Multi-dimensional multiplexing optical fiber sensing chip and spectrum intelligent demodulation method

Abstract

The invention discloses a multi-dimensional multiplexing optical fiber sensing chip and an intelligent spectrum demodulation method, and belongs to the technical field of optical fiber sensing. The chip adopts a unique space division, wavelength division and mode division multi-dimensional multiplexing design, and multi-target synchronous detection is realized by integrating a sensing optical fiber array and a microfluidic detection cell on a chip substrate. The invention further provides an intelligent spectrum demodulation method based on the deep neural network, global spectrum features are extracted through a one-dimensional Transform encoder, local feature aggregation is performed in combination with a graph attention network, and feature optimization is realized by adopting a multi-scale attention fusion mechanism. And finally, concentration detection and substance classification and identification are synchronously realized through a double-branch output structure. The device and the method have the technical advantages of integration, miniaturization and high throughput, high-sensitivity synchronous detection of multiple biochemical quantities is realized, and technical support is provided for clinical early screening, dynamic monitoring and instant decision making.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensor technology, specifically to a multidimensional multiplexed fiber optic sensor chip and a method for intelligent spectral demodulation. Background Technology

[0002] Fiber optic biosensing technology, as an advanced detection method, is based on the interaction between the evanescent field of light waves and the analyte. It infers environmental information (such as refractive index and interface thickness) by measuring changes in characteristic parameters of the light waves (such as wavelength, intensity, and polarization). Compared with traditional biomedical detection equipment, fiber optic sensors have significant advantages such as low cost, label-free operation, high sensitivity, multiplexing, resistance to electromagnetic interference, and good biocompatibility, showing broad application prospects in rapid in vivo / in vitro detection, real-time monitoring, and portable diagnostics. Common technical solutions include fiber structures based on surface plasmon resonance (SPR), side-polished fibers, fiber gratings, and interferometric structures. These solutions enhance sensing performance through different mechanisms. For example, SPR structures utilize metal thin films to excite localized surface plasmon resonance to improve sensitivity, while fiber gratings achieve wavelength-selective sensing through grating period modulation.

[0003] However, most traditional fiber optic biosensing solutions provide only a single sensitive area in space and target only a single biomass or physical parameter in terms of detection mode, resulting in low throughput. To simultaneously acquire multiple biomarkers (such as various proteins and metabolites) or multiple environmental parameters, multiple independent experiments are usually required. This not only reduces detection efficiency but also increases sample consumption, time costs, and operational complexity. This "single biomass-single parameter" design limitation makes it difficult to comprehensively reflect the multidimensional information of pathophysiology (such as the coupled changes of immune markers, metabolites, and microenvironment parameters) in applications such as disease diagnosis, thus affecting the accuracy, sensitivity, and specificity of diagnosis. Furthermore, existing multi-target detection often requires the combination of multiple expensive devices, further limiting its widespread adoption in primary healthcare or rapid on-site testing.

[0004] Therefore, there is an urgent need to develop a multi-parameter synchronous detection fiber optic sensing technology to solve the problems of low detection throughput, low efficiency, and high cost of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a multidimensional multiplexed fiber optic sensing chip and a spectral intelligent demodulation method. Based on space division, wavelength division and mode division multidimensional multiplexing, and through innovative structural design and deep learning algorithm integration, it achieves highly sensitive simultaneous detection of multiple biochemical quantities, providing technical support for clinical early screening, dynamic monitoring and real-time decision-making, and solving the above-mentioned problems in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multidimensional multiplexed fiber optic sensing chip includes a chip substrate, a fiber optic channel, a microfluidic detection cell, a sensing fiber array, a sample inlet, a microfluidic sample inlet channel, a sample outlet, and a microfluidic sample outlet channel, wherein: The chip substrate integrates an optical fiber channel and a microfluidic detection cell using micro-nano fabrication technology, and the surface of the chip substrate is covered with a chip cover plate. The sensing fiber array is set in the fiber channel and passes through the microfluidic detection cell. It is composed of a series of fiber sensing units and is used to excite surface wave resonance and output spectrum. The fiber sensing unit includes a sensing fiber and a heterogeneous thin film waveguide structure integrated on the surface of the sensing fiber. The sample inlet is connected to the microfluidic detection cell via a microfluidic sample inlet channel; The sample outlet is connected to the microfluidic detection cell via a microfluidic sample outlet channel.

