Lidar multi-gas recognition method and system based on deep learning

Abstract

The application discloses a laser radar multi-gas identification method and system based on deep learning, and relates to the technical field of laser radar, and comprises the following steps: a differential absorption laser radar emits laser beams with different wavelengths, receives echo signals after gas absorption, extracts multi-domain features, and adopts multi-objective optimization combined with feature importance evaluation to screen key features; the key features are used to generate a gas image of potential gas, and a gas category of interest is obtained; a graph structure is constructed according to the gas image and gas feature similarity, representation learning of the graph structure is performed by using a graph convolution network, the graph convolution network is enhanced by using the gas category of interest, a gas category branch is constructed, and a gas category is output; a multi-layer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, and linear regression is used to output the concentrations of various gases. The application improves the distinguishing ability of small difference gases, and realizes high-precision identification and concentration inversion of multi-component gases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser radar, more particularly, to a laser radar multi-gas identification method and system based on deep learning. BACKGROUND

[0002] With the acceleration of industrialization and urbanization, air pollution problems are increasingly prominent, especially the emission of multi-component harmful gases poses a serious threat to the environment and human health. Therefore, developing efficient and accurate gas detection technology 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 technology, although have high accuracy, often have limitations such as large equipment size, slow response speed, high cost or difficulty in realizing real-time online monitoring.

[0003] Laser radar technology has great potential in the field of gas detection due to its non-contact, high sensitivity, high spatial and temporal resolution, etc. Differential absorption laser radar can realize quantitative detection of specific gases by analyzing the difference in return signal of laser at different wavelengths on the gas absorption spectrum. However, traditional differential absorption laser radar technology faces challenges in multi-component gas identification, especially in complex mixed gas environment, the absorption spectrum of different gases may overlap, causing cross interference and affecting detection accuracy.

[0004] In recent years, deep learning technology has made breakthroughs in computer vision, speech recognition and natural language processing, etc. Its powerful feature extraction and pattern recognition ability provides a new solution for multi-gas detection. Deep learning model can automatically learn gas features from high-dimensional, nonlinear laser radar spectral data, effectively overcoming the limitations of traditional algorithms. Therefore, how to combine deep learning and laser radar technology to develop efficient and high-precision multi-gas identification method is a technical problem to be solved in the field of environmental monitoring and gas sensing. SUMMARY

[0005] To solve the above technical problems, the present application provides a laser radar multi-gas identification method and system based on deep learning, which realizes high-precision identification and concentration inversion of multi-component gases by improving the discrimination ability of small difference gases.

[0006] The first aspect of the present application provides a laser radar multi-gas identification method based on deep learning, comprising the following steps:

[0007] The differential absorption laser radar emits laser beams of different wavelengths, receives echo signals after gas absorption, performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signals, extracts multi-domain features from the preprocessed echo signals, selects key features by using multi-objective optimization combined with feature importance evaluation, and filters the key features;

[0008] The key features are mapped to generate a gas feature distribution map, the feature clusters are divided by clustering, a gas image of potential gas is generated, and a gas category of interest is obtained based on the gas image;

[0009] A graph structure is constructed according to the gas image and gas feature similarity, representation learning is performed on the graph structure using a graph convolution network, the graph convolution network is enhanced using the gas category of interest, a gas category branch is constructed, and a gas category is output;

[0010] A multilayer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, linear regression is used to output the concentration of each gas, and the gas concentration and gas category are visually displayed.

