Laser radar multi-gas identification method and system based on deep learning

Abstract

The invention discloses a laser radar multi-gas identification method and system based on deep learning, and relates to the technical field of laser radars, and the method comprises the steps: employing a differential absorption laser radar to emit lasers with different wavelengths, receiving echo signals after gas absorption, extracting multi-domain features, and employing multi-target optimization in combination with feature importance evaluation to screen key features; generating a gas portrait of the potential gas according to the key features, and obtaining the type of the interested gas; constructing a graph structure according to the gas portrait and the gas feature similarity, performing representation learning on the graph structure by using a graph convolutional network, enhancing the graph convolutional network by using an interested gas category, constructing a gas category branch, and outputting a gas category; gas concentration branches are built by adopting a multi-layer sensor, final node representation output by the graph convolutional network is used as input, and each gas concentration is output by using linear regression. According to the invention, the distinguishing capability of small difference gases is improved, and high-precision identification and concentration inversion of multi-component gases are realized.

Description

Technical Field

[0001] The present invention relates to the field of laser radar technology, and more specifically, to a laser radar multi-gas identification method and system based on deep learning. Background Art

[0002] With the acceleration of industrialization and urbanization, air pollution is becoming increasingly prominent. The emission of multi-component harmful gases, in particular, poses a serious threat to the environment and human health. Therefore, the development of efficient and accurate gas detection technologies is of great significance for environmental monitoring, industrial safety, and emergency response. Traditional multi-gas detection methods, such as gas chromatography, electrochemical sensors, and infrared spectroscopy, while highly accurate, often suffer from limitations such as bulky equipment, slow response, high cost, and difficulty in achieving real-time online monitoring.

[0003] LiDAR technology, due to its non-contact, high sensitivity, and high spatial and temporal resolution, shows great potential in the field of gas detection. Differential absorption LiDAR can quantitatively detect specific gases by analyzing the differences in echo signals from lasers of different wavelengths on gas absorption spectra. However, traditional differential absorption LiDAR technology faces challenges in identifying multi-component gases, especially in complex mixed gas environments, where the absorption spectra of different gases may overlap, leading to cross-interference and affecting detection accuracy.

[0004] In recent years, deep learning technology has achieved breakthroughs in fields such as computer vision, speech recognition, and natural language processing. Its powerful feature extraction and pattern recognition capabilities offer new solutions for multi-gas detection. Deep learning models can automatically learn gas signatures from high-dimensional, nonlinear LiDAR spectral data, effectively overcoming the limitations of traditional algorithms. Therefore, combining deep learning with LiDAR technology to develop efficient and high-precision multi-gas identification methods is a pressing technical challenge in the fields of environmental monitoring and gas sensing. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a lidar multi-gas identification method and system based on deep learning, which achieves high-precision identification and concentration inversion of multi-component gases by improving the ability to distinguish gases with small differences.

[0006] A first aspect of the present invention provides a laser radar multi-gas identification method based on deep learning, comprising the following steps: A differential absorption lidar is used to emit lasers of different wavelengths and receive echo signals after gas absorption. The echo signals are subjected to noise suppression, differential absorption calculation, and normalization preprocessing. Multi-domain features are extracted from the preprocessed echo signals. Feature selection is performed using multi-objective optimization combined with feature importance assessment to screen key features. Performing feature mapping on the key features to generate a gas feature distribution map, dividing the feature clusters by clustering, generating a gas profile of the potential gas, and obtaining the gas category of interest based on the gas profile; Constructing a graph structure based on the gas profile and gas feature similarity, using a graph convolutional network to perform representation learning on the graph structure, and enhancing the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category; A multi-layer perceptron is used to construct a gas concentration branch. The final node representation output by the graph convolutional network is used as input. Linear regression is used to output the concentration of each gas, and the gas concentration and gas category are visualized.

[0007] In this solution, a differential absorption lidar is used to emit lasers of different wavelengths and receive echo signals after gas absorption. The echo signals are subjected to noise suppression, differential absorption calculation, and normalization preprocessing. Specifically, Obtain environmental parameters, historical pollution data, and pollution emission source information for the target gas identification scenario, construct scenario features, and use the scenario features to perform query matching in the industry-gas association knowledge base using a lightweight classification model; Output the gas presence probability through the matching results, sort the gases according to the gas presence probability, filter the gases that meet the preset probability threshold to generate a probabilistic gas list, query the wavelength corresponding to the strongest absorption peak of the gas based on the probabilistic gas list, and exclude the interference band of the background gas to generate the initial wavelength; Obtaining the detection confidence of a gas in the probabilistic gas list. When the detection confidence continues to be lower than a preset confidence threshold, increasing the auxiliary absorption peak wavelength of the gas. If an unexpected absorption peak in the probabilistic gas list is detected, initiating a full-band scan, locating the unexpected absorption peak through fast Fourier transform, and dynamically adjusting the initial wavelength. Reinforcement learning is introduced to build a reward model for wavelength selection and detection accuracy, and detection priorities are generated based on gas hazard. An optimal adjustment strategy is generated based on the reward model and detection priority, and the wavelength range is adaptively optimized using the optimal adjustment strategy. A tunable laser is used to emit laser light with an adaptively optimized wavelength, and a photodetector is used to receive the echo signal after gas absorption. Time-resolved intensity data is recorded, and noise suppression, differential absorption calculation and normalization processing are performed on the echo signal to generate a pre-processed echo signal.

