基于多维加权融合算法的机车走行部旋转部件健康评估系统设计与应用

Application of Health Assessment System for Rotating Components of Locomotive Running Gear at Baoshen Locomotive Depot

摘要: 针对机车走行部旋转部件健康评估中多源异构数据融合困难、传统单一阈值报警漏报率高的问题，本文从计算机科学与应用的角度，设计并实现了一种基于多维加权融合算法的健康评估系统。系统采用B/S分层解耦架构，后端基于Spring Boot框架，数据库采用MySQL与Redis混合存储方案，日均处理振动、冲击、温度等时序数据约5.2万条。通过数据清洗模块对原始数据进行异常剔除、移动平滑去噪及缺失值插补，提取报警等级、dB有效值、SV趋势、温升极值等7维特征；在决策层，设计了一种基于层次分析法(AHP)的加权融合判决器，将多维特征状态值映射为五级健康等级，并结合“健康等级–维修动作”规则库实现运维建议的自动推理。包神机务段应用结果表明：本系统解决了传统方案对早期轴承剥离、齿轮偏磨的漏报问题，在包含87个历史故障案例的测试集上，故障召回率(Recall)由传统振动报警法的33.3%提升至100%，F1分数由0.50提升至1.0；针对踏面故障，首次预警时间较传统方案提前9天；系统平均决策延迟小于1.5秒，CPU占用率稳定在25%以下，为机车智慧运维提供了可靠的计算机辅助决策支持。

Abstract: Aiming at the difficulties in multi-source heterogeneous data fusion and the high missing alarm rate of traditional single-threshold alarms during the health assessment of rotating components in locomotive running gears, this paper designs and implements a health assessment system based on a multi-dimensional weighted fusion algorithm from the perspective of computer science and applications. The system adopts a layered decoupled B/S architecture. The back end is built on the Spring Boot framework, and a hybrid storage scheme consisting of MySQL and Redis is deployed for the database, processing approximately 52,000 pieces of time-series data such as vibration, shock and temperature every day. The data cleaning module eliminates outliers, performs moving smoothing denoising and missing value imputation on raw data, and extracts seven-dimensional features including alarm level, effective dB value, SV trend, temperature rise extremum, etc. At the decision-making layer, a weighted fusion decision maker based on the Analytic Hierarchy Process (AHP) is constructed to map the state values of multi-dimensional features into five health levels. Combined with the rule base of “Health Level - Maintenance Action”, automatic reasoning of maintenance suggestions is realized. Application results at Baoshen Locomotive Depot demonstrate that the proposed system eliminates missing alarms of early bearing spalling and gear eccentric wear existing in conventional schemes. The fault recall rate increases from 33.3% achieved by the traditional vibration alarm method to 100%. For wheel tread faults, the initial early warning time is 9 days earlier than that of traditional schemes. The average decision latency of the system is less than 1.5 seconds, and the CPU utilization remains below 25%. It provides reliable computer-aided decision support for intelligent operation and maintenance of locomotives.