[0007] Furthermore, the chip substrate is made of either an inorganic material or an organic polymer, wherein: The inorganic material includes glass, silicon wafers, or quartz; The organic polymers include polydimethylsiloxane (PDMS), polymethyl methacrylate (PMMA), polycarbonate (PC), or polytetrafluoroethylene (PTFE).

[0008] Furthermore, the heterogeneous thin-film waveguide structure is composed of a metal thin film and a dielectric thin film. The metal thin film is coated on the surface of the sensing optical fiber, and the dielectric thin film is coated on the surface of the metal thin film to excite surface wave resonance with a strong evanescent field.

[0009] Furthermore, the sensing fiber adopts a tilted fiber grating, the tilt angle of the tilted fiber grating is different at different fiber sensing units, and the thickness of the dielectric film on the surface of the tilted fiber grating is different at different fiber sensing units.

[0010] Furthermore, the sensing fiber adopts a long-period fiber grating, the grating period of the long-period fiber grating is different at different fiber sensing units, and the thickness of the dielectric film on the surface of the long-period fiber grating is different at different fiber sensing units.

[0011] Furthermore, the sensing optical fiber is either a side-thrown optical fiber or an unclad optical fiber, and the thickness of the dielectric film on the surface of the side-thrown or unclad optical fiber is different at different optical fiber sensing units.

[0012] Another objective of this invention is to provide a spectral intelligent demodulation method based on deep neural networks, characterized by comprising the following steps: S1: Multiple multidimensional multiplexed spectra are collected by a sensing chip to construct a multi-label supervised learning dataset, and the training data in the dataset are sequentially subjected to interpolation, normalization, wavelet denoising and data augmentation optimization. The sensing chip is a multidimensional multiplexed optical fiber sensing chip as described in any one of claims 1-6. S2: Construct a one-dimensional Transformer encoder to extract the global spectral features of the multidimensional multiplexed spectrum and output the global feature tensor; S3: Based on the wavelength difference of surface wave resonance of each sensing unit, the spectrum is divided into a preset number of spectral segments, a similarity adjacency matrix between spectral segments is constructed, and the spectral segment features are aggregated based on a graph attention network to output a local feature matrix. S4: Integrate the outputs of all coding layers of the one-dimensional Transformer encoder to form a stable multi-level feature representation, and fuse it with the local feature matrix to obtain a fused feature vector; S5: Based on the fused feature vector, output the concentration information of the analyte through the regression prediction branch, and output the classification information of the analyte through the classification prediction branch.

[0013] Furthermore, the method also includes model training and optimization, employing a two-stage training strategy. It introduces physical consistency data augmentation methods, adversarial training mechanisms, and transfer learning and self-supervised pre-training mechanisms, and uses a hybrid loss function for multi-task supervised learning of regression and classification.

[0014] Furthermore, the hybrid loss function includes concentration mean square error loss. Classification cross-entropy loss Spectrum smoothing regularization term Its expression is: ; Where α, β, and γ are the task weights.

[0015] Furthermore, the evaluation metrics for the regression prediction branch include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The evaluation metrics for the classification prediction branches use AUC value, F1-score, sensitivity, and specificity.