[0008] In this solution, multi-domain feature extraction is performed from the preprocessed echo signal, and feature selection is performed using multi-objective optimization combined with feature importance evaluation to screen key features. Specifically, The time domain features, frequency domain features, time-frequency domain features, and absorption features are extracted from the preprocessed echo signal and regarded as independent modes. Each independent mode corresponds to a subspace, and a multi-objective optimization function is constructed with the goals of maximizing classification accuracy, minimizing feature dimensions, and maximizing inter-modal complementarity. Obtain the intra-modal importance of each feature separately, perform kernel density estimation on all features, calculate the normalized mutual information between each pair of features, generate a symmetric mutual information matrix, use the symmetric mutual information matrix as the weighted directed graph adjacency matrix, iteratively calculate the PageRank value until convergence, and obtain the global influence of each feature as the cross-modal importance; Normalize the intra-modal importance of each feature, and weight the normalized intra-modal importance and cross-modal importance to obtain the final importance; The UMAP dimensionality reduction method is used to map the time domain features, frequency domain features, time-frequency domain features, and absorption features to a low-dimensional manifold space. Latin hypercube sampling is performed in the low-dimensional manifold space, and features are selected according to the final importance and a preset ratio to be forcibly retained in the initial individuals to generate an initial population. Adaptive mutation is performed on the subspace of each independent modality, and binomial crossover is performed by calculating t-SNE similarity. Features with high manifold similarity are preferentially retained. Based on the NSGA-II framework, a modality diversity index is introduced, and the Pareto frontier solution is retained through elite archiving. The top 3 important features of each modality are retained as key features.

[0009] In this solution, the key features are mapped to generate a gas feature distribution map, and the feature clusters are divided by clustering to generate a gas profile of the potential gas, specifically: Standardize the key features and perform manifold learning dimensionality reduction to generate a gas feature distribution map. Perform density clustering based on the gas feature distribution map, calculate the distance from each feature point in the gas feature distribution map to its kth nearest neighbor, draw a sorted distance curve, and select the value corresponding to the inflection point as the neighborhood radius. When the number of feature points within the neighborhood radius of a feature point in the gas feature distribution map is greater than a preset threshold, the feature point is considered a core point. Starting from the core point, recursively merge the density-reachable feature points to form a cluster. Adjacent clusters with a spacing less than a preset distance are merged, and the boundary points are assigned to the cluster closest to the core point. Obtain the feature clusters divided by clustering, calculate the center and dispersion in the feature clusters to characterize the gas fingerprint characteristics, and use feature contribution decomposition to reversely map to the original feature space corresponding to the key features, find the top n features that have the greatest impact on the coordinates in the gas feature distribution map, and generate a gas profile of the potential gas based on the gas fingerprint characteristics and the top n features obtained by feature contribution decomposition.

[0010] In this solution, the gas category of interest is obtained based on the gas profile, specifically: Characteristically encoding the gas image, calculating the degree of matching between the encoded gas image and the standard gas characteristics, retaining gases with a matching degree greater than a preset matching degree threshold as candidate gases based on the matching degree, and obtaining the image feature likelihood of the candidate gas; Based on the current scenario, the initial existence probability of the candidate gas is obtained in the industry-gas association knowledge base as the scenario prior probability. A Bayesian probability model is constructed through the image feature likelihood of the candidate gas and the scenario prior probability to obtain the probability of the candidate gas existing under the gas image, and the gas category of interest is obtained based on the probability.

