ABSTRACT
For robots to perform complicated tasks, they should be capable of withstanding subsystem malfunctions and natural risks. For autonomous underwater vehicles (AUVs) to be safe in accomplishing activities in complicated situations without human aid, they must be able to identify and fix problems. Due to a lack of safety and accuracy, the existing AUV defect diagnosis procedures are inadequate. This paper presents a machine learning (ML) method to diagnose the unexpected faults in AUVs by proposing a reinforced kernel-based artificial neural network (RK-ANN) with the highest safety. Initially, the state-sensor data are gathered as datasets for this investigation and are normalized in preprocessing stage to eliminate the unwanted data. Using a natural autoregressive-based topic modeling technique, we next illustrate how well these patterns may be combined with operator-supplied semantic labels to identify AUV faults and diagnose them using the RK-ANN classifier. RK-ANN considerably reduces the time-consuming load, simplifies the diagnosis procedure, and improves efficiency when compared to standard model-based diagnosis regarding the safety management in AUVs. A demonstration of RK-ANNs on a small quadrotor AUV named “Haizhe” proved its efficacy in the real world. RK-ANN may be used to face the issues of fault identification and elimination in a single-stage diagnostic manner, and its safety is significantly superior to other ML methods.
