Gas concentration identification method and system based on AI model

Abstract

The invention discloses a gas concentration identification method and system based on an AI model, and relates to the technical field of data processing. The method comprises the following steps: acquiring an environment infrared image sequence and a gas concentration time sequence signal; dynamically adjusting the length of the sliding window, and intercepting a window image sequence and a window concentration sequence; dTCWT decomposition is adopted to extract multi-direction sub-band energy features of the infrared image, the multi-direction sub-band energy features are compressed into image feature vectors through 1D-CNN, concentration time sequence feature vectors are extracted through an LSTM network, and joint representation is generated through fusion; outputting a concentration predicted value and danger level probability distribution through a mixed deep learning model; a probability thermodynamic diagram is generated by combining Monte Carlo diffusion simulation, and a leakage source coordinate is accurately positioned by fusing a concentration gradient; triggering a grading response instruction according to the highest risk probability, and generating risk map real-time visualization; the problems that a traditional system is delayed in response, inaccurate in positioning and insufficient in utilization of multi-source data are solved, closed-loop management and control of leakage monitoring, early warning, positioning and disposal are achieved, and the positioning precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing. Specifically, it particularly relates to a gas concentration recognition method and system based on an AI model. Background Art

[0002] In the current field of industrial safety monitoring, especially for high-risk scenarios such as petrochemical and natural gas storage and transportation; in such environments, leakage of combustible / toxic gases (such as CH4, CO2) may trigger major accidents such as explosions and poisonings. Existing monitoring systems generally rely on point gas sensors or static infrared thermal imagers, but the former has a limited coverage range, and the latter is difficult to quantify the concentration change trend. With the development of deep learning and multi-sensor fusion technology, there is still a need to build an integrated monitoring platform with real-time leakage warning, accurate traceability, and intelligent response capabilities to improve the accident prevention and control efficiency and reduce false alarms and missed detections.

[0003] There are problems in the prior art such as response hysteresis and poor adaptability, insufficient utilization of multi-source data, low leakage source positioning accuracy, and rigid risk response mechanisms. Summary of the Invention

[0004] Technical Problems to be Solved In view of the problems in the related art, the present invention provides a gas concentration recognition method based on an AI model to overcome the above-mentioned technical problems existing in the prior related art.

[0005] Technical Solutions To solve the above technical problems, the present invention is realized through the following technical solutions: S1. Collect original signal data to obtain a gas concentration time series signal and a multi-band infrared image sequence; S2. Process the gas concentration time series signal and the multi-band infrared image sequence using a sliding time series window to obtain a window concentration sequence and a window image sequence; S3. Respectively extract features from the window image sequence and the window concentration sequence through a neural network model to obtain an image feature vector and a time series feature vector; Fuse the image feature vector and the time series feature vector to obtain a real-time fusion feature vector; S4. Build a hybrid deep learning model, train the hybrid deep learning model to obtain a converged hybrid deep learning model; Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain a concentration prediction value and a risk level probability; S5. Combine the concentration prediction value with the Monte Carlo model to obtain the leakage source coordinates; execute multi-level risk response instructions according to the leakage source coordinates and the risk level probability; The present invention solves the bottlenecks of traditional gas monitoring systems in terms of real-time performance, accuracy, positioning ability, and intelligence level through a dynamic window mechanism, multi-modal depth feature fusion, hybrid model dual-task prediction, Monte Carlo gradient joint positioning, and probabilistic hierarchical response, providing an end-to-end closed-loop solution for leakage monitoring, early warning, and disposal in high-risk industrial scenarios.

[0006] Preferably, S1 includes the following steps: S11. Collect infrared radiation signals in the environment through a multi-band infrared sensor array to obtain infrared radiation signals; Perform spatio-temporal alignment processing on the infrared radiation signals to obtain a multi-band infrared image sequence; S12. Collect infrared radiation signals and original gas concentration data in the environment through a high-precision gas concentration sensor to obtain a gas concentration time series signal; The present invention collects environmental radiation signals through a dual-band infrared sensor, generates a multi-band infrared image sequence through spatio-temporal alignment, eliminates the position differences of sensors, and ensures the spatio-temporal consistency of images; synchronously obtains the gas concentration time series signal to achieve synchronous acquisition of multi-source data; this design supports dual-gas specific detection and provides a raw data basis for multi-modal fusion with accurate spatio-temporal alignment.

