Heterostructure and composite coating with edge compatible analytics framework for multiparameter optical fiber point sensing

Abstract

A multilayered sensor includes an optical waveguide and a multilayered heterostructure coupled to the waveguide that includes a plurality of concentric thin-film sorbent layers, wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide. Also, method of training an inference model is provided that uses training data that comprises a combination of physics simulated synthetic data for the sensor node and empirical data for the sensor node.

Description

HETEROSTRUCTURE AND COMPOSITE COATING WITH EDGE COMPATIBLE ANALYTICS FRAMEWORK FOR MULTIPARAMETER OPTICAL FIBER POINT SENSINGCROSS REFERENCE TO RELATED APPLICATIONS:

[0001] This application claims priority to U.S. Provisional Patent Application SerialNo. 63 / 707,523, filed on October 15, 2024, and titled "Heterostructure and Composite Coating with Edge Compatible Analytics Framework for Multiparameter Optical Fiber Point Sensing,” the disclosure of which is incorporated herein by reference.STATEMENT OF GOVERNMENT INTEREST:

[0002] This invention was made with government support under grant numbers DE-EE0009632 and DE-NE0009210 awarded by the United States Department of Energy. The government has certain rights in the invention.FIELD OF THE INVENTION:

[0003] The disclosed concept pertains to fiber optic-based sensing, and, in particular, to a fiber optic-based sensor having a heterostructure including a plurality of sensing layers for multiparameter sensing and an edge compatible analytics framework for multiparameter optical fiber point sensing that may be used with the disclosed fiber optic-based sensor.BACKGROUND OF THE INVENTION:

[0004] Fiber optic sensing has emerged as a powerful sensing technology for a wide range of applications, such as energy infrastructure applications. Fiber optic sensing provides numerous advantages including: (1) elimination of electrical wiring, contacts, and power at the sensing location, (2) capability for distributed sensing through either discrete multi-point (i.e. quasi-distributed) or fully continuous (i.e. fully -distributed) sensing modalities, (3) immunity to electromagnetic interference, and (4) improved safety in explosive, flammable, and other potentially hazardous application environments.

[0005] So-called multiparameter sensing refers to sensing systems and modalities where more than one physical, chemical, or biological parameter is measured at the same time, often with a single sensor platform or integrated system. One example of multiparameter sensing is electronic nose technology7. In electronic nose technology7,independent chemical-resistive sensing elements constituting an array are simultaneously monitored to reconstruct signatures of complex gas mixtures. Similar technology that leverages the optical properties of sensing materials is photonic nose technology, which was originally proposed in the context of photonic crystal based thin film structures.

[0006] Recent work in photonic nose technology has shown alternative pathways to multiparameter sensing while retaining the advantages of an optical fiber evanescent wave sensing platform through schemes which include multivariate analysis. These alternative pathways include: (1) multiple sensor elements provided along the length of an optical fiber, (2) multiple independent fiber-based sensors provided in a fiber array / bundle, and (3) multi-wavelength interrogation of a single sensor node.

[0007] FIG. 1 is a schematic diagram illustrating an example sensing arrangement 5 according to pathway (1) above that employs cascaded gas sensing thin films. In particular, sensing arrangement 5 includes an optical fiber 10 having a section 15 wherein cladding 20 of optical fiber 10 has been removed to expose core 25 of optical fiber 10. Three gas sensing gas sensing lay ers 30A, 30B, and 30C are coupled to and surround core 25 in a spaced relationship (i.e.. cascaded) along section 15.

[0008] FIG. 2 is a schematic diagram illustrating an example sensing arrangement 35 according to pathway (2) above that employs a tip-coated multi-core fiber. In particular, sensing arrangement 35 includes a multi-core fiber 40 wherein an individual gas sensing thin film 45 is provided at the end of each core 50.

[0009] FIG. 3 is a schematic diagram illustrating an example sensing arrangement 55 according to pathway (3) above that employs a single composite sensory film. In particular, sensing arrangement 55 includes an optical fiber 60 having a section 65 wherein cladding 70 of optical fiber 60 has been removed to expose core 75 of optical fiber 60. A composite sensory film 80 is coupled to and surrounds core 75. Composite sensory film 80 may be, in one example, a mixed-matrix composite polymer film with functional nanofillers, or, in another example, plasmonic nanoparticle-based nanocomposite in metal oxide or organothioL

[0010] While pathways (1) and (2) above are common and straightfor ard to produce, pathway (3) has only been developed as a proof-of-concept and typically relies on a nanocomposite thin film structure including nanoparticles (NP) / nanocrystals (NC) embedded within a polymeric or ceramic host matrix as thesensing layer, which structure shows several different physical sensing mechanisms with wavelength dependence. In addition, based upon the results reported in prior publications, pathway (3) would require algorithmic discrimination of the raw signals from the sensing layer to resolve the cross-sensitivity between different gas analytes and temperature. In particular, multiple optical physics mechanisms that are associated with electron-phonon scattering, free carrier concentration, and mobility change can be extracted and interrogated separately using multivariate pattern recognition algorithms, such as Principal Component Analysis (PCA) and Support Vector Machines (SVM). Such multivariate gas / temperature sensing signal processing, however, involves additional computational requirements and the potential for correlated responses and incomplete discrimination of key analytes of interest.SUMMARY OF THE INVENTION:

[0011] In one embodiment, a multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes is provided. The sensor includes an optical waveguide and a multilayered heterostructure that is coupled to the optical waveguide. The multilayered heterostructure includes a plurality of concentric thin- film sorbent layers, and wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality7of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

[0012] In another embodiment, a method of making a multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes is provided that includes providing an optical waveguide and forming a multilayered heterostructure as described above on the optical waveguide.

[0013] In yet another embodiment, a method of creating a computing device structured and configured for point sensing based on signals captured by an optical fiber-based sensor node in response to interrogation of the optical fiber-based sensor node is provided. The method includes training an inference model using trainingdata that comprises a combination of physics simulated synthetic data for the sensor node and empirical data for the sensor node, and storing the trained inference model in a processing apparatus of the computing device.

