16
On the third day.
Chen Wen uploaded the final version of the structural schematic to the shared server.
The file name had changed from "V7_WIP" to "V8_Frozen."
Although no one commented on this change in the R&D group chat, the Algorithm Group Leader still took a screenshot and sent it to himself when he saw it.
He opened the interface definition document and quickly flipped to the electrical specifications page.
He checked the sampling rate, bit depth, synchronization protocol, trigger method, and other details item by item against the data interface module in his framework.
"It'll interface fine."
He said to Xiao Zheng beside him.
"The sampling rate is sufficient, and the sync protocol is standard PTP, so we don't need to write custom drivers."
Xiao Zheng leaned over to take a look and asked,
"Boss, when will the vibration array data arrive?"
"Chen Wen said they'll start rolling out the first batch of retest modules tonight, equipped with the new vibration sensors.
We should be able to get the anonymized sampling logs tomorrow morning."
"What about the first version of the annotation framework?"
"Tonight, we'll get the data preprocessing pipeline up and running first.
Once the data arrives tomorrow, we'll feed it straight in."
The Algorithm Group Leader transferred the architecture diagram from the whiteboard into code.
Specifically:
The multi-channel signal time-alignment module, the joint confidence calculation engine, the high-confidence auto-ingestion logic, and the low-confidence manual review queue.
The scaffolding for these four core modules had already been written last week; they just needed the vibration data to be plugged in.
...
At ten o'clock the next morning, the data was ready.
It wasn't the clean signals simulated in a lab, but rather the first batch of vibration array sampling logs collected from real home environments.
The archive alone was over three hundred megabytes, which decompressed into raw time-series data from seventeen households.
Furthermore, each household covered a full twenty-four-hour period.
They first picked a sample that looked like it "should be straightforward."
It was labeled H07, with annotation notes stating: Five-year-old child slipped and fell while running on the living room hardwood floor; confirmed by parents.
Then, they loaded the vibration waveform.
A curve sprang up on the screen, showing stable, low-amplitude background noise for the first few seconds.
It was just the faint vibration of someone walking normally, around 0.05 g.
Then, at the three-second mark, a distinct shock peak appeared.
The vibration spiked to 0.42 g, lasting about 0.15 seconds.
This waveform shape was a textbook "hard contact."
First a rapid rising edge, followed by a brief oscillatory tail.
"This signal is clean,"
Xiao Zheng said.
"It's so much cleaner than the IMU data from the S01.
The S01's accelerometer had a high noise floor, so fall signals were often buried in the background."
"That's the value of an array,"
the Algorithm Group Leader said, zooming in on the waveform details.
"Look at the decay curve of this shock peak. The oscillating tail takes about 0.3 seconds to return to zero.
If an object dropped, the decay would be much faster, probably within 0.1 seconds.
Our system can differentiate based on this."
He then tagged this sample as a "high-confidence fall event."
Subsequently, the joint label was automatically generated: vibration shock peak of 0.42 g.
Added to that was an acoustic spectrum matching the "body impacting ground" template, plus the IMU detecting 0.28 seconds of free-fall acceleration.
The time-alignment error across the three signal channels was seven milliseconds.
Seeing this, he clicked confirm.
The first sample was logged.
Then came the second, the third, the fourth...
At the eleventh one, they hit another roadblock.
It was labeled H12, with annotation notes stating:
Three-year-old child crawling in the living room abruptly shifted from crawling to sitting; parents confirmed nothing abnormal.
On the vibration waveform, a shock peak appeared at the six-second mark: 0.18 g, lasting 0.12 seconds.
Xiao Zheng spoke up,
"Should we label this one?"
The Algorithm Group Leader stared at the waveform.
At 0.18 g, it was lower than the system's preset "suspected fall" threshold of 0.25 g.
Logically, it should be filtered out directly without entering joint evaluation.
But there was an issue.
"Look at this decay curve."
The decay tail lasted 0.28 seconds.
