Abnormally detection based on feature similarity measure of image time series in zinc flotation

Abstract

Aiming at the problem that the anomaly detection process of the flotation process is repeated and the abnormal trend can be captured and the number of abnormal samples is inferior, an anomaly detection method based on the similarity measure between the time series features of the foam image is proposed. The time-series bubble image is collected by the real- time acquisition system in the industrial field, and the size feature of the bubble image is extracted. The log-norm function is used to fit the bubble size distribution method to eliminate the left tilt and large peak difference of the bubble histogram shape, the method of time series piecewise linear representation is then used to generate the trend characteristics of the bubble time series. The probability of abnormal conditions is evaluated by the degree of difference from the trend characteristics of the typical anomalous process foam image, and the Markov distance method is used to measure the similarity to eliminate the influence between different dimensions. Anomaly detection is implemented by using similarity matching with features of time series anomaly feature library. The results of industrial operation data test show that the method can effectively perform online anomaly detection, and successfully divide the abnormal trend data mixed in the experimental process. The effectiveness of the method is verified by the foam image test data set in the industrial field.