[0011] In this solution, a graph structure is constructed based on the gas profile and gas feature similarity, and a graph convolutional network is used to learn the representation of the graph structure. The graph convolutional network is enhanced using the gas category of interest to construct a gas category branch, specifically: Nodes are constructed based on the gas profile of the gas sample, the cosine similarity between nodes is calculated through the k-nearest neighbor graph, the top-k highest similarity edges are retained for each node, and a graph structure is constructed based on the nodes and edges; A graph convolutional network is used to pass messages on the graph structure. A graph attention mechanism is added after each layer of graph convolution to enhance the weighted aggregation of neighbor features through attention and extract common gas features. For each gas category of interest, a class-specific graph convolution layer is independently connected to construct a learnable query vector for each gas category of interest, calculate the correlation between the universal gas features and the query vector, and generate class-aware node representations as class-enhanced features; In the gas category branch, the general gas features and category enhanced features are weighted and summed to obtain the final node representation, the distance between the final node representation and each category of gas samples is calculated, the category corresponding to the minimum distance is selected, and the gas recognition result is output.

[0012] In this solution, a multi-layer perceptron is used to construct the gas concentration branch. The final node representation output by the graph convolutional network is used as input, and linear regression is used to output the concentration of each gas. Specifically, Obtain the final node representation of the graph convolutional network output and perform feature normalization processing. Construct a gas concentration branch based on a multi-layer perceptron. Use the Swish activation function to make predictions in the gas concentration branch and add residual connections between hidden layers. The standardized final node representation is imported into the gas concentration branch to obtain the original concentration prediction, construct a trainable matrix for interference compensation, and correct the original concentration prediction to output the concentration of each category of gas.

[0013] A second aspect of the present invention provides a deep learning-based laser radar multi-gas identification system, which includes: a laser radar transmitting and receiving module, a preprocessing and multi-domain feature extraction module, a feature selection and gas profile construction module, a multi-gas identification module, and a result output and visualization module; The laser radar transmitting and receiving module uses a differential absorption laser radar to transmit lasers of different wavelengths, and receives the echo signal after gas absorption, and dynamically adjusts the laser wavelength to ensure that the absorption spectra of different gases are covered; The preprocessing and multi-domain feature extraction module performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signal, and extracts multi-domain features from the preprocessed echo signal; The feature selection and gas profile construction module uses multi-objective optimization combined with feature importance evaluation to perform feature selection, screen key features, perform feature mapping on the key features to generate a gas feature distribution map, divide the feature clusters by clustering, generate gas profiles of potential gases, and obtain the gas category of interest based on the gas profile; The multi-gas recognition module constructs a graph structure based on gas profiles and gas feature similarities, uses a graph convolutional network to learn representations of the graph structure, and enhances the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category. A multi-layer perceptron is used to construct a gas concentration branch, which uses the final node representation output by the graph convolutional network as input and uses linear regression to output the concentration of each gas. The result output and visualization module outputs the gas concentration and gas type and performs visual display according to the preset equipment.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] A deep learning-based lidar multi-gas identification method achieves high-precision identification and concentration inversion of multi-component gases by integrating differential absorption lidar technology with an improved graph convolutional network. In terms of identification accuracy, this method effectively overcomes the cross-interference problem caused by overlapping gas absorption lines in traditional differential absorption lidar technology through multi-domain feature extraction and a multi-objective optimization strategy. Regarding interference resistance, the combined wavelet denoising and attention-weighted GCN model significantly improves the system's robustness in complex environments. The cross-interference compensation mechanism learned through residual network learning further reduces concentration inversion bias when multiple gases coexist. Regarding real-time performance, the lightweight network design shortens single-shot detection time, meeting the real-time requirements of industrial process control. Furthermore, the introduction of a transfer learning strategy reduces the model's reliance on labeled data, enabling rapid deployment in new scenarios (such as natural gas pipeline leak monitoring) with only a small number of calibration samples, demonstrating significant generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0017] Figure 1 A flowchart of a deep learning-based lidar multi-gas identification method is shown; Figure 2 A flow chart of obtaining gas categories of interest based on gas profiles is shown; Figure 3 A flow chart showing gas identification through the gas concentration branch of the gas category branch is shown; Figure 4 Shown is the block diagram of the deep learning-based lidar multi-gas identification system. DETAILED DESCRIPTION

[0018] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0019] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0020] Figure 1 A flowchart of the deep learning-based lidar multi-gas identification method is shown.

[0021] like Figure 1 As shown, this embodiment provides a lidar multi-gas identification method based on deep learning, including: S101: Using a differential absorption lidar to emit lasers of different wavelengths and receive echo signals after gas absorption, the echo signals are subjected to noise suppression, differential absorption calculation, and normalization preprocessing. Multi-domain features are extracted from the preprocessed echo signals, and feature selection is performed using multi-objective optimization combined with feature importance assessment to screen key features. S102, performing feature mapping on the key features to generate a gas feature distribution map, dividing the feature clusters by clustering to generate a gas profile of the potential gas, and obtaining a gas category of interest based on the gas profile; S103, constructing a graph structure based on the gas profile and gas feature similarity, using a graph convolutional network to perform representation learning on the graph structure, and enhancing the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category; S104: A multi-layer perceptron is used to construct a gas concentration branch, the final node representation output by the graph convolutional network is used as input, linear regression is used to output the concentration of each gas, and the gas concentration and gas category are visualized.