[0007] Preferably, S2 includes the following steps: S21. Calculate the real-time concentration change rate according to the gas concentration time series signal; S22. Obtain the window length according to the real-time concentration change rate and in combination with the dynamic window formula; S23. Intercept the images and concentration data within the time period from the current moment minus the window length to the current moment from the gas concentration time series signal and the multi-band infrared image sequence to obtain a window image sequence and a window concentration sequence; The present invention calculates the concentration change rate in real time, adaptively determines the window length through the dynamic window formula, and intercepts the corresponding time period data to generate a window sequence; significantly improves the response speed to sudden leakage, avoids long-window redundancy during the stable period, and provides an optimized data basis for subsequent feature extraction.

[0008] Preferably, S3 includes the following steps: S31. Perform DTCWT decomposition on each frame of the window image sequence using DTCWT transformation to obtain image DTCWT coefficients; the image DTCWT coefficients include the low-frequency approximation coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients of each frame of the window image sequence; S32. Calculate the energy value of each frame of the window image sequence according to the multi-type coefficients of the image and using the direction energy calculation formula to obtain image sub-band features; S33. Input the image sub-band features into the 33 convolutional kernels in the 1D CNN layer, and obtain the image feature vector through global average pooling; S34. Input the window concentration sequence into the LSTM gated sequence, extract the hidden state of the time step in the window concentration sequence, and obtain the temporal feature vector; S35. Input the image feature vector and the temporal feature vector into the feature concatenation layer for feature processing, and obtain the real-time fusion feature vector; In the present invention, by performing DTCWT decomposition on the infrared image, multi-scale coefficients including low-frequency structures and horizontal / vertical / diagonal direction details are extracted, comprehensively retaining the texture features of the leakage area; by calculating the directional energy to quantify the intensity of each sub-band, image sub-band features are generated; 1D-CNN is used to compress the feature dimension, and the image feature vector is output through global average pooling to avoid loss of spatial information; the temporal dependence relationship of the concentration sequence is mined through LSTM; the image and temporal features are fused to construct a joint representation; significantly enhancing the model's coupling perception ability of the leakage spatial distribution and diffusion dynamics, reducing the false alarm rate.

[0009] Preferably, the said S4 includes the following steps: S41. Construct a hybrid deep learning model; the hybrid deep learning model includes an input layer, a regression branch fully connected layer, and a classification branch fully connected layer; the input layer is used to input the fusion feature vector; the regression branch fully connected layer is used to predict the gas concentration; the classification branch fully connected layer is used for the danger level; S42. Collect historical data; the historical data includes historical fusion feature vectors and historical true labels; the historical true labels include historical predicted gas concentrations and historical danger levels; S43. Use the historical data to combine and optimize the training of the hybrid deep learning model to obtain a converged hybrid deep learning model; S44. Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain the concentration prediction value and the danger level probability; In the present invention, by constructing a dual-branch hybrid model (the regression branch predicts the concentration value, and the classification branch outputs the danger level probability), multi-task collaborative learning is realized; historical fusion features and true labels are collected to construct a training set; the hybrid model is trained and optimized until convergence; the fusion features are input in real time, and the concentration prediction value and the four-level danger probability distribution are output synchronously; this design breaks through the limitations of single-task models, while ensuring the accuracy of concentration prediction, providing a scientific confidence basis for response decisions through probabilistic hierarchical output, and improving the response accuracy.

[0010] Preferably, the said S43 includes the following steps: S431. Set the maximum number of iterations; use historical data to train a hybrid deep learning model in combination with optimization training, and calculate the regression loss and classification loss of the hybrid deep learning model during the training process; S432. Assign weights to the regression loss and classification loss to obtain the total loss of the deep learning model; S433. Use the backpropagation algorithm to calculate the gradient of the total loss with respect to the parameters of the hybrid deep learning model; S434. Use the Adam optimizer to update the parameters of the hybrid deep learning model according to the gradient of the parameters of the hybrid deep learning model to minimize the total loss; S435. Repeat S431, S432, S433, S434. When the maximum iteration coefficient is reached, obtain a converged hybrid deep learning model; In the present invention, weights are assigned to the regression / classification loss to obtain the total loss, which solves the multi-task learning balance problem; the gradient is calculated through backpropagation, and the Adam optimizer is used to iteratively update the model parameters; loop until convergence; this collaborative optimization mechanism significantly improves the generalization ability of the model, improves the training efficiency, and synchronously reduces the prediction errors of the dual tasks.

