Baidu PaddlePaddle easydl x wesken: see how to install the "eye of AI" in bearing quality inspection

With AI The landing of technology in the field of Intelligent Manufacturing ,AI Quality inspection is experiencing rapid development .IDC《 China AI Market analysis of enabled industrial quality inspection solutions 2021》 According to the report , from 2019-2024 year , The compound annual growth rate of the industrial quality inspection software and service market will remain at 30% above .

at present AI Visual inspection in 3C The electronic field is relatively mature , But it is less used in auto parts . Take the bearing industry as an example , The scale of China's bearing industry has ranked third in the world in terms of total bearing volume . However , In the appearance inspection of finished bearings , At present, it basically depends on labor ,AI Quality inspection can only “ Look and sigh ”.

Three difficulties  

The war of bearing visual inspection

Bearing is an important part of modern mechanical equipment , The main function is to support the mechanical rotating body , Reduce the friction coefficient during movement , And ensure its rotation accuracy . In the case of cars , The power of the car 、 reliability 、 Safety and comfort are closely related to it . therefore , It is very important to ensure the yield of factory bearings .

Even though AI Quality inspection in 3C、 Semiconductor and other industries have achieved large-scale applications , But in the field of automotive bearings ,AI The road of visual inspection is a little difficult . Bearings are basic industrial components , Its length can be measured by eddy current technology 、 Thickness, etc , But scratches 、 Bump 、 Black skin 、 The detection of appearance defects such as missing machining basically depends on manual . The reason is , We found that , The bearing industry should realize AI Visual inspection , We need to face the following three problems ：

First of all , Testing is difficult to standardize . Usually , The quality inspector observes with the naked eye 、 Pick out the greasy dirt on the surface by turning the bearing with your fingers 、 Scratches and other defects . in other words , Bearing appearance inspection often Depend on people , To process 、 The recognition of requirements largely depends on the judgment of industry experience , Various types of defect data need to be collected .

second , The detection environment is not friendly . as everyone knows ,3C The electronic manufacturing industry requires the use of dust-free workshops , by AI Quality inspection provides a good environment . However , The production environment of automobile bearing parts industry is much worse , Even if it is claimed to use “ Dust free workshop ” Production of finished bearings , Its environment is only relatively clean and dust-free , But not with 3C The dust-free workshop in the electronic industry is comparable .

Third , It is difficult to collect data . There are many kinds of bearing defects , Such as end face defects 、 Outer diameter defect 、 Inner diameter grinding burn, etc , Some flaws are even difficult for ordinary people to observe with the naked eye . In this case , The traditional vertical camera is difficult to shoot , The camera needs to be photographed by tilting the light source technology , This poses a great challenge to shooting .

in fact , The false alarm rate of visual inspection in the bearing industry is generally required to be controlled within 10% within , And eddy current 、 The false alarm rate of ultrasonic testing is 1%-2%, This also shows the difficulty of visual inspection .

A work of ingenuity  

Control AI Just four steps.

Wesken （ Xiamen ） Intelligent Technology Co., Ltd. is a company specializing in the design of non-destructive testing system for automotive parts 、 production 、 Service company , Facing the problem of bearing appearance defect detection , Embrace actively AI New technology , Finally, based on Baidu PaddlePaddle EasyDL Zero threshold AI The development platform has developed a set of visual inspection solutions for finished bearings , It not only realizes the transformation and upgrading of the company's technology , It also helped it get a big order from renben group, a leading domestic bearing enterprise , Effectively enhance the core competitiveness .

that , How did wesken manage successfully AI This rookie ？

Actually , Wesken is choosing Baidu PaddlePaddle EasyDL Before the platform , Once set up a team to develop algorithm technology , The investment cost is nearly one million , But after more than a year of exploration , Subject to personnel ability and algorithm technology threshold , It is found that the effect of the algorithm is not ideal , It is difficult to meet the accuracy requirements of customers on the production line .

Final , Based on Development cost 、 Processing efficiency 、 Algorithm accuracy 、 Deployment adaptation And so on , Wesken chose Baidu PaddlePaddle EasyDL Model development , Through application analysis 、 Data tagging 、 Model development 、 There are four steps to model deployment , It's done AI Development and application of the model , Not only the cost of model development is reduced compared with the original traditional method 90% above , Wesken also benefited from this , adopt AI Enable quality inspection , The cost of quality inspection is as high as 70% The fall of .

Bearing marking diagram

Application analysis stage ： Combined with defect characteristics , Find a solution . Because the detection target is small , There are many external factors , According to the specific defect location , Size, shape, spacing, etc , Such as end face defects 、 Outer diameter defect 、 Grinding burn of inner diameter 、 The image segmentation model is established based on the texture of ferrule turning tool , Refine defect samples and types .

Data annotation stage ： When marking, it will be difficult to mark because the defect to be detected is too small , Or it is necessary to note that there is a large amount of data 、 Manual marking takes a long time 、 High labor cost （ Usually thousands or tens of thousands of pictures need to be marked ）; And the propeller EasyDL Powerful intelligent annotation function , Manually mark 200 After a picture , Automatically mark the remaining pictures through machine learning , Greatly improved efficiency .

Model development stage ： Baidu PaddlePaddle EasyDL Zero threshold AI The development platform can provide automatic adaptation solutions in combination with business scenarios , It can meet the demands of model training with high accuracy without algorithm foundation . It turns out that , Wesken only uses 800 Piece defect data , The accuracy rate can reach 90% Available models .

Model deployment phase ： Deployment is also an issue that must be considered , Deployment hardware varies , The adaptation workload is too large , The data prediction delay will affect the quality inspection efficiency . Baidu PaddlePaddle EasyDL The output model hardware adapts widely , Export the model with one click through the platform SDK Package can complete the model deployment ; And the model acceleration function it provides , The volume of the model can be compressed without loss of accuracy , Control single picture prediction to 100ms Finish in .

in fact , Based on Baidu PaddlePaddle EasyDL Zero threshold AI Development platform , From data annotation to model validation , Wesken only needs one engineer for one month to eliminate the missed inspection caused by human errors , And according to customer needs , Quickly change test items and modes . The false alarm rate of visual inspection is controlled at 2%-3%, Far below the industry requirements 10%.

“ the 14th Five-Year Plan ” The development plan of intelligent manufacturing is clearly put forward ,2025 The penetration rate of industrial Internet platforms should reach 45%. For a long time to come ,AI+ Industrial Internet will become the main theme of China's manufacturing industry . Accelerate the development of artificial intelligence technology , advance AI Integrated application with industry , From single point equipment to production process , To the enterprise 、 To the industrial area , Baidu PaddlePaddle EasyDL Committed to opening up the industry AI The last mile of landing , Let the manufacturing industry successfully embrace AI.

AI Quality inspection is just a starting point , But it's not the end .

notes ： Flying propeller EasyDL Model capability is provided by Wenxin big model