Battle Motors Uses Plant-Floor AI to Cut Defect Fix Times to 4 Hours

AI on the Line: Predict, Prevent, and Perfect Complex Manufacturing With Battle Motors — Ai4

Ai4

Closed-loop sensor fusion and automated line tuning prevent assembly failures upstream, allowing agile manufacturers to resolve production defects in hours and permanently delete failure modes from design files.

AI-assisted summary of AI on the Line: Predict, Prevent, and Perfect Complex Manufacturing With Battle Motors — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 28, 2025, this presentation features Kelleigh Shankel, Chief Technology Officer at Battle Motors. According to the description, the session explores real-world industrial AI applications that diagnose problems upstream, optimize uptime, and elevate quality on automotive production floors without requiring a total tech stack overhaul.

## TL;DR Kelleigh Shankel details how Battle Motors deploys practical sensor fusion, digital twins, and automated parameter tuning to catch and correct manufacturing defects in real time on heavy-duty commercial truck lines. Grounded in a real-world EV Power Distribution Unit (PDU) case study, she shows how catching anomalies seven hours before end-of-line testing saved $140,000 in scrap, cut field failures by 26%, and reduced issue resolution cycles from days to under four hours. The architecture creates a continuous feedback loop, routing live plant metrics and in-service CAN bus telemetry directly back to design engineers to permanently engineer out recurring failure modes. Shankel emphasizes that nimble, mid-sized OEMs can out-execute bureaucratic legacy giants by proving AI value on tightly bounded components before scaling across the entire factory.

## Atmosphere & style Fast-paced, pragmatic, and refreshingly candid, Shankel delivers an insider's view of physical manufacturing that replaces Silicon Valley buzzwords with blunt shop-floor realities and sharp humor. The session feels like an energetic, no-nonsense masterclass from a seasoned engineering executive who values hard financial ROI and cycle time gains over abstract promises. Her grounded tone carries through the audience Q&A, balancing practical tips on model stress-testing with serious reflections on industrial cybersecurity and testimony before Congress.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) Industrial AI delivers concrete manufacturing value when applied directly to shop-floor operations, shifting quality control from passive alerting to active, sub-hour remediation. - [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) Sensor fusion combining torque, vibration, and electrical signals improves defect detection accuracy by 38% over isolated threshold monitoring. - [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) Automated parameter tuning dynamically recalibrates workstation tooling in under 60 seconds, eliminating production line downtime and engineer wait times. - [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) Streaming automated inspection data back to component vendors accelerates supplier corrective action from weeks down to under 48 hours, driving a 14% drop in incoming defects. - [04:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=288s) Routing assembly line diagnostic data directly to design engineering allows teams to modify CAD models and specifications, permanently deleting failure modes from future product iterations. - [04:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=288s) Deploying continuous telemetry and feedback loops reduced total issue correction time from over three days to under four hours and elevated quarterly first-pass yield from 88% to over 94%. - [08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) AI feedback stacks are portable across diverse manufacturing domains; applying the same system to a purely mechanical line eliminated over $1,000,000 in scrap and prospective warranty costs within 90 days. - [10:37](https://www.youtube.com/watch?v=QazBXHq8DUw&t=637s) Mid-sized OEMs hold a strategic speed advantage over legacy automotive giants by deploying AI on tightly bounded pilot lines without requiring multi-year infrastructure overhauls. - [13:44](https://www.youtube.com/watch?v=QazBXHq8DUw&t=824s) Maintaining industrial AI cybersecurity and model robustness requires deliberately stress-testing systems with aggressive failure simulations rather than relying solely on minor parameter variations. - [14:54](https://www.youtube.com/watch?v=QazBXHq8DUw&t=894s) Standard CAN bus field telemetry enables automated component risk scoring, simultaneously dispatching preemptive fleet service alerts and providing real-world performance data to upstream designers. - [16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) Because government oversight and regulations lag significantly behind manufacturing AI capabilities, heavy-vehicle OEMs must establish proactive internal safety and validation baselines.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) Severe-Duty Trucks, Real-World AI, and the EV PDU Case

Kelleigh Shankel opens by introducing herself as the Chief Technology Officer of Battle Motors, a medium-sized commercial vehicle OEM based in the heartland of America in Ohio. Unlike consumer-facing brands or highway semi-tractors like Tesla's over-the-road vehicles, Battle Motors builds severe-duty vocational trucks designed for demanding everyday municipal and commercial work—such as intracity beverage delivery trailers and 100-foot boom crane trucks that dig telephone poles. With almost 30 years of experience across automotive and heavy manufacturing, Shankel notes she has managed billion-dollar plants, worked on electric trucks long "before they were cool," and weathered plenty of urgent late-night executive emails.

Addressing the audience directly, Shankel emphasizes that her presentation will avoid theoretical, pie-in-the-sky promises about artificial intelligence. Instead, she anchors the discussion in an operational case study from their production line: manufacturing Electric Vehicle Power Distribution Units (EV PDUs). She outlines how an automated AI deployment caught a defect seven hours before end-of-line testing, prevented $140,000 in immediate scrap and rework, reduced field failure rates by 26%, and as she jokes, "made me look like a genius for implementing it." Her takeaway is blunt: it wasn't executive magic; it was practical shop-floor technology.

[01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) The AI "Toys": Sensor Fusion, Digital Twins, and Automated Tuning

Shankel breaks down the primary tools—what she jokingly calls the "toys"—running in Battle Motors' production environment, insisting they serve an active industrial purpose rather than just looking good on executive slide decks. The first layer is **sensor fusion AI**, which merges real-time torque, physical vibration, and electrical signal metrics into a single unified telemetry stream. This allows machine learning algorithms to detect anomalies across correlated signals, yielding a 38% bump in defect detection accuracy over isolated threshold monitoring. In the EV PDU incident, sensor fusion provided the critical "uh-oh moment," spotting signal anomalies long before human assembly technicians could notice.

Next is the **digital twin**, which Shankel clarifies is far from a static, neglected CAD model. Battle Motors feeds live supplier variability data straight into their digital twin environment. When a fresh batch of wire harness connectors arrived with minute deviations in mechanical torque tolerance, the system flagged the variance before the parts were ever assembled onto the chassis. Building on that, their **predictive quality control (QC)** model calculated that if assembly proceeded unchecked, exactly 14 PDUs would fail downstream testing and end up in the scrap heap, avoiding an immediate $87,000 loss. Finally, **automated parameter tuning** closed the loop: instead of halting assembly while waiting for an engineer to wander over to the workstation, the system updated the line's torque driver settings in under 60 seconds, restoring the next unit fully to specification.

```mermaid flowchart TD accTitle: Real-Time Assembly Line AI Feedback Architecture

subgraph DataIngestion ["1. Telemetry Ingestion"] S1["Torque & Vibration"] S2["Electrical Signals"] S3["Supplier Batch Data"] end

subgraph AnalyticsEngine ["2. AI & Digital Twin Engine"] FUS["Sensor Fusion AI<br/>(+38% detection accuracy)"] TWIN["Digital Twin & Predictive QC<br/>(Flags 14 at-risk units)"] end

subgraph ActionFeedback ["3. Real-Time Action & Tuning"] TUNE["Automated Parameter Tuning<br/>(< 60 sec adjustment)"] SUPP["Supplier Corrective Action<br/>(< 48 hr fix)"] DES["Design Engineering Spec Update<br/>(Root-cause elimination)"] end

S1 --> FUS S2 --> FUS S3 --> TWIN FUS --> TWIN TWIN --> TUNE TWIN --> SUPP TWIN --> DES ```

[03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) Measuring Concrete Business Impact Across the Value Chain

Connecting technical capabilities directly to corporate financial and operational returns, Shankel steps through the downstream business metrics generated by the deployment:

* **$140,000 in Scrap and Rework Prevented**: Directly attributed to the seven-hour early detection window produced by sensor fusion catching torque and electrical drift upstream of end-of-line testing. * **26% Reduction in High-Voltage Connector Failures**: Achieved by training predictive QC algorithms on past failure signatures and running incoming supplier lots against those models. * **5% to 8% Cycle Time Improvement**: Generated by automated parameter tuning, which eliminates line stoppages and engineering wait times by dynamically re-calibrating workstation tools in real time. * **14% Drop in Incoming Supplier Defects**: Unlocked by feeding automated inspection trends back to parts vendors; suppliers were provided with empirical deviation profiles, enabling them to execute corrective engineering changes at their own facilities in under 48 hours.

Shankel emphasizes that the true value of AI in physical manufacturing is moving beyond passive alerts to active, sub-hour remediation: "You don't just find the problem, you actually solve it in real time."

