基于遗传算法的多评价标准退化特征选择方法

A Feature Selection Method for Prognosis Based on Genetic Algorithm and Multiple Metrics

机械设备退化特征作为预测模型的输入,其质量直接影响学习模型性能。正确反映退化进展的特征确保预测结果准确性。良好的退化特征集合应该表现出连续增加或减少的趋势,恶化趋势应该与时间存在相关性,同时对监测噪声、退化过程的随机性和操作环境的变化,具有鲁棒性。基于此论文利用遗传算法启发式搜索的特点提出了遗传算法和评价标准(单调性、趋势性、鲁棒性)结合的自动特征选择方法。该方法以评价标准的线性组合作为遗传算法的适应性函数,可以直接反映特征子集对机械退化过程的相关性,引导遗传算法的变异和搜索。通过实验对比表明该算法具有一定的稳定性和有效性,能够在原始特征空间中选择最优特征子集,从而提高预测准确率。

Degradation feature of mechanical equipment directly affects the performance of learning models.The accuracy of prognosis can be precisely reflected by the feature of deterioration.A good degradation feature set should show a trend of continuous increase or decrease,and the deterioration trend should be correlated with time,and it is robust to monitoring noise,degradation process randomness and operating environment changes.Thus,based on the characteristics of heuristic search in genetic algorithm,an automatic feature selection method combining genetic algorithm with evaluation metrics(monotonic,trend and robustness)is proposed.The method takes the linear combination of metrics as an fitness function,which can reflect the correlation between the feature subset and the degradation proceeding,and guide the mutation and search of genetic algorithm.The experimental proves the stability and effectiveness of algorithm and the proposed method can select the best feature subset in the original feature space,thus improving the accuracy of prognosis.