novelones Chapter 22 Underground Breakthrough
Chapter 22 Underground Breakthrough
Chen Hao took two and a half days.
It was half a day faster than the promised three days.
Based on existing underground space channel propagation models, he generated 12,000 sets of simulation data, covering four typical scenarios: subway tunnels, underground parking lots, underground shopping malls, and civil defense projects. Each scenario was further subdivided according to variables such as depth, area, material, and personnel density. The resulting data diversity was even richer than Zuo Cheng had anticipated.
"We ran a pre-training test on this batch of data, and the classifier's accuracy in recognizing underground scenes improved from 78% to 89%," Chen Hao projected the results onto the lab's large screen. "But there's still a gap—an 89% recognition rate means one in ten judgments is incorrect, and parameter mismatches can occur when switching scenes, so performance still can't be improved."
"The gap needs to be filled with real data," Zuo Cheng said.
Han Zhe's reply arrived yesterday afternoon—Lanwan Communications has a subway line base station in the eastern district of Huaxia City, covering three subway stations and two tunnel sections, and can grant seven days of data access. However, there is one condition: data collection requires the Zuocheng team to deploy the collection equipment on-site; Lanwan Communications only provides access to the base station port.
In other words, you have to collect data yourself in the subway.
"Fang Ze and I will go," Zuo Cheng decided. "Chen Hao, you stay in the lab and wait for the data to come back for fine-tuning and training."
On November 26th, Zuo Cheng and Fang Ze carried two boxes of equipment and squeezed into the subway during the morning rush hour.
The acquisition equipment was borrowed from Professor Lin's laboratory, plus a portable embedded acquisition board modified by Fang Ze himself—about the size of a palm, powered by a mobile power supply, capable of continuously recording the raw waveform of radio frequency signals and channel status information at the base station.
They deployed data collection nodes at three subway stations in the eastern district: one on the platform, one at the tunnel entrance, and one in a corner of the station hall. Fang Ze was responsible for hardware debugging, while Zuo Cheng was responsible for data synchronization with the base station.
The whole process was more troublesome than expected.
The electromagnetic environment in subway stations is extremely complex—in addition to base station signals, there are radio waves from train control systems, radio frequency radiation from passengers' mobile phones, electromagnetic leakage from LED screens, and even interference from security scanners. All these noises combined make the signal environment far worse than simulated in a laboratory.
"That's where the difference between simulated data and the real environment lies." Fang Ze squatted in a corner of the platform, adjusting the acquisition board, a thin layer of sweat beading on his forehead. "The model only considered the signal attenuation and reflection caused by the building structure, not these messy environmental noise sources."
"That's why we need real data." Zuo Cheng checked the synchronization status of the base station on his phone. "Simulated data teaches the classifier 'what underground space looks like,' while real data teaches it 'how dirty real underground space is.' Both steps are indispensable."
They spent the entire day on the subway, from the morning rush hour at 7 a.m. to the last train at 10 p.m. Lunch was sandwiches on the platform, and dinner was two rice balls that Fang Ze bought from a convenience store.
Zuo Cheng noticed an interesting phenomenon—Fang Ze was usually quiet and reserved in everyday social situations, but he became much more talkative when working on-site. He would proactively point out signal blind spots at the data collection location, analyze the impact of structural differences between different platforms on signal propagation, and even chat with the subway station security guard for ten minutes to understand the layout of the relay equipment in the tunnel.
"You feel much more comfortable dealing with equipment and the site than with people," Zuo Cheng said.
Fang Ze didn't deny it, he just shrugged: "The equipment doesn't talk nonsense."
At 10:30 p.m., the two of them carried their equipment out of the subway station. A gust of cold wind made Zuo Cheng shiver. It was already winter in Huaxia City at the end of November, and he was dressed too lightly.
Fang Ze pulled a crumpled windbreaker out of his backpack and tossed it to him.
"I brought a spare."
Zuo Cheng took it and put it on without saying thank you. Some things don't need to be said.