[04:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=288s) Closing the Loop: From Field Telemetry to Upstream Design

Shankel shifts from the standalone save to explaining how Battle Motors turned this capability into a continuous operational loop. Data capture does not terminate when a vehicle drives out of the factory doors. Real-time telemetry, warranty records, end-of-line verification tests, and millions of field data points streamed across CAN buses from in-service trucks roll straight back into the central AI models. During the PDU run, sensor fusion flagged the torque drift, predictive QC isolated the impact, supplier AI traced the root cause back to a discrete part vendor lot, and automated tooling tuning adapted before Shankel had even finished her morning coffee.

The overall operational benchmarks showed dramatic shifts across the business: total issue correction time dropped from three or more days down to less than four hours, overall facility scrap rate plummeted by 21%, and first-pass yield escalated from 88% to over 94% within a single quarter—a remarkable achievement in an automotive industry that Shankel characterizes as a "very sleepy, 120-year-old" sector. Crucially, the diagnostic packages generated by the algorithms are routed back to design engineering teams. Rather than managing recurring component patches across long maintenance cycles, design engineers modify the CAD and engineering specifications immediately for future iterations, effectively "deleting" the failure mode from existence.

| Performance Metric | Traditional Baseline | AI-Driven Production | |:--- |:--- |:--- | | **Correction Cycle Time** | 3+ days | < 4 hours | | **First-Pass Yield (Quarterly)** | ~88% | > 94% | | **Scrap Rate** | Baseline | 21% reduction | | **Supplier Corrective Action** | Weeks / Months | Under 48 hours (>1,000% faster) | | **Design Feedback Latency** | Months to years | Real-time / Active production |

[08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) Cross-Line Portability: Replicating the Stack on Mechanical Systems

To demonstrate that the PDU case was not a fluke or limited solely to sensitive electrical circuits, Shankel recounts expanding the identical AI stack across two very different manufacturing lines within Battle Motors. One of these lines was exclusively mechanical, with completely different physical tolerances, suppliers, assembly sequences, and end-of-line inspection protocols. Within 90 days of deployment, the expansion delivered an 18% scrap reduction, a 6% boost in manufacturing cycle time, and eliminated more than $1,000,000 in scrap and prospective warranty expenses.

A significant hidden advantage of this setup is design agility. Battle Motors feeds live manufacturing and field performance data to engineering teams while vehicles are still in pilot or "gamma" production. Shankel points out that legacy OEMs typically require around 18 months of market exposure to surface, isolate, and engineer fixes for early product defects. By cutting that learning loop down to weeks or days, the company eliminates defects before full commercial scale, translating technical uptime into a defensible market advantage.

[10:37](https://www.youtube.com/watch?v=QazBXHq8DUw&t=637s) The Mid-Sized OEM Edge: Scaling Small to Beat Legacy Competitors

Reflecting on her career path—starting as Ford's youngest engineer and their first female product development engineer—Shankel contrasts the agility of a mid-sized OEM against legacy industrial giants. While giants like Ford manage dozens of multi-billion-dollar assembly complexes, their sheer scale creates bureaucratic friction and slowed execution. Battle Motors leverages its leaner footprint to remain "scrappy" and agile, rolling out and validating machine learning applications without multi-year infrastructure overhauls.

She notes that exactly one year prior, she presented the early pilot stages of the PDU monitoring system at the same conference; twelve months later, it governs their full operational footprint. Her primary strategic advice to manufacturing leaders is to resist the temptation to deploy and scale simultaneously across an enterprise: - **Select a tightly bounded proving ground**: Target a single production line, a single component, or one distinct failure mode. - **Instrument and measure obsessively**: "If you cannot prove it, you can't scale it." - **Execute with urgency**: Within a few years, AI feedback loops will be standard operational table stakes across all discrete manufacturing, and late adopters will find themselves perpetually chasing early movers.

She concludes her main remarks by warning the room against letting institutional knowledge decay: "Every day that you're not closing the loop with AI, you're letting your processes relearn the same mistakes... If your factory could talk, it would be saying, 'I told you this last month.'"

---

[13:44](https://www.youtube.com/watch?v=QazBXHq8DUw&t=824s) Q&A: Industrial Cybersecurity and Stress-Testing the Models

An audience member asks about cybersecurity safeguards, inquiring how Battle Motors prevents malicious actors from tampering with manufacturing lines and automated parameter loops. Shankel acknowledges that industrial cybersecurity is the primary reason large OEMs remain terrified of deep AI integration. Recalling past high-profile automotive network breaches that sent shockwaves through the industry, she emphasizes that most vehicle vulnerabilities do not originate in the core software architecture, but rather through accidental digital "open doors" left unpatched within supply chain components on the plant floor.

To counter adversarial attacks and model drift, Shankel explains that Battle Motors rigorously tests its algorithms by actively attempting to induce failure. They discovered that applying minor parameter variations did not sufficiently challenge or reveal flaws in model behavior. To ensure structural resilience, engineering teams deliberately inject aggressive failure scenarios and edge-case anomalies into the simulation loops. Shankel stresses that strict cybersecurity controls, hardware validation gates, and compartmentalized access rules must be enforced at every station on the physical line to prevent unauthorized modifications.

[14:54](https://www.youtube.com/watch?v=QazBXHq8DUw&t=894s) Q&A: Field Telemetry and Real-Time Component Triage

Another attendee asks Shankel to elaborate on how after-sales warranty records and live vehicle telematics are channeled back to design engineers. Shankel clarifies that heavy-duty commercial vehicles operate under fundamentally different lifecycles than consumer passenger cars: Battle Motors trucks stay in grueling municipal and industrial service for an average of 22 years, compared to roughly seven years for light-duty passenger automobiles. Because of this durability mandate, adding thousands of delicate sensors is impractical; instead, they harvest millions of operational data packets directly from the vehicle's standard CAN bus network.

As an example, Shankel describes transmission diagnostics: - Telematics constantly pipe transmission fluid temperatures, housing thermal loads, and shift anomalies (such as gear slippage or mechanical clunking) straight from customer routes back to plant servers. - When an anomalous pattern emerges, the system assigns a quantitative severity score through their automated risk factor model. - If a trend indicates imminent, mission-critical failure, field service engineers receive automated dispatch alerts to inspect the fleet vehicle before an on-road breakdown occurs. - Simultaneously, the identical telemetry package lands on the desks of upstream design engineers, allowing them to adjust transmission cooling parameters, housing geometries, or shift calibrations for the next product cycle while service techs address the immediate customer vehicle.

[16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) Q&A: Navigating the Regulatory Vacuum in Heavy Vehicle AI

The final audience question touches on compliance, asking whether Battle Motors encountered regulatory hurdles when introducing automated AI adjustments into commercial vehicle manufacturing. Shankel smiles and calls automotive regulatory compliance a complex "rat's nest," though she points out that government oversight consistently lags behind technological realities. Currently, there are virtually zero established regulatory compliance frameworks explicitly governing AI in automotive assembly; she observes that until just two years ago, there were only three federal regulations tailored specifically to electric vehicles despite Tesla's multi-decade presence in the market.

Because Battle Motors produces heavy-duty vocational trucks weighing up to 80,000 pounds, Shankel notes that safety and liability cannot wait on slow government standards: the potential risk to human life means internal engineering validation must be rigorous and continuous. To help bridge this gap, Shankel shares that she recently testified before Congress to address the regulatory void surrounding emerging vehicle tech and AI safety, particularly regarding smaller OEMs that operate without uniform technical baselines. She concludes that until public policy matures, high-consequence manufacturers must establish strict internal testing protocols and take responsibility for risk mitigation in real time.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=QazBXHq8DUw&t=0s) Forget the buzzwords — this is where AI gets real. — Ai4

> [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) made me look like a genius for implementing it. — Kelleigh Shankel

> [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) You don't just find the problem, you actually solve it in real time. — Kelleigh Shankel

> [04:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=288s) very sleepy, 120-year-old — Kelleigh Shankel

> [10:37](https://www.youtube.com/watch?v=QazBXHq8DUw&t=637s) If you cannot prove it, you can't scale it. — Kelleigh Shankel

> [10:37](https://www.youtube.com/watch?v=QazBXHq8DUw&t=637s) Every day that you're not closing the loop with AI, you're letting your processes relearn the same mistakes... If your factory could talk, it would be saying, 'I told you this last month.' — Kelleigh Shankel

> [16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) rat's nest — Kelleigh Shankel

## Mentioned Battle Motors, Kelleigh Shankel, Tesla, Ford, United States Congress, EV PDU (Electric Vehicle Power Distribution Unit), CAN bus

## Video description Kelleigh Shankel, Chief Technology Officer, Battle Motors

Abstracts Forget the buzzwords — this is where AI gets real. In the high-stakes world of automotive and complex manufacturing, AI is no longer a future promise. It's here, and it's reshaping how we diagnose problems, maintain uptime, and elevate quality in real time. Join Kelleigh Shankel for a fast-paced look into the cutting-edge AI applications that are revolutionizing the production floor. From predictive diagnostics that spot failures before they happen, to machine learning models that loop quality data directly back into the line, this session reveals how manufacturers are turning complexity into a competitive advantage. You’ll learn: How AI is being used to catch issues upstream before they become scrap, rework, or recalls The latest in predictive maintenance and troubleshooting at the component and system level Real-world case examples where diagnostics + ML feedback loops improved first-pass yield and cycle time How to deploy these tools without overhauling your whole tech stack If you're in automotive, aerospace, or any high-precision production environment, this is your roadmap to building smarter, faster, and with fewer surprises.