[0016] The multidimensional multiplexed fiber optic sensing chip and intelligent spectral demodulation method provided by this invention have the following significant advantages compared to existing technologies: Multidimensional multiplexing of space division, wavelength division, and mode division is achieved based on a single fiber optic sensor array. The highly integrated chip structure effectively reduces the size of the device and facilitates portable deployment. The fiber optic channels and microfluidic detection cells fabricated within the chip substrate work together to support efficient sample flow and detection, significantly reducing manufacturing costs and operational complexity, and offering advantages in integration and miniaturization. By exciting multi-order surface wave resonances at different locations using a heterogeneous thin-film waveguide, the simultaneous sensing of multiple biochemical quantities is achieved, breaking through the traditional limitation of "single biomass - single parameter". The sensing unit array generates multidimensional multiplexed spectra, enabling simultaneous detection of multiple targets and improving detection throughput and efficiency; A demodulation method integrating 1D-Transformer and graph attention network is used to automatically extract spectral features, supporting multi-task outputs such as concentration regression and classification. Combined with physical consistency data augmentation, the model's robustness to interference is enhanced, ensuring high accuracy and reliability in complex environments. It enables low-cost, high-throughput detection in fields such as early disease screening and environmental monitoring, providing key technical support for real-time analysis of multiple parameters and promoting the development of precision medicine and point-of-care diagnosis. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the multidimensional multiplexed optical fiber sensing chip of the present invention; Figure 2 This is a schematic diagram of the sensing fiber array of the present invention, which uses a tilted fiber grating as the sensing fiber. Figure 3 This is a schematic diagram of the multidimensional multiplexing output spectrum of the chip when the tilted fiber grating is used as the sensing fiber in the present invention. Figure 4 This is a flowchart of a spectral intelligent demodulation method based on a deep neural network according to the present invention; In the picture: 1-Chip substrate, 2-Fiber optic channel, 3-Microfluidic detection cell, 4-Sensing fiber array, 5-Inlet, 6-Microfluidic inlet channel, 7-Outlet, 8-Microfluidic outlet channel, 9-Metallic thin film, 10-Dielectric thin film, 11-Surface wave resonance. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0019] This embodiment provides a multi-dimensional multiplexed fiber optic sensing chip, the core of which lies in achieving simultaneous detection of multiple parameters through a unique spatial layout and material combination. For example... Figure 1 As shown, the chip adopts a two-layer stacked structure: the bottom layer is the chip substrate 1, and the top layer is the chip cover. The chip includes: chip substrate 1, fiber optic channel 2, microfluidic detection cell 3, sensing fiber array 4, sample inlet 5, microfluidic sample inlet channel 6, sample outlet 7, and microfluidic sample outlet channel 8. The chip employs a unique multi-dimensional multiplexing design, achieving simultaneous detection of multiple parameters through a single chip. The chip's multiplexing mechanism is based on three core technologies: Spatial multiplexing: Independent addressing of physical channels is achieved by arranging sensing units in different spatial locations; Wavelength division multiplexing: Utilizing the difference in the thickness of a dielectric thin film to generate a wavelength-coded resonant signal; Modular multiplexing: obtaining multidimensional response information by exciting surface wave modes of different orders.

[0020] The chip substrate 1 serves as a support platform, and its size can be flexibly designed according to detection requirements. Inside the substrate, an optical fiber channel 2 and a microfluidic detection cell 3 are integrated using advanced micro-nano fabrication technology. This integrated design ensures a high degree of synergy between optical detection and fluid control. A transparent chip cover is placed over the surface of the chip substrate to facilitate optical observation and isolate it from the influence of the external environment.

[0021] The selection of chip substrate materials takes into full consideration optical performance, chemical stability, and processing feasibility. Substrate materials can be inorganic materials (glass, silicon wafers, or quartz) or organic polymers (polydimethylsiloxane PDMS, polymethyl methacrylate PMMA, polycarbonate PC, or polytetrafluoroethylene PTFE), suitable for high-precision detection and biomedical applications.

[0022] The width and depth of the microfluidic detection cell 3 are in the millimeter range, and the size of the fiber optic channel matches that of standard single-mode or multimode fiber.