[0022] It should be noted that temperature, humidity, air pressure, wind speed, and other data are acquired through meteorological sensors. GPS and geographic information systems are used to determine the location of monitoring points and obtain environmental parameters for the target gas identification scenario. Historical pollution data for the area is queried, and common pollutants are analyzed as historical pollution data. Visible light / infrared cameras or sniffing sensors are used to identify potential emission sources as pollution source information, such as chimneys, vehicle exhaust, and pipeline valves. Scenario features are constructed using environmental parameters, historical pollution data, and pollution source information. A pre-built industry-gas association knowledge base is accessed through a relevant knowledge graph, for example, CH4 and C2H6 for petrochemicals and SO2 and CO2 for coal-fired power plants. A lightweight classification model, such as random forest, is used to perform query matching within the industry-gas association knowledge base for these scenario features. The matching results output the gas presence probability, which is then sorted according to the gas presence probability. Gases that meet a preset probability threshold are screened to generate a probabilistic gas list. Based on this probabilistic gas list, the wavelength corresponding to the gas's strongest absorption peak is queried. For example, the wavelength corresponding to the strongest absorption peak of CH4 is 3.3µm. Interference bands from background gases such as water vapor and carbon dioxide are eliminated to generate the initial wavelength.

[0023] The detection confidence level of each gas in the probabilistic gas list is obtained. If the detection confidence level remains below a preset confidence threshold, the auxiliary absorption peak wavelength of the gas is increased. If an unexpected absorption peak in the probabilistic gas list is detected, a full-band scan is initiated, and the unexpected absorption peak is located using fast Fourier transform (FFT), dynamically adjusting the initial wavelength. Reinforcement learning is used to construct a reward model for wavelength selection and detection accuracy. A detection priority is generated based on the gas's riskiness. An optimal adjustment strategy is generated based on the reward model and detection priority, and the wavelength range is adaptively optimized using this optimal adjustment strategy. A tunable laser, such as a quantum cascade laser or an external cavity diode laser, is used to emit laser light at the adaptively optimized wavelength. An acousto-optic modulator or micro-electromechanical system (MEMS) is used to achieve rapid multi-wavelength polling, ensuring coverage of the absorption lines of different gases and improving detection flexibility. A photodetector receives the echo signal after gas absorption, converts it into an electrical signal, and records the time-resolved intensity data. The echo signal is subjected to noise suppression, differential absorption calculation, and normalization to generate a preprocessed echo signal.

[0024] It should be noted that the time domain features, frequency domain features, time-frequency domain features, and absorption features extracted from the preprocessed echo signal are considered independent modes. Time domain features include signal mean, variance, peak value, and zero-crossing rate. Frequency domain features include fast Fourier transform (FT) to calculate power spectral density, extract the main frequency component, harmonic energy ratio, etc. Time domain features use short-time Fourier transform (SFT) or continuous wavelet transform (CWT) to generate time-frequency spectrograms and extract time-frequency energy distribution characteristics. Absorption features calculate the apparent absorption coefficient at each wavelength based on the Beer-Lambert law. Each independent mode corresponds to a subspace, and a multi-objective optimization function is constructed with the goals of maximizing classification accuracy, minimizing feature dimensionality, and maximizing intermodal complementarity. Maximizing intermodal complementarity is assessed using modal correlation entropy.

[0025] Preferably, the intra-modal importance of the time domain features is obtained based on the Gini importance of random forest, the intra-modal importance of the frequency domain features is obtained by calculating the energy proportion and spectral entropy of each frequency band based on the joint evaluation of energy and entropy, the intra-modal importance of the time-frequency domain features is obtained based on the wavelet coefficient significance test, the absorption peak fitting residual normalization is obtained based on the Voigt linear fit goodness of fit, the absorption cross section is importance weighted, and the importance of the absorption domain features is characterized. In addition, kernel density estimation is performed on all features, the normalized mutual information between each feature is calculated, and a symmetric mutual information matrix is ​​generated. The symmetric mutual information matrix is ​​used as a weighted directed graph adjacency matrix, and a transition probability matrix is ​​defined. The PageRank value is iteratively calculated until convergence, and the global influence of each feature is obtained as the cross-modal importance; the intra-modal importance of each feature is normalized, and the normalized intra-modal importance and cross-modal importance are weighted to obtain the final importance. PageRank global influence The calculation formula for the evaluation is expressed as: ; in represents the damping factor, represents the total number of features, 、 Representation characteristics In the sequence The global influence at the iteration, Representation characteristics The in-degree neighbor set of Indicates that from the feature To Features The transition probability.

