-entropy loss function is not enough to predict the degree of difference between the three probability distributions with a target value between -1 and 1. When using binary classification to solve the ternary classification problem, it is necessary to make corresponding changes in the final output layer. From the previous one node to three, the activation function can choose softmax or sigmoid, the former has the highest accuracy but cannot output probability, the latter can output approximate probability, but the accuracy is lower

Generally, the two are combined to output two sets of [softmaxx 3, sigmod x 3], and then their two respective loss functions are combined together, that is, 6 nodes, 2 loss functions, for small-scale data sets, basically does not work

Images