预测性维护（Predictive Maintenance, PdM）通过提前预测设备的故障或剩余寿命，优化维护策略，减少非计划停机时间和停机损失。实践中，对预测性维护的优化不仅需要提高算法能力，还需要考虑人类决策者的特点，从人机协同的角度进行优化，以避免过于依赖或不信任算法导致过度维修或忽略潜在的故障风险。目前预测性维护系统优化有关研究未能考虑到人类决策者的行为倾向与主观认知因素，为了探讨人与预测性维护系统中的寿命预测算法合作时的决策行为特点，本研究开发了预测性维护决策实验平台，并通过两项实验研究探究了不同场景下算法表现、维护成本及其不确定性对预测性维护决策的影响。研究一探讨了任务参数与算法表现对预测性维护决策的影响。研究一具体对比分析了事后维护成本大小、任务中机器的数量、算法告警成功率三种因素对预测性维护决策行为与主观认知指标的影响。研究一的结果表明，在机器数量不同时，算法表现对预测性维护决策行为的影响存在不同的趋势：在机器数量较多时，参试者对算法建议的服从度随着算法告警成功率的提升而提升；而在机器数量较少时，算法告警成功率的提升反而会降低服从度。此外，事后维护成本大小对决策行为和主观认知的影响在不同场景下较为一致：高维护成本促使决策者采取保守的维护策略，并增加对算法建议的服从度。本研究探究了算法性能、系统复杂性和经济成本因素对预测性维护决策的影响，现实中在设计和部署预测性维护系统时，应综合考虑这些因素，以实现最佳的维护策略。基于研究一关于事后维护成本大小对预测性维护决策影响的结论，研究二进一步探讨了事后维护成本的不确定性对预测性维护决策的影响。研究二通过对比不同损失规则下决策者的决策行为特征与主观认知指标，发现成本不确定性对预测性维护的影响较小；只有在极小概率产生巨大损失这类极端条件下，决策者才会采取较为激进的维护策略，表现出对算法建议的高度依赖或忽视。因此，在实际应用中，需警惕维护人员在面对极端损失风险时的决策偏差。本文的两项研究探究了在不同场景下预测性维护决策者行为和主观认知的变化规律。研究结果为未来的预测性维护系统设计和决策支持提供了参考依据。
Predictive Maintenance (PdM) optimizes maintenance strategies by predicting equipment failures or remaining useful life in advance, reducing unplanned downtime and associated losses. In practice, optimization of PdM not only requires enhancing algorithmic capabilities but also considers human decision-maker characteristics, aiming for human-machine collaboration to avoid over-reliance or distrust of algorithms, which could lead to excessive maintenance or overlooking potential failure risks. Currently, PdM system optimization research often overlooks human decision-makers‘ behavioral tendencies and subjective cognitive factors. To explore decision-making characteristics when collaborating with PdM systems‘ lifespan prediction algorithms, this study developed a PdM decision-making experimental platform and conducted two experimental studies to examine the impact of algorithm performance, maintenance costs, and their uncertainties on PdM decisions in different scenarios.Study One investigated the influence of task parameters and algorithm performance on PdM decision-making. It specifically analyzed the effects of post-maintenance cost size, the number of machines involved in the task, and algorithm alarm success rate on PdM decision behavior and subjective cognitive indicators. The results showed different trends in the impact of algorithm performance on PdM decision-making behavior depending on the number of machines: with more machines, participants‘ compliance with algorithm recommendations increased with higher alarm success rates; with fewer machines, higher alarm success rates decreased compliance. Additionally, the impact of post-maintenance cost size on decision behavior and subjective cognition was consistent across scenarios: high maintenance costs prompted decision-makers to adopt conservative maintenance strategies and increased compliance with algorithm recommendations. The study examined the effects of algorithm performance, system complexity, and economic cost factors on PdM decisions, highlighting the need for comprehensive consideration of these factors in the design and deployment of PdM systems for optimal maintenance strategies.Building on the findings of Study One regarding the impact of post-maintenance cost size, Study Two further explored the effect of post-maintenance cost uncertainty on PdM decision-making. By comparing decision-making characteristics and subjective cognitive indicators under different loss rules, it was found that cost uncertainty had minimal impact on PdM; only under extreme conditions with a very low probability of significant losses did decision-makers adopt more aggressive maintenance strategies, showing high dependence on or disregard for algorithmic recommendations. Therefore, in practical applications, it is crucial to be cautious of decision biases among maintenance personnel when faced with extreme loss risks.The two studies in this paper explore the changes in behavior and subjective cognition of PdM decision-makers under different scenarios. The research findings provide a reference for the design and decision support of future PdM systems.
