A two-dimensional heterojunction array sensing method and device for carbon pollution and high-precision monitoring

Abstract

The application discloses a two-dimensional heterojunction array sensing method and device for carbon pollution and high-precision monitoring. The two-dimensional heterojunction array sensing method comprises the following steps: a tunable light source is used to emit a light wave signal with dynamically changed wavelength, power and modulation frequency to the two-dimensional heterojunction array according to a preset switching sequence; an electric signal generated by the two-dimensional heterojunction array under the joint action of a to-be-detected gas and the light wave signal is collected, and the electric signal is converted into a digital signal in the form of a multidimensional matrix; a hybrid model integrated with a convolutional neural network branch and a recurrent neural network branch is used to extract spatial mode features and dynamic time sequence features respectively, and finally the type and concentration of the gas are output. The application excites the two-dimensional heterojunction array to generate rich characteristic responses through light sequence modulation, and combines a lightweight neural network model, so that the cross-sensitivity problem in gas identification is effectively solved, and the recognition accuracy and response speed in a complex environment are improved.

Description

Technical Field

[0001] This invention belongs to the field of sensor technology, specifically relating to a two-dimensional heterojunction array sensing method and device for high-precision monitoring of carbon pollution synergy. Background Technology

[0002] With the increasing demands for gas recognition accuracy and real-time performance in fields such as industrial safety monitoring, medical breath analysis, and environmental pollution control, the performance of traditional semiconductor gas sensors in complex, multi-component environments is becoming increasingly limited. Traditional metal-oxide-semiconductor sensors mainly rely on the chemical adsorption of gas molecules on the material surface to cause changes in conductivity. However, in practical applications, these devices generally suffer from bottlenecks such as poor recognition specificity, susceptibility to environmental interference, and high power consumption.

[0003] Two-dimensional materials (such as graphene, MXenes, and transition metal chalcogenides) have become a research hotspot in the field of high-performance gas sensing due to their extremely high specific surface area and excellent electrical modulation properties. However, the physical or chemical adsorption characteristics of different gas molecules in a single two-dimensional material often exhibit similarities, leading to insufficient selectivity. Especially when facing mixed gas environments with complex compositions and dynamically changing concentrations, severe cross-sensitivity makes it difficult for sensors to accurately distinguish target gases.

[0004] To improve selectivity, existing technologies typically employ sensor arrays composed of multiple sensing units (i.e., "electronic nose" solutions). These arrays collect an overall response "fingerprint" and analyze it using pattern recognition algorithms. However, the performance of traditional sensor arrays is highly dependent on the differences between physical units, and most only collect static resistance change signals. Due to the lack of effective dynamic control mechanisms, the system acquires limited information dimensions, leading to a significant drop in recognition accuracy when processing signals with highly similar gas components or drastically fluctuating concentrations. Furthermore, existing "electronic nose" technologies often exhibit severe drift phenomena when facing extremely high humidity or complex background interference due to the lack of actively activated dimensions.

[0005] Regarding dynamic control mechanisms, existing research primarily employs thermal excitation (by altering the operating temperature of a microheater) to modify the reaction kinetics of the material surface. However, thermal control suffers from drawbacks such as high thermal inertia, slow response and recovery speeds, high power consumption, and accelerated aging of two-dimensional materials at high temperatures. In contrast, two-dimensional materials possess excellent light absorption and photoelectric conversion efficiency, allowing for faster and lower-power carrier modulation through optical modulation. However, current optical excitation methods mostly rely on steady-state irradiation with a single wavelength, lacking systematic wavelength, power, and frequency conversion sequences for modulation. This prevents the full exploitation of the multidimensional spatiotemporal characteristics inherent in photoinduced gas-sensitive responses.

[0006] Furthermore, in the back-end signal processing stage, deep learning models are typically needed to extract effective features from complex sensor signals. However, existing general-purpose deep learning models are mostly designed for single-dimensional signals, and often employ simple tiling or dimensionality reduction processing for multi-dimensional sensor data. This makes it difficult to effectively analyze the physical correlation features hidden in the photosensitive gas response, resulting in limited accuracy in identifying complex mixed gases.

[0007] Therefore, how to obtain multi-dimensional gas identification information and construct an algorithm architecture that can deeply analyze the related features to achieve high-precision gas identification with biomimetic dynamic sensing capabilities is a key problem that urgently needs to be solved in the current gas sensing field. Summary of the Invention

[0008] The purpose of this invention is to overcome the problems of single information dimension, poor selectivity and difficulty in handling complex mixed gases in existing gas identification technologies, and to provide a two-dimensional heterojunction array sensing method and device for high-precision monitoring of carbon pollution. In this invention, carbon pollution refers to carbon-containing gases and polluting gases.

[0009] To achieve the above objectives, the present invention provides a two-dimensional heterojunction array sensing device for high-precision monitoring of carbon pollution, comprising:

[0010] The light source configuration module emits light wave signals with dynamically changing wavelength, power, and modulation frequency to the two-dimensional heterojunction array according to a preset light source sequence.

[0011] The sensing and detection module includes an optical path structure and a two-dimensional heterojunction array. The light wave signal emitted by the light source configuration module is collimated, defocused, or homogenized through the optical path structure and then projected onto the surface of the two-dimensional heterojunction array to generate a dynamic electrical signal in the gas environment to be measured. The two-dimensional heterojunction array is composed of multiple sensitive units with vertical heterojunction structures. The two-dimensional heterojunction array has at least two different combinations of vertical heterojunction structures.

[0012] The signal acquisition and conversion module acquires dynamic electrical signals and generates a spatiotemporal sequence matrix.

