Kimberly-Clark, BlueScope Pivot AI to Cut Unplanned Mill Downtime

Kimberly-Clark与BlueScope转向AI以减少工厂非计划停机 制造商正在放弃现成软件并构建定制内部模型，警告第三方AI供应商不可持续的年度涨价推高了总拥有成本。 在Woven Capital的Prashant Kumar Bothra主持下，行业领袖Dmitry Belov、Amish Chadha和Amir Pasdar探讨了从被动响应及预防性维护向AI驱动的预测性策略过渡的实践经验。他们深入讨论了将传感器与传统老旧设备集成、平衡操作员信任与仪表板疲劳，以及管理总拥有成本与网络风险等现实运营挑战。专家小组还对比了重型制造与商用机队的投资回报率（ROI）预期，同时评估了自研与采购决策、数据主权以及释放隐藏CMMS数据价值的相关议题。 制造业设备管理正由反应式检修与定期维护转向预测性维护（PdM），通过结合 AI、物联网传感器与实时监测，在设备停机前数周提前发现故障征兆。 (0:00) 盲目加装传感器会增加工业网络安全风险并引发数据饱和，过度翻新老旧工业设备还可能带来额外的停机风险并永久推高总体拥有成本（TCO）。 (20:00) 为避免操作人员产生告警疲劳并确保行动方案明确，预测性维护需依赖人机协同（human-in-the-loop）机制与直观的可视化看板，防止关键预警被埋没。 (22:15) 预测性维护不仅能缓解应急抢修造成的现场压力并预防环保合规罚款，面对大规模设备群时，还可通过出厂时校准的单机声热基线来有效杜绝误报。 (30:00)

Predictive maintenance success in industrial environments depends on pragmatic integration with legacy equipment, actionable alert design, and data ownership, rather than indiscriminately blanketing factory floors with sensors or seeking distant forecasting horizons. Pragmatic Warning Windows and Hidden Financial Drivers (25:00) While predictive maintenance vendors often emphasize multi-week warning capabilities, Amir Pasdar observes that commercial operations rarely require extensive advance notice. In commercial food service, an average lag of 1.5 days occurs before shift workers report broken equipment, making a four-to-five-day warning window more than adequate. Beyond downtime prevention, early alerts mitigate the workplace friction and stress that frontline staff experience when critical tools unexpectedly fail. In heavy industrial environments, predictive monitoring derives its primary financial justification from regulatory risk reduction and contractual compliance. Amish Chadha explains that preventing line breaks—such as web failures on paper drying cylinders that otherwise demand manual recalibration of 150 to 200 set points over several hours—safeguards delivery quotas for major retail clients. Furthermore, predictive sensing catches chemical discrepancies before they trigger costly environmental compliance violations or sustainability penalties. Legacy Asset Constraints and Sensor Overload Risks (20:00) Deploying analytics across existing brownfield facilities requires navigating machinery that predates digital instrumentation. Chadha highlights Kimberly-Clark's Beach Island facility, which operates an 80-year-old analog boiler processing roughly 40,000 liters of water daily. Attempting to comprehensively modernize legacy assets spanning 30- to 50-year economic life cycles causes severe downtime risk and permanently escalates Total Cost of Ownership across ongoing maintenance and capital budgets. Indiscriminately outfitting legacy equipment with hardware also creates technical friction. Chadha warns that each newly installed sensor introduces an additional cybersecurity attack vector and risks flooding operators with confusing, redundant data streams. Rather than coping with data saturation, maintenance teams require a single, dependable operational signal to act decisively when machinery drifts beyond process limits. Human-in-the-Loop Integration and Baseline Signatures (22:15) Predictive algorithms detect anomalous signals beyond human perception, but Dmitry Belov stresses that the decision to halt a line remains tied to equipment criticality and technician judgment. Poor data visualization frequently impairs this collaboration; critical alerts often sit unread dozens of folders deep in dashboards or inboxes, resulting in post-failure recrimination. Presenting prioritized, actionable insights within a shared visual interface ensures operators and reliability engineers make timely intervention calls. Precision is equally essential to prevent alarm fatigue across distributed fleets. Pasdar explains that across 50,000 deployed commercial fryers, false positives would trigger crippling logistical confusion. Henny Penny counteracts this by capturing distinct acoustic and thermal baseline signatures during end-of-line factory testing, enabling diagnostic algorithms to measure an asset's field wear against its own original profile rather than an imprecise fleet-wide average. Commodity Purchases, Core IP, and Data Sovereignty (40:00) The choice between building proprietary tools and buying vendor platforms centers on intellectual property. Chadha argues that standard mechanical behaviors—such as bearing vibration signatures—represent commoditized problems best addressed with commercial off-the-shelf software. Conversely, proprietary operational processes, such as unique chemical balances or wet creping techniques, represent core assets that warrant in-house or dedicated partner development. However, partnering with third-party software vendors carries compounding financial and strategic risks. Chadha reports that enterprise vendors are passing internal AI development costs directly to customers, escalating commercial renewals from an average 30% increase every three years to an annual 30% hike. Simultaneously, Pasdar notes that internal build efforts frequently suffer from high data-science turnover that resets institutional progress, while both panelists caution that relying exclusively on vendor platforms compromises data sovereignty, forcing manufacturers to pay ongoing rent to access their own production history. The panel indicates that predictive maintenance generates viable economic returns only when manufacturers avoid sensor saturation, maintain sovereign ownership of machine telemetry, and focus implementations on high-criticality assets where clear, human-in-the-loop workflows exist to resolve alerts.