[0026] UMAP (Uniform Manifold Approximation and Projection) is used to map time domain features, frequency domain features, time-frequency domain features, and absorption features to a low-dimensional manifold space. UMAP is also used to map high-dimensional features to a low-dimensional manifold space, preserving the topological structure to guide population initialization. Latin hypercube sampling is performed in the low-dimensional manifold space, and features are selected according to the final importance and the preset ratio to be forcibly retained in the initial individuals to generate the initial population; the manifold-guided population initialization speeds up the convergence rate. DE / rand / 1 mutation is performed adaptively on the subspace of each independent modality, and the mutated k-th independent modality subspace is the first individual in the k-th independent modality subspace. The feature vector of the feature Expressed as: ,in Indicates the mode number, Adaptive adjustment according to modal weight, 、 、 Indicates the randomly selected , , Features. And by calculating the t-SNE similarity to perform binomial crossover, high manifold similarity features are retained first. Based on the NSGA-II framework, the modal diversity index MMI is introduced for sorting. , represents the mean vector of the kth modal features. Elite archiving is used to preserve the Pareto frontier solution, retaining the top three most important features of each modal as key features. This approach uses cross-modal importance assessment to improve gas discrimination based on the combined characteristics of time-frequency energy distribution and absorption line shape. The final feature subset includes physically interpretable features such as absorption peak asymmetry, consistent with spectroscopy principles.

[0027] It should be noted that the key features are standardized to eliminate the dimensionality effect, and a low-dimensional coordinate matrix is ​​obtained through manifold learning dimensionality reduction (UMAP dimensionality reduction) to generate a gas feature distribution map. Density clustering is performed based on the gas feature distribution map, and the distance from each feature point to its kth nearest neighbor in the gas feature distribution map is calculated. The sorted distance curve is drawn, and the value corresponding to the inflection point is selected as the neighborhood radius. When the number of feature points contained in the neighborhood radius of the feature point in the gas feature distribution map is greater than the preset threshold, the feature point is regarded as a core point. Starting from the core point, the density-reachable feature points are recursively merged to form a cluster, and the adjacent clusters with a spacing less than the preset distance value are merged, and the boundary points are assigned to the cluster to which the nearest core point belongs. The clustered feature clusters are obtained, and the center and discreteness are calculated in the feature clusters to characterize the gas fingerprint features. The Jacobian matrix analysis is used to perform feature contribution decomposition and reversely map it to the original feature space corresponding to the key features. The top n features that have the greatest impact on the coordinates in the gas feature distribution map are found, and the top n features obtained based on the feature contribution decomposition are used as core features. The gas fingerprint features are combined with the core features to generate a gas portrait of the potential gas.

[0028] Figure 2 A flow chart for obtaining gas categories of interest based on gas profiles is shown.

[0029] According to an embodiment of the present invention, the gas category of interest is obtained based on the gas profile, specifically: S201, encoding the gas image, calculating the matching degree between the encoded gas image and the standard gas characteristics, retaining gases with a matching degree greater than a preset matching degree threshold as candidate gases based on the matching degree, and obtaining the image feature likelihood of the candidate gas; S202, based on the current scenario, obtaining an initial probability of existence of a candidate gas in the industry-gas association knowledge base as a scenario prior probability, and constructing a Bayesian probability model based on the likelihood of the candidate gas image features and the scenario prior probability; S203 , obtaining the probability of the candidate gas existing in the gas image based on the Bayesian probability model, and obtaining the gas category of interest based on the probability.

[0030] It should be noted that the gas profile after calculation is coded With standard gas characteristics The matching degree, the matching degree The calculation is expressed as: Preferably, the standard gas characteristics are extracted from the HITRAN database, including parameters such as absorption line position, line intensity, half-height width, broadening coefficient, and energy level information. A Bayesian probability model is constructed based on the image feature likelihood of the candidate gas and the scene prior probability. The Bayesian probability model is expressed as: ;

[0031] in represents the probability of the candidate gas existing under the gas profile, Candidate gas Gas profile observed The probability density of , preferably modeled by a Gaussian mixture model, Represents candidate gases based on the industry-gas association knowledge base Initial existence probability.

[0032] Figure 3 A flow chart of gas identification through the gas concentration branch of the gas category branch is shown.