[0023] The sensing fiber array 4, as the core detection element, is located within the fiber optic channel 2 and passes through the microfluidic detection cell 3. It consists of a series of fiber optic sensing units and adopts a modular design concept to excite surface wave resonances and output spectra. Each fiber optic sensing unit includes a sensing fiber and a heterogeneous thin-film waveguide structure integrated on the surface of the sensing fiber. Differentiated responses are achieved by precisely controlling the structural parameters of each unit.

[0024] The heterogeneous thin-film waveguide structure consists of a metal thin film 9 and a dielectric thin film 10. The metal thin film 9 (such as gold or silver) is coated on the surface of the sensing fiber, and the dielectric thin film 10 (such as titanium dioxide or indium tin oxide) is coated on the surface of the metal thin film to excite surface wave resonances 11 with a strong evanescent field. The heterogeneous thin-film waveguide is fabricated using magnetron sputtering technology. First, a 20-50 nm thick metal thin film 9 is deposited on the surface of the sensing fiber, and then a dielectric thin film 10 with a thickness of tens to hundreds of nanometers is deposited. This heterogeneous thin-film waveguide structure can effectively excite surface wave resonances of different orders at different wavelengths, generating an enhanced evanescent field for sensing and detection.

[0025] The sample inlet 5 is connected to the microfluidic detection cell 3 via the microfluidic sample inlet channel 6. The sample outlet 7 is connected to the microfluidic detection cell 3 via the microfluidic sample outlet channel 8.

[0026] To achieve simultaneous detection of multiple targets, the core of this chip lies in its multidimensional multiplexing mechanism. This mechanism is achieved by systematically designing differentiated physical parameters for sensing units at different spatial locations within the sensing fiber array. Specifically, spatial division multiplexing, wavelength division multiplexing, and mode division multiplexing work synergistically, jointly encoded in the chip's output spectrum. The sensing fiber within the fiber sensing unit can be one of the following: tilted fiber grating (TFBG), long-period fiber grating (LPFG), side-thrown fiber, or unclad fiber. Each type achieves multiplexing functionality through differentiated parameter design, specifically: Preferably, when the sensing fiber uses a tilted fiber Bragg grating, the tilt angle of the tilted fiber Bragg grating at different sensing units is... The tilt angle varies (e.g., 5°-15°), and the dielectric film thickness h of the heterogeneous thin-film waveguide on the surface of the tilted fiber grating at different sensing units varies (e.g., 10-300 nm). The specific principle is: by adjusting the tilt angle... This allows the comb-shaped spectral envelope of the tilted fiber grating at different sensing units to be distributed in different wavelength ranges. Furthermore, by controlling the thickness h of the dielectric thin film 10, different cladding modes of the tilted fiber grating and different orders of surface wave modes of the heterogeneous thin film waveguide can be coupled at different wavelengths, that is, the cladding mode energy is coupled to the surface wave mode, thereby exciting the surface wave resonances 11 of different orders at different wavelengths, so as to obtain a multidimensional multiplexed output spectrum composed of surface wave resonances of different orders at different wavelengths at the chip output end.

[0027] Preferably, when a long-period fiber grating is used for the sensing fiber, the grating period (e.g., 100-800 μm) of the long-period fiber grating is different at different sensing units, and the thickness h (e.g., 10-300 nm) of the dielectric film 10 on the surface of the long-period fiber grating is different at different sensing units. Specifically, by adjusting the grating period, different cladding modes can be excited at different sensing units using the long-period fiber grating. Furthermore, by controlling the thickness h of the dielectric film 10, different cladding modes of the long-period fiber grating can be coupled with different orders of surface wave modes of the heterogeneous thin-film waveguide at different wavelengths, i.e., cladding mode energy is coupled to the surface wave mode, thereby exciting different orders of surface wave resonances 11 at different wavelengths.