[0013] The data processing module uses a hybrid model of convolutional neural network branches and recurrent neural network branches to extract the spatial pattern features and dynamic temporal features of the spatiotemporal sequence matrix, respectively. After performing nonlinear mapping, it outputs the type and concentration distribution information of the gas to be tested.

[0014] This invention simulates the biological olfactory perception mechanism: a two-dimensional heterojunction array simulates the non-specific cross-response of olfactory receptor cells to gases; a light source configuration module modulates the light according to a preset active light sequence (i.e., a preset light source sequence) to simulate the active olfactory detection behavior of organisms to enhance the interaction between airflow and receptors; and convolutional and recurrent neural network branches simulate the processing mechanism of spatiotemporal neural signals by the olfactory bulb and olfactory cortex. This invention excites the two-dimensional heterojunction array to generate a response signal that evolves dynamically over time through active light sequence modulation, and combines this with a deep learning architecture capable of simultaneously extracting spatial and temporal features, achieving high-precision identification of complex gas components. It is applicable to multi-component environments containing carbonaceous gases and polluting gases.

[0015] In some embodiments, as a preferred embodiment, the light-controlled gas biomimetic recognition device further includes an environmental temperature, humidity and pressure monitoring module, which captures changes in environmental temperature, humidity and air pressure in real time and inputs them into the aforementioned data processing module to assist in baseline drift correction.

[0016] Furthermore, the vertical heterojunction structure is formed by stacking a first two-dimensional material and a second two-dimensional material above the gap between the substrate electrodes; the first two-dimensional material and the second two-dimensional material in each vertical heterojunction structure are different two-dimensional materials, forming two or more different combinations of vertical heterojunction structures, which are then arranged to form the above-mentioned two-dimensional heterojunction array; wherein, the two-dimensional materials include graphene and its derivatives, transition metal chalcogenides, transition metal carbides, transition metal nitrides, transition metal carbonitrides, two-dimensional black phosphorus, hexagonal boron nitride, two-dimensional perovskite materials, metal-organic framework nanosheets, covalent organic framework nanosheets, two-dimensional metal oxide nanosheets and two-dimensional metal nanosheets.

[0017] In some embodiments, as a preferred embodiment, the light source configuration module adopts a multi-channel independently driven solid-state light source array, and the solid-state light source is selected from light-emitting diodes (LEDs), laser diodes (LDs), vertical cavity surface-emitting lasers, infrared thermal radiation sources, superluminescent light-emitting diodes, photonic crystal surface-emitting lasers, tunable semiconductor lasers, and quantum cascade lasers.

[0018] In some embodiments, preferably, the sensing and detection module adopts a hermetically sealed structure. The hermetically sealed structure has the aforementioned two-dimensional heterojunction array and environmental temperature, humidity and pressure monitoring module inside, and a wide-spectrum light transmission window and a gas microchannel that connects the inside and outside on the outside. The optical path structure is located inside or outside the wide-spectrum light transmission window. The optical path structure is equipped with a lens group or optical waveguide element and is configured between the light source configuration module and the two-dimensional heterojunction array. It is used to collimate and defocus or homogenize the modulated light wave signal so that it is projected onto the surface of the two-dimensional heterojunction array without damage and uniformly.

[0019] In some embodiments, as a preferred embodiment, the environmental temperature, humidity and pressure monitoring module includes a temperature sensor, a humidity sensor and a pressure sensor, used to capture changes in environmental temperature, humidity and pressure in real time, and to assist the back-end model in performing gas-sensitive baseline compensation.

[0020] The signal acquisition and conversion module includes a multi-channel cross-group amplifier, an analog filter, a 16-bit synchronous sampling analog-to-digital converter, and a data control unit. The acquired dynamic electrical signals are sequentially passed through the multi-channel cross-group amplifier, the analog filter, and the 16-bit synchronous sampling analog-to-digital converter. The data control unit then arranges the signals according to the channel number and time label to generate a spatiotemporal sequence matrix. The data control unit uses an onboard FPGA or MCU.

[0021] The light-controlled gas biomimetic recognition device of the present invention also includes a communication transmission module and a power supply module. The communication transmission module is connected to the data processing module and uploads the recognition results and device status data of the data processing module to a cloud server via a network communication protocol. The power supply module is used to provide the necessary operating power to the above modules.

[0022] This invention also provides a two-dimensional heterojunction array sensing method for high-precision monitoring of carbon pollution, comprising the following steps:

[0023] (1) Active optical sequence modulation: A tunable light source is used to emit modulated optical wave signals to a two-dimensional heterojunction array. The modulation process is to dynamically change the output wavelength, output power and modulation frequency of the tunable light source in the time domain according to a preset switching sequence, and generate changing photon energy through wavelength switching.

[0024] (2) Multi-bit signal generation and acquisition: A two-dimensional heterojunction array composed of multiple sensitive units with vertical heterojunction structures is constructed. Under the joint illumination of the gas environment to be measured and the modulated light wave signal, each photosensitive unit generates a dynamic electrical signal based on the coupling of photosensitive gas sensing characteristics.

[0025] (3) Spatiotemporal sequence matrix construction: acquire dynamic electrical signals, perform analog-to-digital conversion and pattern arrangement to generate digital signals in the form of multidimensional matrix; wherein, pattern arrangement refers to reconstructing the one-dimensional time-stream digital signal acquired by the front end and converted by analog-to-digital conversion into a multidimensional tensor matrix that is suitable for convolutional scanning, containing spatial dimension, time dimension and parameter dimension of corresponding optical modulation parameters;

[0026] (4) Deep extraction and recognition of hybrid features: The digital signal in the form of a multidimensional matrix is ​​input into the pre-trained neural network model. The neural network model includes a convolutional neural network branch, a recurrent neural network branch and a fusion layer. The convolutional neural network branch is used to extract the spatial pattern features generated by the digital signal at different locations, while the recurrent neural network branch is used to extract the dynamic temporal features of the digital signal as the light sequence changes. The spatial pattern features and dynamic temporal features are nonlinearly mapped through the fusion layer, and finally the type and concentration distribution information of the gas to be tested are output.

