Direct Answer

The sensors collect only voltage, current and temperature — how can the platform report dozens of hazard conclusions? Because "raw data" and "analytical features" are two different things: one set of raw waveforms, after transformation along the five dimensions of statistics, frequency domain, time–frequency, correlation and topology, can be reorganized into hundreds of features — total harmonic distortion (THD), three-phase unbalance, temperature-rise rate, peer-level temperature difference, zero-sequence component, load fluctuation rate... every feature is a mirror that shows one side of a hazard. Feature engineering is the core of "data production"; the sensors are only raw material. The Taiyi intelligent control hub system fixes this five-dimension transformation as the standard process of data production.

A scene: the customer questions "you have only installed a few current and temperature sensors, what can you see?" The engineer opens the platform: on the same transformer, the data page lists more than 300 features — the mean, extreme, standard deviation and rate of change of the three-phase current, the THD and the 3rd/5th/7th harmonic components, the temperature trend rate and peer-level temperature difference, the current–temperature coupling coefficient... Every feature is a "translation": translating the raw waveform into the language of hazards. It is not that the sensors are few; it is that feature engineering multiplies the value of every sensor dozens of times.

1. Statistical Features: First Sketch the "Profile" of the Data

Statistical features are the foundation of feature engineering — they do not change the data itself, only "look at it" from different angles:

Mean — the steady-state level: how large is this current on average? Extremes — the extreme situation: how high is the peak? How low is the trough? Standard deviation — the degree of fluctuation: is the data stable or jumping up and down? Rate of change — the speed of change: is it climbing slowly or jumping suddenly? Cumulative quantity — accumulation over time: how much was consumed in this period, how many times the limit was exceeded?

In one sentence: statistical features answer "what this stretch of data looks like" — the profile comes out first, and anomalies then have a reference.

A set of numbers shows how statistical features "translate": for a certain busbar current, the mean of 800 A states the load level, the extreme of 1200 A states the peak impact, the standard deviation of 15 A states very small fluctuation (the equipment runs steadily), and the rate of change suddenly jumping from 0.2%/min to 8%/min states a state mutation — the same set of data, five statistics tell five stories.

2. Frequency-Domain Features: "Take Apart" the Waveform into Components

A time-domain waveform is a "mixed" whole; a frequency-domain transform splits it into components of different frequencies — harmonic problems rely entirely on this dimension:

Harmonic components of each order — how large are the 3rd, 5th, 7th... harmonics? Which typical loads do they correspond to? THD (total harmonic distortion) — the proportion of harmonics in the total, the core indicator of power quality (GB/T 14549-1993 sets limits exactly around it); Inter-harmonics — components of non-integer multiples of the fundamental, often characteristic signals of variable-frequency drives and arc-type loads.

In one sentence: frequency-domain features answer "which frequency components are hidden in the waveform" — once harmonics and inter-harmonics are taken out, power-quality problems and special loads appear.

A THD example: the THD of a certain variable-frequency drive loop climbed from 8% to 18%. Looking at the waveform alone, "the current has some burrs" does not state how serious it is; once the frequency domain is checked — the 5th harmonic accounts for the largest proportion (the rectification characteristic of the drive), the 7th next, and the THD exceeds the limit of GB/T 14549 — the frequency domain translates "burrs" into "evidence of harmonic violation", and the governance direction (adding filtering) becomes clear immediately.

3. Time–Frequency Features: See "Which Frequency" and "When" Together

A pure frequency domain loses time information — do harmonics exist continuously or appear only momentarily? Time–frequency transforms (short-time Fourier transform, wavelets, etc.) look at "frequency–time" together:

Time-segment spectrum — which frequencies are active in a certain time segment? The harmonic burst at the start moment and the characteristic frequency at the fault moment can be located; Trend frequency band — how does the energy of a certain frequency band change over time? A slowly deteriorating characteristic signal is visible in the energy trend of the band.

In one sentence: time–frequency features answer "which frequency appears at what time" — transient events and slow deterioration cannot escape its eyes.

