214: Chapter 214 Capturing Consciousness at Extremely Low Amplitudes: Echoes Deep Within the Noise!

sentry ship arrival countdown, day 72.

Pangu Laboratory, B2 level, afternoon.

Xiao Yuan finished running the last set of test signals for the 32-channel acquisition system and turned around.

"The baseline noise is 23% higher than our laboratory's environment. The Quantum Radar array on the B2 level and the supercomputer cooling fans are all causing crosstalk in the acquisition channels."

Su Che looked at the fluctuating EEG baseline on the main screen, which was interspersed with regular high-frequency glitches.

"Lower the cutoff frequency of the anti-aliasing filter. Cut out the power frequency and equipment electromagnetic interference first, then perform feature extraction on the neural signals."

"How much should I lower the cutoff?"

"Keep everything below 500 Hz and cut everything above it; the effective frequency band of normal neural activity is within this range."

Xiao Yuan turned back to adjust the parameters. After re-running with the changed filter parameters, the regular glitches on the main screen disappeared, and the baseline became clean, but the signal amplitude was also suppressed.

Academician Chen Shian leaned in to take a look, retrieved the EEG chart from his folder, and placed it on the test bench.

"Chief Engineer Su, look here."

He pointed with his pen tip at the residual θ wave on the chart.

"Su Changshan's EEG signal amplitude here is roughly between two and five microvolts. Our baseline noise level just now is already overlapping with this magnitude."

Su Che pulled up the two sets of data side-by-side on the screen.

On one side was the 32-channel baseline from just now, and on the other was the historical EEG chart of Su Changshan brought by Academician Chen Shian.

It was almost the same.

A real neural signal of two to five microvolts, submerged in two to three microvolts of environmental noise.

"The task of the noise reduction algorithm is to separate these two."

Su Che pulled out the core module of the neural signal noise reduction algorithm and pushed it to Xiao Yuan.

"Use your laboratory's standard EEG database to run a round of algorithm pre-training first. Let the model recognize what a normal neural signal is before identifying the extremely suppressed abnormal signals like those of Su Changshan."

Xiao Yuan took it, flipped through a few lines, and without asking any questions, started working immediately.

On Academician Chen Shian's side, he called over the three accompanying researchers and began to organize the θ wave deep suppression cases from the laboratory database one by one, filtering out the 100 most typical feature sets to feed into the training set.

Academician Li Zhengyang cleared the tabletop in the weapon adaptation area, pulled over a chair, and sat down next to Academician Chen Shian.

"Academician Chen, in those 100 sets of data you have, what is the longest duration of deep suppression?"

Academician Chen Shian didn't look up. "Seven and a half years."

"How long has it been for Su Changshan?"

"He woke up briefly twice this year."

Academician Li Zhengyang put down his cup and didn't ask further.

...

sentry ship arrival countdown, day 71.

B2 level, morning.

The pre-training had been running for nearly a day.

Xiao Yuan fed in the last batch of data in the early hours and received the first version of the model output at 7:00 AM.

He pushed the results to the public large screen.

The algorithm's recognition accuracy for the 100 θ wave deep suppression cases in the training set was 91.4%.

It looked decent, but when Su Che pulled up the confusion matrix, the problem became immediately apparent.

The false negative rate was 12%.

12% of the real, weak neural signals were judged by the algorithm as noise and discarded directly.

"12%." Academician Chen Shian looked at the number. "That is to say, if Su Changshan's residual consciousness signal is emitted, there is a 12% probability that the algorithm will throw it away as noise."

Su Che didn't respond; his hands were already on the keyboard.

High false negatives meant the algorithm's sensitivity threshold was set too conservatively.

He adjusted the threshold, pushing down the barrier for determining "signal presence."

After readjusting and re-running, the false negative rate dropped from 12% to 4%, but the false positive rate rose along with it; 19% of the noise was judged as real signals.

"19% false positive is also not acceptable."