[0027] In step (1), the photon energy preferably changes in a stepwise manner, gradually transitioning from low to high or from high to low. This is achieved through periodic irradiation via the aforementioned light source configuration module, modulated according to a preset active light sequence. Its function and principle lie in the fact that gas molecules with different polarities and functional group structures possess varying physical adsorption energies or chemical bond association barriers on the surface of two-dimensional materials. The stepwise changing spectrum acts like an energy scan. When the photon energy transitions to the desorption activation energy or excitation threshold matching a specific gas molecule, it triggers a sudden response in that gas, thereby inducing differentiated adsorption / desorption kinetic responses at the sensitive interface.

[0028] In step (2), the two-dimensional heterojunction array is composed of vertical heterojunction structures with at least two different combinations of two-dimensional materials. Utilizing the cross-sensitivity characteristics of different two-dimensional material combinations (such as graphene, MXenes, transition metal chalcogenides, etc.) to specific light bands and gas molecules, highly differentiated raw response data is obtained. The complete generation steps of this dynamic electrical signal are as follows: First, a pre-defined sequence of light beams irradiates the heterojunction to excite non-equilibrium electron-hole pairs; second, gas molecules adsorbed on the surface capture or release charge carriers based on their electronegativity, changing the baseline conductivity; next, a stepwise change in the optical parameters breaks and reconstructs the local charge balance on the surface; finally, the photosensitive unit with the vertical heterojunction structure outputs a dynamic current signal containing transient charge-discharge slopes, relaxation time differences, and response values. The function of this mechanism is to expand the traditional static resistance bandwidth into a high-dimensional physical information source containing microscopic adsorption dynamics characteristics.

[0029] The multidimensional tensor matrix constructed in step (3) forms a "dynamic fingerprint" that can characterize gas molecules.

[0030] In step (4), the loss function of the recurrent neural network branch incorporates a calculus constraint term based on the physical boundary; the total loss function L total Defined as:

[0031] ,

[0032] In the formula, L CE For classification cross-entropy loss, L MSEMean square error loss for concentration quantification. This is the penalty term for the physical constraints in the dynamic differential equation, where α is the weighting coefficient. The measured slope of the differential of the acquired dynamic current. This is the prediction function for ideal dynamic evolution.

[0033] As a preferred option, in step (4), the real-time collected temperature, humidity and air pressure are also input into the fusion layer as independent independent variables, and together with the spatial pattern features and temporal features, they participate in the fully connected mapping to eliminate the influence of meteorological background noise on weak electrical signals.

[0034] The present invention has the following advantages over the prior art:

[0035] (1) Unlike the static resistance signal that traditional electronic noses rely on, this invention introduces multiple controlled physical dimensions such as wavelength, power, and frequency through active optical sequence modulation, which greatly improves the "information entropy" output by the two-dimensional heterojunction array and effectively solves the problems of single material recognition limitations and multi-gas cross-sensitivity in complex environments.

[0036] (2) In response to the complex spatiotemporal coupling signals generated by active light modulation, this invention abandons the simple dimensionality reduction algorithm and adopts a hybrid architecture that combines convolutional neural network branches and recurrent neural network branches, which can simultaneously capture the physical distribution characteristics of the array and the adsorption kinetics under light excitation, significantly enhancing the accuracy of distinguishing gases with similar chemical properties.

[0037] (3) Photoinduced carrier modulation is used instead of traditional thermal excitation control, which avoids the high power consumption and thermal inertia delay generated by micro heaters, and enables the system to have a shorter response time and a longer device life. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall structure of the two-dimensional heterojunction array sensing device in Embodiment 1 of the present invention;

[0039] Figure 2 This is a schematic diagram of the optical path component configured between the light source configuration module and the two-dimensional heterojunction array in Embodiment 1 of the invention;

[0040] Figure 3 for Figure 1 A schematic cross-sectional view of a single two-dimensional heterojunction sensitive unit in the two-dimensional heterojunction array shown.

[0041] Figure 4 This is a flowchart of the two-dimensional heterojunction array sensing method for high-precision monitoring of carbon pollution synergy in Embodiment 2 of the present invention;

[0042] Figure 5This is a schematic diagram showing the correspondence between active optical sequence modulation and array dynamic response signal in Embodiment 2 of the present invention;

[0043] Figure 6 This is a schematic block diagram of the multi-channel signal acquisition and pattern arrangement circuit in Embodiment 2 of the present invention;

[0044] Figure 7 This is a diagram of the deep learning model architecture for processing multidimensional spatiotemporal matrices in Embodiment 2 of the present invention;

[0045] Figure 8 The identification results of multi-component VOCs by the active optical sequence modulation scheme in Embodiment 3 of the present invention;

[0046] Figure 9 This is the identification result of multi-component VOCs by the conventional illumination scheme in Comparative Example 3 of the present invention;

[0047] Figure 10 The identification results of the active optical sequence modulation scheme in Embodiment 4 of the present invention for the multi-component monitoring scenario of carbon pollution in flue gas;

[0048] Figure 11 This is the identification result of the traditional illumination scheme in Comparative Example 4 of the present invention for the multi-component monitoring scenario of carbon pollution in flue gas. Detailed Implementation

[0049] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0050] Example 1

[0051] This embodiment provides a two-dimensional heterojunction array sensing device based on the principle of photoinduced gas-sensitive response.

