Rebirth: Starting from Lighting Up the Tech Tree

Chapter 195 Neural Signals

Shen Yiming stared at the jagged waveform on the screen; the coffee in his hand had gone cold.

The waveform danced on the screen like a restless snake. The noise peak on the left was almost as high as the effective signal, and the collected neural electrical signals were submerged in environmental interference, making it impossible for the decoding algorithm to process them.

I have been working continuously for 72 hours.

Three weeks after the joint laboratory was established, progress on its first project was far more difficult than expected. The NX-30 chip's 1024 channels working simultaneously offered strong acquisition capabilities, but large data volume did not equate to high data quality. Neural signals are naturally weak, with brainwave voltages in the microvolt range, while electromagnetic interference in the laboratory often reaches the millivolt level.

Shen Yiming tried bandpass filtering, independent component analysis, and wavelet denoising, but the results were mediocre. Noise was reduced somewhat, but the effective signal was also lost. The signal-to-noise ratio improvement was capped at 30%, and it couldn't be pushed any higher.

Zuo Cheng arrived at the lab at seven in the morning. He pushed open the door and saw Shen Yiming sitting at his workstation with three empty coffee cups on the table.

Didn't you go home last night?

Shen Yiming said he went back, and came back at 2 AM and 4 AM again. Zuo Cheng, this noise problem is more difficult than I thought. Come and take a look.

Zuo Cheng pulled up a chair and sat down, watching the waveform for a few minutes.

Where is the problem?

Shen Yiming explained that the signal source is a neural electrical signal with a bandwidth between 300 Hz and 3000 Hz. However, the power frequency interference in the laboratory is at 50 Hz, and the electromyographic interference is between 300 and 500 Hz. This frequency range overlaps with the low-frequency part of the effective signal. Hardware filtering is not precise enough, and software filtering results in signal loss.

Zuo Cheng didn't speak immediately. He remembered something.

A similar problem arose in last year's Tianqiong satellite data return project. The satellite-to-ground communication link was disturbed by ionospheric signals, with some overlap between the frequency bands and the effective data. These disturbances could not be simply filtered out; an adaptive algorithm had to be used to estimate the disturbance characteristics in real time and then subtract them from the signal. Zuo Cheng had previously developed an improved version using system integration.

He opened the system panel and operated it in his mind.

The panel currently displays 667 points. He selected the cross-branch fusion option between the commercial aerospace branch and the AI ​​branch, found the adaptive noise suppression algorithm in the satellite communication signal processing module, and performed matching analysis with neural signal features.

System notification: Fusion is feasible, consuming 5 points, generating a neural signal adaptive noise suppression algorithm, supporting real-time estimation of interference features, and expected to improve the signal-to-noise ratio by 60%.

Zuo Cheng confirmed. Points dropped from 667 to 662.

He turned off the panel and said to Shen Yiming, "I have an idea. There's a method in satellite communications that doesn't directly filter out a frequency band, but rather builds an interference model in real time, updates it every few milliseconds, and subtracts the estimated interference components from the original signal. This way, noise can be suppressed without losing effective signal strength."

Shen Yiming frowned and said, "I know this idea; it's called adaptive interference cancellation. The problem is that there are too many sources of interference in neural signals—electromyography, power frequency, and the electronic noise from the equipment itself—with significant parameter differences. Can the computational load for real-time modeling handle it?"

Zuo Cheng said that using a low-order autoregressive model to estimate interference has fewer parameters and faster updates. Adding a confidence-weighted algorithm based on signal statistical characteristics allows the algorithm to automatically determine which components are interference and which are valid signals. I'll write a framework for the specific implementation; you can adjust the parameters on it.

Shen Yiming asked, "How long will this take?"

Zuo Cheng said, "I'll give you the framework today."

He sat down at an empty workstation, opened his editor, and started writing code. The details of the algorithms were so clear in his mind that it was as if he had just reviewed them, and every line he wrote was precise.

