← News
Tech for Good Tech for Good Frontiers

The sensor network’s biggest discovery was what it wasn’t measuring

The sensor network’s biggest discovery was what it wasn’t measuring
5 Devices in network
June 2021 To June 2023 Study period
4 Sensors with paired readings
0% CO2 coverage
421,937 sensor readings
205,005 temperature-humidity frames

A two-year sensor network was supposed to track four things about the air inside Grenoble’s Perret Tower. It was actually measuring two—and one of the missing variables, carbon dioxide, appeared in none of the 421,937 recorded frames. That discovery changed the study before its statistical analysis even began.

The finding is a useful warning for anyone trying to make decisions from environmental data. A network can look rich because it produces hundreds of thousands of readings, while the variables researchers hope to compare may never have been measured together. Mamadou Sow’s audit of the Tour Perret LoRaWAN dataset shows why checking what is genuinely present—not just what a project was designed to collect—can be as consequential as choosing the right machine-learning model.

LoRaWAN is a wireless networking standard built to let small, battery-powered devices send modest amounts of data over long distances. It is increasingly used in environmental monitoring, where many devices can gather readings over years. But long-lived sensor networks are assembled from hardware and software that change, transmit on different schedules, and sometimes report different fields. The result is not automatically a coherent, synchronized picture of the environment.

The Science

Sow, affiliated with LIPN-CNRS and Université Sorbonne Paris Nord, examined telemetry from five devices at the Perret Tower in Grenoble, France, recorded between June 2021 and June 2023. The raw archive held 421,937 frames. The initial plan was to analyze temperature, relative humidity, atmospheric pressure, and carbon dioxide together. Before doing that, the study built a coverage audit into the data-extraction pipeline: it checked which fields appeared, how often they appeared, and which devices supplied them.

The audit found that the four variables did not form a four-variable dataset. Carbon dioxide was absent. Pressure came only from a separate Wyres device, while four Elsys EMS sensors supplied the main temperature-and-humidity stream. The devices did not share a synchronized, frame-by-frame record. A complete-case analysis requiring all four readings at once would therefore have produced no usable rows. Trying to align readings across separate devices would risk treating measurements from different sensors and locations as if they were simultaneous observations.

Instead, the analysis focused on the largest consistently co-occurring pair: temperature and humidity from the four Elsys devices. The extraction code searched flexibly through differently structured JSON payloads, parsed timestamps, and removed repeated receptions of the same wireless transmission. That matters because a transmission can be received by several gateways and appear multiple times in raw logs. The pipeline identified duplicates using the device identifier and frame counter, then checked readings against the sensors’ stated physical ranges.

After these steps, the study retained 205,005 complete temperature–humidity frames. That was 95.76% of the 214,076 retained frames in the paper’s post-deduplication, variable-bearing corpus. The distinction between that denominator and the original 421,937 raw frames is important: coverage percentages change depending on whether they refer to the full raw archive or the retained corpus. The paper reports temperature in 93.30% and humidity in 91.15% of raw frames; pressure in 2.15%; and carbon dioxide in none.

Coverage differed sharply by variable

Percent of the 421,937 raw frames containing each variable, as reported in the paper’s coverage table. Pressure and carbon-dioxide coverage are especially limited or absent.

Coverage differed sharply by variable
LabelValue
Temperature93.3 % of raw frames
Humidity91.15 % of raw frames
Pressure2.15 % of raw frames
Carbon dioxide0 % of raw frames

The researchers then used several methods that look for different kinds of data irregularity. A modified Z-score checks whether a single reading is unusually far from the median, using the median absolute deviation—a robust measure of spread that is less distorted by extreme values than the usual mean and standard deviation. Isolation Forest looks instead for unusual combinations of variables: it repeatedly partitions the data and treats points that are easy to isolate as potential anomalies.

Principal Component Analysis, or PCA, summarizes the strongest patterns in a dataset by rotating its axes into new directions of variation. Here, the main direction captures the relationship between temperature and humidity; the second captures departures from that dominant pattern. The pipeline also applies clustering, which groups nearby observations, and a multilayer perceptron (MLP), a small neural network trained to predict humidity using temperature and cyclical time features. If predictions are consistently poor in a new period, that mismatch could help flag sensor drift—a gradual change in a sensor’s calibration.

