Harmful gas sensor detection method based on Fano resonance signal feature extraction

Abstract

The invention provides a harmful gas sensor detection method based on Fano resonance signal feature extraction, which comprises the following steps: acquiring a Fano resonance signal, extracting Fano resonance signal features from the Fano resonance signal, and acquiring spectral information corresponding to the Fano resonance signal; performing comparison of known single gas sample characteristics on the Fano resonance signal characteristics to obtain an initial harmful gas type; analyzing the Fano resonance signal features to obtain derivative features, and intercepting the spectral information according to the Fano resonance signal features to obtain local spectral information; the Fano resonance signal features, the derivative features and the local spectrum information are identified through the deep learning model, the type and the concentration of the harmful gas are obtained, and the type and the concentration of the harmful gas can be effectively identified through the scheme.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, and more specifically, to a method for detecting harmful gases using a sensor based on Fano resonance signal feature extraction. Background Technology

[0002] Traditional gas detection methods, such as electrochemical sensors, semiconductor sensors, and infrared spectroscopy, while meeting basic detection needs to some extent, still face many challenges in practical applications: limited sensitivity, poor selectivity, susceptibility to environmental temperature and humidity interference, slow response speed, insufficient long-term stability, and difficulty in simultaneously detecting multiple gases online. Traditional technologies are particularly inadequate for detecting trace amounts of harmful gases.

[0003] In recent years, gas sensing technologies based on optical principles, especially surface plasmon resonance (SPR) and localized surface plasmon resonance (LSPR) technologies, have gradually become a research hotspot due to their advantages such as high sensitivity, fast response, and strong resistance to electromagnetic interference. However, traditional SPR / LSPR sensors typically rely on the resonant wavelength shift as the sole detection signal, and their sensitivity is limited by the full width at half maximum (FWHM) of the resonance peak. Furthermore, when dealing with complex gas mixtures, a single wavelength shift often fails to achieve accurate differentiation and quantitative analysis of multi-component gases.

[0004] Fano resonance, a unique optical phenomenon originating from quantum interference between discrete and continuous states, manifests as sharp and asymmetrical resonance valley lines in the spectrum. This unique line shape is not only extremely sensitive to minute changes in the refractive index of the surrounding medium, but its parameters also contain rich physical information. In recent years, nanostructure-based Fano resonance sensors have shown great potential in fields such as biomolecule detection and environmental monitoring. However, most current research remains at the stage of using single resonance wavelength shifts for detection, failing to fully utilize the multi-dimensional characteristic information contained in the Fano resonance line shape, resulting in significant limitations in the identification and concentration quantification of complex gas mixtures. Summary of the Invention

[0005] In view of this, the present invention proposes a method for detecting harmful gases based on Fano resonance signal feature extraction, in order to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention proposes a method for detecting harmful gases using a sensor based on Fano resonance signal feature extraction, comprising: Acquire the Fano resonance signal, extract the Fano resonance signal features from the Fano resonance signal, and obtain the spectral information corresponding to the Fano resonance signal; By comparing the characteristics of the Fano resonance signal with those of known single-gas samples, the initial type of harmful gas can be obtained. The characteristics of the Fano resonance signal are analyzed to obtain derived features, and the spectral information is truncated based on the characteristics of the Fano resonance signal to obtain local spectral information; By using a deep learning model to identify the characteristics, derived features, and local spectral information of the Fano resonance signal, the types and concentrations of harmful gases can be obtained.

[0007] Optionally, the Fano resonance signal characteristics include the wavelength, full width at half maximum (FWHM), and optical parameters corresponding to the Fano resonance valley, wherein the optical parameters are reflectivity or transmittance or a combination thereof.

[0008] Optionally, the initial process for obtaining the type of hazardous gas includes: The wavelengths in the Fano resonance signal characteristics are matched with the wavelengths of known single-gas sample characteristics, and the initial type of harmful gas is obtained based on the matching results.

[0009] Optionally, the derived features include: resonance depth, quality factor, resonance area, asymmetry parameter, and offset normalization parameter. The resonance depth is based on the resonance valley depth at the wavelength position in the Fano resonance signal features. The quality factor is the ratio of wavelength to full width at half maximum (FWHM). The resonance area is the product of resonance depth and FWHM. The asymmetry parameter is obtained based on the fitted linear function of the Fano resonance signal. The offset normalization parameter is the ratio of the difference between the wavelength and the baseline wavelength to the FWHM.