在Ai4研讨会上，来自Woven Capital、BlueScope、Kimberly-Clark与Henny Penny的工业领袖指出，预测性维护（PdM）落地的核心瓶颈在于物理工厂的恶劣环境与老旧资产改造风险，而非单纯的算法精度。有效的设备健康策略必须依赖人机协同与闭环执行，并要求企业在区分通用机械与专有工艺的前提下，警惕供应商的成本转嫁并守住数据主权。 实体工厂环境与过度数字化的隐患 (20:00) Woven Capital的Prashant Kumar Bothra指出，资本市场对预测性维护高投资回报率（ROI）的宣传与实体工厂复杂恶劣的真实环境存在显著脱节。Kimberly-Clark全球技术战略负责人Amish Chadha以其最大的Beach Island工厂为例说明，该厂运行着拥有80年历史、日处理约40,000升水的锅炉，主体大部分仍属于模拟系统。由于典型工业资产的经济生命周期长达30至50年，对老旧设备进行颠覆性改造不仅面临不可承受的停机风险，还会永久性推高涵盖维护与资本支出的总拥有成本（TCO）。 Chadha进一步反驳了盲目“全面加装传感器”的设想，并指出两大现实风险：一是每个新增传感器都会在生产网络中引入全新的网络安全攻击入口；二是冗余数据会导致信息过载，工程师真正需要的是单一明确的行动信号，而非50条重复的数据流。 人机协同决策机制与可视化规避疲劳 (22:15) 当预测模型提示存在潜在故障但资深现场技师坚持设备正常运转时，停机决策需在算法与经验之间取得平衡。BlueScope北美运营副总裁Dmitry Belov认为，是否停机必须依据设备的生产关键程度与工厂长期维护战略来判定。目前AI的合理定位是在前端监测超出人类感知能力的微弱异常并提供处置建议，而报警阈值的设定和最终操作评估仍须由人类技术人员把控。 Belov特别强调，数据可视化是设备监测中常被低估的关键环节。预警通知若深埋在层级多达20层的仪表盘目录或未读邮件中，往往会在设备停机后演变为互相推诿责任的“事后诸葛亮”局面。Chadha亦补充指出，操作预警与车间执行必须形成闭环；如果一线操作员在工艺参数越界时不知道具体纠偏动作，现场对整套系统的信任就会迅速瓦解。 重构投资回报：从停机指标到合规价值 (25:00) Henny Penny创新总监Amir Pasdar针对传统ROI认知提出修正，认为预测性维护并不必然需要提前数周发出预警。商用餐饮场景调研显示，由于三班倒员工常推诿报修责任，设备损坏后平均延迟1.5天才会上报；因此提前4至5天的预警窗口已足够安排维修。Pasdar指出，其实际价值往往体现在消除现场员工因设备突发损坏而互相争吵的高压工作状态，保障运营稳定。 在连续化制造领域，提前预警能直接避免灾难性的产线混乱。Chadha介绍，Kimberly-Clark的扬基烘缸（Yankee dryer）一旦发生纸幅断裂，工人必须在没有自动化辅助的情况下，耗费1至2小时手动重新校准150至200个设定点，而计划性维修能确保企业按时向Walmart和Target等零售巨头供货。此外，早期预测还能捕捉漂白剂微量泄漏等隐形隐患，避免巨额环保违规罚款与可持续性审计危机。Belov强调，监测项目必须通过量化美元收益直接服务于订单“准时全量交付”（OTIF）指标，以获得管理层的实质支持。 设备声热基线架构与“暗数据”挖掘 (30:00) Pasdar对比了厂内制造设备与出厂商用设备在技术架构上的差异。工厂内部价值50万美元的激光切割机一旦损坏会致使整线停工，而分布于全球50,000个餐饮网点、单价2万至3万美元的油炸锅若发生频繁误报，则会引发全球范围的运维混乱。为此，Henny Penny在设备下线质检排空炸油并加热元件时，记录每台机器独有的声学与热学特征基线，现场AI模型据此进行个性化偏离对比，而非采用统一的通用均值。 在数据底层建设上，Belov认为当前AI最关键的发展并非单独的边缘传感器，而是整合孤立数据集的基础架构。这包括深入挖掘计算机化维护管理系统（CMMS）历史工单中沉睡的“暗数据”，将过往未被利用的维护记录转化为管理决策所需的完整上下文背景。 自研外采边界、供应商涨价与数据主权 (40:00) 关于预测性维护方案的获取路径，Chadha提出应明确区分商品化问题与核心资产。轴承振动等机械特征在各类设备间高度一致，属于标准商品化问题，企业应直接采购现成方案；而涉及化学配比、湿法起皱技术和纸尿裤吸水率等专有工艺的研发，则必须依赖内部团队或定制化开发。与此同时，Chadha警示企业客户正在承担外部软件厂商整合AI功能的研发转嫁成本，商业软件涨幅已从历史上的每三年30%激增至如今的年均约30%。 Pasdar则提醒，制造业在搭建全自研AI团队时普遍面临技术人员高流失率的风险，核心人才跳槽往往导致项目推倒重来。此外，Chadha严正警告，若将生产和装配线的所有数据直接寄宿于外部供应商平台，企业实质上是在为自身的核心资产“持续支付租金”，制造企业必须坚决捍卫底层数据主权。 综上所述，工业预测性维护正由早期的全盘数字化设想转向务实落地阶段，企业需在控制总拥有成本、挖掘存量数据与维持一线可操作性之间求取平衡。然而，老旧模拟设备与现代数字架构的深度融合难题、高素质AI人才在传统制造领域的流失率，以及商业软件供应商高额的AI成本转嫁，仍是各行业全面普及预测性维护所面临的结构性挑战。

## Background 本内容由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布，记录了一场关于制造业 AI 与设备健康管理的研讨会。根据页面介绍，会议由 Woven Capital、金佰利（Kimberly-Clark）、博思格钢铁（BlueScope）及 Henny Penny 的多位主管专家共同探讨如何通过机器学习、物联网传感器与实时监控预判设备故障、降低计划外停机成本并实现 ROI 落地。 ## TL;DR 研讨会深入剖析了预测性维护在重工业、快消品制造及商用餐饮设备等不同场景下的实际落地路径。专家们指出，成功的 AI 预测性维护并非盲目铺设传感器或追求复杂的算法，而是需要从‘单点故障’和关键业务风险出发，平衡技术成本（TCO）与实际收益。同时，克服跨厂区推广中的数据碎片化、系统集成失控以及建立‘人环回路’（Human-in-the-loop）的操作员信任，是推动预测性维护从试点走向大规模量产的核心关键。 ## Atmosphere & style 整场研讨会呈现出一种极具实战色彩、务实且理性扎实的工业研讨氛围。主持人的提问紧扣行业痛点，促使嘉宾抛开高大上的宣讲 PPT，主动剖析因过度追求技术而失败的试点教训与真实的停机损失。嘉宾之间的互动默契而专业，既有对高尖端 AI 落地难点的深刻反思，又有对一线机械师与操作员经验的极高敬意，展现出工业界在拥抱智能化转型时既充满前瞻性又审慎严谨的务实风范。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=0s) 工业设备维护正在从传统的“事后维修”与“定时检修”向基于 AI 和物联网传感器的“预测性维护”演进，能够在发生严重故障的前数周发出预警。 - [05:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=300s) 关键生产线的计划外停机成本极高（例如金佰利部分生产线停机成本高达每小时 25 万美元），将紧急抢修转化为计划内停机能带来巨大经济价值。 - [05:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=300s) 传统 PLC 和 MES 的单节点自动化改造费用昂贵（约 12 万美元），而现代无线 IIoT 方案不仅部署极快（耗时可缩短至 10 小时），还能大幅降低 CapEx。 - [08:45](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=525s) 过度追求高精尖技术而忽视投入产出比（如为了低成本水泵监控而引入高昂实验室级设备的过度工程）是导致 AI 试点项目失败的核心原因之一。 - [08:45](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=525s) AI 预测模型在跨厂区推广时面临数据碎片化、系统集成爆炸以及操作员信任危机（如误报导致系统被人工关闭）三大核心障碍。 - [14:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=840s) 面对老旧设备，企业应采取分级改造与“爬、走、跑”的渐进式架构，优先挖掘设备自带错误码和历史维修工单，再针对“单点故障”节点外挂传感器。 - [20:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1200s) 盲目铺设传感器会导致无用数据爆炸和更高的网络安全风险，关键在于精准获取关键设备的单一有效信号，而非追求全盘数字化。 - [20:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1200s) AI 在生产现场的最佳定位是辅助决策（Human-in-the-loop），且告警必须具备直观的可视化展示与明确的操作上下文，否则无法在一线建立信任。 - [25:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1500s) 预测性维护的隐性价值体现在降低一线员工心理压力、减少停机重置参数的混乱，以及保障零售交付配额以避免违约赔偿与二次物流成本。 - [30:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1800s) 预测性维护还与环境合规及可持续审计紧密绑定（如及时发现化学品微漏），并通过量化“准时全量交付”（OTIF）指标来证明系统的 ROI。 - [30:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1800s) 分散在现场的大量终端设备（如餐饮设备）若出现假阳性或假阴性告警，会引发大规模的全球运维混乱，需要利用出厂物理签名基线进行精准比对。 - [35:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=2100s) 随着现代化 AI 工具降低开发门槛，预测性维护的核心价值正在从单纯采集传感器数据，转向整合 CMMS 历史工单等深层数据以构建一体化决策环境。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=0s) 预测性维护的三个世界与小组研讨开场 在由 Woven Capital 投资人 Prashant Kumar Bothra 主持的研讨会开场中，他开门见山地总结了制造业设备维护的三种基本模式：**事后维修**（坏了再修）、**定时检修**（不论好坏按计划修），以及**预测性维护**（提前预知并规避故障）。如今，AI、物联网传感器和全天候实时监控技术的爆发，使得设备能够在发生致命故障的前数周就发出预警。然而，预测性维护方案在概念宣讲、投资回报率（ROI）承诺与工厂实际落地的接轨过程中，往往呈现出极大的行业差异。 为了探讨预测性维护在不同制造业场景下的共通性与特殊性，研讨会邀请了三位来自不同工业领域的重磅嘉宾进行自我介绍与年度重点工作分享： * **Dmitry Belov**（博思格钢铁公司/BlueScope Operations Vice President）：负责博思格北美涂层产品业务的运营、维护、可靠性及资本支出（CapEx）。他强调，计划外停机对钢铁这种重资产行业极其昂贵，从被动响应转向基于现代化工具的预测性维护是维持资产高效运转的核心。 * **Amish Chadha**（金佰利公司/Kimberly-Clark 全球技术战略负责人）：负责全公司跨业务部门和区域的现代化技术演进。他认为，传感器和机器学习本身并不新鲜，关键在于如何通过这些技术反观并塑造车间的运营卓越文化与员工行为，从而推动预测性维护从试点走向大规模落地。 * **Amir Pasdar**（Henny Penny 创新总监及智能连接解决方案中心负责人）：Henny Penny 是全球顶尖的商用餐饮设备制造商，客户涵盖麦当劳、肯德基（Yum! Brands）、Chick-fil-A 及汉堡王等连锁巨头。此外，他还在北美食品设备制造商协会（NAFEM）主持 AI、网络安全与物联网连接工作。 ``` [ 制造业设备维护三阶段 ] │ ┌───────────────────────┼───────────────────────┐ ▼ ▼ ▼ 1. 事后维修 2. 定时检修 3. 预测性维护 (Break-Fix) (Scheduled PM) (Predictive) │ │ │ 坏了才去修 定期拆换损耗 实时监控/AI预警 (停机成本极高) (可能浪费工时/资源) (提前数周告警) ``` - **Woven Capital**：丰田（Toyota）旗下的成长型投资机构，总部位于东京，全球化布局流动性、供应链、物流、机器人、自动化及工业 AI 领域。 - **博思格（BlueScope）**：全球钢铁巨头；**金佰利（Kimberly-Clark）**：全球消费品与纸制品巨头；**Henny Penny**：商用餐饮设备龙头。 --- ### [05:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=300s) AI 与传感器技术在生产线上的实际落地场景 讨论进入实战阶段，主持人 Prashant 要求嘉宾抛开宣讲 PPT，分享目前在生产线上**实际运行**的 AI 预测性维护案例。 金佰利的 Amish Chadha 透露，公司已在阀门、转鼓、复卷机及轴承等重型关键设备上大规模部署了振动与电机电流监测模型。在造纸与卫生用品生产线上，最致命的故障特征之一就是“纸幅断裂”（Web Break）。过去，一个轴承在凌晨 2 点突发故障会导致整条生产线瘫痪，**每小时带来高达 25 万美元（$250,000/hr）的停机损失**。如今，通过传感器提前捕获轴承异常振动、能量损耗和微弱噪音，工厂能够将紧急抢修转化为计划内停机。但 Amish 同时强调，机器告警绝不能替代人，“人环回路”（Human-in-the-loop）的操作员经验依旧是生产流程中不可或缺的一环。 博思格的 Dmitry Belov 则从老旧资产改造的角度分享了他们的实践。钢铁行业资产极其沉重，传统方式下，如果要给一台老旧电机安装现代化 PLC（可编程逻辑控制器）并接入 MES 系统，**单个节点的自动化改造报价往往高达 12 万美元**，这让管理者望而却步。而如今的工业物联网（IIoT）技术实现了“近乎零 CapEx”的突破：无需大规模硬绑定接线，只需安装轻量级无线传感器与 Wi-Fi 传输模块，最快 10 小时内即可完成部署，仅需极低的运营支出（OpEx）就能让老旧资产具备实时状态监控能力。 - **停机损失**：金佰利关键生产线停机成本高达每小时 **250,000 美元**。 - **改造对比**：传统 PLC/MES 改造单节点成本约 **120,000 美元**；现代无线 IIoT 部署耗时仅 **10 小时**，大幅降本增效。 --- ### [08:45](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=525s) 失败的 AI 试点项目与过度工程的教训 绝大多数制造企业都经历过“试点很完美、推广就夭折”的 AI 项目。Henny Penny 的 Amir Pasdar 结合自身经历（在其此前的工作中），分享了一个典型的“因沉迷技术而失败”的工程教训。 当时，拥有智能传感器博士背景的 Amir 试图解决水泵“干运转”（Dry Run Pump）的损坏问题。他带领团队精心设计了高精度的硬件，通过测量电压与电流之间的相位差（Phase Shift）和阻抗变化来预测干运转。尽管技术上成功实现了监测，但在落地时却撞了墙：测量所需的 **12 度微小相位差** 必须依赖实验室级别的昂贵设备，而水泵本身的更换成本和停机损失其实并不高。**“我们太沉溺于技术本身了，”** Amir 总结道。因为高昂的技术部署成本远远超过了设备本身的价值与风险折算，该项目在耗时 6 个月后不得不宣告终止。 金佰利的 Amish Chadha 补充了一个关于跨厂区复刻（Cookie-Cutter Deployment）失败的案例。公司曾通过并购引入了一套新的生产资产，原本希望将某个厂区成功运行的“化学废料减排 AI 模型”直接复制过去，结果却大败而归。他总结了 AI 跨厂推广的三大核心障碍： 1. **碎片化的数据（Fragmented Data）**：不同厂区的数据结构与质量天差地别。 2. **失控的系统集成（Integration Explosion）**：即便表面上使用同一种 MES 软件，底层硬件接线和异构报表系统的差异会导致集成复杂度呈指数级爆炸。 3. **操作员信任缺失（Operator Trust Failure）**：如果 AI 系统“狼来了”虚报两次，机械师（Millwright）就会直接手动关闭系统。AI 的控制覆写权（Override Right）必须靠长期准确的表现去“争取”，而不是强行赋予。 ``` [ AI 跨厂推广面临的三大壁垒 ] │ ┌───────────────────────┼───────────────────────┐ ▼ ▼ ▼ 数据碎片化 系统集成爆炸 操作员信任危机 (Fragmented Data) (Integration Explosion) (Operator Buy-In) │ │ │ 异构数据格式 底层硬件接线差异 “狼来了”误报两次 无法通用训练 造成接口费用失控 机械师直接关闭 AI ``` - **失败项目教训**：为了监测低成本水泵而捕捉 **12° 相位差**，导致需要实验室级硬件，ROI 严重失衡。 - **金佰利全球规模**：全球拥有 **96 家工厂（Mills）**，无一完全标准化，极其依赖具备数十年经验的本土操作员（如马来西亚厂区的 Yankee 烘缸操作员）。 --- ### [14:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=840s) 传统老旧设备的改造策略与渐进式架构 在面对生产线上的大量老旧设备（Legacy Equipment）时，企业往往面临两难选择：是在现有控制信号上“打补丁”，还是外挂全新的传感器？ 博思格的 Dmitry Belov 提出了**“双轨并行”策略**：一方面保持短期的运营服务，另一方面从资产全生命周期角度审视，将设备分为“值得维护”、“值得升级”和“必须彻底淘汰”三类。对于使用期长达 30 年的钢铁设备，只要其机械结构完好、能安全生产达标产品，就没有必要进行数百万美元的整体重构。对于低价值或非关键节点设备，最明智的策略甚至就是**“备件就地堆放，坏了直接换新”**，这也是一种合理的维护策略。 Henny Penny 的 Amir Pasdar 则阐述了**“爬、走、跑”（Crawl, Walk, Run）**的渐进式架构理念，强调在加装传感器之前，应优先挖掘现有数据的价值： * **Level 1 数据（爬）**：直接提取设备自带的故障错误码（Error Codes）。 * **Level 2 数据（走）**：结合全球售后维修服务记录（Service Club Notes），分析维修工单与更换备件历史。 * **Level 3 数据（跑）**：在关键风险节点上，针对性地外挂高级传感器。 Amir 借用航空业的冗余设计原理进行了生动对比：飞机对飞控系统和操纵杆都有备份，但起落架却没有备份，因此起落架是必须重点监控的关键节点。同样，在麦当劳的厨房里，一台炸锅可能拥有 4 个油槽（Well），坏掉一个并不影响餐厅营业；但整个系统可能只有 **1 个主油泵（Pump）**，一旦油泵故障，整家餐厅就无法供应薯条。因此，预测性维护资源必须精准倾斜给这类“单点故障”（Single Point of Failure）节点，而不是盲目为所有部件贴传感器。 Amish Chadha 在本段结尾补充道，技术团队在设计方案时切忌脱离实际执行者：“如果在工厂线缆上只采集到了传感器的原始信号，那没有任何意义——信号必须附带对执行者有指导价值的业务上下文（Context）。” ### [20:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1200s) 明确行动逻辑、TCO 权衡与决策归属 在工厂实际运营中，若设备因超越统计过程控制限制（statistical process limit）而发出警报，操作员必须清晰了解后置的逻辑纠偏行动；如果缺乏明确指令，系统便无法建立信任。Kimberly-Clark（金佰利）位于南卡罗来纳州 Beach Island 的工厂（为其全球最大工厂）即是一个典型案例：厂内一台每日煮沸约 40,000 升水、拥有 80 年历史的巨型锅炉至今仍保留大量模拟（analog）部件。若盲目追逐全盘数字化而在所有设备上铺设传感器，不仅会带来极高昂的网络安全风险（cybersecurity vector），还会导致无用数据的爆炸。企业真正需要的是针对特定问题的单一精准信号，而非 50 个冗余数据源；此外，传统工厂通常难以承受停机重构（downtime）带来的巨额成本。 对于“当 AI 预测故障建议停机，而经验丰富的操作员坚持设备无碍时由谁决策”的问题，专家指出这取决于具体情境。数据的采集必须优先服务于关键设备维护策略、备件管理与停机成本。AI 的最佳定位是辅助人类（human-in-the-loop）：在前端由 AI 监测异常并给出可视化建议，再由人机协同定义报警参数。警报通知的呈现方式至关重要，如果信息被深埋于多层文件夹或无人查看的仪表盘中，事后追责便毫无意义。 - 金佰利 Beach Island 工厂配有运行 80 年、日煮沸 40,000 升水的锅炉，保留模拟部件以平衡网络安全与 TCO 成本。 - 盲目铺设传感器会增加网络安全风险并造成数据过载，重点在于获取关键信号而非冗余数据。 - 预测性维护决策并非单一角色（如 Amir）独断，而是依靠终端口操作员、维护人员与可靠性工程师（reliability engineer）共同构筑的环境。 - 告警可视化（Visualization）是极易被忽视但至关重要的环节，能避免提示沉没在邮件或层层目录中。 --- ### [25:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1500s) 预测性维护的 ROI 隐性价值与多维衡量 预测性维护不仅关乎“修好设备”，更核心的是“把握时机”。以 Henny Penny（汉尼潘尼）在餐饮场景的调研为例，餐厅设备发生故障后，由于晚班青少年员工往往选择推迟上报，平均需要 1.5 天才会有人拨打电话报修。考虑到技术人员调度与备件到货时间，设备实际并不需要长达两周的预警期，4 到 5 天的提前量便已足够。不同部件损坏的影响截然不同：加热元件故障可暂时使用备用炸缸，无需即刻派人；但若泵（pump）损坏导致无法炸制薯条，餐厅就不得不紧急停业并在深夜高价召修。客户反馈显示，预测性维护带来的重大隐性价值在于降低餐厅员工的心理压力，维护和谐的工作环境。 在金佰利的纸业制造中（如 Yankee 烘缸等关键设备），计划内停机能显著降低生产线调整的混乱度。重新上线一台设备可能涉及人工手动重置 150 至 200 个设定点（set point），预先计划停机可让管理层根据月度或季度指标自主选择最佳维修时机。此外，预测性维护还能防止违反零售商（如 Walmart 或 Target）的货架上架配额与时限合同，避免产品滞销收回仓库带来的二次物流成本。 - 餐饮设备故障后平均需 1.5 天才有人电话报修，4-5 天的预测预警期即可满足维修调度需求。 - 设备的非显性故障会引发连锁反应：餐厅泵损坏会导致停止供应薯条并触发高成本的夜间紧急维修。 - 预测性维护能大幅降低一线员工的压力（stress level），创造平稳的工作环境。 - 大型纸机停机重构可能需要人工调整 150 至 200 个设定点，计划性停机有助于保障生产配额并避免违约采购风险。 --- ### [30:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=1800s) 环境合规、状态监测与工厂及现场设备的数据互通 预测性维护的隐性成本还体现在财务与合规层面。例如，传感器监测到漂白剂泄漏渗入生产线时，防范的不仅是生产停工，更是后续昂贵的环境排放报告、可持续性合规审计与调查成本。作为纸浆与化学加工企业，金佰利须对婴儿纸尿裤中使用的特定纤维树种进行追踪与合规申报，数据闭环在其中起到关键支撑作用。整体而言，状态监测（condition monitoring）必须能自行产生经济回报，以客户“准时全量交付”（On Time In Full, OTIF）为核心驱动力，并通过在初期建立价值流（value stream）量化投资回报率（ROI），从而在企业内部赢取更多支持。 在 Henny Penny 业务中，工厂内部的制造设备与分布在现场的客户设备在预测模式上存在共通性，但规模与影响截然不同。工厂内部一台半百万美元的激光切割机宕机会直接导致整条产线停摆；而分布在现场的 20,000 至 30,000 美元的小型设备，虽然单台价值较低，但覆盖了多达 50,000 个门店终端。 ``` [工厂端设备] ----> 激光切割机宕机 ----> 产线停摆/直接产能损失 [现场端设备] ----> 50,000+ 餐厅终端 ----> 误报/漏报引发全球性运维混乱 ``` - 漂白剂微渗等隐性故障的监测直接关联环保合规（environmental drainage report）与可持续审计成本。 - 状态监测系统必须通过量化 ROI 证明自身的经济效益，支持“准时全量交付”（OTIF）目标。 - 现场设备规模可达 50,000 个地点，模型若出现假阳性（false positive）或假阴性（false negative）将引发全球范围的运维混乱。 --- ### [35:00](https://www.youtube.com/watch?v=8otdaJp1_Ts&t=2100s) 基线采集、数据闭环与自研/外购 AI 架构决策 针对不同机器的差异，Henny Penny 在生产线末端设置了质量检测环节，在设备首次排油、加热时采集其独特的“物理签名”（signature）。当现场设备运行异常时，系统不仅对比全局平均签名，更对比该机器出厂时的固有基线（baseline）。对于快消品行业，即使纸尿裤是一次性消费品，吸收性异常等客户反馈数据也会形成闭环倒逼研发（R&D）部门排查，分析究竟是 SKU 设计缺陷、生产线校准偏差还是数据遗漏。 ``` [出厂末端质检] ──> 采集设备初始物理签名 (基线 Baseline) │ ▼ [现场实时监控] ──> 异常对比: 固有基线 + 全局平均签名 ``` 在面对自研（In-house build）与购买现成 AI 平台（Turnkey platform）的选择时，成本与内部团队能力是核心考量因素。过去企业往往需要依赖外包团队，耗费大量沟通成本解释业务流程且成功率有限；而如今 AI 工具的普及极大地提升了内部团队的开发效率。最令人兴奋的技术突破并非单纯来自传感器本身，而在于将过去隐藏在 CMMS（计算机化维护管理系统）工单、历史记录等各处的数据打通，构建一体化的数据架构与决策环境，帮助管理层做出最佳决策。 - 设备在出厂前的最后一道质检环节采集排油与加热的固有基线，用于现场异常对比。 - 一次性消费品（如纸尿裤）的终端质量反馈同样会构建闭环，反哺研发与生产线校准。 - 随着 AI 工具门槛降低，企业内部团队自主开发能力的提升正逐步替代传统的高成本外包模式。 - 预测性维护的核心价值正从单纯的传感器数据采集，转向融合 CMMS 历史工单等深层隐藏数据的高效决策环境。 Sorry, something went wrong. Please try your request again. ## Video description Prashant Kumar Bothra — Principal, Woven Capital Amish Chadha — Global Technology Strategy Lead - Office of the CDTO, Kimberly-Clark Dmitry Belov — Vice President, Operations, Coated Products North America, BlueScope Amir Pasdar — Director of Innovation - Center of Intelligent Connected Solutions, Henny Penny Corp Unplanned downtime is one of the costliest problems in manufacturing, and AI is rapidly transforming how plants manage equipment health. This panel will explore how leading manufacturers are using machine learning, IoT sensor data, and real-time monitoring to predict failures before they happen and extend asset life cycles. Panelists will discuss how to integrate predictive maintenance solutions with existing systems, balance accuracy with cost, and build trust with operators and engineers on the factory floor. The conversation will also highlight the vendor landscape—from turnkey platforms to custom AI deployments—and the ROI frameworks that prove value to leadership. Attendees will walk away with a clear understanding of how AI-driven maintenance strategies reduce downtime, improve safety, and unlock resilience across operations. [Source](https://www.youtube.com/watch?v=8otdaJp1_Ts)

[0:00] Prashant Kumar Bothra: The real shift in recent years stems from continuous monitoring, IoT sensor data, and machine learning models that can alert engineers to an impending breakdown weeks before an asset halts. [0:00] Amish Chadha: While sensor technology and machine learning are not strictly new, their true value lies in revealing and improving the operational excellence culture and behaviors on the factory floor. [20:00] Amish Chadha: Every added sensor creates a fresh cybersecurity attack vector. [20:00] Amish Chadha: Data saturation creates confusion; engineers need one clear signal rather than 50 redundant streams. [22:15] Dmitry Belov: Oversaturating operators with unprioritized data leads to burnout, making human-AI collaboration essential. [22:15] Dmitry Belov: Visualization as one of the most underrated yet critical elements of condition monitoring. [25:00] Amir Pasdar: Predictive alerts alleviate high-stress environments where restaurant staff end up fighting over broken machines, preserving workplace calm. [30:00] Dmitry Belov: Condition monitoring must strictly pay for itself by supporting the core objective of On-Time In-Full (OTIF) customer delivery. [35:30] Dmitry Belov: The most exciting frontier in AI is not individual sensors, but the overarching infrastructure that aggregates isolated datasets. [40:00] Amish Chadha: Building today requires specialized skills that are not fungible with legacy technical skill sets. [40:00] Amish Chadha: Housing production and assembly line data exclusively within a vendor's platform means 'paying rent on your asset' rather than truly owning your data.