[0014] In still another embodiment, a computing device structured and configured for point sensing based on signals captured by an optical fiber-based sensor node in response to interrogation of the optical fiber-based sensor node is provided. The computing device includes a processing apparatus having stored therein a trained inference model, wherein the trained inference model has been previously trained using training data that comprises a combination of physics simulated synthetic data for the sensor node and empirical data for the sensor node.BRIEF DESCRIPTION OF THE DRAWINGS:

[0015] A full understanding of the invention can be gained from the following description of the preferred embodiments when read in conjunction with the accompanying drawings in which:

[0016] FIG. 1 is a schematic diagram of a prior art fiber optic-based sensing arrangement that employs cascaded gas sensing thin films;

[0017] FIG. 2 is a schematic diagram of a prior art fiber optic-based sensing arrangement that employs a tip-coated multi-core fiber;

[0018] FIG. 3 is a schematic diagram of a prior art fiber optic-based sensing arrangement that employs a single composite sensory film;

[0019] FIG. 4 is a schematic diagram illustrating an evanescent fiber optic sensing arrangement according to an exemplary embodiment of the disclosed concept that employs a multilayered heterostructure for enabling for optical-based multivariate sensing of an analyte, such as, without limitation, a gas;

[0020] FIG. 5 shows a number of non-limiting exemplary sensing layer materials that may be used for implementation of the sensing layers of the disclosed concept;

[0021] FIG. 6 provides a table of non-limiting exemplary nanocomposite / mixed- matrix sensing layers that may be used to implement one or more sensing layers of the disclosed concept;

[0022] FIGS. 7A-7D show the performance of one non-limiting particular exemplary embodiment of the sensing arrangement of FIG. 4;

[0023] FIG. 8 is a schematic diagram of a data analytics framework for multiwavelength interrogation of a single sensor node, such as, without limitation, the fiber optic sensing arrangement of FIG. 4 from offline physics-informed calibration model training to online edge inference according to an exemplary embodiment of the disclosed concept;

[0024] FIG. 9 provides a table of non-limiting exemplary models that may be employ ed in the data analy tics framework of FIG. 8;

[0025] FIG. 10 shows an exemplary simulation workflow to generate physics-based sensor response data according to an exemplary embodiment of the disclosed concept;

[0026] FIG. 11 is a schematic diagram of an exemplary neural network architecture of the response calibration model generated using the data analytics framework of FIG. 8 that can be used for online inference according to another exemplary embodiment of the disclosed concept; and

[0027] FIG. 12 is a block diagram of an edge computing device according to one particular non-limiting exemplary embodiment of the disclosed concept;

[0028] FIG. 13 shows exemplary simulated single sensor data that may be employed in connection with the disclosed concept;

[0029] FIG. 14 shows exemplary model performance of a particular non-limiting exemplary embodiment of the disclosed concept;

[0030] FIGS. 15A and 15B also show exemplary model performance of a particular non-limiting exemplary’ embodiment of the disclosed concept;

[0031] FIGS. 16A and 16B show exemplary simulated two-sensor data that may be employed in connection with the disclosed concept;

[0032] FIGS. 17A and 17B show exemplary' model performance of another particular non-limiting exemplary' embodiment of the disclosed concept; and

[0033] FIG. 18 shows a non-linear Kalman filter in the form of a thermal-equivalent circuit model for distribution transformers according to another particular nonlimiting exemplary' embodiment of the disclosed concept.DETAILED DESCRIPTION:

[0034] As used herein, the singular form of "a". “an”, and “the” include plural references unless the context clearly dictates otherwise.

[0035] As used herein, the statement that two or more parts or components are“coupled’’ shall mean that the parts are joined or operate together either directly or indirectly, i.e., through one or more intermediate parts or components, so long as a link occurs.

[0036] As used herein, “directly coupled” means that two elements are directly in contact with each other.

[0037] As used herein, the term “number” shall mean one or an integer greater than one (i.e., a plurality).

[0038] As used herein, the term “penetration depth” means the distance that light of a given wavelength can travel into a material or layered structure before its intensity significantly decays. In one particular embodiment, “penetration depth” means the distance that light of a given wavelength can travel into a material or layered structure before its intensity decays by a factor of at least 1 / e.

[0039] As used herein, the term “evanescent wave / leaky mode / lossy mode optical fiber sensor” means an optical fiber where the cladding layer on at least a portion of the optical fiber is removed and replaced with a functional sensing material (in one or more layers) that is optically sensitive to the desired measurand based on evanescent wave / leaky mode / lossy mode light from the optical fiber depending on the real and imaginary refractive index of the layers of the functional sensing material relative to the fiber core and on the incident angles. The functional sensing material may include, for example and without limitation, a material such as an oxide, a perovskite, a polymer, or metalorganic framework material.

[0040] Directional phrases used herein, such as, for example and without limitation, top, bottom, left, right, upper, lower, front, back, and derivatives thereof, relate to the orientation of the elements shown in the drawings and are not limiting upon the claims unless expressly recited therein.

[0041] The disclosed concept will now be described, for purposes of explanation, in connection with numerous specific details in order to provide a thorough understanding of the disclosed concept. It will be evident, however, that the disclosed concept can be practiced without these specific details without departing from the spirit and scope of this innovation.

[0042] As described herein, the disclosed concept provides a vertical and hierarchical multilayered thin film design for use in multi-wavelength interrogation of a singlesensor node in which multiple layers, each having different characteristics, are coupled to and surround an optical waveguide, such as the core of an optical sensing fiber. The vertical and hierarchical multilayered thin film design may be achieved, for example, through layer-by-layer coating methods including, but not limited, to electrostatic and / or capillary self-assembly, covalent bonding, hydrogen bonding, or ligand coordination bonding. In the design of the disclosed concept, each individual layer is considered as a single analyte sensor, such as, without limitation, a gas sensor.

[0043] FIG. 4 is a schematic diagram illustrating an evanescent fiber optic sensing arrangement 85 according to one particular exemplary embodiment of the disclosed concept that employs a multilayered heterostructure 90 for enabling for optical-based multivariate sensing of an analyte. In the non-limiting illustrated embodiment show n in FIG. 4, the analyte is a gas. It will be understood, however, that this is meant to be exemplary only, and that other types of physical and / or biological parameters, such as temperature, pH, magnetic field, glucose, dissolved chemicals, among others, may also be sensed using sensing arrangement 85.