[0028] Preferably, when the sensing fiber is a side-thrown fiber or an unclad fiber, the thickness h (e.g., 10-300 nm) of the dielectric film 10 on the surface of the side-thrown fiber or the unclad fiber at different sensing units is different. Specifically, by controlling the thickness h of the dielectric film 10, different modes of the side-thrown fiber or the unclad fiber at different sensing units can be coupled with different orders of surface wave modes of the heterogeneous thin-film waveguide at different wavelengths, thereby exciting different orders of surface wave resonances 11 at different wavelengths.

[0029] Through the above-mentioned differentiated design, surface wave resonances of different orders are excited at different wavelengths, forming a multidimensional multiplexed output spectrum of space division, wavelength division, and mode division. The gold film thickness of the heterogeneous thin-film waveguide is preferably 20-50 nm, and the dielectric film thickness is preferably 10-300 nm, within which the surface wave resonance signal is strongest.

[0030] This embodiment provides a spectral intelligent demodulation method based on deep neural networks, including: acquiring multiple multidimensional multiplexed spectra through a sensing chip, constructing a multi-label supervised learning dataset, and sequentially performing interpolation, normalization, wavelet denoising, and data augmentation optimization on the training data in the dataset; the sensing chip being a multidimensional multiplexed fiber optic sensing chip; constructing a one-dimensional Transformer encoder to extract global spectral features of the multidimensional multiplexed spectra and outputting a global feature tensor; dividing the spectrum into a preset number of spectral segments based on the wavelength differences of surface wave resonances of each sensing unit, constructing a similarity adjacency matrix between spectral segments, and aggregating the spectral segment features based on a graph attention network to output a local feature matrix; integrating the outputs of all coding layers of the one-dimensional Transformer encoder to form a stable multi-level feature representation, and fusing it with the local feature matrix to obtain a fused feature vector; based on the fused feature vector, outputting analyte concentration information through a regression prediction branch and analyte classification information through a classification prediction branch. The details are described below.

[0031] First, multiple multidimensional multiplexed spectra are collected by a sensing chip to construct a multi-label supervised learning dataset. Then, the training data in the dataset are sequentially subjected to interpolation, normalization, wavelet denoising, and data augmentation optimization. The sensing chip is a multidimensional multiplexed fiber optic sensing chip.

[0032] Under standard experimental conditions (temperature 25±1℃, relative humidity 50±5%), standard sample solutions containing different categories and concentrations of analytes are injected into a multidimensional multiplexed fiber optic sensing chip described in this invention through the injection port (5). Multidimensional multiplexed spectral data are repeatedly collected 3-5 times for each category and each concentration gradient (usually 5-8 gradients are set) to construct a multi-label supervised learning dataset containing analyte category and concentration labels.

[0033] The raw spectral data collected were processed sequentially as follows: Interpolation processing: unify the unequal interval sampling spectra into equal interval data (usually 0.01nm interval). Normalization: min-max normalization is used to map the spectral intensity to the [0,1] interval; Wavelet denoising: Three-level decomposition using the sym8 wavelet basis, followed by soft thresholding for denoising; Data augmentation: Physical consistency enhancement methods are employed, including spectral shifting and intensity perturbation. Physical consistency enhancement methods refer to data augmentation techniques that maintain the physical regularity of the spectrum. Specifically, spectral shifting involves shifting along the wavelength axis within ±0.02 nm to simulate instrument wavelength calibration errors; intensity perturbation randomly adjusts spectral intensity within ±5% to simulate light source fluctuations; baseline drift is achieved by adding a random polynomial baseline to simulate the effects of ambient temperature changes; noise injection adds Gaussian white noise with a signal-to-noise ratio of 30-50 dB to simulate changes in the detection environment spectrum; and the smoothing coefficient in the smoothing regularization term is typically set to 0.01-0.1.