[0:00] Good morning and welcome. I'm Prashan Bhotra, an investor at Woven Capital, and I'm based in Tokyo. We are Toyota's growth stage investment arm and we back companies globally across mobility, supply chain, logistics, robotics and automation, and industrial AI. Just a quick word on the topic before we start. Every plant maintains its machines in one of three ways. You fix things when things break, you fix them on schedule, whether they need it or not, or you try to know the problems in advance. That third one is predictive maintenance and it's what we're here to talk about. What's changed is AI, sensor data, machine intelligence, and monitoring that runs all the time. Ideally, a machine can tell you it's in trouble weeks before it stops. From where we sit at Wilming Capital, we see this space from a lot of angles. the pitch, the claimed ROI, and what survives contact with an actual factory or a plant. And the striking thing is how different it looks from business to business. Same words, completely different problem underneath. And so we have three different manufacturing worlds up here today. A steel maker, consumer products maker, and a kitchen equipment builder. So what I want to get out of the next 45 minutes is what actually carries across the businesses like these and what doesn't. So now I'll hand it over to the panelists. 60 seconds each. Your name, what you own, and what's top of your list this year. You're first. Um, Thank you. up. So my name is Dimitri Belov. I am responsible for the operations in one of our West the part of the United States business, so, uh, coated products of North America. And, um, Under my umbrella is maintenance, reliability, engineering, CapEx, and several other functions. heavily related to the operations. Yeah, predictive maintenance, one of the probably most exciting topics. Thank you. It's very costly when something breaks during the scheduled time. And moving from being reactive to a predictive and using some of the modern tools. to help us to get there. So I think this is probably one of the... the best things we could do to sustain our assets and get as effective as possible as manufacturers. Thank you. Amish Chada, I lead technology strategy at Kimberly-Clark. It's a global remit. I'm responsible for how we adapt to and adopt emerging technology, how we look at the modernization agenda that we need cross-functionally across all of our different business functions and regions. Secondly, what Dimitri said there, you know, the advent of using sensor technology and using AI and machine learning isn't necessarily new or novel. But I think what is important is how it shines a light on the culture that you actually have and the behaviors you actually have on the shop floor from an operational excellence lens. we move from the predictive in pilots onto the shop floor with greater frequency and greater confidence based on the data that we have and what we're working with. Thank you. Thanks, Lord. So, Amir Pastor, I'm the Director of Innovation at Hennypenny and Head of the Center of the Intelligent Connected Solution. We are providing end-to-end digital experience to our customers internally, externally. Hennie Penny, I'm just quick noting that it's a global manufacturer for commercial industrial food equipment for big brands, McDonald's, Chick-fil-A, Burger King, Young Brand. And chances are, if you haven't eaten in those restaurants, your food was prepared in one of our equipment. Also, beside Henny Penny, I am serving as head of the AI, cybersecurity and connectivity for NAFM Association, North America food equipment manufacturers, which most of our members are global. Basically, it's beyond just North America. Great. Thanks, Samir. [5:00] Amish first, can you tell us what's actually running today? rather than what's on the roadmap. Pick one place where you've put AI into a piece of equipment and can you walk us through it? Yeah, absolutely. You know, we've got vibrator and current motor models running on... a good chunk of our assets, particularly heavy equipment, your valves, your drums, your repeaters, your bearings. Each one of them is already giving off signals that are free. Can you hear me now? Perfect. Thank you. We've already got sensors that are giving us particular signatures on bearings, valves, drum repeaters, and so on and so forth. or weeping bleach into the process. Our biggest failure signature is a web break and deploying these sensors and using AI models, we've been able to move from a 2 a.m. emergency failure on the line, which can cost us a quarter million dollars an hour to a planned downtime. inherently with all sorts of other signatures that a bearing is going to fail. You start seeing the energy waste as you see the energy loss and then you hear the hum. But none of that takes away from the human on the loop of the operator on the floor that is critical to the process. Thanks, Amish. And then, Dimitri, what about BlueScope? Yeah, I will probably rephrase that a little bit from my lenses, at least. It's not necessarily what's running under the AI, but it's how do we enhancing the conditional monitoring with the modern tools? Because the concept itself is not novel by any means. How do we... bring all the technologies that's been around for years and make it more accessible for various different assets. I think one of the most exciting things that we have going on is that we've been able to deploy the conditional monitoring to the assets that in the past would require a very high level. capex to be modernized or wired to on the what I would call a level one or the MES basis. So I'm sure here's people coming from somewhat of a manufacturing world, hopefully, so that manage the other assets or different assets. And it's not a surprise that you're coming to a motor with a is how much it's going to cost me to put a modern PLC on it. And then the quote comes up to be $120,000, and you look at him or her and say, "Thank you very much." Right? So, and then those assets are everywhere. It's very, very hard to deploy Um, Something that would be more than... and cost-savvy. I think this current innovations and the technologies that's available to the world allowed us to do that very practically. No CAPEX almost, or zero CAPEX, just the OPEX side of it, and really reap the benefits from the adoption on the floor and in predictive maintenance. So that's what's running. Thanks. We'll get into legacy equipment shortly. Almost everyone has a AI pilot that looked great but never scaled, right? You know, maybe it worked in one plant or on one line but never really made it further. Can you tell us about one of yours and what was the first sign it wasn't going to work? Sure. As we are adding a predictive maintenance to equipment, I can give an example, not from my time at Henny Penny, but... that is The example from detecting the dry run pump actually in different industries. And the issue came because of being in love with the technology. That's the issue usually engineers have. And they can say because I... I have been on both sides of the aisle. I got my PhD in a smart sensor, connected smart sensor, and then I moved to product and basically AI. And my PhD thesis was about basically detecting the failure with impedance measurement and phase shift between voltage and current. And then in the next job, [10:00] I had basically, I said, okay, dry run pump. We can detect it with impedance. And we said, let's go. And I was able to detect it, actually. Our team designed a hardware. The problem was, what is the cost of the detection? And what is the ROI at the end of