[Source](https://www.youtube.com/watch?v=QazBXHq8DUw)

中文简报

Battle Motors利用车间AI将缺陷修复时间压缩至4小时 闭环传感器融合与产线自动化调优在上游预防装配故障，使敏捷制造商能在数小时内解决生产缺陷，并从设计文件中彻底消除失效模式。 Kelleigh Shankel详细阐述了Battle Motors如何在重型商用卡车产线上部署实用的传感器融合、数字孪生与自动化参数调优，以实时捕捉并纠正制造缺陷。结合电动汽车配电单元（PDU）的真实案例，她展示了在下线测试前7小时发现异常如何节省14万美元的报废成本，将现场故障减少26%，并将问题解决周期从数天缩短至4小时以内。该架构构建了一个持续反馈闭环，将车间实时指标和运行中的CAN总线遥测数据直接回传给设计工程师，从源头上彻底消除重复出现的失效模式。Shankel强调，灵活敏捷的中型OEM可以通过先在范围明确的特定零部件上验证AI价值，再逐步推广至整座工厂，从而在执行效率上超越流程官僚的传统行业巨头。 Battle Motors 采用整合扭矩、振动和电信号的多传感器融合 AI，相较单一阈值监控将缺陷检测准确率提升 38%，并在下线测试前 7 小时拦截缺陷，避免了 14 万美元的废品与返工损失。 (1:36) 产线自动化参数调优系统可在 60 秒内动态重新校准工位工具，无需等待工程师现场干预，推动制造循环周期缩短 5% 至 8%。 (3:12) 自动化检测数据向零部件厂商的实时回传，将供应商纠错整改周期从数周缩减至 48 小时以内，促使来料缺陷率降低 14%。 (3:12) 闭环反馈架构将产线问题修复周期由 3 天以上压缩至 4 小时内，全厂废品率下降 21%，季度一次交验合格率从 88% 提升至 94% 以上。 (4:48) 该 AI 反馈架构成功横向拓展至纯机械装配线，在部署后 90 天内消除了超过 100 万美元的废品及潜在质保成本。 (8:45)

Reading summary

Kelleigh Shankel, Chief Technology Officer of Battle Motors, argues that practical shop-floor artificial intelligence—built on sensor fusion, digital twins, and closed-loop parameter tuning—delivers immediate operational gains and design improvements in commercial vehicle manufacturing. Early Defect Interception and Production Tuning (0:04) Rather than relying on post-assembly inspection, Battle Motors implemented an automated AI stack on its Electric Vehicle Power Distribution Unit (EV PDU) line that combines sensor fusion, digital twins, and automated parameter tuning. Sensor fusion aggregates real-time torque, physical vibration, and electrical telemetry into a single stream, which Shankel reports achieved a 38% increase in defect detection accuracy over single-threshold monitoring. In one production run, this setup detected an anomaly seven hours before end-of-line testing. By ingesting live supplier batch data into the digital twin, the system detected slight mechanical torque variances in wire harness connectors before chassis assembly. Predictive quality control models projected that continuing unchecked would cause 14 downstream PDU test failures, preventing an estimated $87,000 in scrap. Automated tuning then updated the workstation's torque driver settings in under 60 seconds, eliminating the line stoppages and engineering delays typical of traditional troubleshooting. Cross-Line Portability and Business Metrics (3:12) Across the initial EV PDU deployment, Shankel attributes $140,000 in prevented scrap and rework, a 26% drop in high-voltage connector field failures, and a 5% to 8% cycle time gain to automated remediation. Sharing empirical deviation profiles with parts vendors reduced incoming supplier defects by 14% and shortened supplier corrective actions from weeks to under 48 hours. Overall plant metrics improved over a single quarter, with first-pass yield rising from 88% to over 94%, total facility scrap falling 21%, and issue correction cycle times shrinking from three or more days to under four hours. To prove transferability, Battle Motors deployed the identical architecture to a purely mechanical production line with distinct tolerances and assembly sequences. Within 90 days, the mechanical deployment reduced scrap by 18%, shortened cycle times by 6%, and eliminated over $1,000,000 in combined scrap and projected warranty expenses. Closing the Loop to Field Operations and Upstream Design (4:48) Because Battle Motors' severe-duty vocational trucks average 22 years of service compared to roughly seven years for passenger cars, the OEM streams operational telemetry directly from standard CAN bus networks rather than adding fragile supplemental sensors. For transmission monitoring, real-time data on fluid temperatures, thermal loads, and shifting behavior feed into a quantitative risk model. High-severity alerts trigger proactive field service dispatches before vehicles break down on customer routes. The same operational data flows back to design engineering teams. While Shankel estimates that legacy OEMs take around 18 months of market exposure to surface and resolve early design flaws, Battle Motors routes diagnostic packages directly to CAD teams during pilot production. Engineers adjust geometries and specifications immediately, removing persistent failure modes before commercial mass production begins. Cybersecurity, Stress Testing, and Regulatory Gaps (13:44) Addressing industrial cybersecurity concerns, Shankel observes that assembly vulnerabilities typically stem from unpatched supply chain hardware on plant floors rather than core software architectures. To safeguard automated parameter controls and prevent model drift, engineers stress-test algorithms by injecting aggressive edge-case failure scenarios into simulation loops, alongside strict hardware validation gates and access rules at each physical station. On the compliance front, Shankel notes that formal government frameworks for AI in automotive assembly remain virtually non-existent, mirroring historic regulatory lags seen with electric vehicles. Because Battle Motors builds vocational trucks weighing up to 80,000 pounds, Shankel—who recently testified before Congress on vehicle AI safety—maintains that manufacturers cannot wait for public policy to catch up and must self-impose stringent validation protocols. Shankel emphasizes that mid-sized manufacturers should start with tightly bounded, measurable proving grounds rather than attempting enterprise-wide rollouts all at once. An unresolved challenge remains the absence of uniform federal safety standards for manufacturing AI, leaving individual OEMs to self-police and establish baseline operational safeguards.

