Portable Emission Measurement System (PEMS)

• Portable Emission Measurement System (PEMS) is a mobile platform integrating advanced gas analyzers and sensors to capture high-frequency pollutant data from vehicles in real-world conditions.
• It utilizes various sensing modalities such as FTIR and NDIR within synchronized data acquisition architectures to enable precise monitoring and regulatory compliance.
• Applications range from basic phone-based diagnostics to embedded machine learning models for transient emission prediction and real-time operational feedback.

A Portable Emission Measurement System (PEMS) is a mobile instrumentation platform designed to directly quantify pollutant emissions from internal combustion engines and vehicles during real-world operation. PEMS enables regulatory compliance verification, on-road emissions modeling, and research into transient engine behavior by capturing high-frequency, multi-gas exhaust data that reflects the complex dynamics encountered outside laboratory settings. Recent deployments range from calibrated FTIR-based laboratory instruments to minimal phone-based setups and sensor-amalgamated digital twins, underpinning both regulatory and scientific applications (Biewer et al., 2021, Shahnavaz et al., 2021, Sundaram et al., 27 Jan 2026).

1. Instrumentation and Sensing Modalities

PEMS consist of specialized gas analyzers, flow measurement units, and supporting transducers that are physically mounted to either a vehicle's exhaust system or integrated into construction equipment. Instrument examples include:

• E9000 Plus Gas Analyzer: Incorporates electrochemical cells for primary regulated species (CO, NO, NO₂, SO₂, O₂, H₂S) and non-dispersive infrared (NDIR) detection for CO₂ and CH₄, with channel ranges and accuracies specified as follows:
  • NO: 0–5000 ppm, ±5 ppm or ±5% reading
  • CO: 0–10,000 ppm, ±5 ppm or ±5% reading
  • CO₂: 0–50% vol, ±3% (<8% CO₂) or ±5% (<50% CO₂)
  • Sample conditioning via Pt100/thermocouple temperature compensation and gauge pressure monitoring (Shahnavaz et al., 2021).
• FTIR Spectroscopy Modules: Used in vehicle PEMS for regulated gases (NOₓ, CO, CO₂, THC), with typical accuracy of ±(1–3)% per species, sampling raw exhaust at up to 5 Hz. Built-in calibration/zero-span routines mitigate sensor drift (Sundaram et al., 27 Jan 2026).

Sampling rates range from 1 Hz (low-cost platforms) to 100 Hz (IMU-enriched studies), with temperature and pressure compensation executed in hardware and/or software pipelines.

2. Data Acquisition Architecture and Signal Processing

PEMS data pipelines interleave multiple temporal streams: raw gas sensor outputs, engine operating conditions (OBD-II, CAN, or direct ECU access), and high-frequency ancillary signals (e.g., inertial, GPS). Architectures include:

• Android-based Minimal PEMS: Combines an off-the-shelf phone, a Bluetooth OBD-II adapter (~$10, ELM327 protocol at 38,400 baud), and an app (LolaDrives) implementing runtime diagnostics via RTLola, sampling OBD PIDs (velocity, mass air flow, fuel rate, oxygen sensor λ) at 1 Hz. Acceleration is derived via finite differencing of velocity, with GPS and ambient temperature furnished by Android system APIs. All streams are timestamped, buffered, and aligned at a 1 Hz grid (Biewer et al., 2021).
• Precision Field and Bench Data Collection: FTIR PEMS on a test vehicle (BMW 530e), recording 146 channels (species, temperatures, pressures) at 5 Hz through on-board acquisition during regulatory drive cycles (urban, rural, motorway), with post-run signal alignment and normalization. For ML-integrated studies, heavy equipment (e.g., excavators) supports simultaneous inertial sensor and tailpipe PEMS recording, with cross-modal synchronization at the start/stop boundary and interpolation onto common timelines (100 Hz grids for sensor–exhaust integration) (Shahnavaz et al., 2021, Sundaram et al., 27 Jan 2026).

PEMS datasets are preprocessed by low-pass filtering (by holding between samples or explicit sliding-window averaging), segment windowing (e.g., 0.25 s with overlap), and rejection of missing-data intervals. Emission vectors are paired window-wise with labeled input features for statistical or learning-based model development.

3. Pollutant Mass-Emission Models and Compliance Metrics

Emission quantification using PEMS relies on integrating high-frequency pollutant concentration and exhaust flow measurements. Canonical computation structures:

• CO₂ (from MAF):

  $$\dot{m}_\mathrm{fuel}(t) = \frac{3600\,\mathrm{MAF}(t)}{\mathrm{AFR}}$$

  $$\dot{m}_{\mathrm{CO}_2}(t) = \dot{m}_\mathrm{fuel}(t) \times \frac{44.01}{12.01}$$

  $$E_{\mathrm{CO}_2}(0,T) = \int_0^T \dot{m}_{\mathrm{CO}_2}(t)\,dt$$

  $$\mathrm{CO}_2\ \left[\frac{g}{km}\right] = \frac{E_{\mathrm{CO}_2}(0,T)}{\mathrm{distance}}$$

  Discretization uses summation at the PEMS sampling rate (Biewer et al., 2021).

• NOₓ (from ppmv + flow):

  $$\dot{V}_{exh}(t) = \frac{\mathrm{MAF}(t)}{\rho_{exh}(t)}$$

  $$\dot{m}_{\mathrm{NO}_x}(t) = [\mathrm{NO}_x]_{\mathrm{ppmv}}(t) \cdot \dot{V}_{exh}(t) \cdot \frac{46}{22.4 \times 10^{-3}}$$

  Resulting mass emissions are integrated and normalized by trip distance.