[0052] like Figure 1As shown, the two-dimensional heterojunction array sensing device 100 of the present invention is composed of a light source configuration module 110, an hermetically sealed structure 120, a sensing and detection module 130, an environmental temperature, humidity and pressure monitoring module 140, a signal acquisition and conversion module 150, a data processing module 160, a communication transmission module 170, and a power supply module 180. The light source configuration module 110 adopts a solid-state light source array driven by 6 independently driven channels. In this embodiment, one ultraviolet LED (center wavelength 365nm), one blue LED (center wavelength 450nm), one green LED (center wavelength 560nm), one red LD (center wavelength 660nm), one near-infrared LD, and one mid-infrared quantum cascade laser (center wavelength 4200nm) are selected. These 6 independent light sources together constitute the fixed light source array of the light source configuration module 110, and their wavelengths cover the ultraviolet to mid-infrared band (the specific light source can be selected according to the requirements). The hermetically sealed structure 120 includes a broadband light-transmitting window 121 (made of materials such as calcium fluoride, zinc selenide, sapphire, or modified infrared quartz) and a gas microchannel 122. The gas microchannel 122 connects the interior and exterior of the hermetically sealed structure 120, placing the sensing module 130 in the environment of the gas to be measured. An optical path assembly composed of aspherical lenses, microlens arrays, or integrating spheres is disposed inside or outside the broadband light-transmitting window 121. This optical path structure is used to shape, expand, and homogenize the discrete beam emitted from the multi-channel light source, ensuring uniform illumination and a stable gas flow field. Figure 2 As shown, the optical path components in this embodiment are configured between the light source configuration module 110 and the broadband light transmission window 121, and sequentially include the following components along the optical path transmission direction: a first plano-convex lens 220, a second plano-convex lens 230, a concave lens 240, a microlens array 250, and a third plano-convex lens 260. The specific working process of this optical path component is as follows: The large divergence angle discrete light emitted by the fixed light source array 210 of the light source configuration module is first collimated by the first plano-convex lens 220, transforming it into a parallel beam. Subsequently, the parallel beam passes sequentially through a focal-free beam-shrinking system composed of a second plano-convex lens 230 and a concave lens 240, compressing the cross-sectional aperture of the beam to precisely match the effective light-passing aperture of the subsequent microlens array 250. Next, the beam-shrinking parallel beam enters the microlens array 250 for spatial frequency domain segmentation and shaping, uniformly dispersing the originally high-energy Gaussian beam. Finally, the uniformly dispersed light passes through the third plano-convex lens 260 for refocusing and secondary collimation, transforming it into a uniformly collimated beam with a flat-top intensity distribution. Ultimately, this uniformly collimated beam passes through the broadband light-transmitting window 121 outside the airtight room, projecting evenly and completely covering the surface of the entire two-dimensional heterojunction array 131. This composite optical path design reduces the energy gradient difference between the center and the edge of the light source, ensuring that all sensitive units in the array receive completely consistent photon excitation energy at the same time.

[0053] The sensing module 130 includes a multi-unit two-dimensional heterojunction array 131, serving as the core feature extraction source for sensing the gas. The environmental temperature, humidity, and pressure monitoring module 140 is fitted near the sensing module 130, inside the hermetically sealed structure 120, or at the entrance of the gas microchannel 122. It captures microclimate parameters and inputs them into the signal acquisition and conversion module 150 to assist in baseline drift correction. The signal acquisition and conversion module 150 is responsible for converting the array's conductivity changes into standardized digital signals. The data processing module 160 integrates a hybrid neural network model specifically designed for analyzing spatiotemporal sequence data. The communication transmission module 170 wirelessly transmits gas concentration information to a cloud server. The power supply module 180 provides the necessary operating power to all the above modules. The device emits modulated light waves (according to a preset light source sequence) through the light source configuration module 110, inducing the two-dimensional heterojunction array to generate a photosensitive gas current. This current is processed by the signal acquisition and conversion module 150, and the data processing module 160 generates the identification result. Finally, the communication transmission module 170 transmits the identification result to the cloud.

[0054] like Figure 3 As shown, the sensitive unit constituting the two-dimensional heterojunction array 131 is constructed on a silicon substrate 310 with a 300 nm thick silicon dioxide layer, and Cr / Au (5 / 50 nm) interdigitated electrodes 320 are fabricated using an electron beam evaporation process. Above the electrode gap, a first two-dimensional material 330 (using P-type WSe2 in this embodiment) and a second two-dimensional material 340 (using N-type MoS2 in this embodiment) are sequentially stacked using a dry transfer process to form a vertical heterojunction structure (the composition of the two-dimensional materials in the vertical heterojunction structure can be selected according to requirements). The two-dimensional heterojunction array 131 achieves "cross-sensitivity" to different gases by changing the combination of heterojunctions (such as graphene / MoS2, MXene / WS2, etc.). At least two heterojunction combinations are used to constitute the two-dimensional heterojunction array 131 to obtain different adsorption / desorption kinetic responses in different spatial arrangements. This configuration utilizes the differences in band arrangement between materials, so that when the array is photoexcited, different sensitive units exhibit distinctly different electronic dynamics characteristics in the adsorption / desorption of the same gas molecules, thus solving the problem of insufficient selectivity of single materials at the physical level.

[0055] Example 2

[0056] This embodiment describes the standard operating procedures for gas identification using the device in Embodiment 1.