[0011] In the scheme, the differential absorption laser radar emits laser beams of different wavelengths, receives echo signals after gas absorption, and performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signals, specifically:

[0012] The environmental parameters, historical pollution data and pollution emission source information of the target gas recognition scene are obtained, the scene features are constructed, and the scene features are used to query and match in the industry-gas association knowledge base using a lightweight classification model;

[0013] The gas existence probability is output through the matching result, the gas existence probability is sorted, the gas generation probability list that meets the preset probability threshold is screened, the wavelength corresponding to the strongest absorption peak of the gas is queried based on the probability gas list, and the initial wavelength is generated by excluding the interference waveband of the background gas;

[0014] The detection confidence of the gas in the probability gas list is obtained, and when the detection confidence continuously falls below the preset confidence threshold, the auxiliary absorption peak wavelength of the gas is increased, if an unexpected absorption peak in the probability gas list is detected, full-waveband scanning is started, the unexpected absorption peak is located through fast Fourier transform, and the initial wavelength is dynamically adjusted;

[0015] A reinforcement learning is introduced to construct a reward model of wavelength selection and detection accuracy, a detection priority is generated based on the gas danger, an optimal adjustment strategy is generated based on the reward model and the detection priority, and the wavelength range is adaptively optimized using the optimal adjustment strategy;

[0016] The adaptive optimized wavelength laser is emitted by using a tunable laser, and the echo signal after gas absorption is received by an optical detector, time-resolved intensity data is recorded, the echo signal is subjected to noise suppression, differential absorption calculation and normalization processing, and a preprocessed echo signal is generated.

[0017] In the scheme, multi-domain feature extraction is performed from the preprocessed echo signal, multi-objective optimization is combined with feature importance evaluation for feature selection, and key features are screened, specifically:

[0018] The time domain features, frequency domain features, time-frequency domain features and absorption features extracted from the preprocessed echo signal are regarded as independent modalities, each independent modality corresponds to a subspace, and a multi-objective optimization function is constructed by maximizing the classification accuracy, minimizing the feature dimension and maximizing the complementarity between modalities as the target;

[0019] The intra-modal importance of each feature is obtained, and the kernel density estimation of all features is performed to calculate the normalized mutual information between each pair of features to generate a symmetric mutual information matrix, the symmetric mutual information matrix is used as the adjacency matrix of the weighted directed graph, and the PageRank value is iteratively calculated until convergence to obtain the global influence of each feature as the cross-modal importance;

[0020] 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;

[0021] The UMAP dimension 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 according to the final importance, a predetermined proportion of features are selected for forced retention to generate an initial population;

[0022] The subspace of each independent modality is subjected to adaptive mutation, and binomial crossover is performed by calculating the t-SNE similarity, high manifold similarity features are preferentially retained, a modal diversity index is introduced based on the NSGA-II framework, Pareto frontier solutions are retained by elite archiving, and the top 3 important features of each modality are retained as key features.

[0023] In the scheme, the key features are mapped to generate a gas feature distribution map, the features are clustered to generate a gas image of potential gas, specifically:

[0024] The key features are standardized and subjected to manifold learning dimension reduction to generate a gas feature distribution map, density clustering is performed based on the gas feature distribution map, the distance from each feature point in the gas feature distribution map to its k-th neighbor is calculated, the sorted distance curve is plotted, and the inflection point corresponding value is selected as the neighborhood radius.

[0025] When the number of feature points contained within the neighborhood radius of a feature point in the gas feature distribution map is greater than a preset threshold, the feature point is regarded as the core point. Starting from the core point, feature points with reachable density are recursively merged to form clusters. Adjacent clusters with a spacing less than a preset distance value are merged, and the boundary points are assigned to the cluster to which the nearest core point belongs.

[0026] The feature clusters of the clusters are obtained, and the center and dispersion of the feature clusters are calculated to represent the gas fingerprint features. The feature contribution decomposition is used to back-map to the original feature space corresponding to the key features. The top n features that have the greatest impact on the coordinates of the gas feature distribution map are found. Based on the gas fingerprint features and the top n features obtained by feature contribution decomposition, a gas profile of the potential gas is generated.

[0027] In this solution, the gas category of interest is obtained based on the gas profile, specifically as follows:

[0028] The gas profile is feature-encoded, the matching degree between the encoded gas profile and the standard gas features is calculated, and gases with a matching degree greater than a preset matching degree threshold are retained as candidate gases, and the profile feature likelihood of the candidate gases is obtained.

