Our Projects

Visual Inspection by Artificial Intelligence Model for Plating Mill

Executive

Objective

the objective of this report is to summarise our recent experiment in using Artificial Intelligence to help metal plating business inspect the products in QC process. This process usually is conducted by a group of well-trained staffs which contributed around 20 – 30% of total staff of the production line. We tried to find the best solution the the problems such as high spending on staffs, inaccurate results from staffs and etc.   

Address of the problems

1.Staff cost for inspection team

Spending on staffs always come with high cost and welfare, moreover when we have a lot of staffs we need to train them regularly which it is time consuming as well. This cost is tend to be higher over time. So having a clear plan in reducing staff cost is a must have for this business.

2. Unreliability of staff inspection skill

Quality control by using human always have inaccurate results due to uncertainty of sight seeing and mismatch opinion. Though the factory tried to implement double layer inspection, there are always mistakes happened to the inspection result.     

3. Product’s Quality Control Level

Inaccurate quality control process lead to reject products from customer and that also has an effect on our a huge loss of the business. With a low margin per unit that the business make on each order, just 5% product rejection could result into a huge lost at the end.

Objective

Proposed Solution

We proposed Visual Inspection Program using Artificial Intelligence as a solution to replace unreliable staff with more accurate Visual Inspection program. Furthermore, A.I. also reduce cost spending on staffs and time  consuming. We believe that our program will be the solution to the problems that most factories have on quality control process.

Implementation Process

  • Problem understanding, propose solution intuition
  • Initial Data collection for Model Training and solution validation ( staff taking pictures of finished product and upload onto A.I. module on-line) for validity check.
  • If validation pass, continue data collection for production as large dataset to slowly improve model accuracy (1st batch 300 pics). Staff need to help inputting inspection result into A.I. and keep feeding.
  • Model Training / Adjusting on large dataset for the best result and test accuracy. Ongoing improvement and adjusting to gain more accuracy along is important.
  • Model and Software delivery – once accuracy level is acceptable, factory will implement it in the production line.

Inspection staffs help training AI by marking color code onto the defecting area on our software. AI will keep learning and adding into AI’s neural network.   

Testing Result and Analysis

This testing round is to test the ability of AI Model in identifying flaw on final product as compared to the result of a group of inspecting staffs. 

  • factory’s staff collected 100 image of both OK and NG products and save as JPG files
  • Three staffs go through the same test and answer as OK or NG. System recorded all the answer for further process.
  • Giving AI model the same set of pictures and let the program answer OK or NG
  • Sum up all the answer and fill it in the data table
  • Calculate the accuracy percentage of an AI program by comparing to majority answers of three staffs

Inspection Result Table (sample)

product staff 1 staff 2 staff 3 AI majority answer staff 1 score staff 2 score staff 3 score AI score
product 1 OK OK NG OK OK 1 1 0 1
product 2 OK NG NG OK NG 0 1 1 0
product 3 NG NG NG NG NG 1 1 1 1
product 4 NG OK OK NG OK 0 1 1 0
total score 2 4 3 2

Next Step

Keep training AI with more dataset to improve accuracy. Next target would be achieving 70% accuracy.