Temporal Drift Research Articles

Dynamics of temporal Kerr cavity solitons in the presence of rapid parameter inhomogeneities: From bichromatic driving to third-order dispersion

Temporal Kerr cavity solitons are pulses of light that can persist in coherently driven, dispersive resonators with Kerr-type nonlinearity. Many studies have shown that such solitons can react to parameter inhomogeneities (e.g., variations in the complex amplitude of the driving field) by experiencing a temporal drift. The vast majority of such studies assume that the inhomogeneity varies slowly across the soliton, leading to the prediction that the soliton drift rate is linearly proportional to the gradient of the inhomogeneity at the soliton position. However, the assumption of a slowly varying inhomogeneity may not hold true in all situations, e.g., when using bichromatic driving or in the presence of third-order dispersion that gives rise to an extended dispersive wave tail. Here we report on theoretical and numerical results pertaining to the behavior of dissipative temporal Kerr cavity solitons under conditions where parameter inhomogeneities vary nonlinearly across the width of the soliton. In this case, the soliton velocity is dictated by the full overlap between its so-called adjoint neutral mode and the parameter perturbation, which we show can yield dynamics that are manifestly at odds with the common wisdom of motion dependent solely upon the gradient of the inhomogeneity. We also investigate how the presence of third-order dispersion and the associated dispersive wave tail changes the motion induced by parameter inhomogeneities. We find that the dispersive wave tail as such does not contribute to the soliton motion; instead, higher-order dispersion yields counterintuitive influences. Our results provide new insights into the behavior of temporal cavity solitons in the presence of parameter inhomogeneities and can impact systems employing pulsed or bichromatic pumping and/or resonators with non-negligible higher-order dispersion.

Signal distortion awakened from optical memory estimated using a calculation method with spatiotemporal separation.

The fidelity of photonic storage and retrieval is an essential criterion in long-distance all-optical network nodes. However, the recovered signals from optical memories based on the photon echo (PE) protocol are accompanied by undesired waveform variation and temporal drift. In this study, we use a numerical calculation method with spatiotemporal separation to investigate the essence of signal distortion. The results show that the asynchronous evolution of the macroscopic population difference and the macroscopic dipole moment with time is responsible for echo signal real distortion caused by phase shifts at the in-phase point of the recorded information. The constructive interference of the dipoles at the moment of reaching the in-phase point induces the photon emission, and this point with a nonspecific phase will be naturally accompanied by waveform changes, a small amount of time advance and delay of the PE signal, which is actually a false signal distortion. Such radiation mechanism of the inhomogeneous broadening media provides a perspective to accurately and correctly recognize the temporal drift and waveform variation of the recovered optical signal.

Calibration of a Stationary Multichannel GPR Monitoring System Using Internal Reflection Measurements

Advanced and extensive processing of ground-penetrating radar (GPR) data, for example, needed for full-waveform inversion approaches, requires a reliable temporal calibration of the system. Usually, the calibration of GPR systems is performed with a known medium between the transmitting and receiving antennas. Thereby, the observed time difference between the expected and measured signal arrival times, termed as time-zero, can be accounted for as a system-specific time delay. For measurement configurations where the antennas are permanently positioned around an object for monitoring purposes, time-consuming additional measurements where parts of the system need to be deinstalled would be required. This is not feasible for the proposed system. Therefore, novel calibration methods for such stationary monitoring systems are required to capture the temporal drift of time-zero. In this article, we present a novel calibration approach that uses internal signal reflections in the measurement system to derive the system-specific time delay without the necessity of knowing the medium between the antennas. We demonstrate that parasitic reflection and coupling signals can be used for accurate in situ calibrations. The presented approach is capable of identifying and correcting for the differences in hardware fabrication, while also correcting the temporal changes in time-zero during experiments. The presented approach is able to reduce the error in time-zero to below 25 ps, enabling high-resolution soil research. The presented approach is characterized by requiring no additional calibration setups or measurements since all the necessary data can be acquired during the original soil measurement.

