More Related Content
Similar to AWS及客戶在AI/ML的數位運行過程中得到的重要經驗與學習 (20)
More from Amazon Web Services (20)
AWS及客戶在AI/ML的數位運行過程中得到的重要經驗與學習
- 1. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
最重要的經驗學習從 AI/ML 商業化的過程
Most Important Lessons Learned
from Applying AI/ML in Real Business
Young Yang, ML Specialist SA
beyoung@amazon.com
- 2. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 3. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 4. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
start
with the
and work
backwards
customer
- 5. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 6. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 7. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Tradition Innovation
Inventory check Stop your
business
Real time
Check out Line up Just walk out
Clerk Value Store operation Customer focus
- 8. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Product life
begins at
Installation
- 9. © 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 10. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Tradition Innovation
- 11. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Tradition Innovation
Initial Building Cost High Fair
Warehouse Space Permanent Elastic
Slot Utilization Fixed Flexible
Routes Static Dynamic
Maintenance Cost High Low
Upgrade Hardware Software
Response to demands Awkward Agility
Reliability Single point of
failure
High availiability
- 12. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Tradition Innovation
Initial Building Cost High Fair
Warehouse Space Permanent Elastic
Slot Utilization Fixed Flexible
Routes Static Dynamic
Maintenance Cost High Low
Upgrade Hardware Software
Response to demands Awkward Agility
Reliability Single point of
failure
High availiability
- 13. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Technology
always comes from
nature
Human
- 14. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 15. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Voice-Enable All the Things
Source: MindMeld
TIME
2005 2010 2015 2020
0B
100B
200B
300B
A Massive shift in voice has already
begun.
• In 2014, voice search traffic was
negligible. Today it exceeds 10% of
all search traffic.
• Virtual assistants exceed 50B voice
searches per month.
• By 2020, over 200 billion searches
per month will be done with voice.
K E Y W O R D
S E A R C H E S
V O I C E
S E A R C H E S
WORLD WIDE SEARCHES
PER MONTH
- 16. © 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Alexa Made Voice the Mainstream UI at Home
“Alexa, call Jane”
“Alexa, order a
pizza”
“Alexa, Start my
TV”
“Alexa, lower the
temperature”
“Alexa, dim the
lights”
- 17. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Your
With customers
takes on-going work
Digital
Relationship
- 18. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 19. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 20. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Earn
of your customers
Trust
- 21. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 22. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
is the key to
optimizing process,
reducing cost and
improving customer
experiences
Data
- 23. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 24. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Data from Assets – The Foundation of Digital Twins
Unable to link
data together
96%
of industrial
state data is not used
Data collected
too
infrequently
39%
of Manufacturers do not
regularly collect data
Data difficult
to access
66%
of industrial companies
find data is difficult to
access
Why?
SCM World/Cisco “Smart Manufacturing & the Internet of Things 2015”
- 25. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
... resulted in this!
Operations (OT) Enterprise (IT)
IT Systems
CRM
Asset Management
ERP
Supply Chain
Finance
Maintenance
Compliance
Shopfloor
Single machine with
multiple components
following different
standards
Complete production
line likely to have
many machines with
different protocols
Challenge: Get
data from OT to IT
and make it usable!
- 26. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Operations (OT)
AWS IoT Helps Customers‘ OT to IT
Factory Machines
Enterprise (IT)
IT Systems
CRM
Asset Management
ERP
Supply Chain
Finance
Maintenance
Compliance
Protocol
conversion
Modbus
conversion
OPC UA
conversion
Gateway
Custom / Proprietary
Protocol
MQTT
- 27. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
ISA-95 in the context of the AWS Cloud
Level 1
Level 2
Line/machine
control
Animation
direct control
Level 3
Level 4
Description
Line/machine
supervision
Manufacturing
Operations
Management
Business
planning &
logistics
MES/
Historian
ERP/PLP/SCM
App/SystemFunction
Line/cell
execution
Business
operations
SCADA/HMI
Supervisory
control
DCS/PLC/RTU
Level 0
Physical
values
Raw data
event signals
I/O Sensor
AWS
Services
Enterprise
apps in the
cloud
Data
ingestion &
analytics
AWS
Greengrass
IoT Device
FreeRTOS
- 28. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
• Intel Atom® Processor
• Gen9 graphics
• Ubuntu OS- 16.04 LTS
• 100 GFLOPS performance
• Dual band Wi-Fi
• 8 GB RAM
• 16 GB storage (eMMC)
• 32 GB SD card
• 4 MP camera with MJPEG
• H.264 encoding at 1080p resolution
• 2 USB ports
• Micro HDMI
• Audio out
• AWS Greengrass preconfigured
• clDNN-optimized for MXNet
• Key Differentiators/Technologies
• Intel cLDNN Library optimized for MXNet
• Intel Deep Learning Deployment Toolkit
AWS DeepLens Specifications
- 29. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
AWS DEEPLENS ARCHITECTURE
Video out
Data out
I N F E R E N C E
D E P L O Y P R O J E C T S
Manage device
Security
Console Project
Management
AWS Cloud
Intel: Model Optimizer
cIDNN and Driver
- 30. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Brainstorming Modeling Teaching
Leverage Amazon experts with decades of ML
experience with technologies like Amazon Echo,
Amazon Alexa, Prime Air and Amazon Go
Amazon ML Solutions
Lab provides ML
expertise
Amazon ML Solutions Lab
- 31. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Defect Detection w/ML
Problem
FST seek yield improvement in their silicon wafer
factories. The current defect detection process is good
but it involve significant human inspection efforts.