[0044] In particular, sensing arrangement 85 includes an optical fiber 95 having a section 100 wherein cladding 105 of optical fiber 95 has been removed to expose core 110 of optical fiber 95. Multilayered heterostructure 90 as described herein is coupled to and surrounds core 110. The non-limiting exemplary multilayered heterostructure 90 of FIG. 4 includes four concentric thin film layers 115, labeled 115A-115D. It will be appreciated, however, that this is meant to be exemplary only, and that multilayered heterostructure 90 may include two, three, or five or more concentric thin film layers 1 15. In addition, each thin film layer 115 is either a functional nanoparticle film (such as a mixed-matrix composite polymer films with functional nanofillers, or, in another example, a plasmonic nanoparticle-based nanocomposites in a metal oxide or organothiolor as described herein) or a sorbent film. In the example multilayered heterostructure 90 shown in FIG. 4, layer 115 A is a functional nanoparticle film and each of layers 115B, 115C, and 115D is a sorbent film (each having different absorbent, adsorbent, and / or light transmission properties as described herein).

[0045] Moreover, sensing arrangement 85 including multilayered heterostructure 90 can leverage a combination of one or more of: (1) molecular sieving and filtering, (2) analyte interaction speci I'icity. and (3) spectral dependence of penetration depthwithin multilayered heterostructure 90 (i.e., the depth that evanescent / leaky mode / lossy mode light penetrates into multilayered heterostructure 90 (bottom to top) depends on the wavelength of the evanescent / leaky mode / lossy mode light). These three tunable and adjustable parameters are highlighted in FIG. 4. In order to realize such a structure in practice, individual gas sensing layers with unique properties can be integrated together using a rational design approach. For example, multiple sorbent or fdter layers with distinct sieving properties can be used sequentially as concentric layers 115 to tailor the interaction of analytes throughout the depth of multilayered heterostructure 90. Similarly, various materials, including mixed matrix or nanocomposite sensing layers, can be judiciously selected to optimize the overall parametric diversity7and spectral response of sensing arrangement 85.

[0046] As FIG. 4 illustrates, as one exemplary embodiment of multilayered heterostructure 90, several determining factors can serve as guiding principles to enable selective multi-gas sensing using an array of sorbent layers 115. The stacking order of each gas sensing layer 115 as an individual gas transducer can first be selected based on the pore size of the sorbent layer 115, which is determined by the kinetic diameter of the target analyte. Sorbent layers targeting small diameter gases should be deposited prior to the sorbent layers targeting larger diameter gases in an increasing order of pore size during bottom-up multilayered structure manufacturing. Secondly, the strength of the adsorption energy7(or bond energy) between sorbent and sorbate could also be structured in a low-to-high order from the bottom of multilayered heterostructure 90, as this determines the kinetics of adsorption and desorption and therefore the response and recovery time of the resulting sensing arrangement 85. Lastly, one possible way to introduce unique spectral dependences and selective responses of each contributing layer 115 is to leverage multilayered heterostructure 90 itself, imposing a leaky engineered waveguide within multilayered heterostructure 90 to enhance interaction of a specific w avelength with a specific layer 115, leading to selective responses of each contributing layer 115. By leveraging the nature of optical penetration depth difference betw een light at different wavelengths within each layer 115, a particular bandwidth for each sensing layer 115 can be sensitized and detected, thus presenting a potential wavelength multiplexing scheme using multilayered heterostructure 90. Thus, in one exemplary embodiment of the disclosed concept, careful engineering is performed to create such a leakyengineered waveguide within multilayered heterostructure 90 that enables molecular sieving and chemical analyte interaction specificity, so that the unique spectral dependence of penetration depth within multilayered heterostructure 90 itself can be leveraged as a heterostructure-induced wavelength-division multiplexing mechanism. The immediate advantage of this aspect of the disclosed concept is that the signal discrimination of the dependence between multiple analytes when using a nanocomposite or mixed matrix approach can be otherwise simplified. Also, in the exemplary embodiment shown in FIG. 4, layer 115A is made of plasmonic metal NPs, which can serve as an optical sensitizer as well as a temperature sensor for temperature self-referencing of the overall multilayer-structured gas sensor.

[0047] Moreover, in particular layers 115 of multilayered heterostructure 90 the lack of penetration of evanescent / leaky mode / lossy mode light of a particular wavelength(s) originating from optical fiber 105 may be fully complete or less than fully complete. In either case, the important thing is that the penetration differences among layers 115 must be sufficient to introduce discrimination between layers 115 and wavelengths. In other words, even if less than complete wavelength specific penetration into one layer 115 vs. the next is achieved, the wavelength discrimination capability of the specific analytes will still be improved as long as there is a greater degree of interaction with one layer 115 vs. the next as a function of the wavelength. Similarly, in particular layers 115 of multilayered heterostructure 90, the molecular sieving may be fully complete or less than fully complete (with some degree of interaction). In either case, the important thing is that the molecular sieving must be sufficient to introduce discrimination between layers 1 15 and analytes. In other words, as long as a specific analyte preferentially makes it deeper into multilayered heterostructure 90 discrimination capability will be enabled. While the ideal would be perfect selectivity, less than perfect selectivity (i. e. , incomplete molecular sieving) will still help improve analyte discrimination.

[0048] The specific material of choice for each of layers 115 of the disclosed concept is dependent upon several factors, including the operating and failure temperature of the infrastructure, humidity, the interfering gases, and the feasibility of fabrication methods transitioning from lab demonstration to commercial-scale production. FIG. 5 shows a number of non-limiting exemplary sensing layer materials that may be used for implementation of layers 115 of the disclosed concept, highlighted according tooperating temperature demonstrated in sensing applications. Such materials include: (i) metal oxides (e.g., TiCh film), (ii) plasmonic metal NP / oxide nanocomposites, (iii) Noble metallic films (e.g., monolithic Au film), (iv) zeolites (e.g, ZSM-5), (v) metal organic frameworks (MOFs) (e.g.. ZIF-8 film), (vi) polymeric mixed-matrix materials (e.g., ITO, NC / PDMS), and (vii) vdW (van der Waals) layers (e.g., M0S2). In addition, a set of non-limiting exemplary' nanocomposite / mixed-matrix sensing layers that may be used to implement one or more layers 115 of the disclosed concept are provided in the table shown in FIG. 6. The materials are metricized in the table of FIG. 6 by their response characteristics, including dominant response wavelength, response FWHM, response time, and sensitivity. It should be noted that since sensitivity' and response time can be dependent upon film-thickness, microstructure and the details of the environmental conditions during sensing experiments, these metrics are for reference purposes and do not necessarily represent the relative sensing performance of one composite material against another for the same target gas analyte.