WiSion: Bolstering MAV 3D Indoor State Estimation by Embracing Multipath of WiFi

Recent years have witnessed a remarkable growth of micro aerial vehicle (MAV) technologies, which are desirable for many applications, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e.g.</i> , warehouse inventory management and home entertainment. These indoor needs significantly improve work efficiency while posing fundamental challenges in the design of MAV state estimation. The state includes a vehicle's position, orientation, and velocity, which are fundamental to guide the motor control and adjust that vehicle's actions in autonomous flight. Existing vision-based solutions only work in well-lit texture-rich environments, while laser-based solutions are limited to MAVs' payload and budget. This paper presents WiSion, a robust and low-cost state estimator that leverages ubiquitous WiFi to estimate six-degree-of-freedom states for MAVs. Our observation is that the multipath of WiFi conceals a wealth of information about a vehicle's state, which helps combat the temporal drift of inertial sensors for smooth state estimation. We realize WiSion by an absolute-relative WiFi sensing module and a WiFi-inertial state estimation module. It works without knowing access points' (APs') positions. We implement the prototype with off-the-shelf products and conduct experiments in indoor venues. Our results show that WiSion achieves the position error of 35.25 cm and the attitude error of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$2.6^\circ$</tex-math></inline-formula> with a maximum linear velocity of 1.74 m/s. Moreover, WiSion can recover APs' positions and is robust to indoor hindrances such as obstacles and multipath.

The FRB 20121102A November rain in 2018 observed with the Arecibo Telescope

ABSTRACT We present 849 new bursts from FRB 20121102A detected with the 305-m Arecibo Telescope. Observations were conducted as part of our regular campaign to monitor activity and evolution of burst properties. The 10 reported observations were carried out between 1150 and $1730\, {\rm MHz}$ and fall in the active period around 2018 November. All bursts were dedispersed at the same dispersion measure and are consistent with a single value of $(562.4 \pm 0.1)\, {\rm pc\, cm^{-3}}$. The rate varies between 0 bursts and 218 ± 16 bursts per hour, the highest rate observed to date. The times between consecutive bursts show a bimodal distribution. We find that a Poisson process with varying rate best describes arrival times with separations ${\gt}{0.1\, {\rm s}}$. Clustering on time-scales of $22\, {\rm ms}$ reflects a characteristic time-scale of the source and possibly the emission mechanism. We analyse the spectro-temporal structure of the bursts by fitting 2D Gaussians with a temporal drift to each sub-burst in the dynamic spectra. We find a linear relationship between the sub-burst’s drift and its duration. At the same time, the drifts are consistent with coming from the sad-trombone effect. This has not been predicted by current models. The energy distribution shows an excess of high-energy bursts and is insufficiently modelled by a single power law even within single observations. We find long-term changes in the energy distribution, the average spectrum, and the sad-trombone drift, compared to earlier and later published observations. Despite the large burst rate, we find no strict short-term periodicity.

Series connection of quantum Hall resistance array and programmable Josephson voltage standard for current generation at one microampere

An accurate current can be realized by a series connection of an invariant voltage and resistance with zero interconnection resistance in the framework of Ohm’s law. In practice, however, accurate current realization is limited by the temporal drift and instability of employed resistors, solid-state voltage sources and interconnection wires against environmental fluctuations, such as temperature or atmospheric pressure. Here, we generated a current of 1 μA, determined by the Josephson voltage and quantized Hall resistance array (QHRA). A programmable Josephson voltage standard (PJVS), 1 MΩ QHRA and ultrastable low noise current amplifier (ULCA) are connected in series to generate and measure a current of 1 μA. The expected current is calculated as the ratio of the Josephson voltage to the QHRA plus the output resistance of the PJVS. The relative difference between the current measured by the ULCA and the calculated current is as small as 40 nA A−1 with a combined measurement uncertainty (k = 1) of 0.13 μA A−1.

A Transfer Learning Approach to Correct the Temporal Performance Drift of Clinical Prediction Models: Retrospective Cohort Study.

Clinical prediction models suffer from performance drift as the patient population shifts over time. There is a great need for model updating approaches or modeling frameworks that can effectively use the old and new data. Based on the paradigm of transfer learning, we aimed to develop a novel modeling framework that transfers old knowledge to the new environment for prediction tasks, and contributes to performance drift correction. The proposed predictive modeling framework maintains a logistic regression-based stacking ensemble of 2 gradient boosting machine (GBM) models representing old and new knowledge learned from old and new data, respectively (referred to as transfer learning gradient boosting machine [TransferGBM]). The ensemble learning procedure can dynamically balance the old and new knowledge. Using 2010-2017 electronic health record data on a retrospective cohort of 141,696 patients, we validated TransferGBM for hospital-acquired acute kidney injury prediction. The baseline models (ie, transported models) that were trained on 2010 and 2011 data showed significant performance drift in the temporal validation with 2012-2017 data. Refitting these models using updated samples resulted in performance gains in nearly all cases. The proposed TransferGBM model succeeded in achieving uniformly better performance than the refitted models. Under the scenario of population shift, incorporating new knowledge while preserving old knowledge is essential for maintaining stable performance. Transfer learning combined with stacking ensemble learning can help achieve a balance of old and new knowledge in a flexible and adaptive way, even in the case of insufficient new data.