Solution
FST would like to partner with AWS to push the yield
envelope with AI/ML. We conducted ML workshop and
hackathon to educate the teams on the latest AWS
technologies. We then worked together to create the
ML silicon defect detection model using tens of
thousands of wafer images for training.
Impact
We finally create a ML defect detection model with
99%+ detection rate (aka., Recall Rate), improved yield
and reduced the human inspection efforts by half.
“ AWS not only is the ML expert with advanced capable
tools but also our partner to show us how to use them to
improve our production operations.
AWS 不只是提供先進機器學習工具的專家,更是教導我們如何
在應用的合作伙伴。
Jason Lin
Chairman, Formosa Sumco Technology Corporation
- 32. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 33. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
- 35. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Quality inspection in manufacturing using
deep learning based computer vision
Speaker Name : 林文寬
Job title : 經理
Company/Org Name : 智邦科技
- 36. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
36
Agenda
智邦導入智慧視覺檢測背景介紹
導入前後架構比較
實例分享
如何應用AWS服務
- 37. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 38. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
智邦科技簡介
台灣新竹市科學工業園區
集團總部 代表著廣納工作夥伴及各方客戶夥
伴們的才智,藉由互動互助的夥伴
關係,發揮出最大的力量。
集智之樹
首次公開發行
1995 年於台灣證券交易所掛牌
上市(股票代碼: 2345)
約 1,395 名員工
員工人數
實收資本額:新台幣 55.28 億元
2017年合併營收: 新台幣364.46億元
(年增率 24%)
經營績效
研發中心據點:
台灣、中國、歐洲、美國
設計研發
慶祝智邦科技
三十週年
- 39. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
產業需求與挑戰
~ 提 升 顧 客 滿 意 度 ~
穩 定 品 質 & 準 確 交 期
慧智 造製
市 場 拉 力
國際大廠積極導入智慧製造,
智邦也要同步升級智造能力。
卡位全球產業鏈
因應少量多樣的生產模式,
智邦要持續精進製造能力。
追求完美零檢出
技 術 推 力
- 40. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
智慧視覺檢測整合智邦供應鏈發展計畫
智慧製造需要大量的資金、人力、技術等做為後盾…
BUT….
智邦能投入大量資源,不代表協力廠也能跟進!
協力廠若不參與,則智造力升級效果有限!
少量多樣的生產,不易蒐集累積資料量!
台灣工廠都面臨缺工問題,品檢人力也是!
AI 模型技術含量高,除了建置更要維運!
- 41. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
執行目標
AOI 判錯率約為 5 % (以總圖片數計算)
• 若以每片 PCBA 角度,幾乎每片都有錯判
• 實際 SMT 之良率約 9x %
每條生產線每天需要 3 個目檢員
及3個隨線維修員
• 目檢員每天需檢查約 50, 000 張圖片,並對不良之
圖片進行標記。
提高 PCBA defect 檢測準確度
減少目檢操作人員
Problem
Goal
智慧解決方案場域導入,資安架構規劃、訓練
- 42. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
智邦資安訓練內容
高階主管的資安危機處理課程
(Cybersecurity-leadership-program, CSLP)
監控與資料擷取系統(SCADA)與工業控制系統
(Industry Control System)安全保護已刻不容緩。
智邦提供訓練課程,教導合作業者學習如何保護
「監控與資料擷取系統(SCADA)與工業控制系統
(Industry Control System)」,包含瞭解網路安
全弱點如何被利用、網路攻擊方式、風險轉移策
略,以及工業控制系統抵擋網路安全攻擊的技巧。
讓合作廠商可快速將學習成果運用在工作環境。
智慧工廠常需做到跨廠區資料傳遞,對於網路
安全攻擊的危機處理能力已是成為管理團隊必備能力。
智邦提供訓練課程將透過說明網路安全風險的相關知識,
建立高階管理人員對於網路安全視野與情勢識別能力,
並設計不同產業的網路安全危機情境處理模擬演練(例如
:公共關係和媒體管理)。預期課後將有助於高階管理
人員,對於網路安全攻擊之應變決策與指揮能力。
監控與資料擷取系統(SCADA)資安防禦實作班
(Cybersecurity-SCADA-Engineer, CSSE)
- 43. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
- 44. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 45. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI影像判別
PASS
NG
END
AOI
Before : AOI Detection Process
Repair 人工複判
RePASSFail
- 46. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
Before - Defect Type
Group 1 : AI容易識別之
Defect type
Group 2 : AI不易識別之
Defect type
此階段資料分
類過亂,未統一
- 47. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
資料分析
• 收集13天的資料,約62萬張。
• 平均1條線,一天約有47,900張圖
片AOI認為是NG圖片
• 在62萬張圖片中,只有1285張圖片
是目檢員判定FAIL 只能降低OP
38% Loading
- 48. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI影像判別
PASS
NG
AI 判別
NG
END
(增加學習)
AOI
After - New AOI Detection Process
Repair
Local ML Server
人工複判
類別判別
Defect type