[0033] According to an embodiment of the present invention, gas identification is performed using the gas concentration branch of the gas category branch, specifically: S301, constructing nodes based on the gas profile of the gas sample, calculating the cosine similarity between the nodes through a k-nearest neighbor graph, retaining the top-k highest similarity edges for each node, and constructing a graph structure based on the nodes and edges; S302, using a graph convolutional network to perform message passing on the graph structure, adding a graph attention mechanism after each layer of graph convolution, and extracting common gas features by enhancing weighted aggregation of neighbor features through attention; S303: For each gas category of interest, a category-specific graph convolution layer is independently connected to construct a learnable query vector for each gas category of interest, calculate the correlation between the universal gas feature and the query vector, and generate category-aware node representations as category-enhanced features; S304: In the gas classification branch, a final node representation is obtained by weighted summing of the general gas features and the category-enhanced features, and the distance between the final node representation and gas samples of each category is calculated. The category corresponding to the minimum distance is selected, and the gas recognition result is output. S305, obtaining the final node representation output by the graph convolutional network and performing feature normalization processing, constructing a gas concentration branch based on a multi-layer perceptron, using a Swish activation function to perform prediction in the gas concentration branch and adding residual connections between hidden layers; S306: The standardized final node representation is imported into the gas concentration branch to obtain the original concentration prediction, construct a trainable matrix for interference compensation, and correct the original concentration prediction to output the concentration of each type of gas.

[0034] It is important to note that in the gas classification branch, three basic GCN layers and a graph attention mechanism serve as a shared backbone to extract common gas features. Each gas category of interest is independently connected to a category-specific GCN layer to construct a category branch. This category branch can be trained with only a small number of labeled samples. A category attention mechanism is introduced to generate category-aware node representations, highlighting the key features of the target gas. Branch fusion generates the final node representation by taking the weighted sum of shared and category features. In the gas concentration branch, a Swish activation function is used to adapt to the continuous-valued output characteristics of concentration regression. Skip connections are added between hidden layers to mitigate gradient vanishing. An interference matrix is ​​added to physically correspond to the degree of overlap in gas absorption spectra. The model is fine-tuned with newly acquired data to adapt to environmental changes, such as seasonal temperature and humidity. Gas categories and concentrations are visualized, triggering alerts when thresholds are exceeded, such as in industrial leak monitoring. Historical detection data is recorded to enable pollution source tracing and gas pollution trend analysis. For example, in VOCs leakage monitoring at refineries, lidar signals containing a mixture of benzene, toluene, and xylene are obtained. Universal features identify the common characteristics of aromatic hydrocarbons, and category branches focus on the 7.8µm absorption peak of benzene and the 9.5µm shoulder peak of toluene. The distance between benzene and toluene is determined by the Mahalanobis distance, and an accurate alarm for benzene leakage is determined based on the distance and benzene concentration, eliminating interference from xylene.

[0035] Figure 4 Shown is the block diagram of the deep learning-based lidar multi-gas identification system.

[0036] A second aspect of the present invention provides a deep learning-based laser radar multi-gas identification system 4, which includes: a laser radar transmitting and receiving module 401, a preprocessing and multi-domain feature extraction module 402, a feature selection and gas profile construction module 403, a multi-gas identification module 404, and a result output and visualization module 405; The laser radar transmitting and receiving module uses a differential absorption laser radar to transmit lasers of different wavelengths, and receives the echo signal after gas absorption, and dynamically adjusts the laser wavelength to ensure that the absorption spectra of different gases are covered; The preprocessing and multi-domain feature extraction module performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signal, and extracts multi-domain features from the preprocessed echo signal; The feature selection and gas profile construction module uses multi-objective optimization combined with feature importance evaluation to perform feature selection, screen key features, perform feature mapping on the key features to generate a gas feature distribution map, divide the feature clusters by clustering, generate gas profiles of potential gases, and obtain the gas category of interest based on the gas profile; The multi-gas recognition module constructs a graph structure based on gas profiles and gas feature similarities, uses a graph convolutional network to learn representations of the graph structure, and enhances the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category. A multi-layer perceptron is used to construct a gas concentration branch, which uses the final node representation output by the graph convolutional network as input and uses linear regression to output the concentration of each gas. The result output and visualization module outputs the gas concentration and gas type and performs visual display according to the preset equipment.

[0037] A third aspect of the present invention provides a computer-readable storage medium, which includes a deep learning-based laser radar multi-gas identification method program. When the deep learning-based laser radar multi-gas identification method program is executed by a processor, it implements the steps of the deep learning-based laser radar multi-gas identification method.

