Research On Multi-sensor Data Classification And Object Detection Method

With the rapid development of the Internet of Things,sensors and other technologies,machine learning has gradually become a key technology for acquiring "hidden knowledge" in sensor data to make it more intelligent.At present,some traditional machine learning technologies have been applied to data collected by sensors in the fields of smart home,environmental monitoring,and factory detection to complete tasks such as classification and object detection,but there are still certain limitations.Therefore,how to improve the accuracy of classification and object detection by applying updated machine learning techniques to the data collected by sensors has important research value.This paper studies the sensor data classification and image-based defect detection in indoor environmental monitoring.The main work are as follows:1.Improved a multiple kernel learning method of sensor data classification in indoor environmental monitoring.For the sensor data with multi-source heterogeneous characteristics,the kernel function in the traditional multiple kernel learning algorithm is mainly based on the recommendations in the literature or experience to select the type and parameters of the kernel function.The improved method in this paper is divided into two steps in determining the type and parameters of the kernel function:firstly,the cross-validation method is used to initially determine a set of optimal basic kernel function types and parameters;secondly,based on the results of the first step,the support vector machine solver is used to train the samples and learn the kernel weight of multiple kernel functions simultaneously to further optimize the combination of kernel functions.The experimental results show that whether it is for the public data set of smart home sensors or the real data set of indoor IoT sensors collected,compared with the average level of single kernel support vector machines and the traditional multiple kernel learning algorithm,the improved multiple kernel learning algorithm achieves higher classification performance and is suitable for data of various scales.2.An image defect detection method of automobile engine parts based on Mask R-CNN is presented.In this method,surface defect datasets of engine parts in PASCAL VOC and MS COCO formats are established for object detection and instance segmentation tasks.Then,in view of the problems of missed detection and low accuracy of Faster R-CNN in the detection of small object surface defects,the idea of transfer learning is used,and only stage-1 training is performed under the weight of the MS COCO pre-training model to improve the training speed and redesign the anchor scales for the small objects.The experimental results show that for the small defects on the surface of engine parts,the optimized Mask R-CNN has improved the training speed and detection accuracy.