Approach to practical application of Deep Learning in manufacturer's production line Masahiro Kashiwagi, Hiroyuki Kusaka, Kiminori Kurosawa and Kenji Nishide 1 1
Applications Service Agriculture Advertisement Education Security Deep Learnig (AI) Financial Retail Manufacture Medical Logistics 2
Productivity improvement ○ Labor cost increase ○ Japan Working-style Reforms Reduction in long working hours Developed country Emerging country “ the Official Website of the Prime Minister of Japan and His Cabinet” http://japan.kantei.go.jp/privacy/terms_e.html Productivity improvement by Deep Learning (AI) 3
Company using AI US 13.6% JPN 5.1% "Ministry of Internal Affairs and Communications" ○ Using AI is still competitive advantage for many companies http://www.soumu.go.jp/english/index.html 4
Fujikura and AI Deep Learning ● Image inspection ● Data analysis ● System Automation ○ In Fujikura, we have actively studied applications of Deep Learning technology in various products 5
Fujikura and AI 6
Fujikura and AI NVIDIA Fujikura America HERE 7
Fujikura and AI ● Image inspection Deep Learning ● Data analysis ● System Automation 8
Products introducing image inspection Coaxial Cable FPC Optical cable Laser diode ○ High accurate image inspection using Deep Learning is very attractive in many products. 9
Fujikura and AI ● Image inspection Deep Learning ● Data analysis ● System Automation ○ Data analysis function is important in sensor system for IoT. 10
Energy Harvesting Sensor System ■ Energy-harvester = Dye-sensitized Solar Cell (DSC) Schematic Diagram Cloud Data Base Server 920MHz 3G/LTE Specified Indoor Sensor Internet Wi-Fi etc. Node Low Power FSN-2001N Radio communication Gateway communication FSN-2000S Browsing Outdoor Sensor Node ○ ○ 64 sensor nodes ○ ○ FSN-2002N-OD ○ ○ Temp.,Humidity,Illuminance,Motion,Pressure ○ ○ Gateway Indoor sensor node FSN-2000S FSN-2001N ⇒Accelerates implementation of IoT !
Fujikura and AI Deep Learning ● Image inspection ● Data analysis ● System Automation ○ Sensor system with data analysis function is good solution, sinze the size of measurement data is large. 12
Fujikura and AI ● Image inspection Deep Learning ● Data analysis ● System Automation 13
Fujikura and AI ● Image inspection Deep Learning ● Data analysis ● System Automation 14
Fiber lasers for material processing High power CW fiber laser High peak pulse fiber laser Cutting Welding Marking 15
Fiber laser cutting 16
Fiber lasers for material processing High power CW fiber laser High peak pulse fiber laser Cutting Welding Marking 17
Fiber laser marking 18
Fiber lasers processing sysem using Deep Learning ○ Adjusting processing conditions ○ Adjusting processing position 19
Fiber laser configuration and components Water-cooled plate Isolator ○ Fiber lasers are made of in-house components. 20
Importance of high accurate image inspection Image inspection Low accuracy Fail Pass Large cost Process 1 Process 2 Process 3 Process 4 Fail Image inspection High accuracy 21
Laser dioede production process Pattern inspection Facet inspection Epitaxy Patterning Cleavage Packaging Testing Coating 22
Facet inspection Microscopic Image Front facet ・ Particle Failure modes ・ Clack ・ Scratch ・・・ 23
Training Images 4 Classes Class1 - Pass 10,000 images Class2 - Failure mode 1 2,000 images Class3 - Failure mode 2 1,000 images Class4 - Failure mode 3 1,000 images 24
Convolutional Neural Network Convolutional Convolutional Convolutional layers layers layers Input image Max pooling Max pooling Max pooling ... ... ... Convolutional Convolutional Fully-coneted layers layers layer Output Max pooling ... ... 25
Training results ○ Image size : 400 x 400 pixcels ○ 15 convolutional layers GPU : NVIDIA GeForce GTX TITAN X Calculation time : about 20 hours 26
Progress of training model Particle Particle Input image (Failure mode) 1st epoch Highlighted region Highlighted region 5th epoch 10th epoch 27
Test results Image size 200 x 200 300 x 300 400 x 400 pixcels pixcels pixcels 5 convolutional layers 95.7% 98.4% 99.1% 10 convolutional layers 95.6% 98.4% 99.2% 15 convolutional layers 97.1% 98.5% 99.6% Skilled worker (2000 x 2000 pixcels) : 97-98% 28
Heatmaps Particle (Highlighted region) Particle (Highlighted region) ○ Highlighted regions are in good agreement with particles. 29
Laser dioede production process Pattern inspection Facet inspection Epitaxy Patterining Cleavage Packaging Testing Coating ○ We have also be developping pattern image inspection using deep learning. 30
Products introducing image inspection Coaxial Cable FPC Optical cable Laser diode ○ The image inspection method is also being applied to other products 31
Sumary ○ We have developed laser diode facet Image inspection method using Deep Learning technology. ○ Test accuracies in image size of 400 x 400 pixcels are ○ Large cost reduction and high productivity would be expected. more than 99%, which are higher than skilled workers ○ We will develop another application of Deep Learning ○ We are recruiting new AI research team staff. technology. ai-info@jp.fujikura.com 32
Recommend
More recommend