[0034] Secondly, a one-dimensional Transformer encoder is constructed to extract the global spectral features of the multidimensional multiplexed spectrum, outputting a global feature tensor. The global feature path uses a one-dimensional Transformer encoder, which can effectively extract the global feature tensor (global spectral features). The encoder has 4-8 layers, preferably 6; too many layers can lead to overfitting, while too few layers result in insufficient feature extraction. Specific implementation parameters include: the input layer performs position encoding using learnable position embedding vectors, with the dimension matching the number of spectral data points; the self-attention mechanism uses scaled dot product attention, with 4-12 attention heads, preferably 8, each head having a dimension of 64, effectively capturing feature information from different subspaces; the feedforward network uses a two-layer fully connected structure, with the intermediate layer having a dimension of 2048 and using the GELU activation function. Layer normalization is applied after each sub-layer to stabilize the training process. Residual connections are used around each sub-layer to avoid gradient vanishing. The output is a global feature tensor. That is, an n×d dimensional feature tensor containing global spectral information, where n is the number of spectral data points and d is the feature encoding dimension.

[0035] Furthermore, based on the wavelength differences of surface wave resonances of each sensing unit, the spectrum is divided into a predetermined number of spectral bands, and a similarity adjacency matrix between spectral bands is constructed. The graph attention network is used to aggregate spectral features and output a local feature matrix. The preset number m of spectral segments is typically 8-12 segments, preferably 10 segments, which achieves the best balance between feature extraction efficiency and computational complexity. The design of the graph attention network for local feature paths focuses on the physical correlation between spectral segments. Node features are initialized using 12 statistical feature vectors of the spectral segments, including mean, variance, skewness, and kurtosis. Edge weights are calculated based on a weighted combination of spectral similarity (cosine similarity) and physical distance (wavelength interval) between spectral segments. The specific formula is as follows: ; in, and They represent the first The and the first The feature vector of each spectral segment node and These represent the center wavelength positions of the two spectral bands. The scale parameter is used to control distance attenuation.

[0036] The graph attention layer employs a 3-layer GAT structure, with each layer containing 8 attention heads, and uses the LeakyReLU activation function. Finally, it outputs a local feature matrix. That is, the m×d-dimensional graph feature matrix, which contains local spectral information.

[0037] The deep learning model is constructed using the aforementioned dual-path feature extraction architecture, which includes a global feature path and a local feature path. This dual-path design preserves global spectral features while making full use of detailed information in local spectral bands.

[0038] Then, the outputs of all coding layers of the one-dimensional Transformer encoder are integrated to form a stable multi-level feature representation. This is then fused with the local feature matrix to obtain a fused feature vector. Feature fusion employs a multi-scale attention fusion mechanism. First, multi-scale pooling is performed on the feature sequence output by the one-dimensional Transformer to extract feature representations at different granularities. Then, a cross-attention mechanism is used to calculate the correlation weights between the two feature sources, achieving adaptive weighted fusion of features. In this structure, the query vector originates from the multi-level features of the Transformer, which contain global context information, while the key and value vectors originate from local features of the graph network, which contain fine-grained structural information. The specific calculation process is as follows: ; in, For global feature representation, For local feature matrices, , , It is a learnable linear projection weight matrix.

[0039] The fused feature vectors undergo a dimensionality transformation layer to be unified into a standard k×d dimensional format (where k is the number of features). Layer normalization is used to stabilize the training process, and residual connections are employed to avoid the vanishing gradient problem. The final feature representation includes both global spectral distribution characteristics and retains detailed information from local spectral bands.

[0040] Finally, based on the fused feature vector The concentration information of the analyte is output through the regression prediction branch. It outputs the classification information of the test object through the classification prediction branch. Based on the fused feature vectors, a parallel dual-branch output architecture is constructed: Regression prediction branch: The network structure uses a three-layer fully connected network with hidden layer dimensions of 256, 128, and 64 respectively. Activation functions: ReLU activation function is used for the first two layers, and linear activation function is used for the output layer. Loss function: Mean squared error loss, with weights initialized to 0.4.