It was almost midnight when Chen Hao got back to the lab. There were three layers of takeout boxes stacked on the table, the computer screen was lit, and the training framework for the classifier was already set up, just waiting for data to be fed in.
Zuo Cheng exported the day's collected data—a total of 4,700 sets, covering the actual channel status of three platforms and two tunnels at different times and with different passenger flows.
"That's enough." Chen Hao glanced at the data volume, pushed up his glasses, and his fingers were already on the keyboard.
I did fine-tuning training all night.
The results came out the next morning.
The classifier's accuracy in recognizing underground scenes is 96.3%.
From 78% to 96.3%, the results far exceeded expectations after two steps: pre-training with simulated data and fine-tuning with real data.
Zuo Cheng immediately embedded the updated classifier into the scene adaptation algorithm and ran a complete set of full-scene simulation tests.
In densely populated urban areas: performance exceeds the benchmark by 38%.
Highways: 41 percent above the benchmark.
Indoor coverage: 33% above the baseline.
Underground space: 31 percent above the baseline.
All four core scenarios exceeded the hard target of 30%.
Zuo Cheng stared at the data on the screen for ten seconds, then slowly leaned back in his chair, putting his hands behind his head.
It's over.
We've finally cracked the toughest bone.
"Beautiful." Fang Ze stood behind him, looking at the screen, his voice calm, but Zuo Cheng could hear a hint of suppressed satisfaction.
Chen Hao didn't say anything, but took off his glasses and wiped them. There was a layer of fog on the lenses from staying up all night.
Zuo Cheng sent a message to the team: "Performance across all scenarios has met the standards, exceeding 30% across the entire line. Next, we will move into the solution development and delivery phase."
Liu Wei replied instantly: "Awesome! I'll upgrade the celebration party to another level!"
Yu Ying's reply came a little later, but it was more informative: "Congratulations. I also have some good news—the reviewers' comments came back today, and both reviewers gave it a 'minor revisions, acceptance' rating. There aren't many suggestions; I've compiled a list. Take a look and see which ones you'll revise and which ones I'll revise."
After minor revisions, it was accepted.
In the peer review process of top journals, "acceptance after minor revisions" is almost equivalent to acceptance—as long as the few minor issues raised by the reviewers are corrected, the paper can be officially published.
Zuo Cheng's lips curled into a smile as he read the message.
The performance of the third stage was up to standard, and the paper published in the top journal received minor revisions. Both major events were resolved on the same day.
The light screen flickered silently in his mind, but he didn't rush to open it.
He picked up his phone and called his mother first.
"Mom, did Dad go for his medical checkup?"
"I went, I went. The report isn't out yet; they said it'll take two days. Your dad complains about the hassle, but he's actually very well-behaved at the hospital; he listens to everything the doctor says."
Zuo Cheng smiled.
"Tell me as soon as the report is out."
"Okay, you go ahead with your work. Oh, by the way, have you bought your train ticket back for winter break?"
"I bought it, Mom, on the 26th of the twelfth lunar month."
"That's good. Your sister said her husband will pick you up from the station."
After hanging up the phone, Zuo Cheng opened the system panel.
[Main Quest Chain - Breaking the Communication Barrier - Stage 3 Progress Update: 82%]
[Remaining tasks: compiling solution documentation, performance test reports, and preparing technical review materials.]
Eighty-two percent. The core technology breakthrough has been completed; what remains is documentation. For Zuo Cheng now, writing documentation is the least suspenseful part.
He closed the control panel, turned on his computer, and created a new document.
The title was: "Blue Bay Communications 5G Network Full-Scenario Intelligent Channel Estimation Algorithm - Technical Review Report".
As his fingers touched the keyboard, a thought flashed through his mind.
After the third stage, the hidden fourth stage begins.
Zhou Henian said that the Sky Dome project is "highly relevant" to his algorithm. The full-scenario deployment required in the third stage is likely just the foundation of Sky Dome.
What is the sky itself?
Satellite communications.
For some reason, Zuo Cheng suddenly thought of those two words. There was no basis for it; it was purely intuition.
But his intuition is rarely wrong.