16
By the time Xiao Zheng dragged the last intermediate sample into the "Confirmed" folder, it was already one in the morning on the third day.
All seventy vibration events marked for "manual verification" had been completely reviewed.
The Data Team Leader worked through the night to produce a statistical report and posted it on the partition of the Algorithm Group Leader's cubicle at nine o'clock the next morning.
Holding a cup of coffee, he scanned through it.
The real distribution was far more skewed than he had anticipated.
The manual verification results for the seventy samples falling into the gray area were as follows:
Only eleven actually belonged to "unexpected falls or loss of balance."
A total of fifty-nine belonged to normal activities.
This meant eighty-four percent were false alarms.
"Nothing we can do about that. Most of the ones stuck in the gray area aren't actually falls,"
the Data Team Leader said.
"Shifting from crawling to sitting, jumping down, throwing toys, adults holding a child and sitting on the floor mat, and so forth.
These make up the vast majority.
The false negatives missed by the system were only eleven."
The Algorithm Group Leader set down his coffee, staring at the number.
"The false negative rate is actually quite low.
The problem lies on the other side.
The system classified a massive amount of normal activities into the 'uncertain' range and pushed them to manual verification.
Even though they didn't end up in the training set by mistake, eighty-four percent of the manual verification workload was eaten up by these normal samples."
"Shouldn't this be an issue with the threshold?"
"Right.
Currently, the vibration threshold for'suspected fall' is set at 0.25 g.
But during normal activities, the impact of shifting from crawling to sitting can reach 0.18 g, and jumping down can reach 0.22 g—both skimming the lower bound of the threshold.
If the system can't decide, it pushes it to manual review."
He flipped to the last page of the report, where there was a distribution histogram.
It showed that
across impact magnitudes from 0.05 g to 0.6 g, normal activities and real falls had a clear separation zone between the two ranges.
Between 0.25 g and 0.35 g was an overlapping zone, but above 0.35 g, every single one was a real fall, with zero false positives.
After looking it over, he said,
"Raise the threshold from 0.25 to 0.35."
Hearing this, the Data Team Leader was taken aback.
"Raise it that high?
Then the real falls between 0.25 and 0.35 will be missed."
"How many are in that range?"
"Of the eleven missed cases, seven fall into this range."
"Seven."
The Algorithm Group Leader did the math and said,
"What proportion of all real fall events does that account for?"
The Data Team Leader flipped back to check.
"Across seventeen households over twenty-four hours, a total of nineteen confirmed falls were recorded.
These seven between 0.25 and 0.35 account for thirty-seven percent."
He pondered for a moment.
"Mm... thirty-seven percent. We definitely can't lose that."
He then picked up a pen and drew a line beneath the histogram.
"Then let's not stick to a single threshold.
Change it to a tiered system.
Above 0.35 g, high confidence, automatically entered into the dataset without pushing to manual review.
Between 0.25 and 0.35 g, keep manual verification, but add frequency-domain features for a secondary screening.
That is, if the dominant frequency is between eight and fifteen Hertz and low-frequency energy accounts for over fifty percent, automatically upgrade it to high confidence.
Whatever doesn't meet this gets pushed to manual review."
"What about below 0.25 g?"
"Continue automatic filtering.
But add an exception rule:
if the acoustic spectrum also detects an anomaly during the same time window and the IMU acceleration exceeds 0.15 g, even if the vibration magnitude is below 0.25 g,
we still have to pull it back for a look."
The Data Team Leader nodded and said,
"So that's like adding two safety nets at the threshold boundary.
If the vibration magnitude is sufficient, it passes directly; if not, audio and the IMU make up for it."
"Right.
Exactly. This way, the manual verification pool should shrink significantly."
The Algorithm Group Leader had Xiao Zheng rerun the seventy cases according to the new rules for validation.
When the results came out that afternoon, the entire Algorithm Group breathed a sigh of relief.
Under the original rules, all seventy cases had been pushed to manual verification, with zero automatically added to the dataset.