[0041] Classification prediction branch: The network structure consists of a one-dimensional convolutional layer (kernel size 3) + attention mechanism + global pooling layer. The attention mechanism uses a multi-head self-attention mechanism with 4 heads. The output layer uses a softmax activation function to output the class probability distribution. The loss function is cross-entropy loss, with weights initialized to 0.3.

[0042] For model training and optimization, a two-stage training strategy is adopted, introducing physical consistency data augmentation methods, adversarial training mechanisms, and transfer learning and self-supervised pre-training mechanisms. A hybrid loss function is used for multi-task supervised learning of regression and classification, and the hybrid loss function includes concentration mean squared error loss. Classification cross-entropy loss Spectrum smoothing regularization term Its expression is: ; α, β, and γ are task weights, which are automatically adjusted to achieve a balance between tasks.

[0043] Spectral smoothing regularization term: .

[0044] The two-stage training strategy includes: a pre-training stage using a large-scale unlabeled spectral dataset and employing a masked spectral reconstruction task for self-supervised learning; and a fine-tuning stage performing supervised training on the target dataset using a progressive unfreezing strategy. Specifically, the masked spectral reconstruction task is implemented by using a random block mask to reconstruct a spectrum of length... One-dimensional spectral data Divided into Each of the following is a non-overlapping local segment containing [number] segments. A series of continuous spectral sampling points, with a high mask ratio. (The preferred range is 60%-75%), and a portion of the segments are randomly selected in a uniform distribution for masking, thereby completing the self-supervised learning task of spectral reconstruction.

[0045] In addition, a performance evaluation and verification mechanism for the model was established. Performance evaluation includes regression performance evaluation and classification performance evaluation. The regression prediction branch uses regression performance evaluation, establishing a complete regression performance evaluation system, including root mean square error (RMSE, requirement <0.05), mean absolute error (MAE, requirement <0.03), and coefficient of determination (R², requirement >0.95). The classification prediction branch uses classification performance evaluation, through comprehensive evaluation of multiple indicators, including AUC value (requirement >0.95), F1-score (requirement >0.90), sensitivity (requirement >85%), and specificity (requirement >90%). For multi-class problems, macro-average and micro-average indicators are calculated. The robustness of the system is comprehensively verified through temperature stability testing (15-35℃), long-term stability testing (continuous operation for 24 hours), and cross-interference testing. A comprehensive performance monitoring system is established to ensure the reliability of the system in practical applications.

[0046] This invention is the first to combine a one-dimensional Transformer with a graph attention network for spectral demodulation, fully utilizing the complementarity of global and local features. Through the thickness gradient design of a heterogeneous thin-film waveguide, it achieves true space-division, wavelength-division, and mode-division multidimensional multiplexing, overcoming the limitations of traditional single-multiplexing methods. This invention not only solves the technical challenge of simultaneous multi-target detection in existing technologies but also opens up new avenues for the in-depth application of fiber optic sensing technology in the biomedical field. Its integrated, intelligent, and high-throughput characteristics make it a promising candidate for applications in precision medicine, point-of-care testing (early disease screening, environmental monitoring), and other fields.

[0047] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A multidimensional multiplexed fiber optic sensing chip, characterized in that, The components include a chip substrate (1), an optical fiber channel (2), a microfluidic detection cell (3), a sensing fiber array (4), a sample inlet (5), a microfluidic sample inlet channel (6), a sample outlet (7), and a microfluidic sample outlet channel (8), wherein: The chip substrate (1) integrates an optical fiber channel (2) and a microfluidic detection cell (3) through micro-nano fabrication technology. The surface of the chip substrate (1) is covered with a chip cover plate. The sensing fiber array (4) is set in the fiber channel (2) and passes through the microfluidic detection cell (3). It is composed of a series of fiber sensing units and is used to excite surface wave resonance and output spectrum. The fiber sensing unit includes a sensing fiber and a heterogeneous thin film waveguide structure integrated on the surface of the sensing fiber. The inlet (5) is connected to the microfluidic detection cell (3) through the microfluidic injection channel (6); The sample outlet (7) is connected to the microfluidic detection cell (3) through the microfluidic sample outlet channel (8).