the day? Because we found out there is a 12 degree, actually, phase shift between these two signals. And you need a kind of lab-grade equipment to do that. And at the end of the day, that pump was not that expensive. And the downtime expense was not that high. So that was a kind of failed project after six months for us. So always basically that's the kind of trade-off between you understanding the... risk, impact, and also what you need to do from the technology standpoint to solve that issue. Thanks. Amish? We had an example, I think a good example of this is we acquired assets through an acquisition some years ago. And looking at a chemical waste reduction type use case, we realized, and we'd done it at one of our plants somewhere, and we thought, oh, this is just cookie cutter, we should be able to take the same thing and deploy it on this new asset, and we should be able to see the same results and we'll claim victory. It doesn't work like that. And one of the things that you learn very, very quickly is that the patterns that repeat, not the code. No single code ever truly is 100% replicable. But the failure modes, they rhyme across the board. And the three things that I've always seen is you've got fragmented data that's one wall, then you've got integrations that explode, because you're not fully well-versed on how everything is wired, no matter how much you may look at it. You could say, I'm using the same big, you know, black box beast of an MES system or surrounding tools, but they're not everywhere with other systems for different reporting needs. And then the third element is the operator needs to be able to buy into it. If the operator is not happy, If you decry wolf twice and the millwright shuts it off, saying, "I don't trust this system." and that's the worst thing that can happen for you. So you need to earn the right to buy into an AI override and, you know, once -- you've got the point where the AI was right and the AI was wrong, then you weigh the law of averages to say, "Okay, you know what? We've now gotten to a point and a mode where we're comfortable that AI override isn't a good place to do what it has to do. How long does it typically take for operator trust to kind of kick in? You're doing this across maybe a POC, maybe one factory or one plant, but... I don't think that there's a... set timeframe that you can go and say that, you know, in six months we need to either know or it doesn't. No one asset is the same. I'll give you context. Kimberly Clark. 96 mils around the world. No standard manufacturing. No one plant, mill or distribution warehouse is the same. We have equipment provided by a range of vendors that supply to almost all of these mills, but they don't supply the same equipment everywhere, right? And you don't have the same operator history everywhere. So if, you know, when you have an operator coming on as a machinist and they're doing their tour of duty and they're learning about a plant, they're assigned to one asset, they're in love with that asset. That asset is their child for a year or two years, maybe three years. 30 years down the line, they understand those bells and whistles better than uh, to go to a mill and to look at the Yankee in Malaysia, he's not going to have the same experience and trust as somebody that was in the mill in Malaysia. for the same amount of time. So, you know, we talked, you heard the talk this morning about a lot of the doomsday predictions and scenarios of, you know, AI is going to take jobs away and so on and so forth. And manufacturing is one of those areas on the line where a lot of people have voiced a lot of concern. You don't see that happening. Thanks. Dimitri and then Amir, this one's for you next. You touched on legacy equipment. Most manufacturing floors predate all of this AI solutions and nobody really wants a multimillion dollar overhaul. So there's usually a fork. Do you bolt new sensors onto the machine or work with what the controls already produce? You know, for example, motor current cycle Which way do you go? How do you decide? What is that choice, whatever you may teach you? Thank you. Well, I think that's a... a short-term approach and it's a long-term approach and i think both of them need to go in parallel so you can't just stop and do a strategy while you need to serve your customers so that's why the shorter term approach which has to be in place. But strategically, you've got to look at every single asset in the lenses of, is it something that is worth maintaining? Is it something that is worth [15:00] something that's worth upgrading, or is it something that is fundamentally needs to be changed? to be able to be a sustainable asset for the future. And I know it's a little generic maybe, but Forcing that conversation helps to deploy the right technology at the right pace, at the right place. So therefore, you're not getting oversaturated. Um, implementation of an initiative, that brings nothing in return. With the legacy equipment, again, depending on how capex heavy it is and what conditions, I always say that, especially from the steel industry, extremely CapEx heavy industry. What's been built 30 years ago, not necessarily bad. If you can maintain it, if you can safely operate it and produce the quality products. So it is a tremendous amount of technologies that are currently available to the operators. to the manufacturers to deploy. As in the past, everything had to be hardwired. So now you put a Wi-Fi box and you connect it in about 10 hours. And speaking back to the adoption of the operators and the maintenance team, when they see how quickly you can turn it around and how little effort it takes for them to understand what's going on, that's like the gold. And if you properly select those pieces of equipment and those assets, you, listening to the users of them. So that's where the success comes. I wish it would be a roadmap, but I don't think it's one. is just sitting with your team and Making sure, I think you just mentioned a really good example of spending a tremendous amount of money on an asset that really doesn't make sense to monitor to begin with. Just put a spare part right next to it and change when it's broke. Right? So that's a strategy too. Nothing wrong with it. Thank you. Amir, anything to add? No, I like your strategy actually. I like crawl, walk, run. That is something we are also doing. So basically, not necessarily you need to add sensors to the equipment to detect the failure or predict the failure. You already have... In most cases, you already have data. Again, I put it level one, level two, level three data points. For example, we collect the data from the error codes coming back from the machine. This is level one to us. We have the information from our service club basically globally that once technicians go to this machine, what was the diagnostic, what was the note, and how they solved the solution, what part they used. Level two could be basically adding more sensors to the system. I go back to the example of you have to look at the impact and risk factor. Like in aviation, the plane, you have a backup for most of the systems. So you have a flight control backup, you have a yoke backup, but you don't have a backup for landing gear. So that is really important. major part that you need to make sure that works. In our case, fryers. Like our concept in aviation, if you go to McDonald's, they have four different wells in their fryers and equipments. If one goes down, they