Case Study: DOE for Soldering

I was asked by a black belt yesterday, who was in charge of a continuous improvement project for a factory in Dominican Republic. According to him, there seemed to be a problem in the soldering process which caused the solder bridges. He wanted to preform DOE to identify a main factor of the defect, and he already created a plan for the DOE.

He explained his plan to me with confidence, but the things didn’t look good.

According to him, the test process detected the solder bridge by turning on the circuit power (short circuit test). The test process collected the test results as binary values (Pass/Fail), and the defect rate was 0.05% or 34 defects per month (based on 3000+ production a month). The soldering process was manual by hands.

According to his DOE plan, the x factors (inputs) were:

  • Temperature of solder
  • Time duration of soldering
  • Size of soldering iron tip

And the y factor (output) was the test result (Pass/Fail) as a binary value.

He already determined the values for the DOE in factorial design.

I listened to him until here and said “No. It wouldn’t be a good way to do”. The blow was the reasons why I said “No”.

DOE for Binary Output (Pass/Fail)

The DOE should be used for the output of continuous value. It doesn’t work very well with the output of binary value. It’s because it couldn’t detect any meaningful change on the output value even with the full factorial design. To detect a meaningful change on the output value, many repeats of each run must be performed because of the defect rate of 0.05%. The total number of tests (repeats times runs) will be huge number. It would be unrealistic in terms of cost and time.

The outputs of repeated runs (i.e., the number of Passes and the number of Fails) can be converted to the ratio (continuous value). But again, to detect a meaningful change on the ratio value (output), many repeats of each run must be performed, and the number of tests (repeats times runs) will be huge number. It would be unrealistic in terms of cost and time.

My suggestion was to measure a continuous value instead of the Pass/Fail binary value. The suggested continuous values were:

  • Thickness of solder
  • Area of solder
  • Resistance of solder

If the DOE uses the continuous value as the y value (output), it can collected much more meaningful information from much less number of tests.

Binary Logistic Regression + Optimization

According to him and his plan, the actual data will be collected from the manufacturing process to save time and cost.

It maybe okay if each condition is well controlled, however, each operator’s condition sounded not under control. According to him, the x factors such as the solder temperature and the length of time was up to the operator.

The first advice to the process and the operators was to conduct MSA (Measurement System Analysis) and to implement a standardized work. After them, he can use the data collected from the manufacturing process, and the data can be used for the binary logistic regression instead of DOE.

The binary logistic regression generates two analytical equations (models):

  • Probability of Pass
  • Probability of Fail

Then, performing optimization of input values using these two analytical equations could give him an insight for improvement.

EVOP (Evolutional Operation)

Another suggestion was to use EVOP (Evolutional Operation) because the EVOP allows to operate manufacturing as normal, and it evolutionary an optimum point of conditions. But the EVOP also needs a continuous value as output.

Combinatorial Testing

The combinatorial test can be used to find a combination of input values which causes the defect.

In any cases, all x factors (inputs) must be well controlled, especially if he uses the data collected from the manufacturing process under operation.

Soldering Process Analysis