Academician Qian Zhenhua poked half his body out from the east workstation. "False positives mean the system will respond frequently to noise. The pilot's Brain-Computer Interface will constantly receive interference commands, and the White Emperor will be a mess on the battlefield."

Su Che placed the two sets of parameter results together.

False negatives and false positives—when one goes down, the other goes up. This is the fundamental contradiction of signal detection.

Relying solely on adjusting thresholds cannot break this deadlock.

He stared at the confusion matrix for nearly three minutes.

"Change the approach."

Su Che pulled up the model architecture.

"Currently, it's a single-stage judgment, deciding the authenticity of the signal at once."

"Change it to a two-stage series connection. The first stage will have high sensitivity—better to capture more than let any slip through. The second stage will be dedicated to false positive cleaning, using temporal features for verification. Real neural signals have temporal continuity, while noise does not."

Xiao Yuan drew a simple diagram of this logic on a draft paper.

"How long should the temporal window for the second stage be set?"

"50 milliseconds. The shortest duration of a neural signal is approximately this magnitude. Any pulse shorter than 50 milliseconds, regardless of amplitude, will be discarded directly by the second stage."

Academician Chen Shian stood up from his chair and walked to the public large screen. "This approach is correct. Our laboratory has done something similar before, but we didn't have a specialized training set for extremely deep suppression."

He turned to the accompanying researchers. "Re-slice that batch of θ wave temporal data into 50-millisecond segments and re-annotate them."

The three researchers divided the work and started cutting, while Xiao Yuan synchronized the model architecture change.

...

Afternoon.

The modified two-stage series model ran the second round of verification.

The false negative rate was 3.1%.

The false positive rate was 4.8%.

Academician Qian Zhenhua silently noted the two numbers in his draft notebook at the east workstation, saying nothing.

Academician Sun Qiwen walked over from the U-shaped array area with his enamel mug, looked down at the screen, and said, "3.1%—that's enough for shooting down planes, but is it enough for saving a life?"

No one answered immediately.

Academician Chen Shian picked up that EEG chart of Su Changshan and stared at it again for a while.

"In Su Changshan's state, the window for residual consciousness signals to appear might be very short, between a few seconds and a dozen seconds."

"If the algorithm happens to trigger the 3.1% false negative rate during this window, this signal will be lost forever."

Su Che reduced the second-stage temporal window from 50 milliseconds to 30 milliseconds.

"Shorten the window, relax the temporal constraints, and let more short-duration pulses pass verification."

Xiao Yuan re-ran it.

The false negative rate dropped to 1.9%.

The false positive rate rose to 7.2%.

"7.2%."

Zhao Mingyuan didn't know when he had walked up next to the public large screen, holding a new bag of goji berries in his hand.

"7.2% is still too much for the pilot. If the White Emperor suddenly receives seven false commands while flying, the pilot will be cursing."

"The signal quality from the pilot is an order of magnitude better than Su Changshan's."

Su Che didn't turn his head. "The amplitude of the pilot's neural signals under high-G maneuvers is tens of microvolts, not a few microvolts. At that magnitude, the first-stage sensitivity can be adjusted back to normal, and the false positives will naturally drop."

"So the parameters you are adjusting now are specifically for Su Changshan?"

No one responded to this.

Zhao Mingyuan stuffed the bag of goji berries back into his pocket and dragged his chair back to his workstation.

...

sentry ship arrival countdown, day 70.

B2 level, early morning.

When Wen Bo came in, half the people in the B2 level were crawling out of folding beds, and there were several empty paper cups on the tables.

One of the young researchers on Academician Chen Shian's side was asleep on the keyboard; Xiao Yuan went to shake him awake and told him to go lie down on a folding bed.

Wen Bo placed a sealed sample box in his hand on the tabletop in front of Su Che.

"A new version of the flexible bio-electrode. The carbon nanotube substrate is doped with a quantum-sensitive coating. The contact resistance is 60% lower than the preliminary version used last time, and the theoretical performance in low-amplitude signal acquisition is better."