[0029] Based on the current scenario, the initial existence probability of candidate gases is obtained from the industry-gas association knowledge base and used as the scenario prior probability. A Bayesian probability model is constructed using the likelihood of the candidate gas profile features and the scenario prior probability to obtain the probability of the candidate gas existing under the gas profile. Based on the probability, the category of gas of interest is obtained.

[0030] In this scheme, 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, and the graph convolutional network is enhanced using the gas category of interest to construct gas category branches. Specifically:

[0031] Nodes are constructed based on the gas profile of the gas sample. The cosine similarity between nodes is calculated using the k-nearest neighbor graph. The top-k edges with the highest similarity are retained for each node. A graph structure is constructed based on the nodes and edges.

[0032] The graph structure is message-passed using a graph convolutional network, and a graph attention mechanism is added after each graph convolutional layer. Neighbor features are aggregated through attention enhancement weighting to extract general gas features.

[0033] For each gas category of interest, a dedicated convolutional layer with a category-specific graph is independently connected. A learnable query vector is constructed for each gas category of interest. The correlation between general gas features and the query vector is calculated, and category-aware node representations are generated as category-enhancing features.

[0034] In the gas category branch, the general gas features and category enhanced features are weighted and summed to obtain a final node representation, the distance between the final node representation and each category gas sample is calculated, the category corresponding to the minimum distance is selected, and a gas recognition result is output.

[0035] In the scheme, a multi-layer perception mechanism is used to build a gas concentration branch, the final node representation output by the graph convolution network is taken as input, and linear regression is used to output the concentration of each gas.

[0036] The final node representation output by the graph convolution network is standardized, a gas concentration branch is built based on a multi-layer perception mechanism, a Swish activation function is used for prediction in the gas concentration branch, and residual connections are added between hidden layers.

[0037] The standardized final node representation is imported into the gas concentration branch to obtain an original concentration prediction, a trainable matrix is constructed to compensate for interference, and the original concentration prediction is corrected to output the concentration of each category gas.

[0038] The second aspect of the application provides a laser radar multi-gas identification system based on deep learning, which comprises: a laser radar transmitting and receiving module, a preprocessing and multi-domain feature extraction module, a feature selection and gas image construction module, a multi-gas identification module, a result output and visualization module.

[0039] The laser radar transmitting and receiving module transmits different wavelength lasers using differential absorption laser radar and receives the echo signals after gas absorption, and dynamically adjusts the laser wavelength to ensure coverage of different gas absorption spectral lines.

[0040] 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.

[0041] The feature selection and gas image construction module uses multi-objective optimization combined with feature importance evaluation to select features, filters key features, maps the key features to generate a gas feature distribution map, divides feature clusters through clustering, generates a gas image of potential gas, and obtains the gas category of interest based on the gas image.

[0042] The multi-gas identification module constructs a graph structure according to the gas image and gas feature similarity, uses a graph convolution network to perform representation learning on the graph structure, uses the gas category of interest to enhance the graph convolution network, constructs a gas category branch, and outputs the gas category; a multi-layer perception mechanism is used to build a gas concentration branch, the final node representation output by the graph convolution network is taken as input, and linear regression is used to output the concentration of each gas.

[0043] The result output and visualization module outputs the gas concentration and gas category and visualizes the same according to a preset device.

[0044] Compared with the prior art, the present application has the following advantages:

[0045] The laser radar multi-gas identification method based on deep learning realizes high-precision identification and concentration inversion of multi-component gas by fusing differential absorption laser radar technology and an improved graph convolution network. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or example descriptions will be briefly introduced as follows.

[0047] Figure 1 A flowchart of the laser radar multi-gas identification method based on deep learning is shown.

[0048] Figure 2 A flowchart of obtaining a gas category of interest based on a gas image is shown.

[0049] Figure 3 A flowchart of gas identification by a gas concentration branch of a gas category branch is shown.

[0050] Figure 4 A block diagram of the laser radar multi-gas identification system based on deep learning is shown. DETAILED DESCRIPTION

[0051] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the drawings needed to be used in the embodiments or example descriptions will be briefly introduced as follows.

