The military-industrial scientific research system of the academic master.

If these rumors about the source of the turbofan 10 technology were only circulated in China, they would essentially not exceed the concept and influence of a small essay.

But it would not be a bad thing if foreign reports, and serious reports at that, were also involved.

In fact, for an engine that may soon enter mass production and service, basic information such as structural design cannot remain unknown for a long time.

After all, you have to issue maintenance manuals and other technical information to grassroots units, and when these things arrive at the grassroots level, it will be impossible to prevent them from leaking as time goes by.

The reason why the outside world's perception of the Turbofan 10 is so biased is mainly because Chang Haonan's movements are too fast.

From determining the design plan, to subsystem design, to component-level testing, bench testing and current installation testing, it was all completed in an intensive period of less than three years.

In particular, the installation tests involving a large scope were even concentrated in the majority of the year from the beginning of 1999 to the present.

Moreover, everything from design to testing was done in one go, with no rework process at all, minimizing the number of insiders.

In other words, for those who have seen what Taihang looks like, let alone retired or retired, they may not even have time to take their annual leave.

The entire project operates almost in a semi-closed environment.

There are naturally fewer opportunities for leaks.

For the outside world, they are also accustomed to the research and development cycle of aircraft engines that often takes ten years or even longer.

The Turbofan 10 has only been three years old and there is no new news. Under normal circumstances, no one would think about tearing it down and starting over. Now that new engines are coming out, unreliable ones are naturally pure guesses, and even if they are reliable, they are completely guessing. Analysts can only shoot arrows first and then draw targets based on the only clue CFM56.

Of course, from a technical perspective, general information such as the overall design plan is actually not that important.

For example, the overall design of the CFM56 is basically known to any ground handler who handles Boeing aircraft.

Even the engines themselves are sold all over the world along with the passenger planes, and you can dismantle them and study them if you want.

But I haven’t seen anyone able to reproduce it exactly as it is.

After all, if it were so easy to build an aircraft engine, then the aircraft engine would no longer need to be the crown jewel of the industry.

But if we look specifically at Turbofan 10, we can take advantage of this information gap——

Since foreigners think that our model is developed from the CFM56 core machine, and their analysis is clear and logical, then we will follow them.

He neither directly admits nor denies it. When asked, he hints three times:

Everyone understands, no need to explain too much, just take a closer look.

In this way, at least there will be a round of werewolf killings within the Western camp - if the Chinese people are so slippery with CFM56, it must be that we have a traitor who sold the relevant technical information.

In the end, even if the matter goes away, or the actual situation of the Turbofan 10 is exposed in a few years, it can still have the effect of destroying the internal relationship between the opponents.

Moreover, the fact that the French approached the Aerospace Power Group for cooperation at this sensitive time point is highly likely to be related to this.

After all, CFM Group was jointly established by Snecma and General Electric.

It's just not clear yet whether the other party wants to investigate the "technical leakage" or has some other purpose...

Thinking of this, Chang Haonan picked up the phone again and dialed the internal number of Zhang Liangping's office...

……

The task he assigned to Zhang Liangping obviously could not be completed within a day or two. Therefore, after dealing with this small emergency, Chang Haonan returned to school as planned and began to develop a specific manifold learning algorithm. .

Compared with the purely theoretical paper he submitted to the Annual Journal of Mathematics, this is the direction he focuses on.

Before the National Day, Chang Haonan had sorted out two basic algorithm ideas, and Yao Mengna and he each chose one to continue researching.

Although the results he conceived on the spot may not be the optimal solution in one step, they are at least representative enough.

The first type is a global idea, which maps adjacent points on the manifold to adjacent points in low-dimensional space during dimensionality reduction, while ensuring that distant points on the manifold are mapped to distant points in low-dimensional space.

The second type is a local idea, which only needs to ensure that close points on the manifold are mapped to adjacent points in the low-dimensional space.

In comparison, the former is more intuitive (of course only relatively intuitive), but the computational complexity is very high, which poses certain challenges to both hardware level and algorithm design.

The local idea is more abstract, and the correspondence between distant points is not clear, but the calculation amount is relatively small, and it seems to be more suitable for the current computer performance.

But this time, it was Yao Mengna who took the initiative to find Chang Haonan a few days later.

However, it is not because the former has constructed the algorithm according to the overall idea.

In other words, I did come up with an algorithm, but found that I had reached a dead end.