[0038] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. In addition, the functional units in the various embodiments of the present invention can all be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the above-mentioned integrated units can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0039] Those skilled in the art will understand that all or part of the steps of the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0040] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A deep learning-based lidar multi-gas identification method, characterized in that: The following steps are involved: A differential absorption lidar is used to emit lasers of different wavelengths and receive echo signals after gas absorption. The echo signals are subjected to noise suppression, differential absorption calculation, and normalization preprocessing. Multi-domain features are extracted from the preprocessed echo signals. Feature selection is performed using multi-objective optimization combined with feature importance assessment to screen key features. Performing feature mapping on the key features to generate a gas feature distribution map, dividing the feature clusters by clustering, generating a gas profile of the potential gas, and obtaining the gas category of interest based on the gas profile; Constructing a graph structure based on the gas profile and gas feature similarity, using a graph convolutional network to perform representation learning on the graph structure, and enhancing the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category; A multi-layer perceptron is used to construct a gas concentration branch. The final node representation output by the graph convolutional network is used as input. Linear regression is used to output the concentration of each gas, and the gas concentration and gas category are visualized.

2. The deep learning-based laser radar multi-gas identification method according to claim 1, characterized in that: A differential absorption lidar is used to emit lasers of different wavelengths and receive echo signals after gas absorption. The echo signals are subjected to noise suppression, differential absorption calculation, and normalization preprocessing, specifically: Obtain environmental parameters, historical pollution data, and pollution emission source information for the target gas identification scenario, construct scenario features, and use the scenario features to perform query matching in the industry-gas association knowledge base using a lightweight classification model; Output the gas presence probability through the matching results, sort the gases according to the gas presence probability, filter the gases that meet the preset probability threshold to generate a probabilistic gas list, query the wavelength corresponding to the strongest absorption peak of the gas based on the probabilistic gas list, and exclude the interference band of the background gas to generate the initial wavelength; Obtaining the detection confidence of a gas in the probabilistic gas list. When the detection confidence continues to be lower than a preset confidence threshold, increasing the auxiliary absorption peak wavelength of the gas. If an unexpected absorption peak is detected in the probabilistic gas list, initiating a full-band scan, locating the unexpected absorption peak through fast Fourier transform, and dynamically adjusting the initial wavelength. Reinforcement learning is introduced to build a reward model for wavelength selection and detection accuracy, and detection priorities are generated based on gas hazard. An optimal adjustment strategy is generated based on the reward model and detection priority, and the wavelength range is adaptively optimized using the optimal adjustment strategy. A tunable laser is used to emit laser light with an adaptively optimized wavelength, and a photodetector is used to receive the echo signal after gas absorption. Time-resolved intensity data is recorded, and noise suppression, differential absorption calculation and normalization processing are performed on the echo signal to generate a pre-processed echo signal.

3. The deep learning-based laser radar multi-gas identification method according to claim 1, characterized in that: Multi-domain feature extraction is performed from the preprocessed echo signal. Feature selection is performed using multi-objective optimization combined with feature importance evaluation to screen key features. Specifically: The time domain features, frequency domain features, time-frequency domain features, and absorption features are extracted from the preprocessed echo signal and regarded as independent modes. Each independent mode corresponds to a subspace. The multi-objective optimization function is constructed with the goals of maximizing classification accuracy, minimizing feature dimensions, and maximizing complementarity between modes. Obtain the intra-modal importance of each feature separately, perform kernel density estimation on all features, calculate the normalized mutual information between each pair of features, generate a symmetric mutual information matrix, use the symmetric mutual information matrix as the weighted directed graph adjacency matrix, iteratively calculate the PageRank value until convergence, and obtain the global influence of each feature as the cross-modal importance; Normalize the intra-modal importance of each feature, and weight the normalized intra-modal importance and cross-modal importance to obtain the final importance; The UMAP dimensionality reduction method is used to map the time domain features, frequency domain features, time-frequency domain features, and absorption features to a low-dimensional manifold space. Latin hypercube sampling is performed in the low-dimensional manifold space, and features are selected according to the final importance and a preset ratio to be forcibly retained in the initial individuals to generate an initial population. Adaptive mutation is performed on the subspace of each independent modality, and binomial crossover is performed by calculating t-SNE similarity. Features with high manifold similarity are preferentially retained. Based on the NSGA-II framework, a modality diversity index is introduced, and the Pareto frontier solution is retained through elite archiving. The top 3 important features of each modality are retained as key features.