[0052] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods have not been described in detail in order to avoid obscuring the present application.

[0053] Figure 1 A flowchart of a laser radar multi-gas recognition method based on deep learning is shown.

[0054] As Figure 1 shown, the embodiment provides a laser radar multi-gas recognition method based on deep learning, comprising:

[0055] S101, a differential absorption laser radar emits laser beams of different wavelengths and receives echo signals after gas absorption, performs noise suppression, differential absorption calculation and normalization preprocessing on the echo signals, extracts multi-domain features from the preprocessed echo signals, selects features by using multi-objective optimization combined with feature importance evaluation, and screens key features;

[0056] S102, the key features are mapped to generate a gas feature distribution map, the feature clusters are divided by clustering, the gas image of the potential gas is generated, and the gas class of interest is obtained based on the gas image;

[0057] S103, a graph structure is constructed according to the gas image and gas feature similarity, representation learning of the graph structure is performed using a graph convolution network, the graph convolution network is enhanced using the gas class of interest, a gas class branch is constructed, and the gas class is output;

[0058] S104, a multi-layer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, linear regression is used to output the concentration of each gas, and the gas concentration and gas class are visually displayed.

[0059] It should be noted that the temperature, humidity, air pressure, wind speed and other data are acquired by the meteorological sensor, the environmental parameters of the target gas identification scene are acquired by combining the GPS and geographic information system to determine the position of the monitoring point; the historical pollution data of the region is queried, and the common pollutants are analyzed as the historical pollution data; the potential emission source is identified as the pollution emission source information, such as a chimney, vehicle exhaust, and pipeline valve, by combining the visible light / infrared camera or sniffing sensor. The scene features are constructed using the environmental parameters, historical pollution data, and pollution emission source information. The industry-gas correlation knowledge base is pre-constructed by accessing the related knowledge graph, for example, petroleum and chemical industry corresponds to CH4 and C2H6, and coal-fired power plant corresponds to SO2 and CO2. The random forest and other lightweight classification models are used to query and match in the industry-gas correlation knowledge base of the scene features. The gas existence probability is output through the matching result, the gas existence probability is sorted, the gas generation probability list that meets the preset probability threshold is screened, the wavelength corresponding to the strongest absorption peak of the gas is queried based on the probability gas list, for example, the wavelength corresponding to the strongest absorption peak of CH4 is 3.3 pm, and the initial wavelength is generated by excluding the interference wavelength of background gases such as water vapor and carbon dioxide.

[0060] The detection confidence of the gas in the probability gas list is acquired, and when the detection confidence continuously falls below the preset confidence threshold, the auxiliary absorption peak wavelength of the gas is increased. If an unexpected absorption peak in the probability gas list is detected, a full-waveband scan is started, the unexpected absorption peak is located by fast Fourier transform, and the initial wavelength is dynamically adjusted. A reinforcement learning is introduced to construct a reward model of wavelength selection and detection accuracy, and a detection priority is generated based on the gas danger. The optimal adjustment strategy is generated based on the reward model and the detection priority, and the wavelength range is adaptively optimized using the optimal adjustment strategy. A tunable laser such as a quantum cascade laser or an external cavity diode laser is used to emit laser with the adaptively optimized wavelength, and a multi-wavelength fast polling is realized by an acousto-optic modulator or a micro-electro-mechanical system to ensure that the absorption spectral lines of different gases are covered and the detection flexibility is improved. The echo signal after gas absorption is received by a photodetector, converted into an electrical signal, and the time-resolved intensity data is recorded. The echo signal is processed by noise suppression, differential absorption calculation, and normalization to generate a pre-processed echo signal.

