交通运输工程与信息学报

【目的】针对铁路机车乘务员历史能力评价与当前作业风险识别相互割裂，以及个体历史能力差异难以有效参与当前风险判定的问题，提出一种历史能力条件化的静态能力-动态风险耦合方法。【方法】基于任务开始前已完成的认知测评、操作绩效和作业隐患记录，构建认知能力、操作技能和作业隐患三维历史能力特征体系。融合层次分析法（AHP）岗位结构权重与熵权法（EWM）样本差异权重，以研究期前12个月的历史安全绩效为外部效标，通过按乘务员分组的5折交叉验证标定组合系数，形成履职能力综合指数（CAI）。构建CAI引导的多模态时序有序风险网络，将任务前确定的历史能力作为个体条件参与当前多模态风险证据的融合与判定，并通过受高风险召回率和假阳性率约束联合标定的能力敏感度权重与动态阈值，实现个体化风险判定。为防止信息泄漏并保证模型评价的有效性，每项任务的CAI仅由任务开始前的历史记录计算，当前任务数据仅在任务结束后用于更新后续任务的能力基线。【结果】重载铁路模拟驾驶案例中，组合CAI与历史安全绩效的折外Spearman相关系数为0.873，高于单独采用AHP和EWM时的0.801和0.788。特征消融结果表明，输入由30维扩展至41维后，基础模型总体准确率由91.1%提高至92.4%，高风险召回率由87.0%提高至90.3%。采用约束优化标定能力敏感度参数和动态阈值后，总体准确率进一步提高至94.5%，高风险召回率提高至94.8%，高风险假阴性率降至5.2%。【结论】所提方法通过外部安全绩效约束构建可解释的历史能力基线，并将其作为个体条件参与当前多模态风险信息的融合与判定，实现历史能力与动态作业风险的跨时间尺度协同建模。频域和非线性特征增强了对操作波动、自主神经调节及信号复杂度的刻画，数据标定的能力调节机制可在控制过度预警的条件下校准风险判定敏感度。【应用】模型输出可为乘务员能力分级、针对性培训和主动风险控制提供量化依据。

【Objective】To address the disconnect between historical competency assessment and current operational risk identification among locomotive drivers， this study proposes a historical-competency-conditioned method coupling static competency with dynamic risk. 【Methods】A three-dimensional historical competency feature system was developed using cognitive assessments， operational performance， and operational hazard records collected before each task. Job-structure weights derived from the analytic hierarchy process (AHP) were integrated with sample-variation weights obtained using the entropy weight method (EWM). Historical safety performance during the 12 months preceding the study was used as an external criterion. The combination coefficient was calibrated through five-fold cross-validation grouped by driver， producing a comprehensive competence index (CAI). Time-domain， frequency-domain， and nonlinear features of operational behavior， physiological fatigue， and the operating environment were extracted from continuous windows of the current task. These features were incorporated into a CAI-guided multimodal temporal ordinal risk network. Individualized risk classification was achieved using conditional attention and jointly calibrated ability-sensitivity weights and dynamic thresholds. The calibration was constrained by high-risk recall and false-positive rate requirements. To prevent information leakage and ensure valid model evaluation， the CAI for each task was calculated exclusively from records available before task commencement. Data from the current task were used only after completion to update the competency baseline for subsequent tasks. 【Result】In the heavy-haul railway simulator study， the out-of-fold Spearman correlation between the combined CAI and historical safety performance was 0.873. This exceeded the correlations obtained using AHP alone (0.801) and EWM alone (0.788). Feature ablation showed that expanding the input dimensionality from 30 to 41 increased the baseline model’s overall accuracy from 91.1% to 92.4%. High-risk recall concurrently increased from 87.0% to 90.3%. After constrained optimization of the ability-sensitivity parameters and dynamic thresholds， overall accuracy further increased to 94.5%. High-risk recall reached 94.8%， while the high-risk false-negative rate decreased to 5.2%. 【Conclusion】The proposed method establishes an interpretable historical competency baseline constrained by external safety performance and incorporates it as an individual-level condition in multimodal risk fusion and classification. It thereby enables cross-timescale collaborative modeling of historical competency and dynamic operational risk. Frequency-domain and nonlinear features improve the characterization of operational fluctuations， autonomic regulation， and signal complexity. The data-calibrated ability-adjustment mechanism further calibrates risk sensitivity while controlling excessive warnings. Model outputs can provide quantitative support for driver competency classification， targeted training， and proactive risk control.