[0010] Optionally, local spectral information can be obtained by truncating the spectral information with the wavelength corresponding to the Fano resonance valley as the center and combining it with the full width at half maximum (FWHM).

[0011] Optionally, the deep learning model employs an encoder-decoder structure.

[0012] Optionally, the process of identifying Fano resonance signal features, derived features, and local spectral information using a deep learning model includes: The Fano resonance signal features, derived features, and local spectral information are integrated into a feature vector; the feature vector is verified and pre-processed; a feature interaction graph is constructed based on the pre-processed feature vector; feature fusion is performed based on the feature interaction graph; and the fused features are processed in depth through a series of sequentially connected residual blocks to obtain the encoded feature vector. Weights are assigned to the encoded feature vectors to obtain a subset of category recognition features and a subset of concentration recognition features. The subset of category recognition features is then processed sequentially with convolution, self-attention mechanism and fully connected layer to obtain discriminant features. The discriminant features are then transformed to obtain the gas category. Based on the gas category, the subset of concentration recognition features is linearly transformed and constrained to obtain the concentration data.

[0013] On the other hand, the present invention also provides a hazardous gas sensor detection system based on Fano resonance signal feature extraction for performing the above-described method.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This method significantly improves detection sensitivity by deeply mining the multi-dimensional linear features inherent in the Fano resonance signal. Traditional methods typically rely solely on a single shift in the resonance wavelength for judgment, while this invention integrates multiple features such as resonance depth, full width at half maximum (FWHM), asymmetry parameters, quality factor, and resonance area. These features exhibit varying sensitivities and response modes to local refractive index changes caused by gas adsorption, collectively forming a richer and more unique optical fingerprint. Through the synergistic analysis and intelligent modeling of multiple features, the system can detect trace gases at low concentrations and types, thereby improving detection accuracy. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] This embodiment proposes a method for detecting harmful gases using a sensor based on Fano resonance signal feature extraction, such as... Figure 1 As shown, it includes: Acquire the Fano resonance signal, extract the Fano resonance signal features from the Fano resonance signal, and simultaneously obtain the spectral information corresponding to the Fano resonance signal; By comparing the characteristics of the Fano resonance signal with known features, the initial type of harmful gas can be obtained. The characteristics of the Fano resonance signal are analyzed to obtain derived features, and the spectral information is truncated based on the characteristics of the Fano resonance signal to obtain local spectral information; The concentration of harmful gases is obtained by identifying the characteristics, derived features, and local spectral information of the Fano resonance signal using a deep learning model.

[0018] Regarding the acquisition of the aforementioned Fano resonance signal, optical information is acquired through a nanostructure capable of inducing Fano resonance. The nanostructure consists of a glass substrate and several integrated basic units arranged on its top surface. The glass substrate is completely transparent to light in the working wavelength band and serves as the carrier for the basic units. Each basic unit on the glass substrate is composed of two parallel nanoscale rods, forming a pair capable of inducing Fano resonance. The rods are made of a high-refractive-index dielectric material, such as silicon or germanium, and are cuboid in shape. The two rods in a pair differ in size, either in length, width, or the aforementioned combination method. This difference is the physical basis for the subsequent complex optical interference. Simultaneously, a small notch is etched in the center of each nanorod to enhance its sensitivity. When a broadband beam of light with a polarization direction aligned with the length direction of the nanorod is perpendicularly irradiated onto the planar structure of the glass substrate... Because the two short rods of the basic unit arranged in the glass substrate are of different sizes, they interact with light, exciting two resonant modes with similar frequencies but very different characteristics in the transmission and reflection spectra. The light waves radiated by these two excited modes interfere as they propagate in space. At a specific light frequency, their waves are exactly out of phase, canceling each other out, resulting in a sharp and steep dip in the intensity of the transmitted light. On a spectrometer, this appears as a sharp and asymmetrical spectral line dip, known as the Fano resonance valley. Simultaneously, the gap in the middle of the nanorod generates an extremely strong, highly localized electromagnetic hotspot around its opening, and this hotspot is completely exposed. When the gas or biomolecules to be measured are exposed to the gap region, they directly change the local optical refractive index at the hotspot, immediately perturbing the Fano resonance valley and causing a precisely measurable shift in its center wavelength, ultimately allowing the acquisition of spectral signals of transmittance or reflectance at different wavelengths. During the measurement process, the basic unit can be directly exposed to the gas to be analyzed. Spectral signals are acquired using the above-described setup, and the wavelength, transmittance, reflectance, and full width at half maximum (FWHM) corresponding to the Fano resonance valley are extracted as Fano resonance signal features. Based on these spectral signals and Fano resonance signal features, subsequent analysis of harmful gases is performed. It should be noted that this invention primarily provides a method for detecting harmful gases using a sensor based on Fano resonance signal feature extraction. The aforementioned structure is merely an illustrative example. It can induce Fano resonance in the sensor probe structure and acquire relevant spectral information and Fano resonance signal features, including the wavelength, FWHM, and either the corresponding transmittance or reflectance, all of which can be effectively analyzed in subsequent sections.