The paper’s central methodological point is not that one detector is best. It is that the techniques are complementary: a value can be ordinary on its own but unusual in combination with another variable, while a gradual drift may stay within physical limits and escape point-by-point checks altogether.

What They Found

Temperature and humidity moved in opposite directions in the data, with a Pearson correlation of (r=-0.662). Correlation describes how two measurements vary together; it does not establish that one causes changes in the other. Still, this relationship gives the researchers a baseline for identifying combinations that do not resemble the network’s usual pattern.

PCA found that the first component accounted for 83.10% of the variance in the standardized two-variable data, with the second accounting for 16.90%. Put simply, most of the observed variation lay along one dominant temperature–humidity axis. The remaining share is not “noise” by definition: it includes departures from that axis, which could reflect real microclimate changes, measurement behavior, or both.

Most variation followed one PCA axis

Share of variance represented by the two principal components in the standardized temperature–humidity dataset.

Most variation followed one PCA axis
LabelValue
PC183.1 % of variance
PC216.9 % of variance

The difference between one-variable and joint anomaly detection was striking. The modified Z-score flagged no readings as univariate outliers, while Isolation Forest identified 2.12% of observations as multivariate anomalies. These methods ask different questions. A temperature or humidity reading can sit within its ordinary range yet still be unusual when paired with the other measurement. The 2.12% should therefore be read as the share flagged by this model under its chosen settings—not as a confirmed count of faulty sensors or bad weather.

DBSCAN, a density-based clustering method, labeled 0.17% of observations as noise and identified 13 clusters after calibration. “Noise” here means observations that did not meet the algorithm’s local-density rule; it is not a verdict that those readings are erroneous. The researchers selected the neighborhood radius using a k-distance plot, a diagnostic that looks for a bend in the distance to each point’s tenth-nearest neighbor. The paper reports a selected radius of approximately 0.071 in standardized space, with a minimum cluster-neighbor count of 10.

Figure 5: k-distance plot (k=minPts=10) used to objectively select the DBSCAN neighborhood radius ε on real-world data. Candidate thresholds are overlaid; the knee of the curve, identified via maximum perpendicular distance to the chord, is retained at ε≈0.071.
Figure 5: k-distance plot (k=minPts=10) used to objectively select the DBSCAN neighborhood radius ε on real-world data. Candidate thresholds are overlaid; the knee of the curve, identified via maximum perpendicular distance to the chord, is retained at ε≈0.071. Source: Mamadou Sow

The machine-learning model’s performance was more modest than its role in the pipeline might suggest. In cross-validation—a procedure that repeatedly trains and tests on separate portions of the data—the MLP achieved an (R^2=0.672) and a root mean squared error of 10.54 percentage points of relative humidity. R-squared summarizes how much of the variation in the held-out values is captured by the model, while RMSE gives a typical scale of prediction error. This is a useful but imperfect prediction, not a precise physical law. The paper proposes large residuals as a possible drift signal; it does not establish that a particular sensor was drifting or report a validated drift event.

Figure 7: Actual Pearson correlation matrix, Temperature–Humidity (n=205,005).
Figure 7: Actual Pearson correlation matrix, Temperature–Humidity (n=205,005). Source: Mamadou Sow

Taken together, the results show why a pipeline needs more than a single outlier filter. The univariate method found no extreme marginal values; the multivariate detector found a subset with unusual pairings; density clustering isolated a smaller set of sparse points; and the predictive model offered a different, time-aware way to examine deviations. These categories overlap conceptually, but they do not mean the same thing, and the paper does not present them as interchangeable counts of bad data.

Why This Changes Things

The most consequential result may be the one that came before all those percentages: the audit prevented the team from analyzing a dataset that did not exist. If the researchers had assumed that temperature, humidity, pressure, and carbon dioxide arrived together, they could have generated an empty dataset, silently discarded most observations, or joined asynchronous readings in a way that implied false simultaneity. Each failure could still produce charts and model output. The problem would be that the output answered a question the sensors had never actually measured.