[0061] It should be noted that the time domain features, frequency domain features, time-frequency domain features and absorption features extracted from the pre-processed echo signal are regarded as independent modalities, wherein the time domain features include signal mean, variance, peak value, zero-crossing rate, etc.; the frequency domain features include fast Fourier transform to calculate power spectral density, extract main frequency component, harmonic energy ratio, etc.; the time-frequency domain features use short-time Fourier transform or continuous wavelet transform to generate time-frequency spectrogram and extract time-frequency energy distribution features, etc.; the absorption features are calculated based on Beer-Lambert law to calculate the apparent absorption coefficient at each wavelength. Each independent modality corresponds to a subspace, and a multi-objective optimization function is constructed with the objectives of maximizing classification accuracy, minimizing feature dimension and maximizing inter-modality complementarity; wherein the inter-modality complementarity is evaluated by modality correlation entropy.

[0062] Preferably, the Gini importance based on random forest is used to obtain the intra-modality importance of the time domain features, the energy and entropy joint evaluation is used to calculate the energy proportion and spectral entropy of each frequency band to obtain the intra-modality importance of the frequency domain features, the wavelet coefficient significance test is used to obtain the intra-modality importance of the time-frequency domain features, and the Voigt line fitting goodness is used to obtain the absorption peak fitting residual normalization, and the absorption cross section is weighted for importance to represent the importance of the absorption domain features. In addition, kernel density estimation is performed on all features, the normalized mutual information between each two features is calculated, a symmetric mutual information matrix is generated, the symmetric mutual information matrix is used as a weighted directed graph adjacency matrix, a transition probability matrix is defined, and PageRank values are iteratively calculated until convergence to obtain the global influence of each feature as the cross-modality importance; the intra-modality importance of each feature is normalized, and the normalized intra-modality importance and cross-modality importance are weighted to obtain the final importance. PageRank global influence The calculation formula of the evaluation is:

[0063] ;

[0064] wherein represents a damping factor, represents the total number of features, , represents the feature , the global influence at the first iteration, the global influence at the second iteration, represents the in-degree neighbor set of the feature , represents the transition probability from the feature to the feature .

[0065] The time domain features, frequency domain features, time-frequency domain features and absorption features are mapped to a low-dimensional manifold space by UMAP (Uniform Manifold Approximation and Projection), and the high-dimensional features are mapped to a low-dimensional manifold space by UMAP, and the topological structure is retained to guide the population initialization. Latin hypercube sampling is performed in the low-dimensional manifold space, and the features are forced to be retained to the initial individuals according to the final importance in a preset proportion, to generate an initial population; the manifold-guided population initialization improves the convergence speed. The DE / rand / 1 mutation is adaptively performed on each independent modal subspace, and the feature vector of the kth independent modal subspace after mutation is expressed as: , wherein indicates the modal number, , , indicates the kth modal feature mean vector, and the Pareto front solution is retained by elite archiving, and the Top 3 important features of each modal are retained as key features. Through cross-modal importance evaluation, the gas discrimination is improved according to the combined features of time-frequency energy distribution and absorption line type, and the final feature subset contains features such as absorption peak asymmetry that can be explained by matter, which conforms to the principle of spectroscopy. ,

[0066] ​​​​​It should be noted that the key features are standardized to eliminate the dimensional influence, the low-dimensional coordinate matrix is obtained by manifold learning dimension reduction (UMAP dimension reduction), the gas feature distribution map is generated, the density clustering is performed based on the gas feature distribution map, the distance of each feature point in the gas feature distribution map to its kth neighbor is calculated, the distance curve after sorting is drawn, and the inflection point corresponding value is selected as the neighborhood radius; when the number of feature points contained in the neighborhood radius range of the feature point in the gas feature distribution map is greater than a preset threshold, the feature point is regarded as a core point, the density-reachable feature points are recursively merged to form a cluster from the core point, adjacent clusters with a distance less than a preset distance value are merged, and the boundary points are classified into the cluster to which the nearest core point belongs; the feature clusters obtained by clustering are obtained, the center and dispersion in the feature clusters are calculated to represent the gas fingerprint feature, and the feature contribution degree decomposition is analyzed through the Jacobian matrix to be reversely mapped to the original feature space corresponding to the key features, the first n features with the greatest influence on the coordinates in the gas feature distribution map are found, the first n features obtained based on the feature contribution degree decomposition are taken as core features, and the gas image of the potential gas is generated by combining the gas fingerprint feature and the core features.