[0019] The raw spectral data of the single gas and harmful gases are acquired, with the horizontal axis representing the wavelength of light and the vertical axis representing the transmittance or reflectance. Under baseline conditions unaffected by harmful gases, a distinctive Fano resonance valley will appear on the spectral curve, serving as the baseline data without harmful gas interference. The spectral data undergoes preprocessing, including filtering and smoothing, and baseline correction. The Fano resonance signal characteristics are then acquired, including the wavelength corresponding to the Fano resonance valley, the full width at half maximum (FWHM), and one or more of the corresponding transmittance or reflectance.

[0020] After acquiring the preprocessed spectral data and Fano resonance signal characteristics, the wavelengths are compared with those in the pre-stored Fano resonance signal characteristics of known types of standard hazardous gases. Gases matching the same wavelength are selected. The Fano resonance signal characteristics of known types of standard hazardous gases are obtained by measuring and analyzing high-purity hazardous gases using a sensor probe. Simultaneously, for each gas type, a single or multiple linear fitting relationship is used between the concentration and the corresponding full width at half maximum (FWHM), transmittance, or reflectance data. This fitted relationship serves as the corresponding analytical model. Based on the identified hazardous gases, the corresponding fitting relationship is used to perform concentration fitting based on the acquired Fano resonance signal characteristics, generating corresponding concentration results. This yields the type and concentration of the hazardous gas in the gas to be analyzed, which is then verified by comparing the spectral data with the known hazardous gas content.

[0021] The above scheme can be used as a simple application to analyze the types and concentrations of harmful gases. It can perform preliminary analysis, roughly determining the concentration and type of harmful gases. While the wavelength position of the Fano resonance can effectively identify the type of harmful gas, its accuracy is insufficient for analyzing the concentration of several harmful gases. In concentration analysis, the concentration of harmful gases is strongly correlated with the aforementioned Fano resonance signal characteristics. Therefore, this strong correlation can be modeled using a deep learning model to effectively identify the coupled concentrations of multiple harmful gases. Simultaneously, the overall spectral characteristics can also be used as auxiliary information input into the deep learning model to help it more effectively identify the types and concentrations of harmful gases. The effective identification method for the above-mentioned harmful gas concentrations can be specifically achieved through encoding and decoding. After acquiring spectral data and Fano resonance signal characteristics, the specific details of using a deep learning model for subsequent harmful gas concentration and type identification are as follows: The raw spectra first undergo a preprocessing process, including filtering and smoothing based on the Savitzky-Golay algorithm. This algorithm, through polynomial least squares fitting within a sliding window, eliminates high-frequency noise while preserving spectral detail, particularly the steep changes at the edges of the Fano resonance valleys. Next, baseline correction is performed using an adaptive iterative weighted least squares method. Through multiple iterations, the slowly varying background in the spectrum is gradually separated from the rapidly changing resonance features, ultimately yielding a corrected spectrum containing only the gas response signal.