精炼阅读摘要

Battle Motors首席技术官Kelleigh Shankel指出，通过在商用车产线上部署传感器融合、数字孪生和自动化参数调谐等AI闭环技术，制造企业能够在车间端实现即时缺陷拦截与设计反哺，显著降低废品率并缩短交付周期。 产线AI闭环：在制造端实现亚小时级缺陷纠偏 (0:04) 在电动商用车动力分配单元（EV PDU）的生产中，Battle Motors利用传感器融合AI将实时扭矩、物理振动与电信号整合为统一数据流，使缺陷检测准确率较单变量阈值监控提升38%。该系统在下线测试前7小时成功预警信号异常；结合将供应商线束微小公差偏差纳入考量的数字孪生系统，预测性质量控制模型推算出后续将有14台PDU面临报废风险，直接避免了87,000美元损失。自动化参数调谐机制随后在60秒内动态调整了工位扭矩拧紧机参数，恢复了装配公差。总体而言，该方案避免了140,000美元的废品与返工开销，并将高压接插件故障率降低26%，产线节拍改善5%至8%。 跨业务协同：从供应链响应到上游工程设计迭代 (4:48) Battle Motors将车间监测与外部供应链及内部设计部门打通，打破了传统制造长周期的反馈滞后。检测到零件偏差后，系统生成的数据画像会反馈至零部件供应商，促使其在48小时内完成工厂端工程变更，推动外协件进厂缺陷率下降14%。在厂区内部，整体缺陷处置周期由原先的3天以上压缩至4小时以内，单季度一次装配合格率（First-Pass Yield）从88%跃升至94%以上，工厂废品率降低21%。现场诊断数据包还会直接同步至设计工程团队，工程师据此直接更新CAD模型与工程规范，在后续试产和量产中从根源消除特定失效模式，避免了传统模式下需要18个月市场投放才能暴露设计缺陷的漫长周期。 技术栈横向移植与中型OEM敏捷落地的战略优势 (8:45) 为验证该体系的通用性，Battle Motors将同一套AI技术栈迁移至一条完全由机械系统构成、装配工序与公差要求截然不同的生产线上。在部署后的90天内，该机械线废品率减少18%，节拍速度提升6%，累计节省了逾1,000,000美元的报废与潜在质保成本。Shankel基于其在大型车企与中型重卡OEM的任职经历强调，大型传统车企庞大的组织层级带来了执行摩擦，而中型车企能够以更轻量、敏捷的架构在数月内全面铺开算法。她建议制造管理者不要盲目全厂铺开，应先聚焦单一产线或失效模式跑通数据验证；她同时警示，AI闭环在数年内将成为离散制造业的基础门槛，缺乏闭环将导致企业持续重复过去的工艺错误。 车联网实况分流、对抗性压力测试与监管缺口应对 (13:44) 针对工业网络安全与模型漂移，Battle Motors采取主动注入极端失效场景与边缘异常数据的策略，通过压力测试倒逼算法鲁棒性，并在各工位部署严格的硬件准入栅栏与访问隔离以抵御网络入侵。在车载运维端，由于重型专用车平均服役寿命长达22年，企业放弃加装脆弱的附加传感器，转而通过车载CAN总线抓取变速箱油温、壳体热负荷及换挡打滑等数百万条底层运行数据。风险评估模型一旦捕捉到关键故障征兆，便自动派发售后服务预警，并将诊断数据反哺设计端以调整冷却方案或壳体结构。此外，面对目前汽车AI组装领域几乎处于空白的政府监管环境，Shankel指出重型专用车单车重达80,000磅且风险极高，企业不能等待政策滞后，必须建立自有的严苛内部验证体系以确保行车安全。 总体而言，演讲展示了AI在复杂工况重载车制造中从单纯被动报警转向主动自适应调优的成效。然而，Shankel也明确指出，当前整个汽车制造业的公共监管框架在AI与电动化应用上存在长期滞后，这迫使生产高风险重型装备的企业必须独自承担安全验证与责任治理的重担，政策缺口与技术快步迭代之间的张力仍待解决。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2025 年 8 月 28 日在 YouTube 平台发布。视频记录了 Battle Motors 首席技术官 Kelleigh Shankel 的演讲，主题聚焦于如何抛开空洞噱头，在严苛的商用重卡及复杂汽车制造中将 AI 转化为实际生产力，涵盖预测性诊断、实时数据反馈闭环以及产线质量优化等落地实践。 ## TL;DR Battle Motors 首席技术官 Kelleigh Shankel 深入拆解了电动重卡高压电源分配单元（EV PDU）产线引入 AI 的真实案例，展示了传感器融合、数字孪生与自动参数调优如何在组装过程中提前拦截缺陷并大幅削减报废成本。演讲进一步阐述了如何将单一产线的闭环方法论横向迁移至机械总装线，并利用下线与售后遥测数据双轨反哺工程设计及现场服务。最后，Shankel 在问答环节探讨了制造端网络安全破坏测试、商用重卡长服役周期下的数据利用，以及面对 8 万磅重卡物理安全与行业监管空白时的责任与考量。 ## Atmosphere & style 整场演讲干练务实、节奏明快，充盈着资深制造业老兵独有的接地气与自信幽默。主讲人以极具实战感的硬核风格，用扎实的产线数字与惊险排障经历取代了浮夸的行业术语，让原本复杂的工业流程复盘听起来扣人心弦。问答环节中，在谈及 8 万磅重卡安全与行业监管滞后时，语调转为严肃而具敬畏感，生动展现了一家敏捷型制造企业在技术前沿扎实突围的魄力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) Battle Motors 聚焦严苛工况的特种商用卡车制造，其落地经验表明工业 AI 必须拒绝空谈，直击降低报废与控制停线损失的商业底线。 - [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) 产线部署的“传感器融合 AI”实时整合扭矩、振动及电气信号，使隐蔽缺陷检出精度提升 38%，比人工质检更早发现异常趋势。 - [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) 动态数字孪生连通供应商来料公差变异数据，结合预测性质控模型，在零部件触碰总装线前即可预测报废并拦截潜在巨额损失。 - [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) 产线自动化参数调优机制可在 60 秒内自适应调整装配公差，无需停线等待人工校准，使生产节拍时间改善 5% 至 8%。 - [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) 端到端闭环将产线排障时间从 3 天以上缩短至 4 小时以内，推动全厂报废率压降 21%，一次良率（FPY）在单季度内从 88% 跃升至 94% 以上。 - [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) 供应商质量 AI 交叉比对入厂检验与出厂测试基准，将零部件质量缺陷的源头整改反馈周期极限压缩至 48 小时之内。 - [07:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=432s) 制造闭环的核心价值在于不仅解决产线眼前的异常，更通过将失效模式数据实时反哺设计端，在物理世界中永久消除该类工程缺陷。 - [08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) AI 制造系统具备极强的模块化迁移能力，短短数天即可跨界复用至纯机械总装线，并在 90 天内降低 18% 报废率并节约逾百万美元。 - [08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) 中小 OEM 凭借极高的敏捷度能够避开大型车企的组织惯性，推行“单点切入验证 ROI 再行横向复用”的务实落地路径。 - [12:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=768s) AI 闭环制造正迅速从前沿竞争优势演变为汽车行业准入门槛，缺乏实时闭环机制的企业本质上是在不断重复已发生过的历史错误。 - [13:44](https://www.youtube.com/watch?v=QazBXHq8DUw&t=824s) 制造与互联车辆的网络安全风险往往来自总装线零部件开放接口，模型防御必须依托严苛的“极限破坏压力测试”来倒逼重构防护韧性。 - [14:54](https://www.youtube.com/watch?v=QazBXHq8DUw&t=894s) 商用重卡长达 22 年的服役周期催生了 CAN 总线遥测数据的双轨分发闭环：下游指导现场服务快速救火，上游同步驱动下一代零件设计改进。 - [16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) 面对自重达 8 万磅重卡的物理破坏力及当前 AI 工业法规彻底空白的现状，前沿主机厂必须自主承担极严苛的安全冗余与风控责任。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) 拒绝纸上谈兵：Battle Motors 的商用重卡制造与 AI 落地实战 作为 Battle Motors 的首席技术官（CTO），Kelleigh Shankel 开宗明义地介绍了公司的定位：他们是一家扎根于美国俄亥俄州心脏地带的中型商用车辆原始设备制造商（OEM）。不同于制造长途干线牵引车的特斯拉，Battle Motors 专注于那些在日常生活中随处可见的严苛工况特种卡车（Severe Duty Trucks）——例如穿梭于城市间的饮料配送半挂车，或是用于立电线杆的百英尺长臂吊车。Shankel 在汽车制造行业摸爬滚打了近三十年，不仅管理过数十亿美元级别的超级工厂，早在“造电动卡车还显得很酷”的年代就投身其中，她还幽默地调侃自己甚至扛过了不少“本可以晚点再发的首席执行官紧急邮件”。 Shankel 强调，本次演讲绝非空谈高大上的制造 AI 理论，而是要解剖一个来自电动车电源分配单元（EV PDU，Power Distribution Unit）的真实产线案例。在这个案例中，AI 提前 7 个小时捕获了潜在缺陷，短短数分钟内为工厂挽回了 14 万美元的损失，将售后现场故障率降低了 26%，甚至“让我因为主导落地而显得像个天才”。但她随即给出了剧透警告：“这不是什么魔法，这纯粹是技术的力量。” - **企业定位**：美国俄亥俄州中型商用车辆 OEM，专注城配与特种作业严苛工况重卡。 - **演讲核心**：拆解电动车 PDU（高压电源分配单元）的真实缺陷拦截案例。 - **直接成效**：提前 7 小时捕获缺陷，避免 140,000 美元报废损失，售后故障率压降 26%。 --- ### [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) 产线 AI 工具箱：传感器融合、数字孪生与自动调优 Shankel 随后盘点了产线上每天都在发挥实效的 AI “装备库”，这些工具绝非仅用来装点汇报幻灯片的摆设。首先是**传感器融合 AI（Sensor Fusion AI）**，系统将装配过程中的扭矩、振动和电气信号数据打包聚合为单一的实时数据流，交由算法识别异常模式。仅仅这一项技术，就将产线的缺陷检出精度直接提升了 38%。在 PDU 案例中，正是传感器融合首先察觉到了不对劲的蛛丝马迹，比产线操作工更早发出了预警。 其次是**数字孪生（Digital Twin）**系统。Shankel 明确指出，他们的数字孪生绝不是静止在电脑里“吃灰”的 CAD 静态模型，而是实时连通了供应商的来料变异数据。当一批扭矩特性存在细微偏差的高压连接器进厂时，数字孪生在零部件尚未触碰总装线之前便触发了警报。紧接着，**预测性质量控制模型（Predictive QC Model）**给出了冷酷的推断：“如果继续按当前参数装配，接下来将有 14 台 PDU 铁定直接报废。”这一预判当场锁定了 87,000 美元的潜在损失。 最令 Shankel 感到兴奋的则是**自动化参数调优（Automated Parameter Tuning）**。系统无需等待工程技术人员慢悠悠走回产线手动校准，AI 在不到 60 秒的时间内就自动完成了扭矩设置的微调。到组装下一台工件时，全部装配参数已完全回归工程公差范围之内，整个调整过程一气呵成。 ```mermaid flowchart LR accTitle: AI-Driven Assembly Line Diagnostic and Tuning Flow subgraph DataCollection["实时感知与预警"] A["多维传感器流<br/>(扭矩/振动/电信号)"] --> B["传感器融合 AI<br/>(检出精度+38%)"] C["供应商来料数据"] --> D["数字孪生系统"] end subgraph DecisionAndAction["预测与自动调优"] B --> E["预测性 QC 模型<br/>(拦截14台报废件)"] D --> E E --> F["自动参数调优<br/>(<60秒完成闭环)"] end F --> G["下一台工件回归标称公差"] ``` --- ### [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) 从技术到商业闭环：质量飞跃与供应链实时协同 将算法能力映射到商业底线上，各模块的投资回报清晰可见。传感器融合带来的 7 小时提前预警，直接化解了 14 万美元的报废与返修成本；预测性 QC 模型将单一异常事件沉淀为通用算法，横向覆盖所有批次与供应商，使高压连接器故障率骤降 26%；实时参数自适应调整使生产节拍时间（Cycle Time）改善了 5% 至 8%，彻底告别了停线数小时甚至数天等待人工排障的窘境。同时，**供应商质量 AI** 通过交叉比对供应商出厂测试基准与厂内入厂检验趋势，发现了深层缺陷规律；工厂将数据即时同步给供应商后，对方在 48 小时内便在源头完成了纠正措施的落地。 这种单点突破随后演进为了可复现的体系化能力。从下线检测（End-of-Line）、售后保修数据到在路运营车辆传回的数以百万计的实时数据点，全部汇聚反哺至底层 AI 模型。在 PDU 案例中，数据流经全链路闭环，使整体排障纠错时间从过去的 3 天以上直接缩短到 4 小时以内，工厂总体报废率压降 21%，季度一次良率（First Pass Yield, FPY）从 88% 跃升至 94% 以上。对于拥有 120 年历史、节奏沉稳甚至略显迟缓的汽车制造业而言，这一数字堪称惊艳。 更深远的价值在于设计端的源头改进。AI 模型将产线与售后数据提炼为洞察包，直接交付给产品设计团队，使工程师能在下一代零部件迭代中彻底消除设计风险。过去需要耗时数月甚至数年的“现场服务-生产-设计”反馈链条，如今被压缩至近乎实时。“设计迭代越快，产品改善就越迅速，客户满意度越高，而我接到的现场求救电话也就越少。” - **排障耗时**：从过去 ≥3 天大幅压缩至 <4 小时。 - **工厂报废率**：整体下降 21%。 - **一次良率（FPY）**：单季度内从 88% 跃升至 94% 以上。 - **供应链纠偏**：从入厂预警到供应商源头整改闭环仅耗时 48 小时。 --- ### [07:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=432s) PDU 实战复盘：从发现漂移到彻底“删除缺陷” 在复盘 PDU 的完整处置过程时，Shankel 还原了惊心动魄的产线现场：当时电动车 PDU 产线正处于生产高峰期，扭矩与电气信号开始出现细微的漂移。这种极度隐蔽的微小渐变，产线质检人员往往难以在下线终检之前凭肉眼或经验察觉。然而，传感器融合 AI 瞬间捕获了异常，预测性 QC 迅速判定这绝非偶发噪声，而是具有统计意义的恶化趋势；供应商 AI 随即调取批次追溯数据，精准锁定了来自特定供应商的某批次连接器。 随后，自动调优机制果断介入，在组装下一台 PDU 时即刻修正了拧紧扭矩参数。潜在的重大质量事故被牢牢控制在仅 14 台工件的范围内，在不到半小时内挽救了 87,000 美元的物料损失。更关键的是后续联动：供应商在 48 小时内关闭了纠正措施（整改速度较以往提升了 1000% 以上），设计工程团队拿到包含完整失效模式的数据包后，第一时间修改了该零部件的工程规范。正如 Shankel 掷地有声的总结：“这已经不是修补问题了，这是将这种失效模式从物理世界中直接删除。” --- ### [08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) 跨产线复用与敏捷破局：中小 OEM 的 AI 竞争优势 PDU 案例的成功验证了一套通用的闭环方法论：检测识别、根本原因诊断、供应链穿透、自动纠偏与设计回溯。Battle Motors 迅速将这套相同的 AI 技术架构迁移至两条截然不同的生产线上——其中一条甚至完全是非电气化的纯机械总装线，拥有完全不同的下线测试标准、零部件公差和失效形态。然而 AI 表现出了极强的通用性，模型仅耗时数天便完成了自适应迁移。在短短 90 天内，新产线报废率下降 18%，节拍时间提升 6%，消除了超过 100 万美元的潜在保修与报废成本。 设计团队现在能够在产品上市的第一年内，甚至在样件试制（Pilot）或 Gamma 试生产阶段，就获得产线实时数据的喂养。这意味着当下一代车型下线时，Battle Motors 已经解决了传统 OEM 通常需要 18 个月才能攻克的工程难题。 谈及与巨头的竞争，Shankel 分享了个人经历：她职业生涯起步于福特汽车，曾是福特产品研发部门的第一位女性工程师，也是当时入职年龄最小的工程师。她坦言，Battle Motors 没有福特那样坐拥十数座数十亿美元级工厂的雄厚资本，但这赋予了他们更加敏捷和具开拓精神（Scrappy）的特质。大企业部署此类系统面临着巨大的组织与规模惯性，而中小 OEM 则能凭借极高的敏捷度迅速推进落地，在细分领域建立起以美元、出勤率和客户信任度衡量的硬核竞争壁垒。 回顾从一年前在大会上分享 PDU 试点，到一年后在全厂所有产线全面推广的历程，Shankel 强调这绝非侥幸，而是坚持“小步快跑、验证 ROI 再规模化”的必然结果。她向在场的制造业同行给出了最核心的建议：“从最小处着手。选定一条产线、一个产品、一个具体的用例。测量一切可量化的指标，如果你无法证明它的价值，你就无法实现扩张。” - **机械产线迁移成效（90天内）**： - 报废率降低 18% - 生产节拍提升 6% - 消除逾 100 万美元的保修与报废风险 - **工程迭代周期**：试制/试产期即时闭环，颠覆传统车企平均 18 个月的排障周期。 - **规模化策略**：严禁在未验证投入产出比的情况下盲目并发铺开，坚持“单点验证 ROI -> 模块化复用扩展”。 ### [12:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=768s) 闭环制造与行动呼吁：别让工厂重蹈覆辙 在演讲的收尾阶段，Kelleigh Shankel 强调了迅速采取行动的紧迫性。她指出，在未来短短几年内，人工智能将成为汽车制造和整车厂（OEM）的行业准入门槛（table stakes）。如果企业现在按兵不动，未来就只能拼命追赶那些早已起步的领跑者。她开玩笑说自己正挡在大家和午餐之间，但仍留下了一句掷地有声的警醒：“只要你一天不利用 AI 实现数据闭环，你的生产流程就在被迫重蹈覆辙，你的员工也在一遍遍犯同样的错误。”如果工厂会开口说话，它一定会抱怨：“我上个月就已经提醒过你了！” 赋予工厂发声的能力，不仅是让系统变得更聪明，更是为了将突发危机转化为可控的日常。借助上游预测与实时闭环，企业能够在短短数分钟内挽回高达 140,000 美元的潜在损失，而对一线团队来说，这只不过是“普通周三的例行公事”。 - **行业趋势预警**：AI 正在迅速从前沿竞争优势演变为主流制造 OEM 的基础准入门槛。 - **痛点核心**：缺乏闭环机制的制造体系，本质上是在不断为已发生过的失误重复买单。 - **成效愿景**：“几分钟内挽回 140,000 美元”将不再是罕见的救火奇迹，而是融入日常的标准化常态。 --- ### [13:44](https://www.youtube.com/watch?v=QazBXHq8DUw&t=824s) 现场问答：严防黑客入侵与产线供应链的安全测试 进入 Q&A 环节，现场观众首先提出了关于网络安全的疑问：在工业制造与互联汽车领域，面对潜在的恶意攻击者，应当采取何种防范举措？Kelleigh 直言不讳地表示，网络安全正是大型主机厂对拥抱 AI 普遍心存顾虑的核心原因。数年前黑客成功入侵车辆系统的事件，曾给整个汽车制造行业带来了剧烈震动。 谈及模型防御与测试时，Kelleigh 透露，如果只是微调参数，根本无法对模型产生足够的冲击来改变其应对模式。Battle Motors 采取的策略是极其严苛的“极限破坏测试”——工程师在模拟环境中主动诱发系统崩溃，全力使其失效（trying to make it fail），以此倒逼系统形成真正稳固的防护机制。她进一步澄清了一个关键认知误区：过去很多车辆被黑客劫持的案例，根源往往不是车辆底盘或核心系统本身的漏洞，而是由于供应链上的某个零部件在生产线上留下了未受保护的“敞开大门”，进而被恶意攻击者顺藤摸瓜。 ```mermaid flowchart TD accTitle: 产线组件到车辆系统的安全渗透与破坏测试防范 subgraph 产线暴露风险 A["供应链零部件入厂"] --> B["装配线检测留有后门隐患"] end subgraph 恶意渗透途径 B --> C["黑客借道产线组件入侵"] C --> D["渗透整车网络系统"] end subgraph 抵御机制与测试 E["极限破坏测试 (试图让模型彻底失效)"] --> F["触发质变级防御响应"] F --> G["全生产流程部署动态安全防线"] end ``` - **安全隐患源头**：实际入侵漏洞多发生在总装线零部件留下的开放接口，而非车辆底层核心架构。 - **模型验证法则**：放弃微小参数调优，转而采取破坏性压力测试来倒逼 AI 行为机制的深层重构。 - **全流程防护**：必须在生产制造的每一个离散环节均嵌入端到端的安全参数校验。 --- ### [14:54](https://www.youtube.com/watch?v=QazBXHq8DUw&t=894s) 现场问答：售后与保修数据如何双轨反哺工程设计与现场服务 第二位观众请 Kelleigh 展开阐述如何将保修（warranty）及售后数据通过 AI 打包，进而反哺给设计工程师。Kelleigh 解释道，商用车制造与普通乘用车截然不同。乘用车的平均设计寿命通常只有 7 年，而商用重卡平均服役时间长达 22 年；由于生命周期极其漫长，车辆不可能无休止地堆砌数千个传感器，必须在耐用性与监测成本之间取得平衡。尽管传感器有限，Battle Motors 每天仍能通过车辆 CAN 总线（CAN bus）捕获数以百万计的运行数据，并实时传输至云端上游。 她以变速箱的潜在故障为例生动地说明了双轨响应机制：当数据流显示出变速箱油温异常、壳体温度过高、换挡打滑或异响顿挫等异常征兆时，风险因子模型（risk factor model）会立刻计算出危险等级并赋予任务关键度（mission critical）。此时，系统会并行触发两条行动管线：一方面，现场服务工程师（FSE）第一时间获取预警并奔赴客户现场检修车辆，防止车辆抛锚停运；另一方面，这些遥测特征与工况数据会同步推送到上游设计工程师的工作台，供其排查设计根因，在下一代零部件与图纸层面斩断缺陷苗头。 ```mermaid flowchart LR accTitle: 商用车 CAN 总线遥测数据双轨闭环流向 subgraph 边缘数据捕获 A["车辆 CAN 总线数据流<br/>(油温/壳温/打滑/顿挫)"] --> B["AI 风险因子评估模型"] end subgraph 双轨协同响应 B --> C["高严重度警报: 任务关键"] C --> D["下游: 现场服务工程师<br/>(主动赶赴现场止损)"] C --> E["上游: 核心设计工程师<br/>(同步溯源重构设计)"] end ``` - **运营生命周期对比**：商用卡车平均服役周期长达 22 年，远超普通乘用车的 7 年周期，对系统长期可靠性要求极高。 - **特征工程与征兆捕获**：重点挖掘变速箱流体温度、物理壳体温度、传动打滑比率等异常信号。 - **双轨闭环并行处理**：现场工程师负责下游救火、保障客户出勤率，设计工程师在上游根除缺陷缺陷，双管齐下。 --- ### [16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) 现场问答：面对监管空白，制造 8 万磅重卡的风险与责任 最后一位观众询问，将整套系统快速落地听起来令人振奋，但在实施过程中是否遇到了行业合规与法规方面的挑战？对此，Kelleigh 感慨地将汽车行业的合规监管形容为一个令人头疼的“烂摊子”（a rat's nest）。她指出，监管体系的一个残酷现实是它严重落后于前沿技术的发展速度：目前针对制造业和汽车领域的 AI 监管规章依然是彻底的“零”；甚至直到两年前，针对电动汽车的规范也总共只有区区三条，与电动车先驱企业特斯拉的存在时长形成了极其讽刺的鲜明对比。 面对法规空白，Battle Motors 无法被动等待外部规范，只能自主开展深层调研并动态调整风控策略。Kelleigh 严肃地强调：“我们制造的是重达 80,000 磅（约 36 吨）的庞然大物，一旦失控，对人类生命安全构成的风险是毁灭性的。”这种对生命的敬畏要求整个团队必须目标高度统一，在生产中实时动态规避一切可能的技术隐患。她还透露，自己今年甚至前往美国国会（Congress）就该议题发表了证词，因为随着越来越多小型 OEM 涉足该领域，监管滞后可能带来难以预估的公共安全隐患。 - **监管与技术脱节**：汽车行业中的 AI 法律法规目前为零，两年前涉及纯电动车的规范也仅有三项。 - **物理安全责任**：Battle Motors 生产的车辆自重高达 80,000 磅，安全冗余与风险容忍度容不得丝毫侥幸。 - **国会听证与政策推进**：CTO 亲自向国会作证，呼吁重视智能重型车及新兴小型主机厂在监管缺位下的潜在安全隐患。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=QazBXHq8DUw&t=4s) 这不是什么魔法，这纯粹是技术的力量。 — Kelleigh Shankel > [01:36](https://www.youtube.com/watch?v=QazBXHq8DUw&t=96s) 如果继续按当前参数装配，接下来将有 14 台 PDU 铁定直接报废。 — Kelleigh Shankel > [03:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=192s) 设计迭代越快，产品改善就越迅速，客户满意度越高，而我接到的现场求救电话也就越少。 — Kelleigh Shankel > [07:12](https://www.youtube.com/watch?v=QazBXHq8DUw&t=432s) 这已经不是修补问题了，这是将这种失效模式从物理世界中直接删除。 — Kelleigh Shankel > [08:45](https://www.youtube.com/watch?v=QazBXHq8DUw&t=525s) 从最小处着手。选定一条产线、一个产品、一个具体的用例。测量一切可量化的指标，如果你无法证明它的价值，你就无法实现扩张。 — Kelleigh Shankel > [12:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=768s) 只要你一天不利用 AI 实现数据闭环，你的生产流程就在被迫重蹈覆辙，你的员工也在一遍遍犯同样的错误。 — Kelleigh Shankel > [12:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=768s) 我上个月就已经提醒过你了！ — Kelleigh Shankel > [12:48](https://www.youtube.com/watch?v=QazBXHq8DUw&t=768s) 普通周三的例行公事 — Kelleigh Shankel > [16:28](https://www.youtube.com/watch?v=QazBXHq8DUw&t=988s) 我们制造的是重达 80,000 磅（约 36 吨）的庞然大物，一旦失控，对人类生命安全构成的风险是毁灭性的。 — Kelleigh Shankel ## Mentioned Kelleigh Shankel, Battle Motors, Ai4, 特斯拉 (Tesla), 福特汽车 (Ford), 电动车电源分配单元 (EV PDU), 传感器融合 AI (Sensor Fusion AI), 数字孪生 (Digital Twin), 预测性质量控制模型 (Predictive QC Model), 自动化参数调优 (Automated Parameter Tuning), 供应商质量 AI (Supplier Quality AI), CAN 总线 (CAN bus), 美国国会 (Congress) ## Video description Kelleigh Shankel, Chief Technology Officer, Battle Motors Abstracts Forget the buzzwords — this is where AI gets real. In the high-stakes world of automotive and complex manufacturing, AI is no longer a future promise. It's here, and it's reshaping how we diagnose problems, maintain uptime, and elevate quality in real time. Join Kelleigh Shankel for a fast-paced look into the cutting-edge AI applications that are revolutionizing the production floor. From predictive diagnostics that spot failures before they happen, to machine learning models that loop quality data directly back into the line, this session reveals how manufacturers are turning complexity into a competitive advantage. You’ll learn: How AI is being used to catch issues upstream before they become scrap, rework, or recalls The latest in predictive maintenance and troubleshooting at the component and system level Real-world case examples where diagnostics + ML feedback loops improved first-pass yield and cycle time How to deploy these tools without overhauling your whole tech stack If you're in automotive, aerospace, or any high-precision production environment, this is your roadmap to building smarter, faster, and with fewer surprises. [Source](https://www.youtube.com/watch?v=QazBXHq8DUw)