[0049] Optical fiber-compatible methodologies that may be used to construct multilayered heterostructure 90 include atomic layer deposition (ALD). radio frequency sputtering, and sequential dip-coating based on heterostructure selfassembly. While the former two methods are attractive options, the latter has the potential for expansion and can offer an exceptional level of thickness and uniformity' control while also imparting useful properties for optical based sensing. Self-assembly on optical fibers for individual chemical sensors has been a developing field of interest. Electrostatic self-assembly, including sequential dip-coating electrostatic self-assembly, is one ty pe of self-assembly process that may be used to fabricate multilayered heterostructure 90. An electrostatic self-assembly procedure starts with a cleaned or pretreated surface (core 110 in the disclosed concept) that is either positively or negatively charged. The procedure then alternately and sequentially adds oppositely charged materials (opposite polyions (i.e., polycations or polyanions)) as new layers in multiple steps, for example by dipping into a solution of the opposite polyions in each step. Polycations and polyanions that are commonly utilized include poly(allylamine hydrochloride) (PAH) and poly(acrylic acid sodium salt) (PAA), respectively, due to their high degrees of pH-dependent ionization. In addition to electrostatic self-assembly as just described, vdW forces, capillary', magnetic andentropic interactions are also viable self-assembly techniques for implementing the disclosed concept. Capillary self-assembly is a bottom-up fabrication technique in which micro- or nanoscale objects are organized into ordered structures by capillary' forces (the forces that arise at liquid-air or liquid-solid interfaces due to surface tension). In capillary self-assembly, small particles (e.g., nanoparticles) are dispersed in a liquid medium. As the liquid evaporates or is withdrawn, liquid-air menisci form between particles and the substrate. Surface tension pulls particles together (or aligns them) to minimize the free energy of the liquid interface. When the liquid fully evaporates, the particles remain locked into ordered arrays or attached to specific sites. It is believed that self-assembled layers have only been demonstrated for either broadband or single wavelength detection in the gas sensing regime.

[0050] One particular exemplary embodiment of sensing arrangement 85 has been builte and tested by the present inventors. In that embodiment, the layers 115 of multilayered heterostructure 90 consist of a Metal-Organic Framework (MOF) ZIF-8 sorbent layer sputter-coated with an islanded Au overlayer. Preliminary' fiber optic transmittance spectral results of that embodiment show tw o spectral sensing response peaks with regards to MOF band-edge absorption and Au plasmonic resonance absorption. In addition, and more specifically, FIG. 7A shows spectral transmittance responses of this particular embodiment to increasing CO2, FIG. 7 A shows the dynamic response magnitude of this particular embodiment for one cycle at both the 2-mIm absorption and Au LSPR peaks, and FIGS. 7C and 7D shows reducing absorptance of the ZIF-8 sensor layer at the two absorption peaks due to desorption of CO2 and N2, respectively, with increasing temperatures.

[0051] Moreover, with respect to multivariate analytics approaches that may be used for multi-wavelength interrogation of a single (point) sensor node (pathway (3) above), several discrimination methods have been studied in the literature. However, an effective approach has not yet been developed in the prior art for extending demonstrated discriminated responses in lab-controllable conditions to the generation of calibrated readouts in the field that account for real-w orld application scenarios. To bridge this gap. the disclosed concept provides a data analytics framework 120 shown schematically in FIG. 8 for creating a response calibration model 125 (a type of inference model) that may be used to implement multi-wavelength interrogation of a single sensor node, such as, without limitation, sensing arrangement 85. For example,data analytics framework 120 as described herein may be used to provide for multigas fiber optic photonic nose response calibration in practice using, for example and without limitation, sensing arrangement 85. In particular, in specific exemplary embodiments, data analytics framework 120 is provided to decode multiparametric responses of an optical fiber-based photonic nose (e.g., using sensing arrangement 85) from offline calibration model training using synthetic data to online model inference using edge devices such as microcontroller units (MCUs)

[0052] More specifically, to address the challenges in (i) acquiring calibration data in“beyond lab” conditions, and (ii) avoiding overfitting response calibration model 125 to the bias of a subset of calibration data, data analytics framework 120 combines physics-simulated data (e.g., simulated time-series sensor data) with empirical data in the training data that is used to train response calibration model 125, and thus employs a physics-informed learning method. Also, algorithmic compatibility with edge hardware / firmware is considered, especially for the purpose of grid-tied energy applications that can benefit from loT-based sensor networks.

[0053] Data analytics framework 120 starts with designing a synthetic training data emulator 130 that synthesizes in-lab experimental data with physics-derived theoretical (simulated) responses. For example, in a multiparametric fiber optic photonic nose, physics-based data can similarly be generated through analytical models for the temperature response, chemisorptive responses to gas that involve electronic exchange, and physisorption responses for room temperature sensing layers. In addition to physics models that describe the interaction between sensing layers and analytes of interest, optical models may also be applied to translate these impacts into simulated optical signatures in an assumed sensor device considering first the impact of the interaction on the optical constants of the constituent sensing materials and subsequently the integration of the sensing materials with optical waveguide device models. A subset of example models that may be employed is summarized in the table of FIG. 9. An exemplary simulation workflow to generate physics-based sensor response data is provided in FIG. 10. One of the simpler examples is to model temperature sensing elements based on plasmonic nanostructures using temperature-dependent Drude free electron behavior, which accounts for modification to the mean scattering time as a function of temperature. In addition, quantifying gas sensing responses of perovskite oxides at high temperaturesexploits defect chemical kinetics and thermodynamics, while pH sensing responses of oxides and plasmonic oxide based nanocomposites considers solution phase chemistry of ions in solution as a function of surface charging behavior for a porous sensing layer medium. Incorporating data generated from such "a priori" physics and optical models into the training dataset, as show n in the bottom portion of FIG. 8, enhances discrimination accuracy of multifunctional and multivariate sensing modalities by filling gaps where experimental data could be constrained.