[0022] After acquiring the corrected spectrum and Fano resonance signal characteristics, to enhance the physical characterization ability of the features, derived features based on core parameters were further calculated. These derived features include resonance depth, quality factor, resonance area, asymmetry parameter, and offset normalization parameter. The resonance depth D was calculated using a relative value method, with the average transmittance or reflectance at the shoulders on both sides of the resonance valley as the background intensity, calculating the relative difference between the valley bottom intensity and the background intensity. The quality factor Q = center wavelength λ / full width at half maximum (FWHM) Γ reflects the sharpness of the resonance; a higher quality factor implies a sharper resonance and higher theoretical sensitivity. The resonance area A is the product of the resonance depth D and the FWHM Γ. The resonance area is proportional to the oscillation energy and provides a holistic measure of the resonance intensity. The asymmetry parameter q was obtained by fitting a Fano linear function; it describes the degree of asymmetry in the resonance valley. Different interactions between gas molecules and plasmon polarons lead to different asymmetry variations. The offset normalization parameter δ = (center wavelength λ - baseline wavelength λ0) / Γ converts the absolute wavelength offset into a relative offset in linewidth units. This normalization process allows for direct comparison of sensors with different initial conditions. These derived features, as supplementary features to the aforementioned Fano resonance signal features, provide a greater knowledge base for further identification of harmful gases.

[0023] Simultaneously, the identification is aided by the local spectral features of the Fano resonance valley. The extraction of these local spectral features employs a dynamic windowing method centered on the resonance valley. The window width is not a fixed wavelength range but is proportional to the measured full width at half maximum (FWHM) Γ, typically set within ±2Γ. The extracted spectral segments are further normalized: on the wavelength axis, the center wavelength λ of the resonance center is used as zero, and the scale is recalculated in units of Γ to align resonance valleys with different linewidths on the horizontal axis; on the intensity axis, normalization is performed in units of resonance depth D to eliminate absolute intensity differences. The normalized spectral segments are discretized into several equally spaced sampling points, forming a one-dimensional vector. This vector completely preserves the local shape information of the resonance valley, including the sharpness of the valley bottom, the asymmetry of the slopes on both sides, and any possible minute secondary structures.

[0024] After multi-scale feature extraction, all features are integrated into a structured feature vector. This vector is organized in an orderly manner according to feature type and scale: first, the core scalar features, then the physically derived features, and finally the local spectral feature vectors. The entire feature vector constitutes a high-dimensional information space, containing multi-level information ranging from microscopic linear details to macroscopic statistical regularities.

[0025] The integrated feature vector described above is encoded using an encoder, transforming this high-dimensional feature vector into a low-dimensional feature representation. Known physical knowledge of Fano resonance is used as training samples to train the network, guiding it to perform efficient searches in a solution space that conforms to physical laws.

[0026] The encoder performs different recognitions for different parameter combinations. For the core scalar feature set (λ, D, Γ), a physical quantity consistency verification unit is used. This unit first checks whether these three parameters meet basic physical constraints: the center wavelength λ should be within the sensor's operating wavelength range; the resonance depth D should be between 0 and 1 (normalized transmittance); and the full width at half maximum (FWHM) Γ should be greater than zero and less than a certain upper limit (determined by the sensor design). Any feature combination that violates these basic constraints triggers a correction mechanism—a reasonable adjustment through a lightweight physical model, rather than a simple rejection. For example, if the input Γ value is negative, the system automatically replaces it with the moving average of the most recent historical measurements and marks the data point as low confidence.

[0027] For the physically derived feature set (Q, A, q, δ), a self-consistency check network is used. This network checks whether the mathematical relationship between the derived features and the original features is consistent. For example, it calculates Q' = λ / Γ and compares it with the input Q value; it calculates A' = D × Γ and compares it with the input A value. If the difference exceeds a threshold, the network generates a correction vector to fine-tune these features to restore their self-consistency. This design ensures that even if small computational errors occur during the front-end feature extraction process, the encoder can automatically correct them, maintaining the inherent consistency of the physical system.

[0028] For local spectral feature vectors, a physical shape attention mechanism is employed. This mechanism first identifies key regions in the spectral vector: the bottom region of the resonance valley, the left rising edge, the right falling edge, and the shoulder plateau region. Then, different attention weights are assigned to each region; these weights are not fixed but dynamically adjusted based on the currently measured physical parameters. For example, when the quality factor Q is high (sharp resonance), the weight of the valley region increases; when the asymmetry parameter q is large, the weight difference between the left and right sides widens. This physically guided attention mechanism enables the network to focus on the most informative spectral region under the current physical state.