Source quotes

[0:04] Kelleigh Shankel: 这不是什么魔法，这纯粹是技术的力量。 [1:36] Kelleigh Shankel: 如果继续按当前参数装配，接下来将有 14 台 PDU 铁定直接报废。 [3:12] Kelleigh Shankel: 设计迭代越快，产品改善就越迅速，客户满意度越高，而我接到的现场求救电话也就越少。 [7:12] Kelleigh Shankel: 这已经不是修补问题了，这是将这种失效模式从物理世界中直接删除。 [8:45] Kelleigh Shankel: 从最小处着手。选定一条产线、一个产品、一个具体的用例。测量一切可量化的指标，如果你无法证明它的价值，你就无法实现扩张。 [12:48] Kelleigh Shankel: 只要你一天不利用 AI 实现数据闭环，你的生产流程就在被迫重蹈覆辙，你的员工也在一遍遍犯同样的错误。 [12:48] Kelleigh Shankel: 我上个月就已经提醒过你了！ [12:48] Kelleigh Shankel: 普通周三的例行公事 [16:28] Kelleigh Shankel: 我们制造的是重达 80,000 磅（约 36 吨）的庞然大物，一旦失控，对人类生命安全构成的风险是毁灭性的。

Source quotes

[0:00] Ai4: Forget the buzzwords — this is where AI gets real. [0:04] Kelleigh Shankel: made me look like a genius for implementing it. [3:12] Kelleigh Shankel: You don't just find the problem, you actually solve it in real time. [4:48] Kelleigh Shankel: very sleepy, 120-year-old [10:37] Kelleigh Shankel: If you cannot prove it, you can't scale it. [10:37] Kelleigh Shankel: Every day that you're not closing the loop with AI, you're letting your processes relearn the same mistakes... If your factory could talk, it would be saying, 'I told you this last month.' [16:28] Kelleigh Shankel: rat's nest

Original transcript

[0:04] My name is Kelly Shankle. I'm the chief [0:05] technology officer of Battle Motors. We [0:08] are a mediumsized commercial OEM. We [0:12] manufacture commercial vehicles um that [0:15] are doing jobs uh that we experience [0:18] every day. So we are not the Tesla over [0:21] the road tractors. We are the intracity [0:23] beverage delivery tractor uh trailers [0:26] and we are um manufacturing 100 foot [0:29] boom cranes that dig telephone poles and [0:31] things like that. So we do the severe [0:34] duty trucks and we manufacture them all [0:37] right in the heartland of America right [0:39] in Ohio. Um I've been in automotive and [0:43] manufacturing for almost 30 years now. [0:47] um running billion-dollar plants, [0:49] building electric trucks before they [0:52] were cool, and yes, even surviving some [0:55] urgent emails from the CEO that could [0:57] have um waited a little bit. So, what [1:01] we're going to talk about today is not [1:02] pie in the sky AI theory as it relates [1:05] to manufacturing. I'm going to walk you [1:07] through a real case from one of our EV [1:10] PDUs. It's a lot of acronym, but one of [1:13] our electric vehicles power distribution [1:15] units. um and how AI caught a defect [1:19] seven hours early, saved us $140,000 in [1:24] minutes, cut field failures by 26%, and [1:28] made me look like a genius for [1:30] implementing it. Um, spoiler alert, it [1:33] wasn't the magic, it was the tech. [1:36] All right, so let's start with the toys. [1:38] The toys are the AI tools that we use [1:41] every day. And yes, they do more than [1:45] look good on a PowerPoint. So, first up, [1:47] sensor fusion AI or all it means is that [1:51] we take torque, our vibration, and our [1:54] electrical signal data, throw it into [1:57] one live stream, and let AI find [1:59] patterns. That bumped up our detection [2:02] accuracy by 38% alone. In our PDU case, [2:07] this was the uh-oh moment. It saw [2:10] trouble before our operators did. Then [2:14] there's the digital twin. And no, this [2:17] isn't a static CAD model just collecting [2:19] dust. Ours is hooked up to our supplier [2:22] variability data. So when a batch of [2:25] connectors showed up with slightly [2:27] different torque behavior, it flagged it [2:30] before they even touched our assembly [2:32] line. [2:34] Then our predictive QC model said, "Hey, [2:37] if you're going to keep going, 14 PDUs [2:40] are for sure headed straight for the [2:41] scrap bin." [2:43] That's $87,000 saved without even [2:46] breaking a sweat. And my favorite, [2:49] automated parameter tuning. The AI [2:52] adjusted the torque setting in under 60 [2:55] seconds. No more having to wait for one [2:58] of my engineers to wander back over to [3:00] adjust the torque settings. AI did that [3:02] in under a minute. By the next unit, we [3:05] were back in spec. So, it was done. It [3:08] was fixed and we moved on. [3:12] All right. So, now let's connect the [3:13] dots between the tools I just walked [3:15] through and the actual business impact. [3:18] So that seven-hour early catch on the [3:20] PDU line that came straight from sensor [3:23] fusion [3:25] that spotted the torque and the signal [3:28] anomalies before our end of line [3:30] testing. That one detection saved [3:33] $140,000 in scrap in our rework. [3:38] the 26% reduction in high voltage [3:40] connector failures that was from [3:43] predictive QC building a model that uh [3:47] from the same events data and applying [3:50] it across all production runs and most [3:52] importantly across supplier batches [3:57] that 5 to 8% cycle time improvement that [4:00] is our automated parameter tuning in [4:02] action instead of waiting hours or days [4:05] to manually adjust processes [4:08] AI tweaked the settings in real time so [4:10] production kept flowing. [4:13] And finally, that 14% drop in incoming [4:16] defects from our suppliers. That's our [4:19] supplier quality AI comparing the [4:21] supplier quality benchmarking test data [4:25] with our own incoming inspection trends. [4:28] It flagged the patterns. We shared those [4:31] with our supplier and they fixed it and [4:35] implemented their corrective actions at [4:36] the source in under 48 hours. [4:40] That's the beauty of AI. You don't just [4:42] find the problem, you actually solve it [4:44] in real time. [4:48] So, we've talked about the tools. We've [4:50] seen the wins. [4:52] And so let's now turn to a one-off save [4:56] in a or this has turned