More than an hour later, he sent the core modules of the framework to Shen Yiming.

Shen Yiming looked at the code and said, "Your confidence-weighted part uses a sliding window for variance estimation. How do you determine the window length?"

Zuo Cheng said, "Let's try 20 milliseconds first, which corresponds to 600 sampling points. That should basically cover a complete neural potential waveform. Run it and see the effect. If it doesn't work, we can adjust it."

Shen Yiming connected the framework to the test data and ran it.

The waveform on the display screen began to change. The jagged noise peaks gradually subsided, and the effective neural potential waveforms slowly emerged, with those sharp action potential signals becoming clearly discernible.

The signal-to-noise ratio started to fluctuate, climbing from 11 decibels to over 20 decibels, and finally stabilizing at 23.7 decibels.

Shen Yiming stared at the number, remained silent for a few seconds, and said, "It has increased by more than 60 percent."

Zuo Cheng asked, "Has it reached the standard of medical-grade signal quality?"

Shen Yiming checked the parameter table and said, "It's reached. The threshold for medical-grade is 20 decibels, and we're currently at 23.7, so there's still some margin."

Zuo Cheng said, "Okay. Let's organize this algorithm into a module and write it into the joint laboratory's technical documentation."

At that moment, Tang Ning walked over from the other side of the lab, carrying a stack of printed labeled data. She glanced at the waveform on the screen and said, "The signal-to-noise ratio has improved?"

Shen Yiming said that it has improved by 63%, reaching the medical grade.

Tang Ning said, "So fast?" She glanced at Zuo Cheng and said, "Is it President Zuo's proposal?"

Zuo Cheng said that Shen Yiming transferred the corps, so the credit should go to him.

Tang Ning placed the data on Shen Yiming's desk and said, "I've compiled the first batch of labeled data, three hundred motion image fragments, each labeled with its corresponding action type. We can start training the decoding model now."

Shen Yiming said, "Wait a minute, let me run the filtering module on the full validation set and confirm that it is stable before proceeding with the next steps."

Zuo Cheng stood up, stretched his shoulders, and asked, "Where's Chen Minghui?"

Tang Ning said he was adjusting the new electrode positioning fixture in the hardware area to improve implantation accuracy.

Zuo Cheng said to send today's progress to Chen Minghui and ask the hardware team to cooperate in the evaluation.

Tang Ning nodded and said, "Okay."

Zuo Cheng walked out of the laboratory and stood in the corridor for a while. The windows outside faced the south-facing campus, gleaming in the sunlight.

He mentally reviewed the progress. The signal quality issue was resolved—that was the first step. Next was the decoding algorithm, then clinical trial application, and finally volunteer recruitment. Each step had variables, but the direction was clear.

At the end of the corridor, Yu Ying pushed open the glass door and walked in, holding two cups of coffee in her hand.

I heard there's been a breakthrough in signal quality?

Zuo Cheng said that the noise has been reduced, and the signal-to-noise ratio has reached medical grade.

Yu Ying handed him a cup of coffee and asked, "How's Shen Yiming?"

Zuo Cheng said he had worked for more than 70 hours straight, but he was still in good condition.

Yu Ying said, "Don't overwork people."

Zuo Cheng said that it was his own fault for not wanting to leave.

Yu Ying said, "That's true." She thought for a moment and said, "The next step is the decoding model?"

Zuo Cheng said, "That's right. We have the signal quality and the data; now it's just a matter of whether the algorithm can push the accuracy up. Our goal is to break through 90%."

Yu Ying said that ninety is a very high threshold.

Zuo Cheng said that they should be able to get there, but it would take some time.

He took the coffee and took a sip. Through the glass wall of the corridor, he could see Shen Yiming adjusting parameters on the screen, completely focused and oblivious to the two people in the corridor.

The signal quality issue was resolved in three weeks; the next challenge is decoding.