don't need to shut down the restaurant, basically they go use the second one. But there is a pump that is just one for each system. And if that goes down, They basically cannot serve fries anymore. So you need to understand the impact of the failure and then decide, do you want to use additional sensor? Does it make sense to add a sensor? Or can you detect it with the existing information you have? But always, again, my recommendation is don't fall in love with the technology. First, understand the impact and then go find the technology that can help you find the resolution for your case. Thank you. Thank you. Just to add, and I think both of them did a great job in responding to that question. You mentioned something absolutely crucial, and that is there's a backup that you need to factor in and build in. And the reality is most of the time when we're thinking about deploying the technology with no domain experience, we're sitting there and we're not designing for the executor. Getting a signature from a sensor on a plant line, the signal is useless. What the signal needs to come with is [20:00] a clear line of action of this has happened, what should happen next logically because this has occurred. And that's where the operator has to help inform that loop and that process. But if the operator doesn't know what action is supposed to be directed to correct something that is going off of a statistical process limit, then... you won't get any trust from it. And I think much of what Dimitri just pointed out about making sure that we're refactoring the TCO cost. If I give you a... contextual example at Beach Island in South Carolina, which is Kimberly Clark's largest plant in the world, we have an 80 year old boiler. that boils about 40,000 liters of water a day. Right? There's a good chunk of that machine today, which is still analog. Because, The layman will say, well, just let's slap sensors over everything in this plant and let's start collecting all of these signals. Every sensor you tap on into a plant, number one is a cybersecurity vector. Number two, what are you going to do with the data? I need one clear signal of X. I don't need 50 of them. question. Right? And are you going to go and overhaul this? And if you overhaul it, how much time is it going to take? Can you afford the downtime? The reality is no. The reality is when you start looking at Questions about dark factories, maintenance, where it is, where we are. There's a TCO consideration for businesses. Even if businesses want to say we're state of the art, we're putting in the best automation technology out there. There are levels, as Amir pointed out, of how far you're willing to go down that path in that journey because it's not a one-time cost. That's now a sustained uplift in your TCO for plant ownership, maintenance, the capex costs associated with it, which will never go back down. And 30, 40, 50 years ago, which is the typical life cycle, the economic life cycle of these assessments, these plants were state of the art. What changed? The technology around us, the landscape changed. The utility of what you had did not. It's still producing what it was designed to produce. Thank you. Very insightful. Actually, speaking of sensors and error codes, Who makes the call, right? The model predicts a failure and says stop the line. the technician or the operator who spent decades with that machine says it's fine. you know, Who actually makes that call? I mean, there's a lot of false positives as well. Um, Thank you. The question, the answer is going to be, it depends, right? It always is. I think to begin with, cannot agree more that you've got to do something with the data that you're collecting. So that's the first layer. Oversaturation with the data is just going nowhere. So people are just going to shut off. But when you get to the point that the data is available, It's been properly prioritized based on the criticality of equipment and the strategy that you're deploying to maintain each asset, spare parts, downtime, cost, so on and so forth. So it is a learning curve for each asset. data set that's coming into the, call it a dashboard of the end user. And that learning curve, is extremely important, and the collaboration is very important between the end user and operator, between the person that maintains it and between somebody, what I would call a reliability engineer, or a team member that really defined what The call should be when the data shows us something. And I think where we've been experimenting with the AI is... At the front end, it's all about the human being looking at it. AI is helping to shape and suggest what it sees, human beings setting up some parameters and deploying the AI to help to monitor it better. or see the abnormalities that we may not see when we were setting up that parameter. Notifications are important. Visualization is probably one of the most important underrated tools when it comes to the conditional monitoring. Everybody, it has somewhere in the email, somewhere in the dashboard that nobody logs into, nobody looks at. When something failed, it's a Monday morning quarterback. Everybody going back, finding it and saying, "It's your fault you didn't look at it." Well, how should I supposed to look at it if it's a 20 folders dip, right? So putting that whole, what I would call an environment together, That's what... [25:00] helps the call to be made at the right time at the right place. saying that Amir is the only person that makes a call and that's it. I think that's a little bit of a stretch. Let's talk about returns and ROI. Predictive maintenance isn't really about fixing things. It's also about timing, right? When that system gives you two weeks of warning or whatever that period is, where does the return actually show up? And Maybe it could be anywhere less obvious, places nobody really puts into the business case. Amir, maybe you first. Sure. So, first we need to understand, we talked about two weeks' notice. So do we need to have a two-week notice? I start from the other side. We did a study basically what's going to happen in the restaurants once the equipment fails. So, and it's a very interesting study because from the moment the failure happened within one of our equipment, it's going to take on average one and a half day for someone to pick up a phone and call someone, say, the system doesn't work. Because in the restaurants, you're dealing with a teenager and third shift. They just wait until tomorrow. Someone else is going to take care of it. to dispatch a technician. Hopefully they have the right part on the truck. First visit, fix it. So now we need to first understand, do we need to two weeks? No. In our case, four or five days is enough. And also we need to understand what is the impact of this failure. As I said, we have different wells within the systems. If one goes down... For example, a heating element. They can use another one. You don't need to dispatch a technician right away. And you don't have, even if you don't dispatch a technician, the downside is they're going to use another part, another will. And if the machine is under warranty, we have to pay basically for the technician to go and just add a heating element, a new heating element. But if the Pomp goes down. Now basically they cannot serve fries anymore. That's a different story. They need to shut down the restaurant. And now we have an emergency call. Basically they have to go in the middle of the night to fix it. One