Warranty Analytics

保修是制造商销售给消费者的产品所附带的合同义务。通常，它规定，如果产品在指定的期限和/或使用级别内出现故障，制造商将通过修复、更换或补偿消费者来补救这种情况。正确处理保修对制造商来说很重要，原因有几个。保修是制造公司最大的成本之一。因此，有效处理保修索赔是一项关键活动，准确预测未来索赔水平和检测索赔模式的新趋势也是一项关键活动(在消费者安全可能因故障而面临风险的情况下尤其重要)。保修索赔数据库可能是制造商拥有的有关其产品在现实世界中的故障频率和原因的唯一数据来源。数据挖掘和文本挖掘技术被用于更好地理解这一点(例如，序列模式挖掘)。保修正越来越多地被用作在市场上提供竞争优势的工具。选择合适的保修政策是一个从制造商的角度考虑成本和利润的优化过程，保修数据分析借鉴了统计学、计算机科学和可靠性工程等领域的知识来解决这些问题。最近的一篇开创性论文(Reliability Meets Big Data：Opportunities and Challenges，Hong&Meeker，2014)指出，随着更多的数据来源(特别是产品传感器捕获的数据)可供分析，可靠性数据分析的性质正在发生变化，并呼吁研究更多地利用这些来源。使用可靠性数据存在许多实际挑战-保修数据不完整且经过正确审查，大部分信息包含在非结构化文本字段中，以及需要考虑的人为因素(未声明的失败产品，声明的未失败产品)。传感器的数据量很大，从孤立的来源合并数据可能很困难。在详细的文献回顾(大约40篇论文)以及与几家公司(卡特彼勒、捷豹路虎、劳斯莱斯)的讨论之后，我确定了我的研究中需要探索的三个领域。进一步自动化保修索赔分析-在行业中，索赔评估通常由保修分析师团队完成，他们手动阅读每个索赔并决定是否付款。由于收到的索赔数量很大，业务规则被用作处理的第一阶段(例如，自动支付某一价值以下的所有索赔)。一些研究着眼于开发文本挖掘工具，以根据技术人员的文本评论对索赔进行分类。其他研究将离群值标记为潜在的欺诈案例。我建议使用有监督的机器学习方法(例如，使用随机森林或神经网络)，对包含保修分析师决策作为标签的历史索赔进行训练。可以执行文本挖掘的一个阶段并将其用作输入特征(例如，识别关键投诉单词)。可以添加强化学习能力，以提高实施后的系统性能。根据产品使用模式提供个性化的保修条款-制造商希望能够在不太可能出现保修索赔的情况下提供更优惠的保修条款。人们可能会将其与保险业相提并论，在保险业，如何付费驾驶计划是一种被接受的定价模式。研究还没有在保修的背景下探索同样的想法。我建议分析传感器数据中的历史使用模式，以了解与故障率的关系，并开发基于风险的定价模型。使用实时传感器数据提高保修索赔预测的准确性-大量研究集中在开发和微调保修预测的统计模型(在生存分析范围内)

A warranty is a contractual obligation attached to a product sold by a manufacturer to a consumer. Typically, it states that if the product fails within a specified period and/or usage level, the manufacturer will remediate the situation, by either repairing, replacing, or recompensing the consumer.Correct handling of warranty is important to manufacturers for several reasons. Warranty is one of the biggest costs to a manufacturing company. Efficient processing of warranty claims is therefore a key activity, as is accurately forecasting future claim levels, and detecting emerging trends in claims patterns (particularly important where consumers safety may be at risk due to faults). Warranty claims databases may be the sole data source a manufacturer has on the frequency and causes of failure of its products in the real world. Data mining and text mining techniques are used to better understand this (e.g. sequential pattern mining). Warranty is increasingly being used as a tool that can provide a competitive edge in the marketplace. Selecting a suitable warranty policy is an optimisation process in which both costs and profits are considered from the manufacturer's perspective.Warranty data analysis draws from the fields of statistics, computer science, and reliability engineering to tackle these issues. A recent seminal paper (Reliability Meets Big Data: Opportunities and Challenges, Hong & Meeker, 2014) observed that the nature of reliability data analysis is changing as more data sources become available for analysis (particularly that captured by a product's sensors), and called for research to make more use of these sources.There are many practical challenges in working with reliability data - warranty data is incomplete and right censored, much of the information is contained in unstructured text fields, and there is a human element to consider (failed products not claimed for, non-failed products that are claimed for). Sensor data is of high volume, and merging data from siloed sources can be difficult.Following a detailed literature review (around 40 papers) and discussions with several companies (Caterpillar, Jaguar Land Rover, Rolls Royce), I have identified three areas to explore in my research. Further automate the analysis of warranty claims - In industry, assessment of claims is generally done by teams of warranty analysts who manually read each claim and decide if payment will be made or not. Due to the volume of claims received, business rules are used as a first stage of processing (e.g. all claims under a certain value are automatically paid). Some research has looked into developing text-mining tools to cluster claims by the technician's text comment. Other research flags outliers as potential cases of fraud. I propose using a supervised machine learning approach, (with for example a random forest or neural network), trained on historic claims that contain a warranty analyst's decision as a label. A stage of text mining could be performed and used as an input feature (e.g. identify key complaint word). Reinforcement learning capabilities could be added to improve system performance following implementation. Provide individualised warranty terms based on product usage patterns -Manufacturers would like to be able to offer more favourable warranty terms where a warranty claim is less likely to arise. Parallels may be drawn with the insurance industry where pay-how-you-drive schemes are an accepted pricing model. Research has yet to explore the same idea in a warranty context. I propose analysing historic usage patterns in sensor data to understand the relationship with failure rates, and developing risk-based pricing models. Improve accuracy of warranty claim forecasting using real-time sensor data - A large amount of research has focused on developing and fine-tuning the statistical models (falling within survival analysis) of warranty forecasting