It was longer than typical "dropped object" decay, but shorter than standard "human fall" decay.
The waveform shape was ambiguous.
"Switching from crawling to sitting—buttocks hitting the ground—also generates an impact. It's just smaller in magnitude."
Then he pulled up the corresponding video frames and continued,
"Look at the child's movement while sitting down.
His buttocks hit the floor first, and then his upper body straightened up.
This process was controlled, not a loss of balance."
"But our vibration sensors can't tell the difference like this."
"Exactly.
Because the magnitude doesn't reach the trigger threshold, and the decay shape is stuck in a gray area.
The system classified it as 'filtered,' which is why I think this entry should be pulled out for manual review."
With that, he fished this entry out of the "auto-filter" queue and pushed it into the manual review pool.
Over the next twenty samples, similar situations arose four more times.
It wasn't a clean dichotomy between "high-confidence auto-ingestion" and "obviously irrelevant direct filtering," but rather a large number of ambiguous events stuck in the gray area.
For instance:
A child jumping down from a sofa—a controlled landing impact.
An adult holding a child and sitting down on a play mat together—an impact waveform resulting from the combined weight of two people.
Or a child throwing a toy onto the floor—similar magnitude, but a different decay profile.
...
The joint confidence engine could handle the clean extremes, but the proportion of gray-area samples was far higher than they had anticipated.
He ran the numbers.
Across the twenty-four hours of data from seventeen households, the vibration array had recorded a total of 317 valid impact events.
Among them, 63 were high-confidence auto-ingested, and 184 were obviously irrelevant and directly filtered.
The remaining 70 were all stuck in the gray area, requiring manual review.
"Twenty-two percent,"
he said to the Data Team Leader.
"The gray area accounts for twenty-two percent.
That's more than double our initial estimate of ten percent."
"That means the workload for manual review will still be massive."
"Right.
And this is just an issue with the vibration channel alone.
Once visual and audio data are integrated, the gray area will only get more complicated after cross-referencing all three channels.
A case filtered out by one channel might be deemed'suspected' by another, making the joint outcome even muddier."
The Algorithm Group Leader leaned back against his chair, staring at the seventy pending events on the screen.
"Let's not rush to adjust the thresholds.
Let's finish manually reviewing all these borderline samples first to see the true distribution.
If most of them turn out to be normal activities, it means the threshold is set too high and can be lowered.
But if most are actually falls that the system missed, it means the vibration array's discriminative power isn't enough, and we'll need to add new feature dimensions."
"Like what?"
"Looking at the frequency-domain features of the impact waveform.
In the time domain, we look at decay profile; in the frequency domain, we look at dominant frequency distribution.
There should be differences in frequency components between a human falling and an object dropping.
Because human bodies are soft, impacts generate more low-frequency components.
Whereas falling hard objects concentrate more high-frequency components."
Then, he opened the signal processing toolbox and performed an FFT on one of the borderline samples.
Once the spectrogram appeared, he narrowed his eyes.
"Look at this,"
he said, pointing at the screen.
"The dominant frequency is concentrated between 8 and 15 Hz, and low-frequency energy accounts for 63 percent.
This is a classic spectral signature of human soft tissue impacting a hard surface.
If a toy hit the floor, the dominant frequency should be above 40 Hz."
"So we can add a frequency-domain evaluation branch."
"Right.
But frequency-domain analysis requires more computational power than time-domain analysis, so we need to see if the S02 has enough compute capacity to run it."
As he spoke, he wrote it down:
Frequency-domain feature extraction module: evaluate computational overhead.
...
Dusk gradually fell.
But in the Algorithm Group's office area, every screen remained lit.
Xiao Zheng was still manually verifying those seventy borderline samples one by one.
With each one confirmed, the training set gained another anchor with ground truth.
Progress was slow—out of the seventeen households' data, only three and a half had been processed on the first day.
However, the model training loss curve finally began to bend downward from its plateau.
...