[0057] like Figure 4As shown, the identification method flow is as follows: First, the light source configuration module 110 emits modulated light wave signals to the two-dimensional heterojunction array 131 (hereinafter referred to as the sensing array in the figure and the following figure) according to the preset light source sequence; then, the signal acquisition and conversion module 150 synchronously acquires the dynamic electrical signal generated by the sensing array under the combined action of the gas to be tested and the modulated light, and performs analog-to-digital conversion and pattern arrangement on the original electrical signal to generate a multi-dimensional digital signal matrix containing spatial and temporal dimensions; finally, the matrix is ​​input into the pre-trained lightweight neural network model in the data processing module 160, and the model extracts features through depth and finally outputs the type and concentration information of the gas to be tested in real time.

[0058] To obtain multi-dimensional signals, one of the core aspects of this invention lies in controlling the emission of light at specific wavelengths to achieve photon energy (E). photon The photon energy exhibits a stepped downward or reverse upward movement. According to the Planck-Einstein relation, the formula for calculating the excited photon energy is:

[0059]

[0060] In the formula, E photon To excite photons, h is Planck's constant, c is the speed of light, and λ is the output wavelength. Within a defined light source sequence, energy gradients are formed from 365 nm (photon energy approximately 3.4 eV), 450 nm (photon energy approximately 2.75 eV), 660 nm (photon energy approximately 1.87 eV), to the mid-infrared 4200 nm (photon energy approximately 0.29 eV). Introducing the principle of photoinduced desorption kinetics, the transient desorption rate constant K of gas molecules under specific illumination is used. des Limited by its unique adsorption activation energy barrier Ea, the approximate Arrhenius modification equation under light intervention can be expressed as:

[0061]

[0062] In the formula, k is the Boltzmann constant, T is the absolute temperature, and K0 is the K0. des E represents the transient desorption rate constant. a This represents the adsorption activation barrier. This formula demonstrates the mechanism of the present invention: when E... photon ≥E a When the exponential term approaches 0, the desorption rate increases rapidly. For example, for strongly adsorbed NO2 molecules (activation energy exceeding 1.5 eV), conventional low-energy light cannot intervene, and the array response maintains the adsorbed state during visible / infrared light scanning. However, when the wavelength is switched to the ultraviolet stage, a violent non-equilibrium desorption current is triggered. By comparing time and wavelength parameters, a physical basis for strictly distinguishing different mixed gases is formed.

[0063] like Figure 5As shown, to enhance the information richness of the sensing signal at the physical level, this embodiment designs a composite optical modulation sequence. This optical modulation sequence has a 20-second cycle and is divided into four characteristic stages: the first stage is the excited state (365nm ultraviolet light), which uses high-energy photons to forcibly generate non-equilibrium carriers, activating high-energy adsorption sites on the material surface; the second stage is the parametric scanning state (alternating 450nm and 660nm pulses, with power modulation waveforms including but not limited to sine waves, square waves, triangular waves, or sawtooth waves), which extracts the dynamic response characteristics of gas molecules through energy gradient perturbations; the third stage is the thermally assisted desorption state (4200nm mid-infrared high-frequency pulses), which uses the photothermal effect to accelerate gas molecule desorption; the fourth stage is the baseline recovery state (dark state or low-power reference light), used to collect the recovery curve and serve as the initial baseline for the next cycle. When the light source configuration module 110 executes this optical modulation sequence, the conductivity of the sensitive unit will undergo complex dynamic evolution as the optical parameters change step by step, since different wavelengths of photon energy correspond to different band excitations. The signal acquisition and conversion module 150 synchronously captures the response current I(t) of N sensitive units in the two-dimensional heterojunction array 131 (i.e., sensor 1...sensor N shown in the figure) at a sampling rate of 200Hz, and arranges them into a multi-dimensional digital signal matrix with N×M spatiotemporal response according to the time axis of the optical sequence (i.e., Figure 5 The spatiotemporal response matrix is ​​shown. This matrix is ​​the "kinetic fingerprint" of the gas, and its information entropy is much higher than that of traditional static resistance change signals.

[0064] Meanwhile, to support the accurate reconstruction of the multi-dimensional matrix, the signal acquisition and conversion module 150 employs a synchronization mechanism for light source and data acquisition. For example... Figure 6 As shown, its core circuit includes a multi-channel transimpedance amplifier array and a 16-bit synchronous sampling analog-to-digital converter (ADC). The circuit locks the ADC sampling clock via a hardware synchronization bus, ensuring nanosecond-level alignment of each current sampling point with the wavelength and power parameters in the light source sequence on the time axis. Furthermore, a fourth-order low-pass filter (or other analog filters) is integrated into the circuit to filter out spike noise generated by light source switching. The acquired data is arranged in patterns by channel number and time stamp via the onboard FPGA or MCU, ultimately generating a structured spatiotemporal sequence matrix that can be directly processed by the algorithm. This circuit design ensures signal integrity and a high signal-to-noise ratio, providing the hardware foundation for high-precision identification.

