Time series data clustering method and system based on dynamic kernel development

Abstract

The invention discloses a time series data clustering method and system based on dynamic nucleus development. The method comprises the steps that S01, an initial core is configured and serves as a starting point of dynamic nucleus division development; s02, obtaining current newly-added time sequence data, stimulating each dynamic core by the current newly-added time sequence data, obtaining corresponding output after each dynamic core responds, selecting the dynamic core with the maximum output as a winning dynamic core, and copying the category of the winning dynamic core to the current newly-added time sequence data; s03, regulating and controlling the splitting opportunity of each dynamic nucleus by using the memory saturation; and S04, clustering the dynamic cores in the updated dynamic core set into different categories according to the centers and the coverage domains of the dynamic cores to obtain clustering results of the current dynamic cores, and returning to the step S02 until the clustering is exited. The method can realize data clustering of dynamic kernel development, and has the advantages of simple implementation method, good robustness, high precision, strong flexibility and the like.

Description

technical field [0001] The invention relates to the technical field of time series data clustering, in particular to a time series data clustering method and system based on dynamic nuclear development. Background technique [0002] With the rapid development of network technology, the sources of data such as video and images are becoming more and more extensive, and it is becoming more and more important to initialize the acquired data information. In terms of security, especially in densely populated areas such as airports and ports, the face information in the surveillance video is more complex and diverse, and most of the face information does not exist in the original database, and most of the information is acquired for the first time , gradually increasing, that is, increasing time series data. In time-series data such as real-time video streams, it is usually necessary to cluster targets in time-series data, such as the clustering of face images in surveillance vide...

AI Technical Summary

Problems solved by technology

However, in practical applications, time-series data is a continuous data stream, and it is impossible to know in advance what kind of samples will be clustered, that is, the category of new data is unknown

[0007] 2) It is necessary to artificially set a priori parameters such as the number of clusters

[0009] 3) The two stages of "learning-application" are independent of each other and cannot be related

Therefore, under the condition of limited learning samples, it is difficult to obtain a reasonable k value that can adapt to the dynamically increasing data through prior knowledge, and the k value set according to the prior cannot be adjusted with the change of data. Once the learning stage is set The mismatch between the fixed k value and the data distribution in the application stage means that the previous training work is all invalid, and the training needs to be restarted, and sometimes the training cost in the early stage is very high

[0011] 4) In the case of a small number of samples, there will be algorithm failure problems

[0012] At present, clustering algorithms cluster samples when a large number of samples are known, but in some clustering tasks, the algorithm can only gradually acquire samples, even when there is only one sample, which makes the current clustering algorithm unable to work normally. Work

[0013] 5) The model structure of the algorithm is fixed, resulting in the inability to self-evolve with the increase of samples, which limits the generalization ability of the application

However, in actual tasks, the category information to be clustered is unknown, and may belong to categories that do not exist in the database. Therefore, clustering algorithms with fixed structures usually cannot give reasonable clustering results when new samples appear.

At the same time, due to the fixed algorithm structure, these algorithms can only solve specific clustering problems. For example, for the face clustering algorithm in surveillance video, if you want to switch to the clustering task for molds on the pipeline, you usually need to modify the algorithm model

[0015] To sum up, in the current clustering method, the sample information changes incrementally with time, but it has nothing to do with the time variable, because it is impossible to predict the samples that will appear in the next moment, and when the algorithm is executed What I had before was only a small amount of local information, and it was impossible to obtain the information of all categories of samples

Therefore, the current clustering method is not suitable for processing a small number of samples without prior knowledge, nor does it have the ability to give reasonable clustering results in real time as the number of samples changes; at the same time, traditional clustering methods are aimed at For incremental clustering problems, usually after adding sample information, all the obtained data will be re-clustered to obtain the clustering results at this time. Although the clustering results at each moment can be given, but To a certain extent, this does not make good use of the knowledge learned before, which will cause a great loss of efficiency

Images