"Mr. Chang, I used the constructed isometric mapping algorithm to perform data point generation optimization tests on the two-dimensional manifold [t, s, X] in the three-dimensional space."

Yao Mengna put a few pieces of paper on Chang Haonan's table:

"For complete surfaces, the efficiency of the algorithm is quite good, and the generation coordinates of the complete S-surface are basically restored."

“But if I dig out a square area with a length and width of π on the two-dimensional manifold, which is equivalent to opening a hole on the surface, which is a very common situation in practical applications, then the generated coordinates will be distorted. , causing the area of ​​the cavity to become larger and become an approximately elliptical area..."

"..."

To put it simply, it just doesn’t work.

"The existence of holes in the manifold means that the subset of Euclidean space equidistant from the manifold is non-convex, and the deviation produced when calculating the shortest path between sample points on the manifold increases..."

The problem discovered by Yao Mengna is also an area that has not yet been studied by Chang Haonan.

Fortunately, the overall idea is relatively intuitive, so he can analyze it on the spot.

"In other words, if you want to use the isometric mapping algorithm, or to expand it, use the global algorithm, then the manifold object must satisfy the conditions of being isometric to a subset of the Euclidean space and the subset being convex."

Chang Haonan paused the ballpoint pen in his hand and finally concluded.

After all, the algorithm itself was optimized and modified by Yao Mengna, so she followed Chang Haonan's ideas this time.

"so……"

Yao Mengna looked troubled:

"Did you know before that this road was dead?"

"Ahem...that's not the case."

Chang Haonan immediately denied:

"I just thought of it after listening to your explanation..."

"Actually, I have been studying how to improve the local linear embedding algorithm (LLE) during this time."

He opened his computer as he spoke, then took out a piece of paper and laid it next to the keyboard:

“The biggest problem with LLE is that the local weights it uses cannot fully reflect the local geometric structure of the high-dimensional manifold. Therefore, for singular or nearly singular systems, a positive number γ needs to be artificially added, but the selection of γ It will greatly interfere with the results..."

"..."

Strangely enough, after writing that paper, Chang Haonan found that his way of thinking seemed to be somewhat different from the past. Specifically, it became more coherent and smooth.

After half an hour of introduction, not only did he not get stuck at all, but even Yao Mengna, who was listening next to him, didn't feel that there was much that he didn't understand.

"so."

Chang Haonan put the pen aside and said in a determined tone:

"Obviously, using multiple sets of linearly independent weight vectors to construct a local linear structure can improve the final embedding result."

Yao Mengna was silent for a while, then nodded:

"Indeed it is."

The two continued to fall into silence.

"Speaking of which, we first started studying manifold learning, it seems to be to...solve the problem of automated inspection of pulsating production lines?"

It was Yao Mengna who broke the silence again.

"Um……"

Chang Haonan nodded:

"However, the content of our current research cannot be directly applied to your topic."

Manifold learning is just a process of data dimensionality reduction, which at best solves one of the many obstacles on the road to automated production.

Hearing this judgment, Yao Mengna sighed:

"Sure enough, my idea may still be a bit too radical..."

That's really radical.

If her plan is fully realized, Factory 112 can directly withdraw the workers and turn it into a black light factory.

However, Chang Haonan then changed the subject:

"However, this improved LLE algorithm can play a role in other fields."

It belongs to the unintentional planting of willows and willows.

"for example?"

Yao Mengna frowned slightly and looked at the dense formulas on the paper in front of her.

"For example...information retrieval, data screening..."

Halfway through, Chang Haonan realized that these did not seem to be specific "applications", but could only be regarded as application-level technologies.

So I thought about it for a while:

"If we want to move closer to specific production links...maybe...equipment status monitoring and automatic fault diagnosis?"

In fact, if this algorithm is promoted, it should still be able to shine in many fields.

But Chang Haonan could only think of his old profession for a while——

For example, an airplane breaks down.

In the past, ground crews could only carry out investigations slowly.

And if the aircraft has enough sensors and can effectively screen and analyze the data fed back by the sensors, then the aircraft's avionics system will determine the specific location and situation of the fault.

Even when the fault is just a clue, it is nipped in the cradle.

There are just some difficulties to overcome in operation.

For example, current sensors are relatively large, and it is difficult to fit a sufficient number into an aircraft.

Regardless, this is clearly a technology with considerable potential.