A time–frequency example: when a certain motor starts, harmonics burst momentarily and fall back within a few seconds — a pure frequency domain sees "high harmonics" and treats it as a continuous problem; the time–frequency domain sees "appears only in the 3 seconds of starting" and immediately identifies it as a starting transient rather than long-term pollution. The three words "at what time" turn a misjudgment into a judgment.

4. Correlation Features: Let Different Measuring Points "Talk"

A single-point feature is a "monologue"; correlation features let multiple measuring points "talk" — hazards are often hidden in relationships:

Temperature-rise–current coupling — is the temperature rising with current at a normal proportion? An abnormal coupling coefficient means deteriorating heat dissipation or increasing contact resistance; Leakage–temperature coupling — does leakage accelerate abnormally as temperature rises? The typical pattern of insulation deterioration; Voltage–load linkage — is the voltage fluctuation strongly correlated with the load mutation? Helps judge whether the voltage-quality problem comes from inside or outside.

In one sentence: correlation features answer "whether the relationship between two quantities is right" — an abnormal coupling relationship often exposes a hazard earlier than a single-point value.

A correlation example: the temperature of a certain loop rose normally with current, but one day the coupling coefficient suddenly doubled — with the same 100 A current, the temperature was higher than usual by a whole margin. Looking only at the temperature "5 degrees higher" might be the environment; the correlation feature directly points out that "the temperature-rise–current relationship is abnormal" — check the cooling fan or contact resistance, and the direction is immediately clear. When the coupling relationship changes, it often alarms earlier than the value itself.

5. Topology Features: Turn Position Information into Judgment Power

Data is not isolated — every measuring point has its position in the distribution topology; topology features turn "position" into judgment power:

Position weighting — the same value has completely different meanings at the incoming line, busbar or end point; weighting by position makes the judgment more accurate; Peer-level temperature difference — a horizontal comparison of the temperature difference of devices at the same level automatically highlights the abnormal device (the peer-level temperature difference is a classic feature of busbar-connection hazards); Cascade tracing — how does a lower-level anomaly propagate to the upper level? Trace along the topology level by level and locate "at which level the source is".

In one sentence: topology features answer "where this value appears and how it compares with its neighbors" — the sense of position upgrades the judgment from "whether the value is right" to "fault location".

A topology example: the temperatures of three feeder cabinets in a workshop rose at the same time — looking at each one "3 degrees high" alone looks like a coincidence; the topology comparison found that all three cabinets take power from the same busbar segment: the problem is not in the three cabinets but in their common upper-level busbar. Tracing level by level, the source (the busbar connection) floats up — this is the power of "the sense of position".

6. From 200 Raw Samples to 500+ Features: The Reorganization Process

The five-dimension transformation is not an abstract concept; on the ground it is a pipeline:

Step 1: raw waveforms into storage — 200–500 sets of raw data (current/voltage/temperature waveforms) are stored first; this is the raw material; Step 2: five-dimension transformation calculation — calculate in batches along the five dimensions of statistics, frequency domain, time–frequency, correlation and topology, dozens of features per dimension; Step 3: feature-library assembly — the transformation results are organized into a feature library by measuring point, by device and by level; the hundreds of features of one measuring point hang on the object model; Step 4: model consumption — the diagnosis model takes data from the feature library rather than reading the raw waveforms directly — what the model sees is "translated language".

In one sentence: raw data is the raw material, feature engineering is the processing, and the feature library is the finished product — the model consumes the finished product, not the raw material.

The feature expansion of one measuring point: the raw waveform of the three-phase current, after the five-dimension transformation — the statistics dimension gives the mean/extreme/standard deviation/rate of change, the frequency dimension gives the THD/3rd/5th harmonics, the time–frequency dimension gives the start-segment spectrum, the correlation dimension gives the current–temperature coupling coefficient, and the topology dimension gives the temperature difference with peer-level devices — one measuring point yields 40+ features, and a transformer with a hundred measuring points yields several hundred features. This is the amplification logic of "200–500 raw samples → 500+ features": the data does not increase; every datum is asked "are you all right" a dozen times — the more features, the more accurate the answer.