crazy feedback I got from the audience, From one of our... big customer asking them what predictive maintenance is it going to bring to you guys. Because uptime, we all know, uptime is going to be higher. They told me it's not just about the visible cost of downtime. It's about our team member in the restaurant that the stress level is going to go up once something is down. And instead of having a calm place in the restaurant, now they are fighting. They have a stress level. We are working in a peaceful environment, not just the obvious downtime that you guys see. Thank you. Amen. I think that's a great example. And you're absolutely right. You hit the nail on the head. Everybody looks first to uptime as the ultimate metric, and it's not the only metric. I talked about the web break as a signature failure that we have. On the Yankee, there are certain elements in a month. We're a listed company. So in a month, when we're looking at our production development quotas, know have certain foreseen events that will occur. So we know that a particular machine will break down after a certain number of hours, and it's just... Tied to how it overheats, it shuts down. And when it does shut down, we know it's going to take about one to two hours to get that machine back online. And depending on which asset we're talking about in which mill around the world, that could be... resetting 150 to 200 set points manually. There's no robot, there's no AI that's taking care of that. The planned downtime does allow us to say, you know what? This is what it's going to take to fix something. You have enough time to figure out, do I want to fix it this month or do I want to fix it next month? It's going to affect my quarter. but where do I want it to have the greatest impact? And you're not going to see the same stress as a customer in the shop that's using your equipment in this case. But where the stress falls in is when we go and we put our products on the shelf in a Walmart or a Target, we have quotas of how much we can put up there and we have time limits as to how long they can be up there. Right. So if we're not meeting our limits and [30:00] contractual obligations that says you put 20 diaper bags on the shelf, they need to be gone in two weeks. Well, if they don't go... then I have to buy them back. I need to take them back into my warehouse. So now I'm worried about restocking quotas around. I'm worried about which markets I've got contractual obligations against, along with the wider issue of, overall morale of what's happening on the line, because people are working around the clock. So it impacts your shift, it impacts everything along the way. you know, the I would argue that the hidden cost there as well is there's elements there that the CFO is actually the sponsor of. When I have a sensor that's telling me that I've got more bleach going in and weeping into the line, that is something that I'm moving away from stuff that is impacting obvious downtime on the shop floor to the hidden costs of something that could have become an environmental drainage report, which would carry more compliance costs and more sustainability costs and issues that we would be investigated or audited for. as a result. So at every level, it starts feeding your sustainability and environmental agenda as well. And we're a pulp and paper chemical process company. So, there's a lot of environmental regulations that were besetting. To give you an example, and everybody takes this for granted, having worked for this company, I look at a diaper as a very sophisticated piece of technology, which I never would have done before, but we have to report on the specific species of fiber that exist. in our products and what we're doing to replace that in the ecosystem around us. So you can just go and say, if we overbleed something or something doesn't work, what do we do with that data loop? And how do we track that? So, way back. I think it's as simple as at least my belief that conditional monitoring has to pay for itself. So if it doesn't, it shouldn't be in place. We look at it from the standpoint of what are we here for, or why are we doing business to begin with. serving our customers on time in full. That's what drives all the decisions to how we deploy the technologies to maintain our assets. And that comes exactly back to the criticality of the equipment and the strategies for maintaining the reliability of the equipment. I think one of the very important things uh, pieces of that approach that is sometimes underrated when people start the implementation, of these technologies, they forget to set up a value stream around counting how much value it brings in the dollar signs, understanding the deployment not just cost but the benefit on a day-to-day basis, really helps to next step or next phase Because when you see the return on investment paying back, You can report it out, you can share with the team. The success always brings more people and more excitement from different levels within the organization. And that's how we... have been doing this and how we're thinking about it going forward. It only makes sense when it pays back. Yeah. I think this is interesting. Okay. Amir, Thank you. you know, Henny Penny builds equipment that, you know, works years in customers kitchen, right? So you're watching machine health in your own factory as well as equipment health at your customer sites out in the field. So how much do those two actually overlap in terms of, you know, data, the models that the AI models that or the solutions that you're looking at, you know, is it the same team that's overlooking both? Thank you. So good question. Definitely we are not using our own equipment in the plant, but in many cases, the same concept basically how the predictive maintenance works. I go from the impact side. In our factory basically once let's say the big laser cutter that we have it goes down, we have to stop the line and you can count how many fryers, how many holding cabinets we are making and this is the downtime and that's the cost to the company. Once we talk about the equipment in the field, we are not talking about the half a million dollar equipment, no. You're talking about the 20, 30 grand basically equipment, but at the end of the day, you are not talking about the one location. 100 locations, 1,000 locations. You are talking about... [35:00] 50,000 locations. If my model doesn't behave as we expect, now we have a false positive, false negative in 50,000 locations. And think about the chaos we are going to make globally, actually. so That's one part of it. So from the data standpoint, not necessarily the overlap, but I want to touch on something. Once we talk about the predictive maintenance, I believe you mentioned it at the beginning. The pattern is same, but not necessarily the data is going to stay same from machine to machine. So what we need to do is to find the baseline for this machine. We have a quality check at the end of the line, and that is the place basically as a company, we can find the signature of this machine once they drain the oil, once they heat the element basically. It's a little bit different from machine to machine. And we are going to capture that signature. And once we compare this machine in the field, we are not going to just compare it with the average signature. We can also compare it to the baseline this machine had from the get-go. Thanks. Amish, your product is the opposite, right? You use once and then it's