[0065] Furthermore, to effectively extract the multidimensional features of the signal, another core aspect of this invention lies in employing a deep learning network model with a dual-branch hybrid architecture tailored for photosensitive gas characteristics, simultaneously capturing the physical distribution characteristics of the spatiotemporal sequence matrix and the adsorption kinetics under photoexcitation. For example... Figure 7As shown, firstly, for an array containing N sensor units, taking a single-cycle T time sampling step size and S light modulation states as an example, the dimension of the input tensor reconstructed by the acquisition module is strictly defined as... The three-dimensional matrix. Convolutional neural network branch 410 uses a two-dimensional convolutional kernel, specifically for scanning the spatial dimension of the response matrix and identifying the spatial response pattern features between different heterojunction sensitive units in the matrix; recurrent neural network branch 420 uses a bidirectional long short-term memory network, which can also be replaced with a long short-term memory network, gated recurrent unit, bidirectional gated recurrent unit, simple recurrent unit, or independent recurrent neural network, etc., depending on the computing power configuration, to focus on analyzing the dynamic transient process generated by the switching of light sequences. At the algorithm model iteration level, in order to accurately capture features, this embodiment adds the aforementioned physical boundary-based calculus constraint term to the loss function of deep learning. The total loss function L total Defined as:

[0066] ,

[0067] In the formula, L CE For classification cross-entropy loss, L MSE The mean square error loss for concentration quantification is the last term in the formula. This is a penalty term for the constraints of the dynamic differential physics (where α is the weighting coefficient). The measured slope of the differential of the acquired dynamic current. (For ideal dynamic evolution prediction function). Force the model to learn the dynamic slope of current change. This allows for the extraction of the "kinetic fingerprint" of gases. The physical essence of the dynamic slope is the transient adsorption rate constant characterizing the gas molecules in the photoexcited state. Since different gas molecules have different excitation energy barriers under specific wavelengths of light, this derivative term constitutes the core feature for distinguishing gases with similar chemical polarities. The adsorption / desorption rates of different gas molecules under specific light energies are unique. This architecture enables the deep mining of gas molecule dynamics characteristics from complex spatiotemporal signals, outperforming traditional algorithms that reduce the dimensionality of data.

[0068] The spatial features extracted by the convolutional neural network branch 410 and the temporal features extracted by the recurrent neural network branch 420 are concatenated in the fusion layer 430 to form a comprehensive feature vector. This vector is finally passed through the fully connected layer 440 and the Softmax / linear activation function to output the species probability and concentration estimate of the target gas.

[0069] In addition, the model needs to be pre-trained with a large number of samples and supports transfer learning during actual deployment. It can complete the weight fine-tuning with a small amount of field data, effectively overcoming the pain point of poor individual consistency of two-dimensional materials while ensuring real-time performance at the edge.

[0070] For data fusion of the environmental temperature, humidity and pressure monitoring module 140, the real-time collected temperature, humidity and air pressure can be input into the fusion layer 430 as independent scalar variables, and together with the sensing feature vectors extracted by the deep learning network, they participate in the fully connected mapping to eliminate the influence of meteorological background noise on weak electrical signals.

[0071] Example 3

[0072] This embodiment provides a specific implementation scheme for applying the technology of the present invention to a multi-component VOCs gas identification scenario.

[0073] This embodiment employs a two-dimensional heterojunction array composed of four two-dimensional materials: MoS2 / WSe2, graphene / MoS2, MXene / WS2, and Bi2O3 / MXene. In the active optical sequence modulation step, the light source configuration module is set to periodically scan and output according to a specific energy gradient: a single cycle takes 20 seconds, sequentially outputting beams with center wavelengths of 660nm (lasting 5 seconds), 450nm (lasting 5 seconds), and 365nm (lasting 10 seconds). A mixed gas with cross-interference characteristics—formaldehyde (5ppm), acetone (10ppm), and ethanol (20ppm)—is introduced into the hermetically sealed structure 120 through a gas microchannel 122. For data processing, this embodiment rigorously collects the dynamic transient current slope generated by the array at the instants of the aforementioned gradient optical parameters, arranges it into a three-dimensional tensor containing spatial and temporal sequences, and inputs it into a pre-trained neural network model for recognition and computation.

[0074] Test results are as follows Figure 8 As shown, the recognition accuracy rates of formaldehyde, acetone, and ethanol in this embodiment are 96.3%, 97.8%, and 97.1%, respectively. The quantitative relative error of each component after separation is reduced to within 4.5%, and the characteristic response output time is shortened to 20 seconds.

[0075] Comparative Example 3

[0076] This comparative example provides a traditional illumination scheme for multi-component VOCs gas identification scenarios and tests it.

[0077] This comparative example uses a two-dimensional heterojunction array constructed from the two-dimensional material combination as described in Example 3. An LED light source with a wavelength of 450 nm was used as the excitation source. A mixed gas with cross-interference characteristics was introduced: formaldehyde (5 ppm), acetone (10 ppm), and ethanol (20 ppm). The change in static resistance amplitude after the heterojunction array reached adsorption steady state was extracted, dimensionality reduction was performed using principal component analysis, and the data were then input into a support vector machine for classification.

[0078] Test results are as follows Figure 9As shown, the accuracy rates of this scheme in identifying formaldehyde, acetone, and ethanol are 56.7%, 64.2%, and 72.1%, respectively. Due to mutual interference, the relative errors in the quantification of each component are all greater than 35%, and the time for a single feature extraction (to reach steady state) is approximately 231 seconds.

[0079] As can be seen from Example 3 and Comparative Example 3, although VOCs molecules such as formaldehyde, acetone, and ethanol have similar polarities, their activation energy barriers for detaching from the surface of the heterojunction material differ objectively. The constant wavelength light source in Comparative Example 3 simultaneously excites these molecules, causing aliasing of electrical signals. However, Example 3, employing the technical solution of this invention, constructs an "energy sieve" through active optical sequence modulation: as the light wavelength crosses from 660nm (low energy) to 365nm (high energy), the photon excitation energy gradually increases; when the energy just crosses the activation energy barrier of a certain VOCs gas, it triggers non-equilibrium desorption of the gas molecules, thereby generating a rapidly changing current at a specific node on the time axis. This physical dynamic mechanism, which decouples aliased signals that were originally overlapping in space and amplitude in the time dimension through sequential staggered decoupling, combined with the temporal capture and extraction capabilities of deep learning, eliminates the boundary ambiguity problem in the feature space, significantly improves the separation and recognition accuracy of gases with similar polarities, and greatly shortens the feature response time.