[0011] Preferably, the S5 includes the following steps: S51. Obtain the current concentration field according to the real-time concentration distribution. Combine the current concentration field with the concentration prediction value, and simulate the leakage diffusion process through the Monte Carlo model to obtain a diffusion probability heat map; S52. Calculate the concentration gradient according to the current concentration field; combine the concentration gradient with the diffusion probability heat map to obtain the leakage source coordinates; S53. Set multi-level danger response instructions; S54. Execute the multi-level danger response instructions according to the danger level probability; S55. Perform image fusion and annotation processing on the original infrared image, heat map, and leakage point coordinates to obtain a dynamic risk map; The dynamic risk map is displayed in real time on the monitoring interface, and leakage treatment is carried out according to the displayed leakage point coordinate position; The present invention passes through.

[0012] A gas concentration recognition system based on an AI model for implementing the above-mentioned gas concentration recognition method based on an AI model, including a multi-modal data acquisition module, a dynamic window preprocessing module, a multi-modal feature fusion module, a hybrid intelligent prediction module, and a leakage source localization, risk response, and visualization module; The multi-modal data acquisition module uses a multi-band infrared sensor array and a high-precision gas concentration sensor to synchronously collect infrared radiation signals and raw gas concentration data in the environment, and generates a multi-band infrared image sequence and a gas concentration time series signal through spatio-temporal alignment; The dynamic window preprocessing module adaptively determines the sliding window length based on the gas concentration change rate calculated in real time, applying the dynamic window formula; and outputs a window image sequence and a window concentration sequence associated with the current moment, providing time-aligned and length-adaptive data slices for subsequent analysis. The multimodal feature fusion module extracts image subband features including low-frequency approximation and directional details by performing DTCWT decomposition on the input window image sequence, and generates an image feature vector through directional energy calculation and 1DCNN combined with global average pooling; processes the window concentration sequence using an LSTM network to extract a temporal feature vector; and splices and fuses these two heterogeneous feature vectors to obtain a real-time fusion feature vector. The hybrid intelligent prediction module inputs the real-time fusion feature vector into a two-branch hybrid deep learning model, which includes a regression branch for predicting future gas concentration values and a classification branch for predicting the probability distribution of different risk levels; the model is trained with historical data and optimized using weighted total loss and the Adam optimizer until convergence. The leakage source location, risk response and visualization module is responsible for accurately locating the leakage source, executing hierarchical response instructions and generating a dynamic risk map; based on the predicted gas concentration values, combined with the current concentration field data, uses the Monte Carlo model to simulate a large number of possible leakage diffusion paths to generate a diffusion probability heat map; calculates the spatial gradient of the current concentration field; accurately locates the coordinates of the most likely leakage source by fusing the heat map probability and concentration gradient information; triggers preset corresponding-level risk response instructions according to the risk level probability output by the prediction module; fuses and annotates the original infrared image, diffusion probability heat map and leakage source coordinates to generate a dynamic risk map that intuitively displays the risk situation, and displays it in real time on the monitoring interface to guide the leakage treatment operation.

[0013] Beneficial effects The present invention has the following beneficial effects: By means of a dynamic window mechanism, multimodal deep feature fusion, dual-task prediction of a hybrid model, Monte Carlo gradient joint location and probabilistic hierarchical response, the present invention solves the bottlenecks of traditional gas monitoring systems in terms of real-time performance, accuracy, location ability and intelligent level, and provides an end-to-end closed-loop solution for leakage monitoring, early warning and disposal in high-risk industrial scenarios.

[0014] The present invention dynamically adjusts the sliding window length through the real-time concentration change rate, shortens the window when the concentration surges to quickly capture mutation signals; and extends the window during the flat period to accumulate sufficient historical data, significantly improving the response speed and stability of the system to sudden leakage events.

[0015] The present invention extracts multi-directional texture energy features by using DTCWT decomposition, compresses them into high-dimensional vectors through 1DCNN, and retains spatial structure information. The LSTM network is used to extract long-term dependence features of concentration sequences and capture diffusion dynamics. By fusing image and time-series features, a joint representation is constructed to enhance the model's perception ability of the "spatiotemporal coupling characteristics" of leakage events and reduce the misjudgment rate.