This concern reaches well beyond one tower. Cities are adding networks to track heat, air quality, moisture, noise, and other local conditions. Those measurements can inform building management, public-space design, climate adaptation, and public-health decisions. But sensor counts alone do not guarantee useful evidence. A large archive may contain substantial duplication, gaps, device-specific variables, and changes in how payloads are encoded. An audit that reports coverage by variable and device makes those weaknesses visible while there is still time to revise the analysis.

The dataset also illustrates why “more variables” is not always better. Pressure was present in 9,071 raw frames, but only on the separate Wyres stream; the authors excluded it from the synchronized temperature–humidity analysis rather than force an alignment. In the coverage table, this is 2.15% of all raw frames. The abstract’s 4.24% figure uses a different denominator—the 214,076 retained frames—so the two percentages describe the same count in different ways. Making denominators explicit is part of data quality, not bookkeeping trivia.

The study’s multi-method approach also offers a more nuanced picture of sensor reliability. A physical-range check can catch impossible values but not every implausible relationship. A marginal outlier detector can miss a point that is ordinary in each variable separately but strange as a pair. A density method can identify sparse observations without knowing whether they are sensor failures or rare conditions. And prediction residuals may reveal a slowly changing mismatch, but a model can also be wrong because the environment changed in a way it did not learn.

Those distinctions matter for practical deployments. Treating every flagged reading as defective could erase rare but real events. Treating every unflagged reading as trustworthy could miss subtle degradation. The sensible role of an automated pipeline is to sort observations into useful review categories, preserve provenance, and guide maintenance or investigation—not to replace judgment about what happened in the physical world.

Figure 8: Scree plot of eigenvalues (bivariate PCA: T, H, real data). PC1 alone captures 83.10% of total variance.
Figure 8: Scree plot of eigenvalues (bivariate PCA: T, H, real data). PC1 alone captures 83.10% of total variance. Source: Mamadou Sow

The implementation is another part of the paper’s contribution. Sow describes a C++17 pipeline with a flexible JSON search, a coverage audit during extraction, deduplication, and methods tailored to the two-variable dataset. For PCA, a two-dimensional correlation matrix has a direct algebraic solution, so this particular calculation does not need a general-purpose iterative solver. For DBSCAN, the paper uses a uniform grid to limit neighborhood searches, addressing the memory and computational demands that can arise when clustering more than 200,000 observations. These engineering choices are relevant where monitoring systems need to process large archives or run with constrained computing resources.

But a fast implementation does not make the findings universal. The analysis concerns one deployment, one tower, four Elsys sensors, and one two-year period. A correlation observed there may not hold in another building, climate, or sensor configuration. Nor does the study show that the flagged points correspond to known instrument failures or verified atmospheric events. It demonstrates a way to audit and interrogate data, not a universal calibration standard for urban sensors.

What’s Next

The next test is whether this workflow helps operators make better decisions outside a single retrospective analysis. That could mean applying the same coverage-first extraction to other LoRaWAN deployments, comparing results across buildings and sensor models, and checking whether the anomaly categories match maintenance logs or independently calibrated reference instruments. Such comparisons would reveal which flags tend to signal hardware problems, which capture unusual but real conditions, and which depend heavily on a model’s thresholds.

The proposed drift detector especially needs a longer evaluation. The MLP’s cross-validated prediction error shows that temperature and time contain useful information about humidity in this dataset, but an imperfect model can produce residuals for many reasons. A convincing drift alarm would need to distinguish progressive sensor bias from weather patterns, changes in building use, missing data, and other shifts in the underlying environment. Ground-truth calibration checks over time would help establish that distinction.

Coverage auditing can also become more informative than a one-time report. If a device stops sending humidity, changes its payload format, or begins reporting a new field, an automated monitor could alert operators before a downstream analysis silently changes. Publishing clear denominators, device-level coverage, and the rules used to exclude or retain data would let other teams reproduce the analysis and decide whether its dataset fits their own question.

The lesson is both modest and consequential: before asking a sensor network what the air is doing, ask what the network actually recorded. That discipline cannot repair missing measurements or prove that an anomaly is a malfunction. It can prevent a more basic mistake—building confident conclusions on variables that were never jointly observed—and make the next generation of urban monitoring more trustworthy.