Multisite evaluation of prediction models for emergency department crowding before and during the COVID-19 pandemic.

To develop a machine learning framework to forecast emergency department (ED) crowding and to evaluate model performance under spatial and temporal data drift. We obtained 4 datasets, identified by the location: 1-large academic hospital and 2-rural hospital, and time period: pre-coronavirus disease (COVID) (January 1, 2019-February 1, 2020) and COVID-era (May 15, 2020-February 1, 2021). Our primary target was a binary outcome that is equal to 1 if the number of patients with acute respiratory illness that were ED boarding for more than 4h was above a prescribed historical percentile. We trained a random forest and used the area under the curve (AUC) to evaluate out-of-sample performance for 2 experiments: (1) we evaluated the impact of sudden temporal drift by training models using pre-COVID data and testing them during the COVID-era, (2) we evaluated the impact of spatial drift by testing models trained at location 1 on data from location 2, and vice versa. The baseline AUC values for ED boarding ranged from 0.54 (pre-COVID at location 2) to 0.81 (COVID-era at location 1). Models trained with pre-COVID data performed similarly to COVID-era models (0.82 vs 0.78 at location 1). Models that were transferred from location 2 to location 1 performed worse than models trained at location 1 (0.51 vs 0.78). Our results demonstrate that ED boarding is a predictable metric for ED crowding, models were not significantly impacted by temporal data drift, and any attempts at implementation must consider spatial data drift.

The Bayesian Time Machine: Accounting for temporal drift in multi-arm platform trials.

Multi-arm platform trials investigate multiple agents simultaneously, typically with staggered entry and exit of experimental treatment arms versus a shared control arm. In such settings, there is considerable debate whether to limit analyses for a treatment arm to concurrent randomized control subjects or to allow comparisons to both concurrent and non-concurrent (pooled) control subjects. The potential bias from temporal drift over time is at the core of this debate. We propose time-adjusted analyses, including a "Bayesian Time Machine," to model potential temporal drift in the entire study population, such that primary analyses can incorporate all randomized control subjects from the platform trial. We conduct a simulation study to assess performance relative to utilizing concurrent or pooled controls. In multi-arm platform trials with staggered entry, analyses adjusting for temporal drift (either Bayesian or frequentist) have superior estimation of treatment effects and favorable testing properties compared to analyses using either concurrent or pooled controls. The Bayesian Time Machine generally provides estimates with greater precision and smaller mean square error than alternative approaches, at the risk of small bias and small Type I error inflation. The Bayesian Time Machine provides a compromise between bias and precision by smoothing estimates across time and leveraging all available data for the estimation of treatment effects. Prior distributions controlling the behavior of dynamic smoothing across time must be pre-specified and carefully calibrated to the unique context of each trial, appropriately accounting for the population, disease, and endpoints.

TDTMF: A recommendation model based on user temporal interest drift and latent review topic evolution with regularization factor

This paper constructs a novel enhanced latent semantic model based on users’ comments, and employs regularization factors to capture the temporal evolution characteristics of users’ potential topics for each commodity, so as to improve the accuracy of recommendation. The adaptive temporal weighting of multiple preference features is also improved to calculate the preferences of different users at different time periods using human forgetting features, item interest overlap, and similarity at the semantic level of the review text to improve the accuracy of sparse evaluation data. The paper conducts comparison experiments with six temporal matrix-based decomposition baseline methods in nine datasets, and the results show that the accuracy is 31.64% better than TimeSVD++, 21.08% better than BTMF, 15.51% better than TMRevCo, 13.99% better than BPTF, 9.24% better than TCMF, and 3.19% better than MUTPD ,which indicates that the model is more effective in capturing users’ temporal interest drift and better reflects the evolutionary relationship between users’ latent topics and item ratings.