[0067] Figure 2 A flowchart for obtaining a gas category of interest based on a gas image is shown.

[0068] According to an embodiment of the present application, the gas category of interest is obtained based on the gas image, specifically:

[0069] S201, feature encoding is performed on the gas image, the matching degree of the encoded gas image and the standard gas feature is calculated, the gas with a matching degree greater than a preset matching degree threshold is retained as a candidate gas based on the matching degree, and the image feature likelihood of the candidate gas is obtained.

[0070] S202, the initial existence probability of the candidate gas is obtained based on the industry-gas association knowledge base in the current scene as the scene prior probability, and a Bayesian probability model is constructed through the image feature likelihood of the candidate gas and the scene prior probability.

[0071] S203, the probability of the existence of the candidate gas under the gas image is obtained based on the Bayesian probability model, and the gas category of interest is obtained based on the probability.

[0072] It should be noted that the matching degree of the encoded gas image and the standard gas feature is calculated, and the calculation of the matching degree is as follows: , preferably, the standard gas features are extracted from the HITRAN database, including absorption line position, line strength, half-width, broadening coefficient, and energy level information, etc. A Bayesian probability model is constructed by the profile feature likelihood of the candidate gas and the scene prior probability, and the Bayesian probability model is expressed as:

[0073] ;

[0074] wherein represents the probability of the existence of the candidate gas under the gas profile, represents the probability density of the observation of the gas profile when the candidate gas exists, preferably modeled by a Gaussian mixture model, represents the initial existence probability of the candidate gas based on the industry-gas association knowledge base.

[0075] Figure 3 A flowchart for gas identification by the gas concentration branch of the gas category branch is shown.

[0076] According to an embodiment of the present application, gas identification is performed by the gas concentration branch of the gas category branch, specifically:

[0077] S301, a node is constructed according to the gas profile of the gas sample, the cosine similarity between nodes is calculated by k-nearest neighbor graph, the edges with the top-k highest similarity are retained for each node, and a graph structure is constructed based on the nodes and edges;

[0078] S302, the graph structure is passed through a message using a graph convolution network, a graph attention mechanism is added after each layer of graph convolution, neighbor features are aggregated by attention enhancement weighting, and general gas features are extracted;

[0079] S303, for the gas category of interest, a 1-layer category-specific graph convolution layer is independently accessed, a learnable query vector is constructed for each gas category of interest, the relevance of the general gas features and the query vector is calculated, and a category-aware node representation is generated as a category-enhanced feature;

[0080] S304, in the gas category branch, the general gas features and the category-enhanced features are weighted and summed to obtain the final node representation, the distance between the final node representation and each category gas sample is calculated, the category corresponding to the minimum distance is selected, and the gas identification result is output;

[0081] S305, the final node representation output by the graph convolution network is obtained for feature standardization processing, a gas concentration branch is constructed based on a multi-layer perception mechanism, a Swish activation function is used for prediction in the gas concentration branch, and residual connections are added between hidden layers; ​​

[0082] S306, import the final node representation after standardization processing into the gas concentration branch, 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.

[0083] It should be noted that in the gas category branch, 3 layers of basic GCN layers and graph attention mechanism are used as a shared backbone to extract general gas features, and each gas category of interest independently accesses 1 layer of category-specific GCN layer to construct a category branch. The category branch can be trained only with a small amount of labeled samples, and a category attention mechanism is introduced to generate a category-aware node representation. The key features of the target gas are highlighted through the attention mechanism, and the weighted sum of the shared features and the category features is obtained through branch fusion to generate the final node representation. In the gas concentration branch, the Swish activation function is used to adapt to the continuous value output characteristics of the concentration regression. The skip connection is added between the hidden layers to alleviate the gradient disappearance. The interference matrix can be physically corresponding to the degree of overlap of the gas absorption spectrum. The model is fine-tuned combined with newly collected data to adapt to environmental changes, such as seasonal temperature and humidity effects. The gas category and concentration are visualized, and when the threshold is exceeded, an early warning is triggered, such as industrial leakage monitoring. The historical detection data is recorded to realize pollution tracing and gas pollution trend analysis. For example, in the VOCs leakage monitoring of a refinery, laser radar signals containing benzene, toluene, and xylene are obtained. The general features of aromatic hydrocarbons are identified, the category branch focuses on the 7.8µm absorption peak of benzene and the 9.5µm shoulder peak of toluene, the Mahalanobis distance is determined for benzene and toluene, and the benzene leakage is accurately alarmed according to the distance and benzene concentration, and the interference of xylene is excluded.