[0054] In addition, a response calibration model such as response calibration model125 includes a loss function that is a measure of how far the model’s predicted response is from the true reference value during calibration. The response calibration model is trained (or fitted) by minimizing this loss function so that it best maps raw' sensor signals to predicted response values. Also, a response calibration model such as response calibration model 125 also typically employs a dimensionality reduction algorithm to transform raw' sensor signals into a smaller set of features that still capture the meaningful information, and classifier algorithms to provide categorical recognition. The fidelity of the features extracted from dimensionality reduction / classifier algorithms can be improved according to the disclosed concept by implementing a physics-informed loss function where part of the physics-simulated data (e.g., simulated time-series sensor data) generated by synthetic training data emulator 130 is assigned to a regularizer of the loss function to avoid overfitting the model with empirical data limited to the otherwise biased lab-controlled environment. This is shown in the top portion of Fig. 8.

[0055] Also, referring again to the bottom portion of FIG. 8, intelligent selection of the extracted multivariate features can play a critical role in maximizing information content and limit systematic cost and / or complexity. Two directions are considered in this step. The first direction is from the hardware perspective, wherein the optimal wavelengths for each parametric measurement can be determined based on a principal component (PC) loading plot from the multi -wavelength data, and used to guide the design excitation / detection schemes that are selected around these w avelengths. The second direction is from the software / algorithmic perspective, wherein the PC scores of new spectral data can be calculated using the previously extracted principal components and serve as the sensor response values for later calibration. Both cases minimize the required data and thus reduce the computational complexity' for thefollowing steps, the computing requirements at the edge, and the overall cost of the sensor device hardware.

[0056] Furthermore, still referring to the bottom portion of FIG. 8, assuming that the long-term drifts in the light source and detection circuitry are properly self-referenced such that any observed transients are attributed to the specific sensing layers, the single sensor node sensor system (feature-guided hardware and feature-inferred responses) can then be exposed to two additional scenarios sequentially to generate training data for response calibration model 125, namely drifting datasets for longtime duration within the laboratory setting and any transient data observed during field testing. These additional datasets can be integrated within the framework through a selected ensemble algorithm to leam the two important scenarios and build response calibration model 125.

[0057] In the exemplary embodiment, response calibration model 125 is implemented using a neural network, such as a recurrent neural (RNN). In one particular exemplary embodiment, response calibration model 125 is implemented using a Long Short- Term Memory (LSTM) network architecture trained by drift and transient data sequentially to account for time-dependent dynamics as shown in FIG. 11. Response calibration model 125 in this particular embodiment includes drift compensator LSTM cell 130 a block 135 of stacked LSTM cells 140, which adds, on the basis of an RNN structure, the concept of additional input connections and output connections to the network internal state. The result is a cell that is capable of learning how to control the network memory, or the state, improving the capability to remember longer term information and more complex time dependencies and temporal features. As seen in FIG. 11, drift compensator LSTM cell 130 includes a forget gate 145, an input gate 150 and an output gate 155. Forget gate 145 is composed of a scaling function and a multiplicative operation, making it responsible to control the state growth. Input gate 150 and output gate 155 control reading and writing to the memory, where the tanh activation function learns the features as the self-connections regulates the flow of information. Comparatively with the general gate descriptions, a simplistic interpretation is to consider that in a time step, the amount of remaining drift is stored in the cell memory and controlled by forget gate 145, then drift changes are calculated and added to memory by input gate 150, and output gate 155 decides how to compensate for the drift. Once drift compensator LSTM cell 130 is built,sensors can be deployed in the field. This field dataset will be compensated for the sensory material -caused drift before it is applied as the training set for the final response calibration model. Drift compensator LSTM cell 130 outputs a time series that still contains sequential and temporal features; to address that, block 135 of stacked LSTM cells 140 is used to capture time dependencies that arise from external environmental changes and interactions. A dense layer 160 is attached in sequence with block 135 to help better retain static patterns and handcrafted features. In the exemplary embodiment, dense layer 160 is a Dense Neural Network (DNN) in which inputs are the output of the LSTM (the block 135 of stacked LSTM cells 140) and the response of previous times (not necessarily in the same stage of the DNN) to help better retain static patterns and handcrafted features. The specific architecture of the DNN can be optimized accordingly, for example, a modified ID CNN or encode- decoder (i.e., an autoencoder) such that the past data can be an input to latent space. In addition, in one exemplary embodiment, the DNN has multiple stages and the inputs of the DNN are not limited to a single stage of the DNN, but rahter the inputs can be received at different stages of the DNN. The number of LSTM cells 140 and layers of the network is a hyperparameter and should be optimized accordingly.

[0058] In the exemplary embodiment, data analytics framework 120 is edge compatible and is structured and configured to perform data analysis directly on edge devices close to where the sensor data is generated (rather than sending all data to a central cloud or data center). FIG. 12 thus provides a block diagram of an edge computing device 165 according to one particular non-limiting exemplary embodiment of the disclosed concept. As seen in FIG. 12, edge computing device 165 is coupled to a sensor node 170 through a photodetector 175 and an operational amplifier 180. Sensor node 170 is structured and configured for multi-wavelength interrogation, and may be, for example and without limitation, sensor arrangement 85. Edge computing device 165 includes a processing apparatus 185 that comprises a processor 190 and a memory 195. Processor 190 may be, for example and without limitation, a microprocessor (pP), a microcontroller, or some other suitable processing device, that interfaces with memory 195. Memory 195 can be any one or more of a variety of types of internal and / or external storage media such as, without limitation, RAM, ROM, EPROM(s), EEPROM(s), FLASH, and the like that provide a storage register, i.e., a non-transitory machine readable medium, for data storage such as inthe fashion of an internal storage area of a computer, and can be volatile memory or nonvolatile memory. Memory 195 has stored therein a number of routines (comprising computer executable instructions) that are executable by processor 190, including routines for implementing response calibration model 125 of the disclosed concept as described herein. As such, edge computing device 165 takes photocurrent- converted voltages from operational amplifier 180 as input data and in real time infers multiparameter readouts from such inputs.

[0059] Moreover, the present inventors have done certain work to prove the viability of using a combination of experimental and simulated sensor data according to the disclosed concept. In that work, the application of long short-term memory (LSTM) networks as described herein was applied to synthetic time-series data generated to reflect the performance of single and two sensor models, as described below.