Quantum-correlation-based free-space optical link with an active reflector

We present a quantum-correlation-based free-space optical(FSO) link over 250 m using an outdoor active reflector 125 m from the transceiver station. The performance of free-space optical communication can be significantly degraded by atmospheric turbulence effects, such as beam wander and signal fluctuations. We used a 660 nm tracking laser to reduce atmospheric effects, by analyzing the fast beam wander and slow temporal beam drift, using this information to correct the quantum channel alignment of the 810 nm signal photons. In this work, the active reflector consisted of a mirror, a 6-axis hexapod stage, and a long-range wireless bridge. The slow drift of the beam path due to outdoor temperature changes was steered and controlled using wireless optical feedback between the receiver units and the active reflector. Our work provides useful knowledge for improved control of beam paths in outdoor conditions, which can be developed to ensure high quality quantum information transfer in real-world scenarios, such as an unmanned FSO link for urban quantum communication or retro-reflective quantum communication links.

Sea ice surface temperatures from helicopter-borne thermal infrared imaging during the MOSAiC expedition

The sea ice surface temperature is important to understand the Arctic winter heat budget. We conducted 35 helicopter flights with an infrared camera in winter 2019/2020 during the Multidisciplinary Drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition. The flights were performed from a local, 5 to 10 km scale up to a regional, 20 to 40 km scale. The infrared camera recorded thermal infrared brightness temperatures, which we converted to surface temperatures. More than 150000 images from all flights can be investigated individually. As an advanced data product, we created surface temperature maps for every flight with a 1 m resolution. We corrected image gradients, applied an ice drift correction, georeferenced all pixels, and corrected the surface temperature by its natural temporal drift, which results in time-fixed surface temperature maps for a consistent analysis of one flight. The temporal and spatial variability of sea ice characteristics is an important contribution to an increased understanding of the Arctic heat budget and, in particular, for the validation of satellite products.

An Adaptive Temporal-Concept Drift Model for Sequential Recommendation

Recently, owing to the great advances in Web 2.0 and mobile devices, various online commercial services have emerged. Recommendation systems play an important role in dealing with abundant product information from massive numbers of online e-commerce transactions. Providing an accurate recommendation at the correct time to customers can contribute to a surge in business success. In this paper, an adaptive temporal-concept drift learning-based recommendation system, ATCRec, is developed for precisely tackling the sequential recommendation problem. We embed sequences of items into the latent spaces and learn both general preferences and sequential patterns concurrently via a recurrent neural network. Specifically, ATCRec captures dynamic changes in the temporal and concept drift contexts by modifying the gate units in a traditional recurrent neural network. The proposed model provides a unified and flexible network structure to learn and reveal the opaque variation of user preferences over time. We evaluate the robustness and performance of ATCRec on two real-world datasets, and the experimental results demonstrate that ATCRec consistently outperforms existing sequential recommendation approaches on various metrics. This indicates that integrating users' temporal information and concept drift variation through time are indispensable in improving the performance of recommendation systems.

An Adaptive Temporal-Concept Drift Model for Sequential Recommendation

Recently, owing to the great advances in Web 2.0 and mobile devices, various online commercial services have emerged. Recommendation systems play an important role in dealing with abundant product information from massive numbers of online e-commerce transactions. Providing an accurate recommendation at the correct time to customers can contribute to a surge in business success. In this paper, an adaptive temporal-concept drift learning-based recommendation system, ATCRec, is developed for precisely tackling the sequential recommendation problem. We embed sequences of items into the latent spaces and learn both general preferences and sequential patterns concurrently via a recurrent neural network. Specifically, ATCRec captures dynamic changes in the temporal and concept drift contexts by modifying the gate units in a traditional recurrent neural network. The proposed model provides a unified and flexible network structure to learn and reveal the opaque variation of user preferences over time. We evaluate the robustness and performance of ATCRec on two real-world datasets, and the experimental results demonstrate that ATCRec consistently outperforms existing sequential recommendation approaches on various metrics. This indicates that integrating users' temporal information and concept drift variation through time are indispensable in improving the performance of recommendation systems.