[0080] Example 4

[0081] This embodiment provides a specific implementation plan for applying the technical solution of the present invention to a multi-component monitoring scenario of carbon pollution in flue gas.

[0082] This embodiment employs a two-dimensional heterojunction array composed of six two-dimensional materials: MoS2 / MXene, BN / MoS2, MoS2 / WSe2, graphene / MoS2, MXene / WS2, and Bi2O3 / MXene. In the active light modulation step, the light source configuration module is set to output precisely according to a specific 30-second single-cycle cyclic sequence: the first stage of this sequence is a deep cleaning period, outputting a mid-infrared pulse with a center wavelength of 4200nm (lasting 5 seconds), utilizing the infrared thermal effect to force the desorption of high-concentration CO2 and water molecules occupying the sites; the second stage is a baseline recovery period, where the light source is turned off and the system is in a dark state (lasting 5 seconds); the third to fifth stages are step-energy excitation periods, sequentially outputting visible to ultraviolet modulated beams with center wavelengths of 660nm (lasting 5 seconds), 450nm (lasting 5 seconds), and 365nm (lasting 10 seconds). Simulating flue gas emission conditions, a mixed gas with a relative humidity of 85% and containing 12% CO2 was introduced, along with a multi-component pollutant gas containing NO (30ppm), SO2 (10ppm), NH3 (10ppm), and benzene (2ppm). For data processing, the transient current slope generated at specific optical parameter abrupt changes was collected, and simultaneously acquired micro-meteorological array data was used as a background compensation tensor. Both were input into a pre-trained neural network model for nonlinear intercept correction and recognition calculations.

[0083] Test results are as follows Figure 10 As shown, the identification accuracy of the four target gases NO, SO2, NH3 and benzene in this embodiment reached 91.5%, 85.2%, 87.8% and 84.3%, respectively. At the same time, the relative error of the quantitative test of CO2 was controlled within 5%, and the relative detection error of the quantitative test of the above four trace components was less than 10%.

[0084] Comparative Example 4

[0085] This comparative example provides a traditional thermal control scheme for multi-component monitoring of carbon pollution in flue gas and tests it.

[0086] This comparative example uses a two-dimensional heterojunction array constructed with the same two-dimensional material combination as in Example 4. Instead of optical sequence modulation, thermal excitation control was achieved using an integrated microheater. The flue gas test conditions were simulated identically to those in Example 4, introducing a mixed gas with a relative humidity of 85% and containing 12% CO2, and injecting a multi-component pollutant gas containing NO (30 ppm), SO2 (10 ppm), NH3 (10 ppm), and benzene (2 ppm). The characteristic sequence of the resistance amplitude change of the two-dimensional heterojunction array during the temperature variation cycle was extracted, dimensionality reduced using principal component analysis, and input into a support vector machine for testing and evaluation.

[0087] Test results are as follows Figure 11As shown, after introducing a background gas of high concentration CO2 and high humidity water vapor, the active sites on the surface of the two-dimensional heterojunction array experience severe acceptor saturation due to thermal drift and the strong masking effect of interfering components on this type of array material. The qualitative identification accuracy of this method for trace components NO, SO2, NH3, and benzene is as low as 53.4%, 38.5%, 35.1%, and 12.6%, respectively. The relative error for quantitative detection of each trace pollutant component is as high as 68%–100%, making it impossible to effectively detect trace components.

[0088] As can be seen from Example 4 and Comparative Example 4, Example 4 of the present invention significantly improves the anti-interference ability under strong carbon and high humidity interference, surpassing the traditional thermal excitation control scheme. The thermal control of Comparative Example 4 suffers from severe thermal inertia, failing to quickly remove large amounts of H2O and CO2 occupying active sites. Furthermore, high concentrations of CO2 cause drift in the static resistance of the sensor array, which is the main reason for the high quantitative deviation of conventional algorithms. In contrast, the active optical sequence pulses used in Example 4 of the present invention enable rapid desorption of H2O and CO2, providing more adsorption sites for trace amounts of target pollutants such as NO and SO2. The present invention extracts the transient current slope of the optical parameter jump; this slope has specific characteristics, effectively mitigating the influence of baseline drift, thereby improving the discrimination accuracy of complex gases.

[0089] This invention organically combines a two-dimensional heterojunction array, dynamic light modulation, and a lightweight deep learning model to construct a novel gas recognition paradigm. It effectively solves the pain points of existing technologies and has significant practical value and broad application prospects.

Claims

1. A two-dimensional heterojunction array sensing device for high-precision monitoring of carbon pollution, characterized in that: The two-dimensional heterojunction array sensing device includes: The light source configuration module emits light wave signals with dynamically changing wavelength, power, and modulation frequency to the two-dimensional heterojunction array according to a preset light source sequence. The sensing and detection module includes an optical path structure and a two-dimensional heterojunction array. The light wave signal emitted by the light source configuration module is collimated, defocused, or homogenized through the optical path structure and then projected onto the surface of the two-dimensional heterojunction array, generating a dynamic electrical signal in the environment of the gas to be measured. The two-dimensional heterojunction array is composed of multiple sensitive units with vertical heterojunction structures arranged together. The two-dimensional heterojunction array has at least two different combinations of vertical heterojunction structures. The signal acquisition and conversion module acquires dynamic electrical signals and generates a spatiotemporal sequence matrix. The data processing module uses a hybrid model of convolutional neural network branches and recurrent neural network branches to extract the spatial pattern features and dynamic temporal features of the spatiotemporal sequence matrix, respectively. After performing nonlinear mapping, it outputs the type and concentration distribution information of the gas to be tested.