[0016] The present invention designs a hybrid model dual-task collaborative prediction structure, accurately predicts future gas concentration values through a regression branch, and outputs a multi-level risk probability distribution through a classification branch. Through joint optimization of a weighted loss function, quantitative prediction of concentration and risk probabilistic classification are synchronously achieved, providing a dual basis for response decision-making.

[0017] The present invention realizes accurate positioning of the leakage source, intelligent multi-level response and visualization. A probability heat map is generated through a Monte Carlo diffusion model to quantify the leakage possibility at each location. Through gradient probability fusion positioning, the concentration gradient is multiplied by the heat map, and the maximum value point is taken as the leakage source coordinate. By fusing the physical diffusion model and real-time field data, the positioning anti-noise performance and accuracy are greatly improved. By triggering preset instructions according to the highest probability danger level, the response strategy and risk confidence level are accurately matched. The infrared image, heat map and leakage coordinates are dynamically fused to generate a risk map, which visually displays the position of the leakage point and the diffusion trend in real time, assisting personnel in quickly locating and treating, and improving the emergency response efficiency.

[0018] Of course, it is not necessary for any product implementing the present invention to achieve all the above advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the drawings required for describing the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic flowchart of a gas concentration recognition method based on an AI model according to the present invention; Figure 2 It is a schematic flowchart of the operation of a convergent hybrid deep learning model in a gas concentration recognition method based on an AI model according to the present invention; Figure 3 It is a schematic flowchart of intercepting an image sequence and a concentration time-series segment of a corresponding length in a gas concentration recognition method based on an AI model according to the present invention; Figure 4 It is a schematic diagram of the modules of a gas concentration recognition system based on an AI model according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the invention with reference to the accompanying drawings in the embodiments of the invention. Obviously, the described embodiments are only a part of the embodiments of the invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the invention without creative efforts shall fall within the scope of protection of the invention.

[0022] In the description of the present invention, it should be understood that the terms "open hole", "upper", "lower", "top", "middle", "inner", etc. indicating the orientation or positional relationship are only for the convenience of describing the invention and simplifying the description, rather than indicating or implying that the components or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention.