Group 2
Defect
type
Group 1
Fail
RePASS
- 49. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 50. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
圖片分析
目前標示的地方,人眼
很難判定是否Defect
AOI提供出來的座標圖與實
際之零件位置,並無法相
對應,通常是一個小框,
即使往外擴展,也不一定
會包含到完整之零件包含
更多之零件
標示錯誤,應該標直向
却標成橫向
Fig.1 Fig.2 Fig.3
- 51. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI’s ROI ≠ AI model’s ROI
Fig.1 Fig.2 Fig.3
*ROI (Region Of Interest)
- 52. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
Relabeling
• 針對前面所發現的問題,先將範圍限縮在容易識別之特徵類別C02(缺件),C08(零件位
移),C09(零件立碑),C13(零件反白),C14(側立)
• 因為立碑跟缺件特徵很像,最後將上述併成C02,C08,C13,C14四類
• 針對這四類重新label
反白 側立
Pass Pass
缺件 零件位移
- 53. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
初步驗證結果
正確率
高達98.98%
- 54. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
重新定義瑕疵種類
- 55. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C02 : 缺件
- 56. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C03 : Marking
- 57. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C04 : 極反
- 58. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C08 : 移位
- 59. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C09 : 立碑(可視為C02缺件)
- 60. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C11 : 引脚變形
- 61. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C13 : 反白
- 62. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - C14 : 側立
- 63. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - S01 : 短路
- 64. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - S02 : 錫少
- 65. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
瑕疵案例 - S07 : 空焊
- 66. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
主要解決的問題 – 五大類
C02:缺件
C08:移位
C09:立碑
C13:反白
C14:側立
01
C11:引腳變
形
04
C03 :
Marking
02
C04 :極性反
05
S :錫的問題
03
- 67. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI錯誤顯示設定定義
系統錯誤不良原因及代碼 AOI設備不良原因
不良代碼 不良原因 AOI錯誤顯示 AOI 錯誤顯示包含不良原因
C02 缺件 MISSING 缺件、移位、立碑、側立
C11 引腳變形 LEADFAIL 引腳變形
C03 錯件 WRONG_PARTS 錯件
C04 極性 POLARITR 極性反
S02、S03、S06、S07 錫少、空焊 INSUFFICIENT SOLDER 空焊、錫少、冷焊
S01 短路 BRIDGE 短路
C13 反白 UPSIDE DOWN 反白
說明:
1、AI針對AOI輸出不良原因僅接受一種錯誤代碼。
2、原AOI設備輸出偏移(SHIFT)、立碑(TOMBSTONE)均以缺件(MISSING)代表設備輸出。
3、原AOI設備輸出空焊(VOIDFAIL)、錫少(INSUFFICIENT SOLDER)以錫少(INSUFFICIENT SOLDER)代表設備輸出。
- 68. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI錯誤顯示設定圖示
- 69. S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 70. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
使用AWS服務進行AI Model training
Object detection
CNN
Data Analysis
- 71. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
EC2 Type
- 72. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AMI Type
- 73. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
Using GPU Instances
- 74. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
AOI Machine
Collect
Cleaning & Label
Training on AWS p2.8xlarge
(8 GPU)
Training Testing
Object Detection
Classification
Testing on AWS g4.4xlarge
(1 GPU)
Data
Augmentation
Defect
Code
●
●
●
●
●
●
●
●
●
●
Pass Fail
- 75. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
目前驗證結果
- 76. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
CONFIDENTIAL
Lessons Learned
Different inspection outcome by Human and Machine
Data imbalance
Efforts on relabeling
Label verified by single operator
Finding the right tool/ algorithm
Accton & AOI platform co-operation
- 78. © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
Thank you!
S U M M I T © 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Tim Lin
tim_lin@accton.com