Improving the electrical and structural stability of highly piezoresistive nickel–carbon sensor thin films

Abstract. The family of sputter deposited granular metal-based carbon-containing sensor films is known for their high sensitivity transforming force-dependent strain into electrical resistance change. Among them nickel–carbon thin films possess a gauge factor of up to 30, compared to only 2 for traditional sensor films of metal alloys. This high sensitivity is based on disordered interparticle tunneling through barriers of graphite-like carbon walls between metal–carbon particles of columnar shape. Force and pressure sensors would benefit a lot from the elevated piezoresistivity. A disadvantage, however, is a disturbing temporal creep and drift of the resistance under load and temperature. This contribution shows how to stabilize such sensor films. A significant stabilization is achieved by partially replacing nickel with chromium, albeit at the expense of sensitivity. The more chromium used in these NixCr1−x-C layers, the higher the optimum annealing temperature can be selected and the better the electrical stabilization. A good compromise while maintaining sensitivities well above the standard of 2 is identified for films with x=0.5 to 0.9, stabilized by optimized temperature treatments. The stabilizing effect of chromium is revealed by transmission electron microscopy with elemental analysis. The post-annealing drives segregation processes in the layer material. While the interior of the layer is depleted of chromium and carbon, boundary layers are formed. Chromium is enriched near the surface boundary, oxidized in air and forms chromium-rich oxide sub-layers, which are chemically very stable and protect against further reactions and corrosion. As a result, creep and drift errors are greatly reduced, so that the optimized sensor coatings are now suitable for widespread use.

Getting the nod: Pediatric head motion in a transdiagnostic sample during movie- and resting-state fMRI.

Head motion continues to be a major problem in fMRI research, particularly in developmental studies where an inverse relationship exists between head motion and age. Despite multifaceted and costly efforts to mitigate motion and motion-related signal artifact, few studies have characterized in-scanner head motion itself. This study leverages a large transdiagnostic public dataset (N = 1388, age 5-21y, The Healthy Brain Network Biobank) to characterize pediatric head motion in space, frequency, and time. We focus on practical aspects of head motion that could impact future study design, including comparing motion across groups (low, medium, and high movers), across conditions (movie-watching and rest), and between males and females. Analyses showed that in all conditions, high movers exhibited a different pattern of motion than low and medium movers that was dominated by x-rotation, and z- and y-translation. High motion spikes (>0.3mm) from all participants also showed this pitch-z-y pattern. Problematic head motion is thus composed of a single type of biomechanical motion, which we infer to be a nodding movement, providing a focused target for motion reduction strategies. A second type of motion was evident via spectral analysis of raw displacement data. This was observed in low and medium movers and was consistent with respiration rates. We consider this to be a baseline of motion best targeted in data preprocessing. Further, we found that males moved more than, but not differently from, females. Significant cross-condition differences in head motion were found. Movies had lower mean motion, and especially in high movers, movie-watching reduced within-run linear increases in head motion (i.e., temporal drift). Finally, we used intersubject correlations of framewise displacement (FD-ISCs) to assess for stimulus-correlated motion trends. Subject motion was more correlated in movie than rest, and 8 out of top 10 FD-ISC windows had FD below the mean. Possible reasons and future implications of these findings are discussed.

Level 2 processor and auxiliary data for ESA Version 8 final full mission analysis of MIPAS measurements on ENVISAT

Abstract. High quality long-term data sets of altitude-resolved measurements of the atmospheric composition are important because they can be used both to study the evolution of the atmosphere and as a benchmark for future missions. For the final ESA reprocessing of MIPAS (Michelson Interferometer for Passive Atmospheric Sounding) on ENVISAT (ENViromental SATellite) data, numerous improvements were implemented in the Level 2 (L2) processor Optimised Retrieval Model (ORM) version 8.22 (V8) and its auxiliary data. The implemented changes involve all aspects of the processing chain, from the modelling of the measurements with the handling of the horizontal inhomogeneities along the line of sight to the use of the optimal estimation technique to retrieve the minor species, from a more sensitive approach to detecting the spectra affected by clouds to a refined method for identifying low quality products. Improvements in the modelling of the measurements were also obtained with an update of the used spectroscopic data and of the databases providing the a priori knowledge of the atmosphere. The HITRAN_mipas_pf4.45 spectroscopic database was finalised with new spectroscopic data verified with MIPAS measurements themselves, while recently measured cross-sections were used for the heavy molecules. The Level 2 Initial Guess (IG2) data set, containing the climatology used by the MIPAS L2 processor to generate the initial guess and interfering species profiles when the retrieved profiles from previous scans are not available, was improved taking into account the diurnal variation of the profiles defined using climatologies from both measurements and models. Horizontal gradients were generated using the ECMWF ERA-Interim data closest in time and space to the MIPAS data. Further improvements in the L2 V8 products derived from the use of the L1b V8 products, which were upgraded to reduce the instrumental temporal drift and to handle the abrupt changes in the calibration gain. The improvements introduced into the ORM V8 L2 processor and its upgraded auxiliary data, together with the use of the L1b V8 products, lead to the generation of the MIPAS L2 V8 products, which are characterised by an increased accuracy, better temporal stability and a greater number of retrieved species.