4. The deep learning-based laser radar multi-gas identification method according to claim 1, characterized in that: The key features are mapped to generate a gas feature distribution map, and the feature clusters are divided by clustering to generate a gas profile of the potential gas, specifically: Standardize the key features and perform manifold learning dimensionality reduction to generate a gas feature distribution map. Perform density clustering based on the gas feature distribution map, calculate the distance from each feature point in the gas feature distribution map to its kth nearest neighbor, draw a sorted distance curve, and select the value corresponding to the inflection point as the neighborhood radius. When the number of feature points within the neighborhood radius of a feature point in the gas feature distribution map is greater than a preset threshold, the feature point is considered a core point. Starting from the core point, recursively merge the density-reachable feature points to form a cluster. Adjacent clusters with a spacing less than a preset distance are merged, and the boundary points are assigned to the cluster closest to the core point. Obtain the feature clusters divided by clustering, calculate the center and dispersion in the feature clusters to characterize the gas fingerprint characteristics, and use feature contribution decomposition to reversely map to the original feature space corresponding to the key features, find the top n features that have the greatest impact on the coordinates in the gas feature distribution map, and generate a gas profile of the potential gas based on the gas fingerprint characteristics and the top n features obtained by feature contribution decomposition.

5. The deep learning-based laser radar multi-gas identification method according to claim 1, characterized in that: The gas category of interest is obtained based on the gas profile, specifically: Characteristically encoding the gas image, calculating the degree of matching between the encoded gas image and the standard gas characteristics, retaining gases with a matching degree greater than a preset matching degree threshold as candidate gases based on the matching degree, and obtaining the image feature likelihood of the candidate gas; Based on the current scenario, the initial existence probability of the candidate gas is obtained in the industry-gas association knowledge base as the scenario prior probability. A Bayesian probability model is constructed through the image feature likelihood of the candidate gas and the scenario prior probability to obtain the probability of the candidate gas existing under the gas image, and the gas category of interest is obtained based on the probability.

6. The deep learning-based laser radar multi-gas identification method according to claim 1, characterized in that: A graph structure is constructed based on the gas profile and gas feature similarity. A graph convolutional network is used to learn the representation of the graph structure. The graph convolutional network is enhanced using the gas category of interest to construct a gas category branch. Specifically, Nodes are constructed based on the gas profile of the gas sample, the cosine similarity between nodes is calculated through the k-nearest neighbor graph, the top-k edges with the highest similarity are retained for each node, and a graph structure is constructed based on the nodes and edges; A graph convolutional network is used to pass messages on the graph structure. A graph attention mechanism is added after each layer of graph convolution to enhance the weighted aggregation of neighbor features through attention and extract common gas features. For each gas category of interest, a class-specific graph convolution layer is independently connected to construct a learnable query vector for each gas category of interest, calculate the correlation between the universal gas features and the query vector, and generate class-aware node representations as class-enhanced features; In the gas category branch, the general gas features and category enhanced features are weighted and summed to obtain the final node representation, the distance between the final node representation and each category of gas samples is calculated, the category corresponding to the minimum distance is selected, and the gas recognition result is output.

7. The deep learning-based laser radar multi-gas identification method according to claim 6, characterized in that: A multi-layer perceptron is used to construct the gas concentration branch. The final node representation output by the graph convolutional network is used as input, and linear regression is used to output the concentration of each gas. Specifically: Obtain the final node representation of the graph convolutional network output and perform feature normalization processing. Construct a gas concentration branch based on a multi-layer perceptron. Use the Swish activation function to make predictions in the gas concentration branch and add residual connections between hidden layers. The standardized final node representation is imported into the gas concentration branch to obtain the original concentration prediction, construct a trainable matrix for interference compensation, and correct the original concentration prediction to output the concentration of each category of gas.

8. A deep learning-based laser radar multi-gas identification system, characterized in that: Implementing the deep learning-based laser radar multi-gas identification method according to any one of claims 1 to 7, the system includes a laser radar transmitting and receiving module, a preprocessing and multi-domain feature extraction module, a feature selection and gas profile construction module, a multi-gas identification module, and a result output and visualization module; The laser radar transmitting and receiving module uses a differential absorption laser radar to transmit lasers of different wavelengths, and receives the echo signal after gas absorption, and dynamically adjusts the laser wavelength to ensure that the absorption spectra of different gases are covered; The preprocessing and multi-domain feature extraction module performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signal, and extracts multi-domain features from the preprocessed echo signal; The feature selection and gas profile construction module uses multi-objective optimization combined with feature importance evaluation to perform feature selection, screen key features, perform feature mapping on the key features to generate a gas feature distribution map, divide the feature clusters by clustering, generate gas profiles of potential gases, and obtain the gas category of interest based on the gas profile; The multi-gas recognition module constructs a graph structure based on gas profiles and gas feature similarities, uses a graph convolutional network to learn representations of the graph structure, and enhances the graph convolutional network using the gas category of interest to construct a gas category branch and output the gas category. A multi-layer perceptron is used to construct a gas concentration branch, which uses the final node representation output by the graph convolutional network as input and uses linear regression to output the concentration of each gas. The result output and visualization module outputs the gas concentration and gas type and performs visual display according to the preset equipment.

Images