[0023] Embodiment 1 Please refer to Figure 1 、 Figure 2 、 Figure 3 The present invention discloses a gas concentration recognition method based on an AI model, including the following steps: S1. Collect the original signal data to obtain the gas concentration time series signal and the multi-band infrared image sequence; The S1 includes the following steps: S11. Collect the infrared radiation signal in the environment through a multi-band infrared sensor array (4.26μm CO2 band + 814μm methane band) to obtain the infrared radiation signal; Perform spatio-temporal alignment processing on the infrared radiation signal to obtain the multi-band infrared image sequence; S12. Collect the infrared radiation signal and the original gas concentration data in the environment through a high-precision gas concentration sensor to obtain the gas concentration time series signal; S2. Process the gas concentration time series signal and the multi-band infrared image sequence by using a sliding time series window to obtain the window concentration sequence and the window image sequence; The S2 includes the following steps: S21. Calculate the real-time concentration change rate according to the gas concentration time series signal; the calculation formula is as follows, ; Wherein, r c represents the real-time concentration change rate, c represents the gas concentration, t represents the time, represents the concentration c The derivative of time t (i.e., the change speed of the concentration with time); S22. Obtain the window length according to the real-time concentration change rate and in combination with the dynamic window formula. The dynamic window formula is as follows: ; Wherein, L w represents the window length, 50, 30, and 10 respectively represent the window lengths under different real-time concentration change rates, and 0.1% / s and 0.5% / s respectively represent different predicted concentration change rate thresholds; S23. Intercept the images and concentration data within the time period from the current moment minus the window length to the current moment from the gas concentration time series signal and the multi-band infrared image sequence to obtain the window image sequence and the window concentration sequence; S3. Respectively perform feature extraction on the window image sequence and the window concentration sequence through a neural network model to obtain an image feature vector and a time series feature vector; Fuse the image feature vector and the time series feature vector to obtain a real-time fusion feature vector; S3 includes the following steps: S31. Perform DTCWT decomposition on each frame of the window image sequence by using the DTCWT transform to obtain the image DTCWT coefficients. The image DTCWT coefficients include the low-frequency approximation coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients of each frame of the window image sequence. The low-frequency approximation coefficients retain the main energy and overall structural information of the image, similar to a blurred version of the image. The horizontal detail coefficients contain the edge and texture information of the image in a specific direction α. The vertical detail coefficients contain the edge and texture information of the image in a specific direction 90 degrees α, and the diagonal detail coefficients contain the edge and texture information in a specific direction 45 degrees + α; S32. According to the multi-type coefficients of the image and using the direction energy calculation formula, calculate the energy value of each frame of the window image sequence to obtain the image sub-band features. The direction energy calculation formula is as follows: ; Wherein, E α represents the energy value calculated in the direction α (the energy values of the image in different directions together constitute the image sub-band features), x , y represents the pixel coordinates in each frame of the window image sequence, D α ( x , y ) represents the DTCWT coefficient value at the coordinate ( x , y ) in the α direction; S33. Input the image sub-band features into the 33 convolutional kernels in the 1D CNN layer, and obtain the image feature vector through global average pooling; S34. Input the window concentration sequence into the LSTM gated sequence, extract the hidden states of the time steps in the window concentration sequence, and obtain the temporal feature vector; the LSTM gated sequence contains 128 hidden units; S35. Input the image feature vector and the temporal feature vector into the feature concatenation layer for feature processing, and obtain the real-time fusion feature vector; S4. Construct a hybrid deep learning model, train the hybrid deep learning model, and obtain a converged hybrid deep learning model; Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain the concentration prediction value and the risk level probability; The S4 includes the following steps: S41. Construct a hybrid deep learning model; the hybrid deep learning model includes an input layer, a regression branch fully connected layer, and a classification branch fully connected layer; the input layer is used to input the fusion feature vector; the regression branch fully connected layer is used to predict the gas concentration; the classification branch fully connected layer is used for the risk level; S42. Collect historical data; the historical data includes historical fusion feature vectors and historical true labels; the historical true labels include historical predicted gas concentrations and historical risk levels; S43. Use the historical data to optimize and train the hybrid deep learning model in combination to obtain a converged hybrid deep learning model; The S43 includes the following steps: S431. Set the maximum number of iterations; use the historical data to optimize and train the hybrid deep learning model in combination, and calculate the regression loss and classification loss of the hybrid deep learning model during the training process; S432. Assign weights to the regression loss and classification loss to obtain the total loss of the deep learning model; S433. Use the backpropagation algorithm to calculate the gradient of the total loss with respect to the parameters of the hybrid deep learning model; S434. Use the Adam optimizer to update the parameters of the hybrid deep learning model according to the gradient of the parameters of the hybrid deep learning model to minimize the total loss; S435. Repeat S431, S432, S433, S434. When the maximum iteration coefficient is reached, obtain a converged hybrid deep learning model; S44. Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain the concentration prediction value and the risk level probability; if the probability of the first-level risk is 0.05, the probability of the second-level risk is 0.15, the probability of the third-level risk is 0.68, and the probability of the fourth-level risk is 0.12; then it is determined as the third-level risk at this time; S5. Based on the predicted concentration value and combined with the Monte Carlo model, obtain the coordinates of the leakage source; execute multi-level hazard response instructions according to the coordinates of the leakage source and the probability of the hazard level; The S5 includes the following steps: S51. Obtain the current concentration field according to the real-time concentration distribution. Based on the current concentration field and combined with the predicted concentration value, simulate the leakage diffusion process through the Monte Carlo model to obtain a diffusion probability heat map; the formula of the Monte Carlo model is as follows, ; Where, P ( x , y ) represents the diffusion probability heat map (i.e., the probability of diffusion occurring at the position ([[]] x , y ) at [[[]] t ), N represents the total number of diffusion simulations performed based on the fluid mechanics model, path i represents the [[[]] i th diffusion simulation path, represents the indicator function (if the path path i passes through the position ([[]] t , x , y ) at [[[]] i then the value is 1, otherwise it is 0), and the path ; δ Where, x C( x , y ) represents the concentration gradient, , respectively represent the first-order partial derivatives of the concentration of each monitoring point in the current concentration field in the [[[]] x, y direction; Based on the concentration gradient and combined with the diffusion probability heat map, obtain the coordinates of the leakage source; the calculation formula is as follows, ; Where, ([[]] x s , y s ) represents the coordinates of the leakage source, argmax() represents finding the point ([[]] x , y ) that makes the subsequent function value the largest, represents the coordinate point ([[]] x , y) the product of the gradient in the heat map of diffusion probability and the probability of the coordinate point ( x , y ); S53. Set multi-level danger response instructions; for example, the first-level danger response instruction is to activate the local audible and visual alarm; the second-level danger response instruction is to start the ventilation system + SMS alarm, the third-level danger response instruction is to cut off the power supply of non-explosion-proof equipment; the fourth-level danger response instruction is to trigger the fire sprinkler + cloud platform broadcast; S54. Execute multi-level danger response instructions according to the danger level probability; S55. Perform image fusion and annotation processing on the original infrared image, heat map, and leakage point coordinates to obtain a dynamic risk map; The dynamic risk map monitoring interface is displayed in real time, and leakage treatment is carried out according to the displayed leakage point coordinate position.

