In the era of Industry 4.0, on the one hand, everyone is often shocked by the extremely advanced fully automatic production lines of industry leaders. On the other hand, they often lament that their own companies cannot keep up with the times due to limited funds, and their operating efficiency is not high. In fact, with the right methods, you can also help your own business speed up and increase efficiency.
What are the production models in the manufacturing industry? What's the point?
In order to help everyone understand how to use advanced analysis methods to help enterprises speed up and increase efficiency, let's talk about how to use the AI platform in JMP software to improve the level of enterprise operations and achieve the desired results through a real case in actual work. Unexpected effects. I hope it will be of some reference and help to you.
Has your business encountered such a problem?
The high level flow chart below represents the manufacturing process of some electronic companies. in:
- A- 'produce the product' represents the process of manufacturing a product through a certain technology;
- B-'Electrical test process' represents testing various electrical parameters of products through large-scale automatic testing equipment;
- C-'Human inspection and binning' represents the process of manual judgment (including appearance defects) and classification;
- D-'SPC system' represents the enterprise's existing statistical process control system;
- E-'Call for action' represents the relevant measures required to respond to the SPC alarm;
- F-'move to next' means that if the SPC result is normal, it can be released to the next step.
The order of information flow is that the results of B-'Electrical test process' (hereinafter referred to as station 
are transmitted to C-'human inspection and binning' (hereinafter referred to as station C). Station C manually performs the inspection based on the results of station B and its own judgment. To further classify different substandard product types, station C further transmits the classification data of different substandard product types to station D as input to the SPC system and makes judgments for the next step of product disposal.
The problems exposed in the process are very obvious. The manual judgment of station C not only occupies a large amount of human resources and introduces the factor of personnel fatigue, but also the processing cycle is naturally limited by manual operation time. Recalling the core competitiveness of manufacturing companies we just mentioned, this is tantamount to extending the delivery cycle, increasing product control risks and increasing corporate operating costs. So, is there some possibility of using a computer to illustrate the steps to achieve this? If so, how?
Start research and test your hypothesis
Since station B generates a lot of electrical test data, our hypothesis is whether the data generated by station B can be used to predict the manual classification results of station C. In fact, many of the electrical parameters of station B, such as signal-to-noise ratio, are affected by the accuracy of inductors, resistors, capacitors, etc. during assembly. Therefore, theoretically, the test results of station B can predict assembly to a certain extent. appearance defects. To this end, we first selected the test data of station B in the past six months. It should be noted here that only the electrical defective product data of station B was selected (the reason is that for this case, the defective product data is the valuable information that needs to be extracted. As long as one of the electrical tests at station B fails, it is a defective product). Each row in the table represents the information of a single product. Use the 'Join table' function in the JMP software to combine the data at station B (Table B ) and the historically manually classified data (Table C) of site C are connected together through Product ID.
Product ID in Table B is serial data, Test1 to test N are project data, including the pass/fail results of each electrical test at station B. Reject Type in Table C is project-type data, which contains the results of manual classification at station C. (Due to the huge amount of data, some screenshots are as follows)
Click the join function below Table to enter the following window, select Product ID as the primary key to connect Table B and Table C.
In this way, we combine the two tables together. In order to facilitate everyone's understanding, the test items of Test19 and later are hidden. Please pay attention to: At this time, there is a Reject type column, which is the manual classification result of station C.
The next step is to make predictions, so which method should we choose?
Why choose neural networks?
Relying on JMP's powerful 'Predictive modeling' platform, we can access algorithms that are commonly used in the industry. (Although JMP itself does not include algorithms like Tensor flow, etc., JMP can directly call the output results of R or Python. Interested friends can consider it for their own projects, but this does not affect the utility of JMP, because You will save a lot of time writing programs). Such as k-Nearest neighbor, Naive-Bayes, Bootstrap Forest (Decision Tree) and Neural networks.
In this case, we chose to use the neural network algorithm. The main reason is that the neural network algorithm has won many awards in the World Algorithm Competition in recent years. People are more and more interested in learning this algorithm, so the application This case may be more instructive for everyone's practical work. Secondly, because the amount of data in electronics companies is generally huge (components are getting smaller and smaller), neural network algorithms will improve their accuracy as the amount of data increases, maximizing the various possible combinations of linear or nonlinear approximations. (Like a chessboard being laid out), but other algorithms tend to hit a ceiling in accuracy, although they may be useful when dealing with small amounts of data. Third, other algorithms have many usage restrictions and assumptions. For example, Naive-Bayes assumes that all prediction variables are independent of each other. Once this assumption does not hold in actual cases, the effect will be greatly reduced.
Predictive modeling, validation analysis
After selecting the neural network, we will officially enter the analysis. First, we need to set the ratio of training data and validation data. Based on experience and the characteristics of the project, here we select 75% as training data and the other 25% as validation data (of course you can also use it according to your own Choose other ratios depending on the characteristics of the project, such as 80% training data and 20% verification data. There is no hard and fast rule on the ratio itself). After selecting Analyze ->Predictive Modeling->Make Validation Column in the JMP software menu, the table will automatically add a column of validation according to 75% Training Set and 25% Validation Set, as shown below:
Note that there is an extra last column of Validation at this time.
