Development and Implementation of a Reliable Decision Fusion and Pattern Recognition System for Object Detection and Condition Monitoring
A monitoring task of production system (bucket-wheel excavator) is investigated for the development and realization of a multisensor-based monitoring system. The objective of the monitoring system is to obtain in real time reliable decisions on the presence of target objects (large stones) in the transported material during the production process to avoid disturbances or failures of the transportation process. Due to the complexity of the considered production system, different physical effects are used for the development of the multisensor-based monitoring system. The measured signals are acquired using different sensors (five acceleration sensors, two load cells, and a laser scanner). Due to the inevitable and varying time shift between the stimulations of the individual sensors, each signal is individually subjected to preprocessing, feature extraction, and classification process. The proposed monitoring system consists of three modules: acceleration, laser scanner, and decision fusion modules. For the acceleration module which uses acceleration signals of five different acceleration sensors, two detection approaches are developed. The first approach (STFT-SVM) is based on Short-Time Fourier Transform (STFT) as feature extraction tool, Support Vector Machine (SVM) for the classification, and a novel decision fusion process to fuse the individual decisions. The second approach (CWT-SVM) is based Continuous Wavelet Transform (CWT) as feature extraction tool, Support Vector Machine (SVM) for the classification, and a rule-based decision fusion process to fuse the individual decisions. Both approaches are trained, validated, and tested using real industrial data. The developed approaches show strong improvements in detection and false alarm rates. Due to the implementation complexity and the high number of false alarms of the STFT-SVM approach in comparison to the CWT-SVM approach, the CWT-SVM-based approach is chosen for the development of the overall monitoring system. The Laser scanner module which processes the laser scanner signal consists of prefiltering, filtering, validation, and classification process. The module is validated, and successfully tested on real industrial data. The decision fusion module fuses the decisions of both detection modules in order to obtain a final reliable decision. Three fusion techniques are investigated, which are OR-logic, Bayesian Combination Rule (BCR), and the new developed decision fusion technique Basic Belief Fusion (BBF). Due to the characteristics of the considered application, the OR-Logic is chosen to perform the fusion task. For the online realization, the weightometer module is added to avoid false alarms which could be caused by acceleration module. Additionally modifications and simplification processes are performed in order to overcome the hardware limitations The proposed monitoring approach is developed for online and real time implementation, and it achieves high detection rate, with minimum false alarms rate, thus the production process disturbance is minimized.