Reliability Prediction Based on SVR and Improved SCSO

Reliability prediction can provide basis for the identification of potential improvement area, cost control, and mission reliability assessment, etc. However, for complex equipment, there are many reliability influencing factors and incomplete knowledge of failure causes, which lead to a significant disparity between the predicted outcomes and the real values for traditional reliability prediction methods. To address the above issues, this research paper introduces an approach that utilizes the Support Vector Regression (SVR) model and Sand Cat Swarm Optimization (SCSO). To begin with, the sliding window technique is employed on the historical reliability data to generate time series samples, with the 5 adjacent data as sample data, and the sixth as label of the sample, and train SVR model on these samples; Second, the SVR model parameters are optimized using the ISCSO algorithm to obtain the optimal combination of parameters. In the testing stage, firstly, historical reliability data was used to predicted future data by the model, and the predicted data are then added to the sequence to form new samples, thus old data are discarded and new data are predicted continuously to realize continuous reliability prediction. Finally, the algorithm proposed in this paper is validated on a diesel engine reliability dataset. The algorithm proposed in this paper demonstrates its superiority through the Normalized Root Mean Squared Error (NRMSE) evaluation. The NRMSE of SVR-ISCSO is 8.86E-05, showcasing a remarkable 99.24% year-on-year decrease compared to the standard SVR. Additionally, it exhibits a 5.86% year-on-year decrease compared to SVR-SCSO, further validating the effectiveness of the proposed algorithm.