[0060] With respect to the single sensor model, synthetic data was generated to simulate the response of an idealized sensor with a linear equilibrium response to concentration (c) and linear response constant (a) as follows:where the sensor response (s) can generically represent the measured response of some sensing system (e.g., intensity from a transmissive or absorptive sensor, backscattered intensity, wavelength shift). A finite response time is introduced by simulating the time-resolved sensor response to a dynamically changing c(t) with the 1st order ODE model as follows:which can be approximately used to generate the time series as follows, using a finite- difference approximation:With no loss of generality, the sensor response and concentration are assumed to be normalized between 0 and 1. and a is unity. The time spacing (At) is arbitrarily taken as 0. 1 s. To simulate a varying concentration gas environment, and to not bias the dataset with a structured or periodic time series, the concentration is randomly selected between 0 and 1 every 2 seconds, using the built-in rand function in Matlab. For the simulated dataset, r is set to 0.5 s. An example of this simulated single sensor data is shown in FIG. 13.

[0061] An LSTM model was trained on the synthetic sensor data in Python using thePyTorch package. The dataset was comprised of a total series of 75,000 points, generated from the equation above. Each point included a concentration value (c, generated via a random schedule as described herein) and a simulated sensor response at the same time point (s). With a sequence length of 25, the sensor response value was used as the input layer and the gas concentration was used as the output layer. The goal was to train a model that, given some past time-series of the sensor response, is able to correctly estimate the current value of the concentration. The hyperparameters of the network were tuned through rial and error to determine the best results, using 4 stacked LSTM layers and with a hidden size of 200 (number of units in the LSTM layers). The model was trained over 200 epochs with a learning rate of 0.005. The dataset was split into test and training data, using 80% of the data (60k points) for training and 20% of the data (15k points) for testing. The result of the model applied to the test dataset is shown in FIG. 14. A median error of 0.44% was observed in the test dataset relative to the true concentration values. A larger mean error was observed, as the most significant deviations were from large outliers immediately after the concentration changed.

[0062] An interesting capability of the LSTM generated response is that it is able to quickly identify the current gas concentration, demonstrating a very fast effective response time. This can be seen in FIGS. 15A and 15B. FIG. 15A is comparison of the LSTM-generated prediction of gas concentration vs. actual, and FIG. 15B is a comparison of the model sensor response to the simulated concentration during the same time window.

[0063] With respect to the two-sensor model, a two-sensor synthetic dataset was generated using two sensors (si and s2) with differing levels of sensitivity to two target gases concentrations as follows:using the same dynamic response of the single sensor model as follows:Model parameters were defined so that the two sensors have markedly different response to the two target gases, with sensor 1 being more responsive to gas 2 (an = 1, ai2 = 10) and sensor 2 being more responsive to gas 1 (a2i = 8, 022 =2). The synthetic dataset was generated as described above. Gas concentrations were randomized after 60 seconds, with TI = 10 s and 12 = 30 s, using a time step of 100 ms for the discretized model, as above. An example of this simulated single sensor data is shown in FIGS. 16A and 16B.

[0064] An LSTM model was trained on the synthetic sensor data in Python using thePyTorch package, as above. The dataset was comprised of a total series of 50,000 points. Each point included two concentration values (cl and c2, generated via a random schedule) and a simulated sensor response at the same time point (si and s2). With a sequence length of 100 (10 s), the sensor response values were used as the input layer and the gas concentrations were used as the output layer. The hyperparameters of the netw ork w ere tuned through rial and error to determine the best results, using a 4 stacked LSTM layers and with a hidden size of 200 (number of units in the LSTM layers). The model w as trained over 200 epochs with a learning rate of 0.005. The dataset w as split into test and training data, using 75% of the data (37.5k points) for training and 25% of the data (12.5k points) for testing. The result of the model applied to the test dataset is shown in FIGS. 17A and 17B. A median error of 0.24% was observed for gas 1 predictions and 0.77% was observed for gas 2 predictions. As with the single sensor model, the largest error was observed at the transition points between gas species; the larger median error for the second sensor may be expected given its larger time constant.

[0065] Furthermore, another exemplary embodiment of edge-compatible sensor data analytics of the disclosed concept provides a system for estimating the top-oil temperature (TOT) (i.e., the temperature of the transformer insulating oil measured at or near the top of the transformer tank) and hot spot temperature (HOT) (i.e., the estimated maximum temperature inside the transformer windings, at the most thermally stressed point) of a distribution transformer using a fiber-optic based temperature sensor (such as an FBG temperature sensor) and a trained non-linear Kalman filter (another ty pe of inference model), w herein the non-linear Kalman filteris trained using a combination of physics-simulated data (e.g., simulated time-series sensor data) and empirical data as described herein. For example, a non-linear Kalman filter requires parameters like the noise covariance matrices (Q for process noise, R for measurement noise). In one embodiment, those parameters may be learned based on they can also be learned from a combination of physics-simulated data (e.g., simulated time-series sensor data) and empirical data as described herein. In another exemplary7embodiment, the non-linear Kalman filter may be implemented using a neural network, where the neural network is trained a combination of physics- simulated data (e.g., simulated time-series sensor data) and empirical data as described herein. The non-linear Kalman filter in the form of a thermal-equivalent circuit model for distribution transformers is shown in FIG. 18. The trained non-linear Kalman filter of this embodiment may be coded into a microcontroller (such as a standard ESP32 microcontroller module with 4 MB flash memory’) and executed with SRAM (such as 520 KB SRAM size). Such an implementation is similar to the implementation show n in FIG. 12, except that sensor node 90 would be the temperature sensor and the trained non-linear Kalman filter would be stored in memory 195.

[0066] While certain of the exemplary embodiments described herein include a multilayered optical waveguide-based sensor that is a gas sensor, it will be understood that this is meant to be exemplary' only. The disclosed concept may also be used to other parameters, such as, without limitation, magnetic field or electric field.

[0067] While specific embodiments of the invention have been described in detail, it will be appreciated by those skilled in the art that various modifications and alternatives to those details could be developed in light of the overall teachings of the disclosure. Accordingly, the particular arrangements disclosed are meant to be illustrative only and not limiting as to the scope of disclosed concept which is to be given the full breadth of the claims appended and any and all equivalents thereof.

Claims

What is claimed is:

1. A multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes, comprising: an optical waveguide; and a multilayered heterostructure coupled to the optical waveguide, wherein the multilayered heterostructure includes a plurality of concentric thin-film sorbent layers, and wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

2. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure includes at least three concentric thin-film sorbent layers.

3. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure is: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

4. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure is at least structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure.

5. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure is at least structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure6. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure is at least structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

7. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure is at least structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, and stmctured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure.

8. The multilayered optical waveguide-based sensor according to claim 5, wherein for the at least one of the analytes a strength of the adsorption energy is different for each of the sorbent layers.

9. The multilayered optical waveguide-based sensor according to claim 8, wherein the plurality of thin-film sorbent layers includes at least a first layer and a second layer, wherein the first layer is positioned closer to the optical waveguide than the second layer, wherein for the at least one of the analytes the strength of the adsorption energy in the first layer is less than the strength of the adsorption energy in the second layer.

10. The multilayered optical waveguide-based sensor according to claim 4, wherein the plurality of thin-film sorbent layers includes at least a first layer and a second layer, wherein a pore size of the first layer is smaller than a pore size of the second layer.

11. The multilayered optical waveguide-based sensor according to claim 1 , wherein the plurality of analytes comprises a plurality of gases.

12. The multilayered optical waveguide-based sensor according to claim 1, wherein the multilayered heterostructure further comprises a layer of plasmonic metal nanoparticles.

13. The multilayered optical waveguide-based sensor according to claim12, wherein the layer of plasmonic metal nanoparticles is positioned directly on the optical waveguide.

14. The multilayered optical waveguide-based sensor according to claim13, wherein the layer of plasmonic metal nanoparticles is structured to serve as an optical sensitizer and a temperature sensor for temperature self-referencing of the multilayered optical waveguide-based sensor.

15. The multilayered optical waveguide-based sensor according to claim 1, wherein the optical waveguide comprises a fiber-optic member.

16. A method of making a multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes, comprising: providing an optical waveguide; forming a multilayered heterostructure on the optical waveguide, wherein the multilayered heterostructure includes a plurality of concentric thin-film sorbent layers, and wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent light / leaky mode / lossy mode originating from the optical waveguide.

17. The method according to claim 16, wherein the forming the multilayered heterostructure comprises forming the multilayered heterostructure using layer by layer self-assembly.

18. The method according to claim 17, wherein the self-assembly is electrostatic self-assembly.

19. The method according to claim 17, wherein the electrostatic selfassembly comprises sequential dip coating electrostatic self-assembly.

20. The method according to claim 17, wherein the self-assembly is capillary self-assembly.

21. The method according to claim 16, wherein the multilayered heterostructure includes at least three concentric thin-film sorbent layers.

22. The method according to claim 16, wherein the multilayered heterostructure is: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent light / leaky mode / lossy mode originating from the optical waveguide.

23. The method according to claim 16, wherein the multilayered heterostructure is at least structured to enable molecular sieving and fdtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure.

24. The method according to claim 16, wherein the multilayered heterostructure is at least structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure25. The method according to claim 16, wherein the multilayered heterostructure is at least structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

26. The method according to claim 16, wherein the multilayered heterostructure is at least structured to enable molecular sieving and fdtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, and structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure.

27. The method according to claim 24, wherein for the at least one of the analytes a strength of the adsorption energy is different for each of the sorbent layers.

28. The method according to claim 27, wherein the plurality of thin-film sorbent layers includes at least a first layer and a second layer, wherein the first layer is positioned closer to the optical waveguide than the second layer, wherein for the at least one of the analytes the strength of the adsorption energy in the first layer is less than the strength of the adsorption energy in the second layer.

29. The method according to claim 23, wherein the plurality of thin-film sorbent layers includes at least a first layer and a second layer, wherein a pore size of the first layer is smaller than a pore size of the second layer.

30. The method according to claim 16, wherein the multilayered heterostructure further comprises a layer of plasmonic metal nanoparticles.

31. The method according to claim 30, wherein the layer of plasmonic metal nanoparticles is positioned directly on the optical waveguide.

32. A method of creating a computing device structured and configured for point sensing based on signals captured by an optical fiber-based sensor node in response to interrogation of the optical fiber-based sensor node, comprising:training an inference model using training data that comprises a combination of physics simulated synthetic data for the sensor node and empirical data for the sensor node; and storing the trained inference model in a processing apparatus of the computing device.

33. The method according to claim 32, wherein the physics simulated synthetic data is generated using one or more physics models that describe interactions between the sensor node and analytes of interest and one or more optical models that translate the interactions into simulated optical signatures of the sensor node.

34. The method according to claim 32. wherein the method is for creating a computing device structured and configured for multiparameter point sensing based on signals captured by the optical fiber-based sensor node in response to multiwavelength interrogation of the optical fiber-based sensor node, and wherein the inference model is a response calibration model.

35. The method according to claim 34, wherein the sensor node comprises a multilayered optical waveguide-based sensor.

36. The method according to claim 34. wherein the response calibration model employs a loss function and wherein the training comprises assigning at least part of the physics simulated synthetic data to a regularizer of the loss function to avoid overfitting the response calibration model with the empirical data.

37. The method according to claim 34, wherein the response calibration model is configured to extract a plurality of multivariate features from raw sensor data generated by the sensor node for making a plurality of parametric measurements, and wherein the plurality of multivariate features are selected by determining optimal wavelengths for each parametric measurement based on a principal component (PC) loading plot from multi-wavelength data.

38. The method according to claim 37, further comprising using the optimal wavelength to design excitation and detection schemes for the sensor node.

39. The method according to claim 34. wherein calibration of the response calibration models includes calculating principal component (PC) scores of new spectral data using previously extracted principal components and using the PC scores as sensor response values.

40. The method according to claim 34, wherein the training data includes one or more long-time duration drifting datasets for the sensor node and transient data observed during field testing of the sensor node.

41. The method according to claim 33. wherein the response calibration model is implemented using a neutral network.

42. The method according to claim 41, wherein the neutral network is a recunent neural network sequentially coupled with a dense neural network.

43. The method according to claim 42, wherein the recurrent neural network comprises a Long Short-Term Memory' (LSTM) network architecture.

44. The method according to claim 43. wherein the LSTM network architecture is sequentially trained using drift and transient data.

45. The method according to claim 44, wherein the LSTM network architecture includes drift compensator LSTM cell coupled to a block of stacked LSTM cells.

46. The method according to claim 45, wherein the drift compensator LSTM cell includes a forget gate, an input gate, and an output gate, wherein the forget gate includes a scaling function and a multiplicative operation making the forget gate responsible to control state growth, wherein the input gate and the output gate control reading and writing to memory'.