[0084] Figure 4 A block diagram of a deep learning-based laser radar multi-gas recognition system is shown.

[0085] The second aspect of the present application provides a deep learning-based laser radar multi-gas recognition system 4, which comprises: a laser radar transmitting and receiving module 401, a preprocessing and multi-domain feature extraction module 402, a feature selection and gas image construction module 403, a multi-gas recognition module 404, and a result output and visualization module 405.

[0086] The laser radar transmitting and receiving module transmits different wavelengths of laser light using differential absorption laser radar and receives the echo signal after gas absorption, and dynamically adjusts the laser wavelength to ensure coverage of the absorption spectrum of different gases.

[0087] 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.

[0088] The feature selection and gas image construction module adopts multi-objective optimization combined with feature importance evaluation to select features, screens key features, performs feature mapping on the key features to generate a gas feature distribution map, divides feature clusters through clustering, generates a gas image of potential gas, and obtains a gas class of interest based on the gas image;

[0089] The multi-gas recognition module constructs a graph structure according to the gas image and gas feature similarity, uses a graph convolution network to perform representation learning on the graph structure, uses the gas class of interest to enhance the graph convolution network, constructs a gas class branch, and outputs the gas class; a multilayer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, and linear regression is used to output the concentration of each gas;

[0090] The result output and visualization module outputs the gas concentration and gas class, and performs visual display according to a preset device.

[0091] The third aspect of the present application provides a computer readable storage medium, the computer readable storage medium comprises a deep learning based laser radar multi-gas recognition method program, when the deep learning based laser radar multi-gas recognition method program is executed by a processor, the steps of the deep learning based laser radar multi-gas recognition method are realized.

[0092] In several embodiments provided in the present application, it should be understood that the disclosed method and system can be implemented in other ways. The system embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division mode, 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 or direct coupling or communication connection between the various components shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms. In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware or in the form of hardware plus software functional unit.

[0093] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, the foregoing program can be stored in a computer readable storage medium, and the program executes the steps including the above-mentioned method embodiments when executed; and the foregoing storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disc or an optical disc and various storage medium capable of storing program codes.

