Baidu PaddlePaddle easydl helps manufacturing enterprises with intelligent transformation

As industrial 4.0 Coming of age , How to use the sharp sword of artificial intelligence , Realize the transformation and upgrading of traditional production methods , Standing at the tide of a new round of industrial revolution , It has become a problem that every industrial manufacturing enterprise has to think about . Industry has a lot of data accumulation , Industrial production 、 Quality testing 、 Management and other links are continuing 、 A lot of 、 Quickly generate data , It is the blue ocean of artificial intelligence applications . The present , Represented by machine vision AI technology , Is being widely used in 3C Electronics 、 Food manufacturing 、 Automobile parts manufacturing and other fields , Including defect detection 、 Production environment, safety and other functions ,AI In the process of industrial intelligent transformation, high hopes are also placed on .

In the field of industrial quality inspection AI application

Manufacturing is inseparable from quality inspection . The products we are looking for , They all pass through the industrial quality inspection link before they leave the factory smoothly .

There is a strong demand for quality inspection in all walks of life

Quality inspection requires high precision , Accounting for% of the total labor cost of the factory 40%. for instance , Bearing defect detection target in industrial quality inspection , It could be a small scratch , Or it could be a small gap . In this case , Defective visual senses are not intuitive . In the whole process of human testing , It takes a lot of time 、 High manpower investment . The efficiency of quality inspection directly affects the production and delivery efficiency of enterprises . therefore , The intelligent empowerment of industrial quality inspection has become a cost saving , The inevitable trend of increasing production capacity .

Industrial quality inspection direction , Through the evolution of three processes ：

You can see from the above picture that , In the deep learning stage , It can gradually solve the problem of complex defect shape 、 Goals with complex environments , A wider range of problems can be solved , More scenes covered . But the threshold of deep learning is high , stay AI The development phase , Higher requirements for raw data , The requirements for developers are also higher .

that , Is there a simple way to get started and ensure efficient quality inspection ？

Car parts AI Quality inspection pain and difficulty

The current case enterprise comes from the solution provider in the direction of industrial bearing quality inspection —— Wesken , Have deep business scenarios and technical accumulation in the direction of bearing quality inspection ; But in AI Algorithm field , Lack of deep enough technical precipitation . In the process of intelligent empowerment , Defect detection problems encountered by enterprises , Mainly including internal material inspection 、 Size / Shape and position detection and appearance defect detection .

Defect types and solutions commonly used by enterprises

Internal material inspection ： Including material composition 、 Porosity and hardness test , At present, the main technology used is EM Electromagnetic technology and ultrasonic technology to do relevant testing ;

Size / Shape and position detection ： For example, check the diameter of the bearing 、 Whether the height and wall thickness meet the requirements , The current is through 3D Laser and micro magnetic field technology ;

Appearance defect detection ： Including scratches on the surface 、 Bump 、 Internal rust . Traditional machine vision cannot solve the problem of relative irregularity for the time being , Including defect location , Usually rely on manual detection .

The enterprise also tried to set up an algorithm team to do AI Development , Solve the problem of intelligent detection , But the algorithmic personnel needed to form an algorithmic team 、 The time cost of investing in the research process 、AI The cost of machines such as servers that need to be invested in training is accumulated , It is estimated to reach a million . The core appeal of the enterprise is to reduce the investment in the early exploration stage , utilize AI Enable appearance defect detection scenario , So as to improve the efficiency of the overall quality inspection link , The above is the demand background of the enterprise users .

As a solution provider of deep ploughing bearing quality inspection for many years , In the process of intelligent transformation, they encountered the following main problems ：

First , How to reduce the cost investment in the business exploration stage ？

second , How to accurately mark defects , So as to provide high-quality training data ？

Third , How to collect data of various defects , Make up for the pain point of less defective samples ？

Fourth , How to adapt various hardware , Simple and efficient deployment ？ How to ensure prediction efficiency ？

So how does the enterprise user pass Flying propeller EasyDL Solve the problem one by one , And get high returns ？

be based on Flying propeller EasyDL Create a finished bearing visual inspection solution

First , The defects of automobile bearing are analyzed , So as to preliminarily determine the need for application Flying propeller EasyDL What kind of model .

Combined with defect characteristics , Select the applicable task type

Determine the use based on defect analysis Flying propeller EasyDL Object detection and image segmentation model . Next, start data preparation around the defect detection of the end face → model training → Model deployment .

Data preparation

The defect to be tested is too small , Marking is difficult , At the same time, mark a large number , High labor costs . stay Flying propeller EasyDL On the annotation interface , Provide many zoom in or zoom out tools for targets with small defects , You can zoom as needed for accurate annotation . Here's the picture ：

In the face of a large amount of data , Intelligent annotation function can be adopted . Start smart annotation after a few annotations , It can intelligently analyze the marked pictures , Then mark the remaining pictures with one click . Take this enterprise as an example ,200 It takes time to mark the picture manually 3 Hours , The remaining 600 Smart tagging of a picture takes only 1 Hours .

model training

The sample size of some defects is small , How to improve data utilization ？ through Flying propeller EasyDL Data enhancements , One picture can be derived from more than one picture , Improve data utilization . meanwhile , Through the automatic hyper parameter search strategy , The utilization of data training in relatively complex scenes , At the same time, the optimal parameter combination based on this scenario can be produced , Achieve higher model accuracy . If the target detector is too small , You can choose a small target detection algorithm . The enterprise in this scenario , choice 800 A defect picture , No code training gives an accuracy rate of 90% Available models .

Model deployment

The problem encountered by enterprises is that the overall prediction delay will directly affect the quality inspection efficiency . Use EasyDL Provided model acceleration function , Compress the model volume without loss of accuracy , Reduce prediction delay . The enterprise compresses the model and deploys it in T4 Server , Single picture prediction can be made in 100ms Finish in .

meanwhile , Flying propeller EasyDL The output model hardware adapts widely , One click export of the platform to adapt to the mainstream hardware SDK The package completes the model deployment . For businesses , There is no need to do additional hardware adaptation , Significant savings in work costs .

Final , The enterprise built a platform based on Flying propeller EasyDL Visual inspection solution for finished bearings . be based on Flying propeller EasyDL Machine learning detection algorithm , Use the industrial camera to take pictures of the bearings on the production line , The geometric parameters of the bearing are obtained by the sensor and drawn into an image , The server at the production site carries out image classification and detection , Judge whether the appearance quality of the bearing meets the requirements , It can detect the scratch of the bearing 、 Bruise 、 Bruise 、 Chipping 、 Rust and other defects .