[0024] Embodiment 2 Please refer to Figure 4 , a gas concentration recognition system based on an AI model for implementing the above-mentioned gas concentration recognition method based on an AI model, including a multi-modal data acquisition module, a dynamic window preprocessing module, a multi-modal feature fusion module, a hybrid intelligent prediction module, and a leakage source localization, risk response, and visualization module; The multi-modal data acquisition module uses a multi-band infrared sensor array and a high-precision gas concentration sensor to synchronously collect infrared radiation signals and raw gas concentration data in the environment, and generates a multi-band infrared image sequence and a gas concentration time series signal through spatio-temporal alignment; The dynamic window preprocessing module adaptively determines the sliding window length based on the real-time calculated gas concentration change rate by applying the dynamic window formula; outputs a window image sequence and a window concentration sequence associated with the current moment, providing time-aligned and length-adaptive data slices for subsequent analysis; The multi-modal feature fusion module performs DTCWT decomposition on the input window image sequence, extracts image sub-band features including low-frequency approximation and directional details, and generates an image feature vector through directional energy calculation and 1DCNN combined with global average pooling; uses an LSTM network to process the window concentration sequence and extracts a time series feature vector; after splicing and fusing these two heterogeneous feature vectors, a real-time fusion feature vector is obtained; The hybrid intelligent prediction module inputs the real-time fusion feature vector into a two-branch hybrid deep learning model, which includes a regression branch for predicting future gas concentration values and a classification branch for predicting the probability distribution of different danger levels; the model is trained with historical data and optimized using weighted total loss and the Adam optimizer until convergence; The leakage source location, risk response and visualization module is responsible for accurately locating the leakage source, executing hierarchical response instructions and generating a dynamic risk map; based on the predicted gas concentration values, combined with the current concentration field data, using the Monte Carlo model to simulate a large number of possible leakage diffusion paths, generating a diffusion probability heat map; calculating the spatial gradient of the current concentration field; accurately locating the coordinates of the most likely leakage source by fusing the heat map probability and concentration gradient information; triggering the preset corresponding level of danger response instructions according to the danger level probability output by the prediction module; fusing and annotating the original infrared image, diffusion probability heat map and leakage source coordinates to generate a dynamic risk map that intuitively displays the risk situation, and real-time displaying it on the monitoring interface to guide the leakage treatment operation.

[0025] In the description of this specification, the description referring to terms such as "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0026] The preferred embodiments of the invention disclosed above are only used to help illustrate the invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principle and practical application of the invention, so that those skilled in the art can well understand and utilize the invention.

Claims

1. A gas concentration recognition method based on an AI model, characterized in that, It includes the following steps: S1. Collect the original signal data to obtain the gas concentration time series signal and the multi-band infrared image sequence; S2. Use a sliding time series window to process the gas concentration time series signal and the multi-band infrared image sequence to obtain the window concentration sequence and the window image sequence; S3. Respectively extract features from the window image sequence and the window concentration sequence through a neural network model to obtain the image feature vector and the time series feature vector; Fuse the image feature vector and the time series feature vector to obtain the real-time fusion feature vector; S4. Construct a hybrid deep learning model, train the hybrid deep learning model to obtain a converged hybrid deep learning model; Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain the concentration prediction value and the risk level probability; S5. Combine the concentration prediction value with the Monte Carlo model to obtain the leakage source coordinates; Execute the multi-level hazard response instruction according to the leakage source coordinates and the risk level probability.