into a [4:59] repeatable superpower and closing the [5:02] loop. So here's the play. Everything [5:04] that happens at the end of our line [5:06] testing, every warranty data and [5:09] millions and millions of data points [5:12] coming directly from our vehicles in the [5:14] field today, all roll up into our model [5:18] and all of that is transmitted and goes [5:20] back into our AM AI model running those [5:23] tools that I mentioned previously. In [5:26] the PDU case, the sensor fusion caught [5:28] the torque drift. Our predictive QC [5:31] modeled it. supplier AI traced it back [5:34] to [5:36] uh the specific batch [5:38] um and uh automated parameter tuning [5:42] said cool fixed it all before I finished [5:44] my morning coffee and that's where the [5:47] KPIs I shared before look so good [5:50] correction time dropped from 3 days or [5:52] more to under four hours in totality [5:56] scrap rate fell 21% since we started [6:00] implementing this and first pass yield [6:02] jumped from 88% to over 94% in an entire [6:07] quarter. That's staggering for um [6:11] complex manufacturers, especially in [6:13] automotive. We're a very sleepy [6:15] industry. We're 120 years old. So that's [6:18] staggering. [6:20] And uh this is the part that I really [6:22] like that data doesn't just stop at our [6:25] assembly line. The model packages all of [6:28] the insights and hands them off to my [6:31] design engineers so they can bake the [6:34] fix into the next iteration of the [6:36] component. That way we won't we won't [6:38] just prevent it from happening will [6:41] eliminate the risk on the next [6:43] iteration. [6:45] Before that kind of feedback loop be [6:47] between field service production data [6:50] and design took months if not years. Now [6:54] it's almost in real time. [6:56] And the faster the design learns, the [7:00] faster the product improves, the happier [7:02] the customer and the fewer hate calls I [7:05] receive from the field asking to fix it. [7:12] So this is where everything that we've [7:14] been talking about really comes [7:15] together. So we are mid production on [7:19] our EV PDUs. torque and electrical [7:22] signal drift start creeping in and it's [7:25] the again it's the kind of subtle shift [7:27] a human may miss ended [7:31] um until the final end of line but [7:34] sensor fusion sees it right away [7:36] predictive QC runs the data and says [7:39] nope this isn't noise this is now a [7:41] trend [7:43] supplier AI pulls up the batch history [7:45] directly from our supplier and boom we [7:48] now know it's tied to a specific batch [7:50] of connector coming in from our [7:52] supplier. [7:54] Automated tuning kicks in, adjusts the [7:56] torque settings on the very next unit, [7:58] and just like that, the problem is [8:00] contained to 14 PDUs, and that's $87,000 [8:04] saved in under a half hour of time. [8:08] First pass yield for the quarter goes up [8:10] 6%. Supplier corrective action closes in [8:13] 48 hours. And that is um over a,000% [8:17] faster than our previous average. And [8:19] that was just done in one quarter. And [8:22] here's where that loop I just talked [8:24] about really kicks in. Engineering gets [8:26] that data package, sees the connector [8:28] issue, and we changed the spec on the [8:31] component um immediately. That failure [8:34] mode is now eliminated from any future [8:37] builds. That's not fixing the problem. [8:39] That's deleting it from existence. [8:45] So why does the PDU example matter so [8:47] much? [8:49] Because once you get the system working, [8:52] the detection, the diagnosis, the [8:55] supplier tie-in, and the automated fix, [8:58] plus the design feedback, you can plug [9:00] this in anywhere, anywhere in [9:01] manufacturing. [9:03] We took that exact stack and deployed it [9:05] in two very different lines. [9:10] In just 90 days, scrap went down 18%. [9:15] cycle time went up 6% and over a million [9:18] dollars in warranty scrap was eliminated [9:22] in just 90 days. [9:25] And one of those lines was building [9:26] something entirely different from a PDU. [9:28] It wasn't even an electrical component. [9:30] It was a mechanical one. So the end of [9:32] line testing was completely different. [9:35] Different suppliers, different [9:37] tolerances, different failure modes. But [9:39] the AI didn't care. The models adapted [9:42] in days and not months. [9:45] And here's the hidden win. The our [9:48] design teams are now getting the live [9:51] production fed data before a product [9:54] even finishes its first year in getting [9:57] to the market. Our engineers are able to [10:02] make real-time design tweaks sometimes [10:04] while the model is still in pilot or [10:06] even gamma production. [10:09] That means by the time the next version [10:11] hits our assembly line, we've already [10:14] solved problems that most OEMs will take [10:17] around 18 months to solve. And that's [10:20] industry data. And that is the AI [10:22] advantage. We have faster fixes, smarter [10:26] products, and that's a competitive edge [10:28] you can measure in dollars, uptime, and [10:32] customer trust. [10:37] So here's the bottom line. [10:40] AI in manufacturing isn't something you [10:42] plan for someday. It's here. It is [10:46] running in our plants now. Now I am a [10:49] mediumsized [10:50] OEM. I uh actually started out my career [10:54] at Ford. I was uh Ford's first um female [10:58] engineer in uh product development and I [11:01] was their youngest engineer ever hired. [11:03] Um, but we are not a Ford. I don't have [11:06] a uh, you know, I don't have a dozen [11:09] billion dollar plants. Um, but that [11:12] makes us scrappy. We are agile and we [11:15] can evolve and this is being deployed. [11:19] Um, and it's easier for me to deploy as [11:21] a medium-siz OEM than it is for somebody [11:24] to say a Ford. Now, is for doing this? [11:27] Absolutely. But the scale, as you can [11:30] imagine, is a lot higher. So what I am [11:34] saying and I will take the advantage of [11:36] being the small guy and taking those [11:39] competitive wins where I can get them. [11:41] Um it is using AI to help our help solve [11:46] our problems and also take that data in [11:49] real time to solve problems upstream to [11:51] even eliminate the risk. It's catching [11:54] problems before they exist, solving them [11:57] in hours, not days, and feeding that [12:00] knowledge back into the design so they [12:02] never return. [12:04] The PDU case started as