[0029] Then, a multi-scale feature fusion module effectively interacts features at different scales. This module first constructs a feature interaction graph, where nodes represent feature groups at different scales, and edges represent the physical relationships between features. Strong connections are established between scalar feature nodes and derived feature nodes; medium-strength connections are established between scalar feature nodes and spectral feature nodes; and connections are established between derived feature nodes based on their physical correlation. The topology of this interaction graph is not fixed but dynamically adjusted according to the physical state of the input features, although the weights of its connections are pre-set.

[0030] Building upon the interaction graph, this module implements a message passing mechanism with physical constraints. Each node, when receiving information from neighboring nodes, performs a physical plausibility filter: only accepting information compatible with the current node's physical state. For example, when a wavelength node receives information from a linewidth node, it checks whether this information matches the dispersion relation allowed by the current wavelength value. This filtering mechanism prevents physically impossible feature combinations from propagating in the network.

[0031] Feature fusion is performed using a gated attention mechanism, where the interaction weights mentioned above are added to the corresponding attention weight calculation through a weighted sum. Different weights are assigned to the features to generate fused features. The mechanism also records which original features are associated with the fused features.

[0032] Then, hierarchical feature refinement is used for deeper processing. The initially fused feature representations are progressively refined through a series of residual blocks. Each residual block contains two fully connected layers and a specific activation function design: wavelength-related neurons use the Softplus activation function to simulate the logarithmic response characteristics of optical frequencies; intensity-related neurons use the Sigmoid activation function to ensure the output range is within a physically reasonable range; and linewidth-related neurons use the ReLU activation function to guarantee non-negative output. These customized activation functions incorporate domain knowledge into the network structure, improving the interpretability and physical consistency of the feature representations. The refinement process progressively removes irrelevant information and noise, strengthening the signal components relevant to the gas detection task. Finally, the encoded results for the above features are generated.

[0033] The decoder decodes the encoder's output. After obtaining a single feature vector from the encoder, the core task of the decoder is to decompose this unified feature representation into outputs specific to different tasks. The decoder employs a multi-stream parallel architecture to specifically handle two gas detection tasks: species identification (classification task) and concentration quantification (regression task).

[0034] The first layer of the decoder is a task-aware feature allocation network. This layer does not perform feature transformation but learns how to assign different parts of the encoded features to different task flows. For the input encoded feature vector, two parallel gating vectors are designed to process the encoded vector through a gating mechanism. Each gating vector is constrained to a value between 0 and 1 by a sigmoid function. Two task-specific feature subsets are calculated by using the gating vectors as weights. The feature subset for the category recognition task is the element-wise product of the original encoded features and the category recognition gating vector, while the feature subset for the concentration regression task is the element-wise product of the original encoded features and the concentration regression gating vector.

[0035] Then, the feature subset of the category recognition task is processed through a category recognition structure, consisting of three layers. The first layer is a local discriminative network, which uses a set of one-dimensional convolutional kernels to scan the features and extract local patterns. The second layer is a global relational network, which uses a self-attention mechanism to calculate the interrelationships between feature dimensions and discover the cooperative variation patterns between different physical parameters. The third layer is a high-order interaction network, which learns nonlinear combinations of features through multiple fully connected layers to form high-order discriminative features. The output of the high-order discriminative features is converted into probability values ​​containing various gases using a Softmax function, and results greater than a certain probability threshold are selected as the category results.

[0036] Then, a concentration regression structure is used to estimate the output concentration of various possible gases. A multi-head regression network is employed. Based on the aforementioned category results, corresponding heads are activated, each head specifically predicting the concentration of one gas. The concentration estimate for each gas is calculated through a linear transformation and then activated using a ReLU activation function to ensure non-negativity. Several constraints are added to the output layer. First, a non-negativity constraint: all concentration estimates must be greater than or equal to zero. Second, an upper bound constraint: the concentration estimate for each gas cannot exceed the sensor's detection limit for that gas. Third, a sparsity constraint: L1 regularization encourages sparsity in the concentration vector, as only a few gases typically coexist in practical applications. Furthermore, for the category results, wavelength-filtered category results can be selected as reference information for activating the head of the concentration regression result, and the category identification structure is removed to reduce computational complexity.