In-Flight Relative Radiometric Calibration of a Wide Field of View Directional Polarimetric Camera Based on the Rayleigh Scattering over Ocean

The directional polarimetric camera (DPC) is a Chinese satellite sensor with a large field of view (FOV) (±50° both along-track and cross-track) and a high spatial resolution (about 3.3 km at nadir) that operates in a sun-synchronous orbit. It is a difficult task to calibrate the in-flight relative radiometric variation of the sensors with such a wide FOV. In this study, a new method based on Rayleigh scattering over the ocean is developed to estimate the radiometric sensitivity variation over the whole FOV of DPC. Firstly, the theoretical uncertainty of the method is analyzed to calibrate the relative radiometric response. The calibration uncertainties are about 2–6.9% (depending on the wavelength) when the view zenith angle (VZA) is 0° and decrease to about 1–3.8% when VZA increases to 70°. Then, the method is applied to evaluate the long-term radiometric drift of the DPC. It is found that the radiometric response of DPC/GaoFen-5 over the whole FOV is progressively drifting over time. The sensitivity at shorter bands decreases more strongly than longer bands, and at the central part of the optics decreases more strongly than the marginal part. During the 14 months (from March 2019 to April 2020) of operational running in-orbit, the DPC radiometric responses of 443 nm, 490 nm, 565 nm, and 670 nm bands drifted by about 4.44–23.08%, 4.75–16.22%, 3.86–9.81%, and 4.7–16.86%, respectively, from the marginal to the central part of the FOV. The radiometric sensitivity has become more stable since January 2020. The monthly radiometric drift is separated into the relative radiometric part and the absolute radiometric part. The relative radiometric drift of DPC is found to be smoothly varying with VZA, which can be parameterized as a polynomial function via VZA. At last, the temporal radiometric drift of DPC/GaoFen-5 is corrected by combining the relative and absolute radiometric coefficients. The correction is convincing by cross calibration with MODIS/Aqua observation over the desert sites and improving the aerosol retrievals. The Rayleigh method in this study is efficient for the radiometric sensitivity calibration of wide FOV satellite sensors.

Drift characteristic analysis of additive manufactured Ag NPs-PEDOT:PSS flexible temperature sensor

Additive manufactured flexible sensors have prospect applications in wearable devices, healthcare, soft robotics, and metaverse scenarios. However, there is still performance gaps between the emerging additive manufactured flexible sensors and traditional subtractive manufactured rigid sensors. Temporal drift characteristic, as one of the important performance parameters of sensors, was evaluated and analyzed with inkjet printed Ag NPs-PEDOT:PSS flexible temperature sensor, and the potential influence mechanism were explained from the glass transition of the substrate, the swelling effect of water molecules, and melting effect of silver nanoparticles perspective, thereby providing solutions for bridging the performance gaps of additive manufactured flexible sensors.

Semiquantitative Classification of Two Oxidizing Gases with Graphene-Based Gas Sensors

Miniature and low-power gas sensing elements are urgently needed for a portable electronic nose, especially for outdoor pollution monitoring. Hereby we prepared chemiresistive sensors based on wide-area graphene (grown by chemical vapor deposition) placed on Si/Si3N4 substrates with interdigitated electrodes and built-in microheaters. Graphene of each sensor was individually functionalized with ultrathin oxide coating (CuO-MnO2, In2O3 or Sc2O3) by pulsed laser deposition. Over the course of 72 h, the heated sensors were exposed to randomly generated concentration cycles of 30 ppb NO2, 30 ppb O3, 60 ppb NO2, 60 ppb O3 and 30 ppb NO2 + 30 ppb O3 in synthetic air (21% O2, 50% relative humidity). While O3 completely dominated the response of sensors with CuO-MnO2 coating, the other sensors had comparable sensitivity to NO2 as well. Various response features (amplitude, response rate, and recovery rate) were considered as machine learning inputs. Using just the response amplitudes of two complementary sensors allowed us to distinguish these five gas environments with an accuracy of ~ 85%. Misclassification was mostly due to an overlap in the case of the 30 ppb O3, and 30 ppb O3 + 30 ppb NO2 responses, and was largely caused by the temporal drift of these responses. The addition of recovery rates to machine learning input variables enabled us to very clearly distinguish different gases and increase the overall accuracy to ~94%.