one small pilot. [12:08] Uh I actually spoke at this conference [12:11] last year. We were in the middle of [12:13] deploying the PDU case. Literally 12 [12:15] months later, we've now deployed it [12:17] across our entire manufacturing line. [12:19] That's how quickly it can evolve. [12:23] That wasn't luck. That was starting [12:25] small. That was proving out the ROI and [12:30] then scaling it. Not trying to deploy [12:33] and not trying to scale simultaneously. [12:36] If you take nothing else from today, [12:38] take this. Start small. Pick that one [12:41] line, one product, one use case. Measure [12:45] everything. If you cannot prove it, you [12:48] can't scale it. and move fast because AI [12:52] will be table stakes in manufacturing [12:55] and OEMs in just a few years. And if you [12:58] wait until then, you'll be chasing the [13:00] people like me who started it now. [13:04] So, let me leave you with this. Um, [13:06] because I know that I'm standing in [13:07] between you and lunchtime. [13:10] Um, every day that you're not closing [13:12] the loop with AI, you're letting your [13:15] processes relearn the same mistakes. [13:17] You're letting your people relearn the [13:20] same mistakes over and over again. If [13:22] your factory could talk, it would be [13:24] saying, "I told you this last month. I [13:26] did." So, let's give it a voice. Let's [13:29] make it smarter. And let's make sure [13:31] that the next time somebody saves [13:34] $140,000 [13:35] in minutes, that that's you. It's just [13:39] another Wednesday. So, thank you guys [13:41] very much for the time. Appreciate it. I [13:44] guess from a point of view of like cyber [13:45] security, how uh what kind of safeguards [13:47] are uh preventative steps can you take [13:49] in order to avoid bad actors um you know [13:52] in your industry? [13:54] >> Oh yes. Um that is the main topic and [13:56] why big OEMs are scared of AI. Um and [13:58] there is a massive risk right um hackers [14:02] hacked uh somebody uh several years ago [14:04] and that literally sent a shock wave [14:06] throughout the OEMs. Um it is modeling [14:10] and it's really just testing those [14:12] models. um trying to make it fail. When [14:16] we were putting it back into the models [14:18] and changing the parameters slightly, [14:19] that wasn't enough of a shock to get the [14:22] model to behave differently, to actually [14:23] change the behaviors. We actually had to [14:26] try to make it fail. And when we were [14:28] able to do that, that was enough to [14:30] trigger enough of a shift to make a [14:33] bigger difference than when we tried to [14:34] just tweak it slightly. Um and cyber [14:37] security uh parameters have to be in [14:40] effect at every single step. um how the [14:43] hackers were able to take over vehicles [14:46] was actually an open door that somebody [14:48] left in a component on the line. It was [14:51] not the vehicle itself. [14:54] >> You mentioned using warranty and after [14:56] sales information and then AI to package [14:59] that for design engineers. Would you [15:00] mind expanding on that? [15:02] >> Sure, absolutely. um you know our [15:04] vehicles um are again since I'm a [15:08] commercial OEM I don't have uh thousands [15:11] of sensors it it's always a balance [15:13] because my trucks are in service for 22 [15:15] years on average not seven um like light [15:18] duty folk or light duty cars so um mine [15:21] have to last a long time but like I said [15:24] I capture millions of data coming across [15:27] my canvas uh and shoving it upstream so [15:30] let's say we are seeing a potential [15:33] failure in our transmissions. I'm going [15:35] to see that in the data coming in from [15:37] the field. I'm going to see trends in [15:39] transmission temperature, fluid [15:41] temperature. I'm going to see [15:43] transmission um housing temperature. I'm [15:45] going to see uh shifts slipping, [15:48] clunking, things like that. That data is [15:51] fed in real time both to my field [15:54] service engineers to try and parse out, [15:56] you know, the why and then also to the [15:58] design engineer simultaneously to say, [16:01] hey, um, this is, you know, and and all [16:04] of that's fed into my risk factor model. [16:06] And so they actually uh assign a [16:09] severity to it that says, hey, this is [16:12] mission critical. This has a potential [16:14] of high failure. Um, so the field [16:16] service engineers get called to the [16:18] trucks to try to avoid the failure for [16:20] the customer. Meanwhile, the design [16:21] engineers are trying to problem solve [16:23] upstream to try and solve it [16:25] simultaneously. [16:26] Yep, good question. [16:28] >> Question over here. So, I don't know [16:31] what regulations [16:32] um apply, but number one, inspirational [16:36] that you guys were able to do this. [16:37] Number two, when you talk, it sounds so [16:41] amazing and quick. I don't know if it [16:43] was quick to put this all in place. So, [16:45] did you have any regulatory challenges [16:47] as you went through this process? [16:50] >> Oh, [16:52] regulatory compliance in our in my [16:54] industry. That is a rat's nest. But, um, [16:58] the good and bad thing about regulatory [17:00] compliance in anything automotive is [17:03] that it's behind. It can't keep up with [17:06] technology. There are no regulatory [17:09] compliance regulations around AI. [17:12] zero. [17:14] Until two years ago, there were three [17:18] related to electric vehicles. Three. [17:22] And how long has Tesla been around? So, [17:25] um, what we had to do was go out and do [17:27] our own research. Um, and a lot of it [17:30] because it is evolving so fast, we had [17:33] to shift on the fly to, uh, basically [17:37] mitigate risk because we make 80,000lb [17:40] vehicles. That's a lot of risk to human [17:42] beings, right? So, we had to do a lot of [17:44] that in real time. Um, and it it no, it [17:47] wasn't easy, but it took everybody on [17:49] the team being aligned with that goal [17:52] and working with lawmakers. I actually [17:55] addressed Congress, unfortunately, this [17:57] year talking about this issue [17:59] specifically because there's a huge need [18:02] um because there's all sorts of smaller [18:04] OEMs and god knows what's going on [18:06] there, right? And that can get pretty [18:08] scary pretty quick. Um but regulatory [18:11] compliance in automotive is severely [18:13] behind technology. Very good question [18:16] though.