[0094] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A method for laser radar multi-gas recognition based on deep learning, characterized in that, The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the following steps: The method comprises the According to the gas image and gas feature similarity, a graph structure is constructed, a graph convolution network is used for representation learning of the graph structure, and the graph convolution network is enhanced using a gas class of interest to construct a gas class branch, and output a gas class, comprising: constructing a node according to the gas image of the gas sample, calculating the cosine similarity between 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; using a graph convolution network to perform message passing on the graph structure, adding a graph attention mechanism after each layer of graph convolution, and extracting general gas features by attention-enhanced weighted aggregation of neighbor features; for a gas class of interest, independently access a 1-layer class-specific graph convolution layer, construct a learnable query vector for each gas class of interest, calculate the relevance of the general gas features and the query vector, and generate a class-aware node representation as a class-enhanced feature; in the gas class branch, the general gas features and the class-enhanced features are weighted and summed to obtain the final node representation, the distance between the final node representation and each class gas sample is calculated, the class corresponding to the minimum distance is selected, and the gas recognition result is output; A multi-layer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, and linear regression is used to output the concentration of each gas, and the gas concentration and gas class are displayed visually. 2.The deep learning-based LIDAR multi-gas recognition method of claim 1, wherein, A differential absorption laser radar is used to emit laser beams of different wavelengths and receive echo signals after gas absorption, and the echo signals are subjected to noise suppression, differential absorption calculation and normalization preprocessing, specifically: Obtain the environmental parameters, historical pollution data and pollution emission source information of the target gas identification scene, construct scene features, and use the scene features to query and match in an industry-gas association knowledge base using a lightweight classification model; Output the gas existence probability through the matching result, sort according to the gas existence probability, filter the gas generation probability list that meets the preset probability threshold, query the wavelength corresponding to the strongest absorption peak of the gas based on the probability gas list, and exclude the interference waveband of the background gas to generate an initial wavelength; Obtain the detection confidence of the gas in the probability gas list, and if the detection confidence continuously falls below a preset confidence threshold, add the auxiliary absorption peak wavelength of the gas, and if an unexpected absorption peak is detected in the probability gas list, start full-waveband scanning, locate the unexpected absorption peak through fast Fourier transform, and dynamically adjust the initial wavelength; Introduce reinforcement learning to construct a reward model for wavelength selection and detection accuracy, and generate a detection priority based on the gas danger, generate an optimal adjustment strategy based on the reward model and the detection priority, and use the optimal adjustment strategy to adaptively optimize the wavelength range; A tunable laser is used to emit laser beams of adaptively optimized wavelengths, and an optoelectronic detector is used to receive echo signals after gas absorption, record time-resolved intensity data, and perform noise suppression, differential absorption calculation and normalization processing on the echo signals to generate preprocessed echo signals. 3.The deep learning-based LIDAR multi-gas identification method of claim 1, wherein, Based on the gas image, a gas class of interest is obtained, specifically: The gas image is feature-encoded, the matching degree of the encoded gas image and the standard gas feature is calculated, the gas with a matching degree greater than a preset matching degree threshold is retained as a candidate gas according to the matching degree, and an image feature likelihood of the candidate gas is obtained; An initial existence probability of the candidate gas is obtained based on the current scene in an industry-gas association knowledge base as a scene prior probability, a Bayesian probability model is constructed through the image feature likelihood of the candidate gas and the scene prior probability, a probability of existence of the candidate gas under the gas image is obtained, and a gas class of interest is obtained based on the probability. 4.The deep learning-based LIDAR multi-gas identification method of claim 1, wherein, A multi-layer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, and linear regression is used to output the concentration of each gas. Specifically, The final node representation after standardization processing is input into the gas concentration branch, the original concentration prediction is obtained, a trainable matrix is constructed for interference compensation, and the original concentration prediction is modified to output the concentration of each gas class. The deep learning-based laser radar multi-gas identification method of any one of claims 1-4 is implemented, and the system comprises a laser radar transmitting and receiving module, a preprocessing and multi-domain feature extraction module, a feature selection and gas image construction module, a multi-gas identification module, a result output and visualization module; 5. A deep learning based laser radar multi-gas identification system, characterized by, The laser radar transmitting and receiving module transmits different wavelength lasers using differential absorption laser radar and receives the echo signals after gas absorption, and dynamically adjusts the laser wavelength to ensure coverage of different gas absorption spectral lines; 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 image construction module uses multi-objective optimization combined with feature importance evaluation for feature selection, filters key features, maps the key features to generate a gas feature distribution map, divides feature clusters through clustering, generates a gas image of potential gas, and obtains a gas class of interest based on the gas image; The multi-gas identification module constructs a graph structure based on the gas image and gas feature similarity, uses a graph convolution network to perform representation learning on the graph structure, and uses a gas class of interest to enhance the graph convolution network, constructs a gas class branch, and outputs the gas class; a multi-layer perception mechanism is used to construct a gas concentration branch, the final node representation output by the graph convolution network is used as input, and linear regression is used to output the concentration of each gas; The result output and visualization module outputs the gas concentration and gas class, and visualizes the output according to a preset device. ​

Images