Chapter 1107 I have three methods, three!
It is of course impossible to say that Chang Haonan can completely understand this technology in just a few dozen minutes.
It’s just that he is better at refining the essence of the problem.
With this sentence just now, the subsequent research direction has been basically determined.
That is, error modeling is combined with adaptive compensation in the cutting part, while floating clamping is used in the clamping part to minimize the post-processing deformation caused by the fixed area.
Of course, even so, the accuracy of picking is still not as good as the handwork of a master.
But this is a necessary step to allow the aerospace industry to develop to real scale.
In fact, Yang Weihua also raised concerns in this regard.
Chang Haonan’s answer was the same as in the previous meeting at Haofa Group -
You can carefully select from a batch of products, distinguish batches with higher consistency, and install them separately. to the same engine.
This seemingly wild idea is almost subversive for Yang Weihua, who has a background in mechanical processing.
He stopped packing things and froze in place for a long time:
“After I entered the industry, the first thing my master taught me was standardized processing... Now it seems , but it’s a bit behind the times..."
The tone was full of emotion.
Given Yang Weihua’s age, he should have entered the industry in the early 1980s, which happened to be the second round of large-scale upgrading of the Chinese machinery industry.
It's been about twenty years now.
You know, in the early years when our country's processing accuracy was not good enough, this method of selecting qualified products from a batch of products was implemented.
It is through the efforts of their generation that a relatively complete standardized processing system has been established.
Now we have to turn back.
The classic server is now open.
"Standardization, of course, is correct."
Chang Haonan also felt the other party's emo and replied with a smile:
"But for truly high-demand production, no matter how precise it is, There will always be errors in the production method, so in terms of processing results, there is no standardization in the strict sense. "
"In this case, we might as well change our thinking and implement standardization in the production process field. , using production efficiency to make up for the lack of product consistency to a certain extent..."
In this era, most people's understanding of "high-level production" is relatively simple.
High precision and good replaceability.
Still the same thing as in the 1980s.
But in fact, let alone mechanical manufacturing, even in the field of semiconductor processing, there are random errors.
There is no shame in special selection.
The key is to be able to figure it out...
After some enlightenment, Yang Weihua left the conference room thoughtfully.
But Wei Yongming seemed to have no intention of leaving.
Even the laptop on the table was confiscated.
"Yongming, is there anything else?"
Chang Haonan also stopped when he saw this and sat back on the chair.
He and Wei Yongming are similar in age, and they are old friends who have known each other for a long time, so they call each other more casually in private.
And Wei Yongming did not talk nonsense and went directly to the topic:
"I just carefully analyzed the clamping method proposed by Lao Yang. If it is refined into a multi-objective optimization problem, then the decision-making Although there are only three variables, the objective function is completely more...especially for workpieces with relatively large bending and torsion degrees, there must be at least four or more." Chang Haonan did not speak, but nodded and said yes.
This natural look almost made Wei Yongming lose his confidence.
However, after calming down, he continued to speak:
"But... multi-objective optimization using genetic algorithms is generally based on Pareto control. As the number of targets increases, the number of targets in the population increases. The number of non-dominated solutions will increase exponentially, causing the algorithm's search capability to deteriorate rapidly..."
"I have previously tested the non-inferior sorting genetic algorithm's solution to a four-dimensional objective optimization problem... Even after many rounds of adjustments, In the end, it is all divergence...or a solution with poor accuracy, and the calculation time is very exaggerated. It may not matter to our group, but when it comes to implementation, the production side may not have such high-level computing power support..."
Wei Yongming said, opening a PDF file from the computer, then turning the screen towards Chang Haonan.
The latter took a brief look and found that it was a fairly typical grid search problem.
It seemed like some kind of random test.
“That’s true.”
Chang Haonan pushed the computer back:
“From an intuitive geometric point of view, an optimization problem with an M-dimensional objective, It is equivalent to dividing each dimension of the target space into r grids. Assuming that the non-dominated solutions to the problem are uniformly distributed, then the boundary of the Pareto front of the grid constitutes the upper bound of the number of particles. ”
He said and drew a three-dimensional diagram on the paper.
"Obviously, for an optimization problem with M objectives, M*r^(M-1) solutions are needed to form the Pareto front. For an optimization problem with 5 objectives, even if we set r to a relatively low 25. It still requires nearly a million solutions to describe the basic Pareto front... It is normal for some unexpected calculations to cause divergence, or for several months to be unable to provide results."
< br>Wei Yongming breathed a sigh of relief.
Obviously, it’s not that his question is too simple, it’s just that Chang Haonan seems to be mentally prepared:
“Then the next research direction is...”
Unexpectedly, Chang Haonan shook his head directly.
"Not sure yet."
Wei Yongming was immediately surprised.
However, before he could organize his words, Chang Haonan continued:
“I have initially thought of three technical routes, but I don’t know yet which direction is more suitable... ”
“……”
The former wiped the sweat that had just appeared on his forehead and swallowed back the words that were already on his lips:
"Tell me, if it doesn't work, we can all try..."
< br>This time, Chang Haonan did not write anything more, but took out a notebook that he had prepared long ago:
“First of all...the individual objective function value can be calculated according to a certain proportion. After zooming in or out, comparing with other individuals is to relax the Pareto dominance relationship, so that the advantages and disadvantages of some non-dominated individuals can be compared, and the search ability of the algorithm can be enhanced. "
Wei Yongming's thinking also changed. Very quickly:
"Using this idea can improve the differential evolution algorithm...but it seems that it can easily lead to individuals falling into cyclic dominance?"
"So it needs to be improved, and the differential evolution algorithm is not particularly suitable for this kind of problem...In short, we We will study it in detail later.”
Chang Haonan waved his hand, and then turned the notebook to the next page:
“Performance evaluation indicators can also be used as the objective function to guide the search instead of the Pareto dominance relationship. Process... But this type of algorithm can only get one solution every time it runs, requires multiple calculations to get the optimal solution set, and can easily fall into interval optimality."
"In the end, pair search is not considered. Goals that have little impact on the process or final optimization results, that is, reducing the dimensionality of the objective function. I am personally optimistic about the prospects of this technical route, but there are many problems that need to be overcome mathematically..."
Hearing this, Wei Yongming couldn't help but complain in a low voice:
"This is a good thing..."
However, Chang Haonan, who has a sharp ear, caught it:
"You can't say that."
"After all, you still have to consider the user's feelings... If it is too complicated at the algorithm level, you will inevitably encounter obstacles in future promotion..."
(End of this chapter)