gone. No, I wouldn't say that. Even in the consumer industry, the field data and everything else that you see all comes back to the same. If I have a diaper that has absorbency issues, that's customer feedback as a loop that's going back to the R&D process, which is adjusting and investigating and figuring out why. What went wrong? Was it the SKU that wasn't right? Was there a quality control issue that we didn't configure for on the line? got to go back to the R&D team. Why did it occur? Was it something that was missed? And if it was missed, again, what in the pattern needs to be adjusted? What did we not account for or was there something that was off in terms of calibration? You know, Was it the data that we didn't have or what did we miss as part of that? So it does still come back. The loop still exists. And I would argue that it exists in every industry. Thank you. So, So a lot of AI applications for manufacturing warehouses come across a desk every year and most of the demos look great. What we can't see from the outside is what happens after deployment. when the tool meets a real plan. You three see that part When you're choosing between a ready-made AI platform versus building something yourselves, what's your decision rule? Well, I mean, probably, first of all, the cost. That's a big factor. And also... If we can do it internally... I'm a big proponent of doing it internally. One of the things that is... makes me extremely excited about this technology is we are Increasing the efficiency and bringing... a lot of new capabilities to our internal teams that before had to be completely outsourced. So with the AI, development with the tools that we have We can... process so much more and develop so much more by ourselves. that In the past, you would have to bring somebody from outside. to develop it for you Teach them what your process is. Try to explain them. what you're looking for, and then five times out of ten, you're probably not going to succeed because you're on the time constraint, the project, so on and so forth. My thinking and one of the realizations even out of this conference, how much different teams and how much people can do now themselves. And connecting those dots and building that environment Um, that doesn't just include sensors, right, but includes the AI tools that bring all that data together, including the data that sits deep inside of everybody's CMMS in the work orders. historical that nobody right now has a very good access to. So when I was reading about this panel and thinking about what is the most exciting pieces. To me, it's not the sensor and not the data is coming from the sensor. It's about the infrastructure and the environment where you can bring all the data from everywhere that in the past was completely hidden, together to make the best managerial decision possible. And I think a lot of people in that journey, a lot of people are going there. and The world of the possibilities with the AI is, [40:00] very, very, very good. Love that answer. - Don't add. I think fundamentally it's... Build versus buy versus partner. Right. Um, I think everybody, to some extent, likes the idea of building, but building things is hard. And one of the things you realize as you're driving through this is that what it takes to build today is not the same as what it used to take to bill yesterday. And many of the skill sets that we have and that are up and coming in the world of AI and data science and so on and so forth, they're not as fungible. In fact, they're not fungible. and you have people that are specialized and trained in very specific techniques and tool sets, which are new and novel, but they don't know any of the old stuff. So what do you do with the people that only know the old stuff, and how do you re-envisage their roles? I think the experience has taught us, and we've played around with this. We were historically, four years ago, we were historically a buy shop, and then we decided we wanted to start building in some build muscle, right? We needed to bring new talent in. We needed to poach some talent. We needed to give them enough opportunities to build and prove that we had the capability to build. But over time, we realized as well, you know what? There are things that are a genuine commodity out there, right? Buy the commodity. A bearing vibration sounds very similar, almost the same in many different machines, irrespective of what the asset is. listen to that or the sensor that's going to let you listen for that. But the things that are unique and specific to you, so in our case, and there's a pulp and paper company, the absorbency level, the wet crepping techniques, you know, how we deal with the chemical balances, those are things that are specific to our DNA, specific to our products and our R&D. Build custom for that. And you don't have to build it alone. You can partner with different players to come up with a custom unique the wider infrastructure that Dimitri was talking about. But I think, you know, at the end of the day, you also need to factor in the TCO of buy at the moment. I say this unabashedly. We're having to cover the R&D costs of literally any vendor out there that we're working with across our tech stack that is now embedding AI features or doing research on new products that are AI-based, because that's where the wider trend is going. So... I don't have a choice in that matter. I don't get to say I'm only going to pay you for the features that I use. You buy the whole product and you use what you use. what used to be a three-year 30% escalation on your commercial tech stack is now going to be an annual 30% cost that you're going to have to bear one way or the other. And it doesn't matter how many times a year or every two years you go through an app rationalization process to say, what systems, what vendors, what platforms, what products am I going to use? You're still going to come back to square one of... Where should I be building? Where should I be buying? Where does it make sense for me to partner to offset cost? That's my take on it. Thank you. I want to add something to that. One pattern I have seen actually, as I'm serving at NAFM, working with different manufacturers and also working with the giant restaurant chains. Thank you. Thank you. If this is not a core value to your business, And you want to do it internally. The catch is what I have seen as a pattern, you build up your AI team, you build up your predictive maintenance team, and the turnaround within these teams, Mostly it's high. You build a team, after two, three years, someone leaves the team, and now you start from scratch. I have seen it multiple times. So that is a kind of cautious, and make sure if you want to do it in-house, you build a team in a right way. Otherwise, we partner with someone who can help you. Do that. One thing to add to that around the build scenario, and even the partner scenario to be aware of, is... When you're using a vendor's platform, Be careful about what's happening with your data. It's your data sovereignty. So if you just rely on the vendor and you say, I'm going to house all of my data, my production data and my line data in your tool, now you're paying rent on your asset. You're not owning it. You're paying the vendor a lump sum subscription fee and you're paying rent on top of it. So be mindful of that. Great. Thank you for the answers. I believe we're at time, but I want to also turn to the audience if there's any questions. Questions?