stjorthotic Chapter 52 Breaking Through the Wall
Chapter 52 Breaking Through the Wall
Yu Ying's MATLAB code was sent to Zuo Cheng's email in the early hours of that day.
More than three thousand lines, with very well-written comments—this was Yu Ying's style; she always kept everything she handed over to others neat and tidy. Zuo Cheng spent two hours reading it through and extracted the mathematical core of the cyclic stationary feature detection.
The core idea is not complicated—communication signals naturally possess periodic stationary characteristics; for example, carrier frequency, symbol rate, and frame structure all produce periodic characteristic peaks in the second-order statistics of the signal. By calculating the cyclic autocorrelation function of the signal, the target signal can be separated from noise at extremely low signal-to-noise ratios, while simultaneously identifying the modulation type and occupied bandwidth of the signal.
"Elegant." That was the assessment Zuo Cheng gave himself after reading the mathematical derivation.
However, the problem is also obvious—the computational cost of the cyclic autocorrelation function is on the order of the square of the signal length. Yu Ying's MATLAB simulation takes four hundred milliseconds to run one spectrum sensing cycle on a PC, while the computing power of the embedded platform of the Tianqiong terminal is less than one-tenth that of a PC. Direct porting would require more than four seconds for a single sensing operation, far exceeding the real-time requirement of fifty milliseconds.
The difference is almost two orders of magnitude.
Zuo Cheng closed the MATLAB code, opened a blank document, and began writing his own solution.
He doesn't need to copy Yu Ying's implementation; what he needs is to "translate" the algorithm into a form that can run on an embedded platform.
This is precisely the capability that his fusion-grade blade "embedded intelligent channel processing" has given him—three years of equivalent engineering experience have made him intimately familiar with every inch of computing power on the embedded platform.
The first challenge was to improve computational precision. Full-precision calculation of cyclic autocorrelation requires floating-point operations, but the fixed-point arithmetic unit of an embedded platform is eight times faster than floating-point. Zuo Cheng derived a 16-bit fixed-point conversion scheme, converting all critical multiplication and accumulation operations to fixed-point, keeping the precision loss within 0.5 dB, which is perfectly acceptable for spectrum sensing.
This cut reduced the computational workload by 60%.
The second challenge lies in the scale of the data. Cyclic autocorrelation requires a two-dimensional scan of the entire signal, but not all cyclic frequencies are meaningful—the carrier frequency and symbol rate of the satellite signal are known, and a finite number of candidate cyclic frequencies can be pre-calculated, allowing detection only at these frequency points.
This cut removed another 70%.
With two cuts, the total computational load was reduced to 12% of the original. Four hundred milliseconds multiplied by 0.12 equals forty-eight milliseconds.
It's just right, stuck at the 50-millisecond mark.
But "just enough" is not Zuocheng's standard.
He reviewed the embedded platform's hardware manual again and found a third optimization point—the platform's DMA controller supports data prefetching, allowing the data for the next frame to be moved from the ADC to the buffer while the CPU is calculating the current frame. This overlapping of calculation and data movement further reduces the equivalent processing time by 20%.
Forty-eight milliseconds multiplied by 0.8 equals thirty-eight milliseconds.
A surplus of 24 percent. That's enough.
Zuo Cheng wrote the entire optimization plan into a twelve-page technical document, clearly outlining every step from mathematical derivation to engineering implementation. By the time he finished, it was already dawn, and he could hear the footsteps of early risers jogging past his window.
Instead of going to sleep immediately, he sent the document to Fang Ze and Liu Wei.
Fang Ze replied ten minutes later—he hadn't slept either, he'd been adjusting the simulation platform's parameters.
"I've reviewed the proposal. I need to verify the 0.5 dB accuracy loss in fixed-point quantization. Give me two days."
"good."
Liu Wei replied half an hour late: "Brother Cheng, I tried a similar approach to the DMA prefetch optimization in the LoRa project before, and it's confirmed to be feasible. I'll write that part of the driver code."
Zuo Cheng looked at the two messages and the corners of his mouth curved slightly.
The team is growing. Six months ago, he was the only one who could write these optimization plans, but now everyone can take over a part of it.
Two days later, Fang Ze's simulation results came out—in a simulation environment with 120 satellite signals, the fixed-point scheme achieved a spectrum sensing accuracy of 91.3% and a false alarm rate of 0.7%. The accuracy exceeded the target of 90%, and the false alarm rate was far below the upper limit of 5%.
Looking at the number, Zuo Cheng finally allowed himself to breathe a sigh of relief.
The core algorithmic problem of the spectrum sensing front-end has been solved.
The system light screen lit up.
[Technology Tree Perception: Host Completes "Embedded Cyclic Stable Spectrum Perception" Technical Solution]
This solution integrates the capabilities of blade "spectrum sensing and management" and "embedded intelligent channel processing (fusion level)".
[Trigger passive effect: Skill boost ×1.2]
[Development efficiency of the spectrum sensing front-end module is expected to improve by 20%]
It wasn't a new blade, nor a new task; it was merely the triggering of a passive effect. But Zuo Cheng knew what this 20% efficiency improvement meant—the spectrum sensing front-end had gone from being the most lagging module to potentially being the first to complete.
He called Yu Ying that evening.
"Kongkong, your cyclic stability scheme was a huge help. The embedded optimization scheme for the core algorithm has been developed, and simulation verification has passed, achieving an accuracy of 91%."
There was a two-second silence on the other end of the phone.
"Ninety-one percent?" Yu Ying's voice held a hint of uncertainty. "My MATLAB simulation, even under ideal conditions, only achieves ninety-three percent. You got ninety-one percent using fixed-point arithmetic on an embedded platform?"
"The loss of accuracy in fixed-point refining is controlled to 0.5 dB, which has a negligible impact. The main way to reduce computational load is through candidate cycle frequency pre-screening and DMA pipeline."
Yu Ying remained silent for a few more seconds.
"Brother, you know what? My advisor spent five years researching cyclic stationary detection, but they couldn't solve the real-time problem. You engineered it in just two days."
"It wasn't just two days," Zuo Cheng said. "It's all the experience I've accumulated on embedded platforms over the past year, plus the algorithmic foundation you gave me. Without either one, it wouldn't have been possible."
"That's impressive too." There was something in Yu Ying's tone that Zuo Cheng rarely heard—not admiration, but genuine recognition between peers. "By the way, if my advisor knew about this result, he'd probably like to collaborate with you on a paper. Are you interested?"
"Yes. But we'll have to wait until the project's confidentiality period expires."
"Okay, I'll mention it to him first."
After hanging up the phone, Zuo Cheng leaned back in his chair and stared at the ceiling for a while.
He suddenly recalled Lu Mingyuan's retelling of Zhou Henian's words—"This young man's vision extends beyond the ground terminal."
Now he vaguely understood the meaning of those words.
Spectrum sensing is not just a module of a ground terminal. It is a capability—the ability to enable communication systems to "see" the electromagnetic environment. This capability can be used on ground terminals, drones, vehicle-to-everything (V2X) systems, and in any scenario that requires real-time spectrum sensing.
What Zhou Henian saw was not a module, but a possibility.
Zuo Cheng closed his eyes. The technology tree grew quietly deep within his consciousness, its thirteen leaves swaying gently in the dim light.
It's still early. But the direction is already clear.