Compliance with regulations (e.g., EU RDE) is validated via both drive structural metrics (segment minimums, ambient/altitude windows) and raw pollutant thresholds (e.g., NOₓ ≤ 168 mg/km) (Biewer et al., 2021).

4. Integration with Data-Driven Modeling and Advanced Architectures

PEMS ground-truth data form the cornerstone for supervised learning and system identification in both regression and predictive control domains:

• Direct Emission Estimation: Simultaneous inertial and PEMS data enables windowed label–feature pairing; for example, 0.25 s overlap windows, with pollutant means as targets and inertial statistics (mean, IQR, peaks of specific axes across sensors) as features. Random Forest regression achieved $R^2$ of 0.94 (CO, CO₂) and 0.91 (NOₓ) on test sets, outperforming other machine learning baselines (Shahnavaz et al., 2021).
• Latent Space Predictive Architectures: JEPA (Joint Embedding Predictive Architecture) leverages PEMS FTIR data, encoding a history of regulated species and controls into a compact latent space ($D=50$). Predictive inference occurs by mapping past emissions and scheduled engine inputs into future emission trajectories, optimized via compounded MSE, regularization, and covariance-penalty terms. Latent architectures outperform LSTM by ∼28% (WMSE) on transient emission prediction, with marked improvement in NOₓ/CO₂ ramp and peak fidelity (Sundaram et al., 27 Jan 2026).

Model optimization for embedded deployment employs structured pruning (up to 30% parameter culling) and quantization (bfloat16 standard), reducing model size and inference latency with minor error increases. INT8 quantization appears suboptimal without hardware-specific kernel and fine-tuning adjustments.

5. Operational Constraints, Calibration, and Performance

PEMS accuracy and practical reliability hinge on rigorous calibration and validation procedures:

• Electrochemical and NDIR sensors are factory-calibrated and zero-checked before field deployment. FTIR-based units undergo zero/span routines before and after each drive to correct baseline drift. Quoted channel accuracies are typically ±(1–3)% (FTIR) or as detailed in device manufacturer tables (Shahnavaz et al., 2021, Sundaram et al., 27 Jan 2026).
• Minimal PEMS (OBD-based, e.g. LolaDrives) are classified as "un-calibrated" and intended for indicative use only, with lab comparisons indicating ±5% error for CO₂ and ±15% for NOₓ relative to benchtop references—provided onboard sensors are themselves calibrated (Biewer et al., 2021).
• Phone-based monitoring (Snapdragon 660, 2018) with ~30 streams/50 outputs and 1 Hz processing yields <5% CPU usage, <150mA power, ~50MB memory, and end-to-end UI update of <100 ms (Biewer et al., 2021).

Key regulatory alignment includes drive time (90–120 min), ambient/altitude compliance, and segment distance allocation per EU 2017/1151. Violations are flagged via runtime triggers and post-analysis.

6. Representative Applications and Comparative Results

PEMS are deployed for both regulatory compliance and method development:

• Phone-Based RDE Monitoring: LolaDrives on two real-world Audi A6 45-TDI (Euro 6d-TEMP) RDE runs recorded NOₓ emission violations (214 mg/km vs. EU limit 168 mg/km) and procedural test errors (driving dynamic exceedance). All UI signals (violation tracks, segment failures) coincided with ground-truth review (Biewer et al., 2021).
• Heavy Equipment Field Estimation: Windowed regression using paired inertial–PEMS data on excavator operations yielded high-fidelity predictions of transient CO, NOₓ, and CO₂ emissions, supporting real-time operational feedback (Shahnavaz et al., 2021).
• Latent Dynamic Model Generalization: JEPA, trained on large-breadth FTIR PEMS datasets (BMW 530e, bench and road), demonstrated enhanced transient generalization—tracking NOₓ peaks within ±0.05 s, correcting LSTM under-smoothing/lag in rapid torque ramps, and retaining performance post-pruning for embedded application (Sundaram et al., 27 Jan 2026).

Summarized test outcomes are as follows:

Platform	NOₓ [mg/km]	CO₂ [g/km]	Verdict
LolaDrives, Drive 1	214	183	NOₓ violation
LolaDrives, Drive 2	99	205	Segment dynamics violation

Source: (Biewer et al., 2021), Table aggregated.

7. Limitations and Future Directions

PEMS deployments face technical and operational challenges:

• Uncalibrated, minimal PEMS lack traceable reference quality but enable democratized in-the-wild emission diagnosis and large-scale, crowd-sourced analytics.
• High-cost FTIR and NDIR PEMS require continuous calibration and periodic maintenance to retain accuracy, especially on prolonged campaigns.
• Temporal alignment between high-frequency ancillary sensors and exhaust measurements is nontrivial and prone to error without rigorous timestamping and preprocessing.
• Emission dynamic modeling from limited ECU signals constrains observability. Expansion to additional sensory inputs (e.g., manifold pressure, exhaust temperature, fuel rate) is recommended for future model refinement (Sundaram et al., 27 Jan 2026).

Notably, integration of PEMS data with latent representation learning and embedded model compression opens pathways toward real-time, adaptive emission control in both conventional and hybrid powertrains (Sundaram et al., 27 Jan 2026).

• "RTLola on Board: Testing Real Driving Emissions on your Phone" (Biewer et al., 2021)
• "Automated Estimation of Construction Equipment Emission using Inertial Sensors and Machine Learning Models" (Shahnavaz et al., 2021)
• "A Latent Space Framework for Modeling Transient Engine Emissions Using Joint Embedding Predictive Architectures" (Sundaram et al., 27 Jan 2026)