47. The method according to claim 45, wherein the drift compensator LSTM cell is structured and configured to output a time series that contains sequential and temporal features, and wherein block of stacked LSTM cells is structured and configured to capture time dependencies that arise from external environmental changes and interactions.

48. The method according to claim 34, wherein the sensor node comprises a multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes, comprising: an optical waveguide; and a multilayered heterostructure coupled to the optical waveguide, wherein the multilayered heterostructure includes a plurality of concentric thin-film sorbent layers, and wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of the multilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

49. The method according to claim 32. wherein the inference model is a nonlinear Kalman filter.

50. The method according to claim 49, wherein the trained nonlinear Kalman filter is implemented using a neural network.

51. The method according to claim 49, wherein the sensor node comprises a temperature sensor and wherein the trained Kalman filter is configured for estimating the top-oil temperature (TOT) and hot spot temperature (HOT) of a distnbution transformer.

52. A computing device structured and configured for point sensing based on signals captured by an optical fiber-based sensor node in response to interrogation of the optical fiber-based sensor node, comprising: a processing apparatus having stored therein a trained inference model, wherein the trained inference model has been previously trained using training data that comprises a combination of physics simulated synthetic data for the sensor node and empirical data for the sensor node.

53. The computing device according to claim 52, wherein the physics simulated synthetic data is generated using one or more physics models that describe interactions between the sensor node and analytes of interest and one or more optical models that translate the interactions into simulated optical signatures of the sensor node.

54. The computing according to claim 52, wherein the computing device is structured and configured for multiparameter point sensing based on signals captured by the optical fiber-based sensor node in response to multi-wavelength interrogation of the optical fiber-based sensor node, and wherein the inference model is a response calibration model.

55. The computing device according to claim 54, wherein the sensor node comprises a multilayered optical waveguide-based sensor.

56. The computing device according to claim 54, wherein the sensor node comprises a multilayered optical waveguide-based sensor for sensing parameters relating to a plurality of analytes, comprising: an optical waveguide; and a multilayered heterostructure coupled to the optical waveguide, wherein the multilayered heterostructure includes a plurality of concentric thin-film sorbent layers, and wherein the multilayered heterostructure is one or more of: (i) structured to enable molecular sieving and filtering of the plurality of analytes within the sorbent layers of the multilayered heterostructure, (ii) structured such that at least one of the analytes will interact differently with different sorbent layers of themultilayered heterostructure, and (iii) structured to exhibit spectral dependence of penetration depth within the multilayered heterostructure of evanescent / leaky mode / lossy mode light originating from the optical waveguide.

57. The computing device according to claim 54, wherein the response calibration model employs a loss function and wherein the response calibration model has been previously trained by assigning at least part of the physics simulated synthetic data to a regularizer of the loss function to avoid overfitting the response calibration model with the empirical data.

58. The computing device according to claim 54, wherein the response calibration model is configured to extract a plurality of multivariate features from raw sensor data generated by the sensor node for making a plurality of parametric measurements, and wherein the plurality of multivariate features were selected by determining optimal wavelengths for each parametric measurement based on a principal component (PC) loading plot from multi-wavelength data.

59. The computing device according to claim 54, wherein the response calibration models was calibrated by calculating principal component (PC) scores of new spectral data using previously extracted principal components and using the PC scores as sensor response values.

60. The computing device according to claim 54, wherein the training data includes one or more long-time duration drifting datasets for the sensor node and transient data observed during field testing of the sensor node.

61. The computing device according to claim 54, wherein the response calibration model is implemented using a neutral network.

62. The computing device according to claim 61, wherein the neutral network is a recurrent neural network.

63. The computing device according to claim 62, wherein the recurrent neural network comprises a Long Short-Term Memory (LSTM) network architecture.

64. The computing device according to claim 63, wherein the LSTM network architecture was sequentially trained using drift and transient data.

65. The computing device according to claim 64, wherein the LSTM network architecture includes drift compensator LSTM cell coupled to a block of stacked LSTM cells.

66. The computing device according to claim 65, wherein the drift compensator LSTM cell includes a forget gate, an input gate, and an output gate, wherein the forget gate includes a scaling function and a multiplicative operation making the forget gate responsible to control state growth, wherein the input gate and the output gate control reading and writing to memory.

67. The computing device according to claim 65, wherein the drift compensator LSTM cell is structured and configured to output a time series that contains sequential and temporal features, and wherein block of stacked LSTM cells is structured and configured to capture time dependencies that arise from external environmental changes and interactions.

68. The computing device according to claim 52, wherein the inference model is a nonlinear Kalman filter.

69. The computing device according to claim 68, wherein the trained nonlinear Kalman filter is implemented using a neural network.

70. The computing device according to claim 68, wherein the sensor node comprises a temperature sensor and wherein the trained Kalman filter is configured for estimating the top-oil temperature (TOT) and hot spot temperature (HOT) of a distribution transformer.

71. The computing device according to claim 52, wherein the computing device is an edge compatible microcontroller unit.

72. The method according to claim 32. wherein the computing device is an edge compatible microcontroller unit.

73. The computing device according to claim 52, wherein the physics simulated synthetic data is simulated time series data for the sensor node.

74. The method according to claim 32, wherein the physics simulated synthetic data is simulated time series data for the sensor node.

75. The computing device according to claim 54, wherein the sensor node comprises an evanescent wave / leaky mode / lossy mode optical fiber sensor.

76. The method according to claim 34, wherein the sensor node comprises an evanescent wave / leaky mode / lossy mode optical fiber sensor.

77. The method according to claim 45, wherein the dense neural network is structured to receive as inputs the outputs of the block of stacked LSTM cells and previous responses of the response calibration model.

78. The method according to claim 77, wherein the dense neural network has multiple stages and wherein the dense neural network may receive the inputs at different stages of the dense neural network.

79. The method according to claim 77, wherein the dense neural network comprises a ID CNN or an autoencoder.

80. The computing device according to claim 65, wherein the dense neural network is structured to receive as inputs the outputs of the block of stacked LSTM cells and previous responses of the response calibration model.

81. The computing device according to claim 80, wherein the dense neural network has multiple stages and wherein the dense neural network may receive the inputs at different stages of the dense neural network.

82. The computing device according to claim 80, wherein the dense neural network comprises a ID CNN or an autoencoder.

Images