A soft sensor for the surface cleanliness level of ultra-smooth optical component based on the Genetic-algorithm and the Related Vector Machine

We presents a series of the key technologies to address the disadvantages of optical surface flaws interference and high-fallout rate when detecting the surface cleanliness of ultra-smooth optical components. We designed a movable clamping fixture and 3-D electric control detection platform for multi-size standby detected optical components. We adopted the Genetic-algorithm to detect the convex of the object to acquire those individually sealed areas which approximately show real particulates in the digital image. We subsequently adopted the Related Vector Machine (RVM) method to recognize particulates. The multi-parameters extracted from geometric space, grey-level space and Gabor-field space separately compose a character vector used for model recognition. The Supported Vector Machine (SVM) and RVM methods are compared using the same samples. Experiments show that RVM possesses better accuracy and requires fewer training samples even if dealing with 19-D problems than other conventional model recognition methods. The research also indicates that the soft sensor system model meets the requirements of surface cleanliness level detection in engineering fields employing ultra-smooth optical components.