After running under the new rules,
twenty-three were automatically included with high confidence—all of which were real falls, with zero false positives.
The manual verification pool shrank to seventeen cases.
Among them, three were ultimately confirmed as real falls, fourteen as normal activities, and the remaining thirty were filtered out directly.
The manual verification workload dropped from seventy to seventeen.
The workload was slashed by seventy-six percent.
"And the twenty-three automatically included cases had zero false positives."
Xiao Zheng double-checked and continued,
"This is even more accurate than when annotators relied purely on visually inspecting video footage."
"Of course it is."
The Algorithm Group Leader pulled up the training dashboard and said,
"Annotators watching the footage will agonize over 'does this count as a fall?', but with the combined three-way judgment of vibration, audio, and the IMU, physical signals don't second-guess themselves."
Then, he clicked "Start Training."
This time, the model did not keep him waiting long.
The loss curve dropped by twelve percent in just the first two epochs, more than the total of the previous two weeks combined.
By the sixth epoch, the rate of decline began to slow, ultimately settling at a point thirty-one percent lower than before.
He stared at the curve for a long time.
It wasn't because the number was particularly stunning.
After all, a thirty-one percent improvement wouldn't turn many heads in an academic paper.
But this came from real household data, not from parameter tuning.
That made it quite impressive.
Then, he said to Xiao Zheng,
"Go ahead and run the test set."
The test set consisted specifically of those edge cases that had caused the greatest disagreement among annotators.
For example, lying prone and remaining motionless, being covered by a blanket, or sitting on the floor leaning back against a sofa.
Previously, the model's confusion rate on these samples was nearly forty percent.
But once the new model finished running, the confusion rate dropped to nineteen percent.
"Not bad! That cut it right in half."
The Data Team Leader leaned in to look, remarking with emotion,
"Even though the prone posture one is still the easiest to confuse, at least it can differentiate it now."
Algorithm Group Leader Zhou clicked open a specific sample.
It was the exact one that had vexed Xiao Zheng for forty minutes in the beginning.
A boy lay prone on the carpet, face down, motionless for twelve seconds, before turning over to continue reading.
Old model: labeled "suspected fall," confidence sixty percent, annotators disagreed.
New model: labeled "normal lying down," confidence eighty-three percent.
Judgment criteria:
Vibration array had no impact peak, IMU had no freefall phase, acoustic spectrum showed no anomalies.
All three signals remained silent; combined confidence determined it as "non-fall."
"Now our model has finally learned.
The old model would start raising alarms whenever it saw a child lying still.
But now, the new model first checks if there's vibration on the floor, sound in the air, or acceleration from the body.
If none of these are present, then it isn't a fall."
However, he didn't let the team celebrate.
Instead, he added a note in the training log:
"Current model confusion rate on boundary samples remains close to twenty percent.
Primary remaining issues:
Multi-person in-frame scenarios, namely signal superposition when an adult falls while holding a child;
Soft surfaces, such as excessive vibration dampening during falls on carpets/mattresses leading to missed detections;
And scenarios where a child is completely still and silent after falling.
For such samples, the vibration signal may be sufficient to trigger the threshold, but lacking auxiliary audio and IMU signals, the combined confidence will downgrade to medium and still require manual verification.
..."
He then saved the log and closed the panel.
In fact, it wasn't just Xiao Zheng; everyone across the entire office area hadn't left their desks for these three days.
Him included.
Fortunately, heaven rewards the diligent.
Their hard work had paid off.
The loss curve had dropped, the confusion rate was halved, and the joint annotation framework combining vibration, audio, and IMU was fully validated.
Though issues like twenty percent boundary samples, missed detections on soft surfaces, and multi-person superposition scenarios remained,
there was nothing they could do about it for now.
After all, problems always outnumbered answers.
He then opened the list of the next batch of data awaiting processing.
Resolving 317 impact events was merely the first task.
There were still 720,000 hours of data waiting to be fed in.
...