He paused. "But it hasn't been tested on a human body, so I don't know what it feels like when attached."

Su Che opened the sample box.

Inside was a translucent thin-film electrode, thinner than the preliminary version from last time, with bionic fan-shaped bifurcations on the edges to fit the curve of the scalp.

Academician Chen Shian walked over, put on sterile gloves, took out the electrode, and peered at it under the light for a while.

"Low contact resistance is a good thing, but will the quantum-sensitive coating have ionic contamination in an environment with sweat?"

Wen Bo said, "I tested it in a constant-temperature drying chamber; I haven't tested it in a sweat environment."

"It must be tested." Academician Chen Shian put the electrode back into the box. "The human scalp is not a constant-temperature drying chamber. It is normal for pilots to sweat under high-stress combat conditions, and for Su Changshan..."

He didn't finish his sentence and took off his gloves.

Su Che pulled up the parameter table of the electrode characteristics on the screen.

"First, conduct an ionic contamination simulation experiment. Use physiological saline to simulate the sweat environment and run a 4-hour continuous acquisition to see if there is any drift in signal stability."

Academician Chen Shian arranged for a researcher to prepare the physiological saline immersion test.

Wen Bo didn't leave; he stood nearby, squeezing the sealed sample box.

"Chief Engineer Su, our materials institute has another plan: using graphene oxide instead of the quantum-sensitive coating. The tolerance to ionic contamination is better, but the contact resistance is a bit higher—about 15% higher."

Su Che thought for a moment. "Run both plans. Whichever has better signal stability in a sweat environment, use that one."

Wen Bo put the sample box into his work jacket pocket and went out to make a call to coordinate the materials for the other plan.

At the other end of the workstation, Zhao Mingyuan pulled out some old code for swarm coordination, looked at the signal interface definition of the Brain-Computer Interface, and occasionally glanced toward the center of the B2 level.

He hadn't spoken the whole time.

Until before noon, the physiological saline immersion test had been running for two hours. The signal of the version with the quantum-sensitive coating began to show slight drift—the amplitude was not large, but it continued to accumulate.

Xiao Yuan pushed the drift curve to the public large screen.

"It started drifting after two hours. It will likely drift out of the effective range by four hours."

"Then use the graphene oxide version. Compensate for the high contact resistance problem with algorithms."

Su Che added a gain compensation layer to the noise reduction model, performing a correction amplification on the signal amplitude after acquisition to make up for the amplitude loss caused by the resistance.

After reviewing the changes, Academician Chen Shian offered no objections.

Xiao Yuan fed the parameters of the new version of the electrode into the acquisition system and ran a baseline test again.

This time the baseline was much cleaner, without drift, and without the trend of ionic contamination.

Su Che combined the three components—the two-stage series noise reduction model, the gain compensation layer, and the new electrode parameters—and ran the first complete system joint debugging.

The test signals used were the 100 sets of historical data brought by Academician Chen Shian, which included three sets of extremely deep suppression samples closest to Su Changshan's state.

The results came out.

For the three sets of extremely deep suppression samples, the algorithm identified the signals in all cases; not a single one fell into the false negative category.

Academician Chen Shian didn't speak; he picked up that EEG chart of Su Changshan from the tabletop, folded it, and put it into his folder.

The B2 level was quiet for a while.

Su Che saved the results of this joint debugging into the encrypted storage and marked it with a time stamp.

On the bottom right of the main screen, the sentry ship arrival countdown: 70 days.

The algorithm was running, but it wasn't time to finalize the version yet.

The bidirectional interaction part—the machine-to-human signal feedback path—hadn't been touched yet.

The pilot sits inside the White Emperor.

How to feed the deep-space situational awareness perceived by the White Emperor into the pilot's nervous system?

Making him truly "feel" it rather than "see" it—these are two completely different problems.

Su Che closed the joint debugging log window.

He spread out the technical framework for the next stage on the main screen.