[0037] Compared to existing technologies, this method significantly improves detection sensitivity by deeply mining the multi-dimensional linear features inherent in the Fano resonance signal. Traditional methods typically rely solely on a single shift in the resonance wavelength for judgment, while this invention integrates multiple features such as resonance depth, full width at half maximum (FWHM), asymmetry parameters, quality factor, and resonance area. These features exhibit varying sensitivities and response modes to local refractive index changes caused by gas adsorption, collectively forming a richer and more unique "optical fingerprint." Through the synergistic analysis and intelligent modeling of multiple features, the system can achieve low-concentration detection of trace gases, thus improving detection accuracy.

[0038] Traditional optical sensors are susceptible to interference from factors such as fluctuations in ambient temperature and humidity, light source stability, and mechanical vibration. This invention, by introducing a series of physically guided feature engineering and network design mechanisms, significantly suppresses the influence of non-target factors on the final detection results. Even under harsh operating conditions with varying temperature and humidity and background gas interference, the system maintains stable and reliable detection performance.

[0039] This invention achieves high-precision simultaneous quantitative analysis of gas types and concentrations. Through an innovative multi-stream decoder architecture with task decoupling, the system can process both gas type identification and concentration quantification in parallel and collaboratively. Specifically, the concentration regression stream optimizes the mapping relationship between features and concentration values, ensuring the overall accuracy of the analysis.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting harmful gases using a sensor based on Fano resonance signal feature extraction, characterized in that, include: Acquire the Fano resonance signal, extract the Fano resonance signal features from the Fano resonance signal, and obtain the spectral information corresponding to the Fano resonance signal; By comparing the characteristics of the Fano resonance signal with those of known single-gas samples, the initial type of harmful gas can be obtained. The characteristics of the Fano resonance signal are analyzed to obtain derived features, and the spectral information is truncated based on the characteristics of the Fano resonance signal to obtain local spectral information; By using a deep learning model to identify the characteristics, derived features, and local spectral information of the Fano resonance signal, the types and concentrations of harmful gases can be obtained.

2. The method according to claim 1, characterized in that, The characteristics of the Fano resonance signal include the wavelength, full width at half maximum (FWHM), and optical parameters corresponding to the Fano resonance valley, where the optical parameters are reflectivity or transmittance or a combination thereof.

3. The method according to claim 2, characterized in that, The initial process for obtaining the type of hazardous gas includes: The wavelengths in the Fano resonance signal characteristics are matched with the wavelengths of known single-gas sample characteristics, and the initial type of harmful gas is obtained based on the matching results.

4. The method according to claim 2, characterized in that, The derived features include: resonance depth, quality factor, resonance area, asymmetry parameter, and offset normalization parameter. The resonance depth is based on the resonance valley depth at the wavelength position in the Fano resonance signal features. The quality factor is the ratio of wavelength to full width at half maximum (FWHM). The resonance area is the product of resonance depth and FWHM. The asymmetry parameter is obtained based on the fitted linear function of the Fano resonance signal. The offset normalization parameter is the ratio of the difference between the wavelength and the baseline wavelength to the FWHM.

5. The method according to claim 1, characterized in that, By using the wavelength corresponding to the Fano resonance valley as the center and combining it with the full width at half maximum (FWHM), the spectral information is truncated to obtain local spectral information.

6. The method according to claim 1, characterized in that, The deep learning model employs an encoder-decoder structure.

7. The method according to claim 1, characterized in that, The process of identifying Fano resonance signal features, derived features, and local spectral information using deep learning models includes: The Fano resonance signal features, derived features, and local spectral information are integrated into a feature vector; the feature vector is verified and pre-processed; a feature interaction graph is constructed based on the pre-processed feature vector; feature fusion is performed based on the feature interaction graph; and the fused features are processed in depth through a series of sequentially connected residual blocks to obtain the encoded feature vector. Weights are assigned to the encoded feature vectors to obtain a subset of category recognition features and a subset of concentration recognition features. The subset of category recognition features is then processed sequentially with convolution, self-attention mechanism and fully connected layer to obtain discriminant features. The discriminant features are then transformed to obtain the gas category. Based on the gas category, the subset of concentration recognition features is linearly transformed and constrained to obtain the concentration data.

8. A hazardous gas sensor detection system based on Fano resonance signal feature extraction, characterized in that, Used to perform the method described in any one of claims 1-7.

Images