2. The multidimensional multiplexed fiber optic sensing chip according to claim 1, characterized in that, The chip substrate (1) is made of either an inorganic material or an organic polymer, wherein: The inorganic material includes glass, silicon wafers, or quartz; The organic polymers include polydimethylsiloxane (PDMS), polymethyl methacrylate (PMMA), polycarbonate (PC), or polytetrafluoroethylene (PTFE).

3. The multidimensional multiplexed fiber optic sensing chip according to claim 1, characterized in that: The heterogeneous thin-film waveguide structure is composed of a metal thin film (9) and a dielectric thin film (10). The metal thin film (9) is coated on the surface of the sensing fiber, and the dielectric thin film (10) is coated on the surface of the metal thin film (9) to excite surface wave resonance (11) with a strong evanescent field.

4. The multidimensional multiplexed fiber optic sensing chip according to claim 1, characterized in that: The sensing fiber adopts a tilted fiber grating. The tilt angle of the tilted fiber grating is different at different fiber sensing units, and the thickness of the dielectric film (10) on the surface of the tilted fiber grating is different at different fiber sensing units.

5. A multidimensional multiplexed fiber optic sensing chip according to claim 1, characterized in that: The sensing fiber adopts a long-period fiber grating. The grating period of the long-period fiber grating is different at different fiber sensing units, and the thickness of the dielectric film (10) on the surface of the long-period fiber grating is different at different fiber sensing units.

6. The multidimensional multiplexed fiber optic sensing chip according to claim 1, characterized in that: The sensing fiber is either a side-thrown fiber or an unclad fiber, and the thickness of the dielectric film (10) on the surface of the side-thrown fiber or the unclad fiber is different at different fiber sensing units.

7. A spectral intelligent demodulation method based on deep neural networks, characterized in that, Includes the following steps: S1: Multiple multidimensional multiplexed spectra are collected by a sensing chip to construct a multi-label supervised learning dataset, and the training data in the dataset are sequentially subjected to interpolation, normalization, wavelet denoising and data augmentation optimization. The sensing chip is a multidimensional multiplexed optical fiber sensing chip as described in any one of claims 1-6. S2: Construct a one-dimensional Transformer encoder to extract the global spectral features of the multidimensional multiplexed spectrum and output the global feature tensor; S3: Based on the wavelength difference of surface wave resonance of each sensing unit, the spectrum is divided into a preset number of spectral segments, a similarity adjacency matrix between spectral segments is constructed, and the spectral segment features are aggregated based on a graph attention network to output a local feature matrix. S4: Integrate the outputs of all coding layers of the one-dimensional Transformer encoder to form a stable multi-level feature representation, and fuse it with the local feature matrix to obtain a fused feature vector; S5: Based on the fused feature vector, output the concentration information of the analyte through the regression prediction branch, and output the classification information of the analyte through the classification prediction branch.

8. The spectral intelligent demodulation method based on deep neural networks according to claim 7, characterized in that: This method also includes model training and optimization, employing a two-stage training strategy. It introduces physical consistency data augmentation methods, adversarial training mechanisms, and transfer learning and self-supervised pre-training mechanisms, and uses a hybrid loss function for multi-task supervised learning of regression and classification.

9. The spectral intelligent demodulation method based on a deep neural network according to claim 8, characterized in that, The hybrid loss function includes concentration mean square error loss. Classification cross-entropy loss Spectrum smoothing regularization term Its expression is: ； Where α, β, and γ are the task weights.

10. The spectral intelligent demodulation method based on a deep neural network according to claim 7, characterized in that: The evaluation metrics for the regression prediction branch include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The evaluation metrics for the classification prediction branches use AUC value, F1-score, sensitivity, and specificity.

Images