2. The two-dimensional heterojunction array sensing device according to claim 1, characterized in that, The two-dimensional heterojunction array sensing device also includes an environmental temperature, humidity and pressure monitoring module, which captures changes in environmental temperature, humidity and air pressure in real time and inputs them into the data processing module to assist in baseline drift correction.

3. The two-dimensional heterojunction array sensing device according to claim 2, characterized in that, The vertical heterojunction structure is formed by stacking a first two-dimensional material and a second two-dimensional material above the gap between the substrate electrodes. The first two-dimensional material and the second two-dimensional material in each vertical heterojunction structure are different two-dimensional materials, forming two or more different combinations of vertical heterojunction structures, which are then arranged to form the two-dimensional heterojunction array. The two-dimensional materials include graphene and its derivatives, transition metal chalcogenides, transition metal carbides, transition metal nitrides, transition metal carbonitrides, two-dimensional black phosphorus, hexagonal boron nitride, two-dimensional perovskite materials, metal-organic framework nanosheets, covalent organic framework nanosheets, two-dimensional metal oxide nanosheets, and two-dimensional metal nanosheets.

4. The two-dimensional heterojunction array sensing device according to claim 3, characterized in that, The light source configuration module adopts a multi-channel independently driven solid-state light source array, which is selected from light-emitting diodes, laser diodes, vertical cavity surface-emitting lasers, infrared thermal radiation sources, superluminescent light-emitting diodes, photonic crystal surface-emitting lasers, tunable semiconductor lasers, and quantum cascade lasers.

5. The two-dimensional heterojunction array sensing device according to claim 4, characterized in that, The sensing and detection module adopts a hermetically sealed structure. The hermetically sealed structure contains the two-dimensional heterojunction array and the environmental temperature, humidity and pressure monitoring module inside, and a wide-spectrum light transmission window and a gas microchannel that connects the inside and outside outside. The optical path structure is located inside or outside the wide-spectrum light transmission window.

6. The two-dimensional heterojunction array sensing device according to claim 5, characterized in that, The environmental temperature, humidity, and pressure monitoring module includes a temperature sensor, a humidity sensor, and a pressure sensor; the signal acquisition and conversion module includes a multi-channel cross-group amplifier, an analog filter, a 16-bit synchronous sampling analog-to-digital converter, and a data control unit; the acquired dynamic electrical signals are sequentially passed through the multi-channel cross-group amplifier, the analog filter, and the 16-bit synchronous sampling analog-to-digital converter, and then arranged by the data control unit according to the channel number and time label to generate a spatiotemporal sequence matrix; the data control unit is an onboard FPGA or MCU.

7. The two-dimensional heterojunction array sensing device according to claim 6, characterized in that, The two-dimensional heterojunction array sensing device also includes a communication transmission module and a power supply module; the communication transmission module is connected to the data processing module and uploads the data to the cloud server.

8. A two-dimensional heterojunction array sensing method for high-precision monitoring of carbon pollution synergy, characterized in that, Includes the following steps: (1) Active optical sequence modulation: A tunable light source is used to emit modulated optical wave signals to a two-dimensional heterojunction array. The modulation process is to dynamically change the output wavelength, output power and modulation frequency of the tunable light source in the time domain according to a preset switching sequence, and generate changing photon energy through wavelength switching. (2) Multi-bit signal generation and acquisition: A two-dimensional heterojunction array composed of multiple sensitive units with vertical heterojunction structures is constructed. Under the joint illumination of the gas environment to be measured and the modulated light wave signal, each photosensitive unit generates a dynamic electrical signal based on the coupling of photosensitive gas sensing characteristics. (3) Spatiotemporal sequence matrix construction: dynamic electrical signals are collected, analog-to-digital conversion and pattern arrangement are performed to generate digital signals in the form of multidimensional matrix; the pattern arrangement refers to the reconstruction of the one-dimensional time-flow digital signal collected by the front end and converted by analog-to-digital conversion, based on the spatial coordinate mapping of the corresponding sensitive unit, the band parameter label of the optical signal and the time axis mark, to construct a multidimensional tensor matrix suitable for convolutional scanning, which includes spatial dimension, time dimension and parameter dimension of corresponding optical modulation parameters; (4) Deep extraction and recognition of hybrid features: The digital signal in the form of the multidimensional matrix is ​​input into the pre-trained neural network model; the neural network model includes a convolutional neural network branch, a recurrent neural network branch and a fusion layer. The convolutional neural network branch is used to extract the spatial pattern features generated by the digital signal at different positions, and the recurrent neural network branch is used to extract the dynamic temporal features of the digital signal as the light sequence changes. The spatial pattern features and dynamic temporal features are nonlinearly mapped through the fusion layer, and finally the type and concentration distribution information of the gas to be tested are output.

9. The two-dimensional heterojunction array sensing method according to claim 8, characterized in that, In step (1), the photon energy changes in a stepwise manner, gradually transitioning from low to high or from high to low, and is periodically irradiated; in step (2), the two-dimensional heterojunction array is composed of vertical heterojunction structures with at least two different combinations of two-dimensional materials.

10. The two-dimensional heterojunction array sensing method according to claim 8, characterized in that, In step (4), the loss function of the recurrent neural network branch incorporates a calculus constraint term based on the physical boundary; the total loss function L total Defined as: ， In the formula, L CE For classification cross-entropy loss, L MSE Mean square error loss for concentration quantification. This is the penalty term for the physical constraints in the dynamic differential equation, where α is the weighting coefficient. The measured slope of the differential of the acquired dynamic current. This is the prediction function for ideal dynamic evolution.

Images