Classification of non-stationary data stream with application in sensors for insect identification

Abstract

Applications such as intelligent sensors should be able to collect environment information and make decisions based on input data. An example is an under-development low-cost sensor to detect and classify insects in their species using a laser beam and machine learning techniques. This sensor is an important step towards the development of intelligent traps capable of attracting and selectively capturing insect species of interest such as disease vectors or agricultural pests, without affecting the beneficial species. The data gathered by the sensor constitutes a stream with non-stationary characteristics in which the main information is the insect wing beat frequency and it is influenced by environmental conditions like temperature, humidity and atmospheric pressure. The sensor classification algorithm should be capable of identifying concept drifts in absence of labels in the test phase, differently from the current techniques. Furthermore, being an embedded system, the sensor has memory and processing constraints. Thus, the main objective of this work is the classification of non-stationary data streams without the need of immediate availability of true class labels for the last classified instance. This classification should be efficient in terms of memory and processing to be embedded in a sensor.