7. Feature Quality Determines Model Credibility: Three Quality Gates

No matter how many features there are, if the quality fails, it is all wasted effort. Three quality gates keep watch:

Quality gate 1: physical meaning — every feature must be able to state "what physical state it represents" (temperature-rise rate = insulation/heat-dissipation health); unstateable "black-box features" would rather be absent than wrong; Quality gate 2: stability — under normal conditions the feature value should be stable; a feature fluctuating too much will misjudge the normal state as an anomaly; check the baseline stability before going online; Quality gate 3: discrimination — the feature must clearly distinguish "normal/abnormal" (the fault-state value is separated from the healthy-state value); a feature without discrimination fed to the model only adds noise.

In one sentence: the three feature-quality questions — can it be stated? can it stay stable? can it be separated? If all three pass, the feature is the model's nutrition; if any one fails, it is the model's noise.

The cost of a bad feature: a feature with no physical meaning, unable to stay stable, and unable to separate normal from abnormal, fed to the model — at best adds noise and slows training; at worst misjudges the normal state as a fault, and one false alarm is enough to make the site lose trust in the system. So the quality gate is not "frosting on the cake"; it is the "admission ticket": every feature in the library must pass the three questions — better fewer and accurate than many and messy.

8. The Four-Step Landing Method of Feature Engineering

Step 1: list the diagnostic targets — which hazards should the platform report (overheating/harmonics/unbalance/insulation deterioration)? The target list comes first; Step 2: set the feature list — which features correspond to each target (selected from the five dimensions), one feature table; Step 3: compute the feature baseline — after going online, run for a period first and compute the normal interval of every feature as the reference for judging deviation; Step 4: build quality monitoring — regularly check the stability and discrimination of the features; recompute or replace a feature when it degrades.

In one sentence: targets → list → baseline → monitoring; with the four steps done, feature engineering changes from "knows how to compute" to "usable" — and data production is truly closed.

9. Back to FEXLINK: How Feature Engineering Lands

To land this logic on products: the Taiyi intelligent control hub system takes feature engineering as the core of data production — the raw waveforms collected by the sensing layer enter the platform and, after the five-dimension transformation of statistics, frequency domain, time–frequency, correlation and topology, are reorganized into hundreds of features; the feature library hangs on the object model for diagnosis, early-warning and optimization models to use. The sensors decide "whether it can be collected"; feature engineering decides "whether it can be understood" — with the same batch of sensors, the better the feature engineering, the more and the more accurate the hazard conclusions; this is exactly the meaning of "data production". During selection, ask "how many features can your platform compute for one measuring point" — the one answering "dozens to hundreds, divided into five dimensions" is the one that has truly turned data into an asset; the one answering "just voltage, current and temperature" shows that feature engineering is not yet its core.

Scope and Limitations

  • The conclusions in this article are intended for LV distribution and electrical-safety sites. Standard numbers (GB/T 14549-1993) refer to their currently effective editions, and product parameters follow the latest FEXLINK selection handbook and product knowledge base.
  • Quantitative criteria (protection zone, discharge and residual voltage, earth resistance, thresholds and grading) hold under the preconditions stated in this article. Where the site voltage level, earthing arrangement, environmental category or load profile differs, the figures must be recalculated. Specific configurations involving Taiyi control hub should be verified against actual circuit parameters.
  • Specific product models and parameters are subject to the latest selection handbook and product knowledge base; this article does not constitute selection or procurement advice.
  • Standard numbers cited here refer to the currently effective editions; when a standard is revised, the related quantitative conclusions must be re-checked.
  • Deployment and implementation for specific projects must be carried out by qualified engineers based on site survey and design drawings.

Frequently Asked Questions

How 200

The sensors collect only voltage, current and temperature — how can the platform report dozens of hazard conclusions