Click on the embedded formula in the Validation column as follows, and you can see that JMP actually uses the Random Category function.
According to the table after JMP automatically assigns the Validation column, we select Analyze > Predictive Modeling > Neural. The Neural startup window is shown below:
Then we come to the Model Launch console. The most troublesome thing here is how to set the number of hidden layers (in this example, only the 1st hidden layer is set to reduce the complexity of the model).
The empirical formula is set as follows, where N represents the number of input layer nodes, N2 represents the number of 1st hidden layer nodes, and M represents the number of Output layer. After calculation, the number of 1st hidden layer (N2) is 31. The learning efficiency and penalty coefficient use JMP default values. You can refer to the following hierarchical arrangement diagram:
Here I share with you an empirical formula for your reference.
Number of Hidden=N2
Number of Input=N
Number of Output=M
Number of Hidden Points: N2 < Product count - 1
Number of Inputs: N < Product count- 1
N2=sqrt(N+M+1)+a (a=1 to 10)
Plugging in the values in this example, we get N2 as 31.
N=734 M=6
N2=31
The final calculation result is as follows, the Generalized R Square of Validation is greater than 0.8.
Run ROC Curve: Predictive and Specialized Modeling > Partition Models > Partition Platform Options > ROC Curve.A value of 1 in the ROC (Receiver Operating Characteristic) curve represents a perfect fit of the model, and a value of 0.5 represents that the model is almost unable to distinguish differences between groups. Judging from this figure, the areas of all curves (different colors represent different types of defective products) are greater than 0.8, which is in line with expectations.
Interpret results and expand applications
After bringing the Validation results into the calculation (as shown in the table below), you will find that the neural network algorithm can achieve more than 70% agreement with manual classification for key types of substandard products (A, B, C), and for C The degree of match between categories is as high as 95%. The production environment based on this case can be used for daily key SPC control input to assist product monitoring. In other words, when the electrical test at station B ends, if an abnormality occurs, the control chart will alarm in time.
Actual
|
Predicted Count
|
|
Reject Type
|
Reject A
|
Reject B
|
Reject C
|
Reject D
|
Reject E
|
Reject F
|
Total Count
|
Match Rate
|
Reject A
|
786
|
228
|
53
|
25
|
16
|
6
|
1114
|
71%
|
Reject B
|
144
|
1178
|
62
|
3
|
9
|
4
|
1400
|
84%
|
Reject C
|
30
|
14
|
947
|
3
|
2
|
1
|
997
|
95%
|
Reject D
|
104
|
34
|
2
|
25
|
14
|
1
|
180
|
14%
|
Reject E
|
53
|
11
|
4
|
9
|
30
|
0
|
107
|
28%
|
Reject F
|
64
|
2
|
0
|
9
|
2
|
3
|
80
|
4%
|
It seems that the coincidence rate of Reject D/E/F type substandard products is low. why? After investigation, it was found that Reject D was mainly defects related to foreign objects on the product, Reject E was mainly defects related to product labels, and Reject F was mainly defects related to the outer packaging of the product. Theoretically, these three types of defective products are not caused by the assembly technology of the previous station, and it is difficult to accurately judge them using the electrical parameters of station B. Therefore, at the end of the process, the coverage of automatic appearance equipment inspection was expanded from random inspection to 100% coverage of all products, so that SPC monitoring of Reject D/E/F categories can be omitted (because SPC monitoring is very important) The purpose is to determine whether good products have problems through monitoring. Since 100% of products can be automatically inspected, there is no need to conduct SPC monitoring for these three types of defects at station C).
The interesting phenomenon is that some of the classification results of Class D, Class E and Class F neural networks are classified into other substandard product types with higher accuracy (in this case, Class A). Does that mean that we should believe it more? How about checking the factors of manual misjudgment (fatigue, unclear vision, etc.) based on the model judgment results? Imagine if a product has both A defect and one of D/E/F defects, the artificial intelligence will naturally make a judgment based on the visual results that are easiest to find, but in fact it may be the more important Class A defect. . I believe everyone will have different answers in their own workplaces, but in any case, the cost savings and benefits brought to enterprises by streamlining processes have far exceeded those trivial purchases of automation equipment or professional tools to improve production process capabilities. Cost expenditure.
Therefore, based on the above analysis results, the cancellation of station C not only shortened the detection time of critical quality defects, but also saved a lot of labor costs (in the later stage, the model can be refreshed regularly and connected to the SPC system software, and the model will update the data regularly. To correct the model, the update cycle can be done every day). Such a result that kills two birds with one stone is welcome to both engineers and management.
Of course, the result you see now is not achieved overnight, but is the result of our repeated attempts, practice, and continuous optimization and improvement. It condenses the strength and hard work of the team. I hope we have gone through it in the dark. Detours can illuminate the path of your work practice and provide you with some navigation and direction.