2. The gas concentration recognition method based on an AI model according to claim 1, characterized in that, The S1 includes the following steps: S11. Collect the infrared radiation signal in the environment through a multi-band infrared sensor array to obtain the infrared radiation signal; Perform spatio-temporal alignment processing on the infrared radiation signal to obtain the multi-band infrared image sequence; S12. Collect the infrared radiation signal and the original gas concentration data in the environment through a high-precision gas concentration sensor to obtain the gas concentration time series signal.

3. The gas concentration recognition method based on an AI model according to claim 1, wherein, The S2 includes the following steps: S21. Calculate the real-time concentration change rate according to the gas concentration time series signal; S22. Combine the real-time concentration change rate with the dynamic window formula to obtain the window length; S23. Intercept the images and concentration data in the time period from the current time minus the window length to the current time from the gas concentration time series signal and the multi-band infrared image sequence to obtain the window image sequence and the window concentration sequence.

4. The gas concentration recognition method based on an AI model according to claim 1, wherein The S3 includes the following steps: S31. Perform DTCWT decomposition on each frame of the image in the window image sequence by using DTCWT transform to obtain the image DTCWT coefficients; The image DTCWT coefficients include the low-frequency approximation coefficient, the horizontal detail coefficient, the vertical detail coefficient, and the diagonal detail coefficient of each frame of the image in the window image sequence; S32. Calculate the energy value of each frame of the image in the window image sequence according to the multi-type coefficients of the image and use the direction energy calculation formula to obtain the image sub-band feature; S33. Input the image sub-band feature into the 33 convolutional kernels in the 1DCNN layer and obtain the image feature vector through global average pooling; S34. Input the window concentration sequence into the LSTM gated sequence and extract the hidden state of the time step in the window concentration sequence to obtain the time series feature vector; S35. Input the image feature vector and the time series feature vector into the feature splicing layer for feature to obtain the real-time fusion feature vector.

5. The gas concentration recognition method based on an AI model according to claim 1, characterized in that, The S4 includes the following steps: S41. Construct a hybrid deep learning model; The hybrid deep learning model includes an input layer, a regression branch fully connected layer, and a classification branch fully connected layer; The input layer is used to input the fusion feature vector; The regression branch fully connected layer is used to predict the gas concentration; The classification branch fully connected layer is used for the risk level; S42. Collect historical data; the historical data includes historical fusion feature vectors and historical true labels; the historical true labels include historical predicted gas concentrations and historical hazard levels; S43. Use the historical data to combine with an optimized training hybrid deep learning model to obtain a converged hybrid deep learning model; S44. Input the real-time fusion feature vector into the converged hybrid deep learning model to obtain a concentration prediction value and a hazard level probability.

6. The gas concentration recognition method based on an AI model according to claim 5, wherein, The S43 includes the following steps: S431. Set the maximum number of iterations; use the historical data to combine with an optimized training hybrid deep learning model, and calculate the regression loss and classification loss of the hybrid deep learning model during the training process; S432. Assign weights to the regression loss and classification loss to obtain the total loss of the deep learning model; S433. Use the backpropagation algorithm to calculate the gradient of the total loss with respect to the parameters of the hybrid deep learning model; S434. Use the Adam optimizer to update the parameters of the hybrid deep learning model according to the gradient of the parameters of the hybrid deep learning model to minimize the total loss; S435. Repeat S431, S432, S433, S434. When the maximum iteration coefficient is reached, obtain a converged hybrid deep learning model.

7. A gas concentration recognition method based on an AI model according to claim 1, characterized in that, The S5 includes the following steps: S51. Obtain the current concentration field according to the real-time concentration distribution. Combine the current concentration field with the concentration prediction value, and simulate the leakage diffusion process through the Monte Carlo model to obtain a diffusion probability heat map; S52. Calculate the concentration gradient according to the current concentration field; combine the concentration gradient with the diffusion probability heat map to obtain the leakage source coordinates; S53. Set multi-level hazard response instructions; S54. Execute the multi-level hazard response instructions according to the hazard level probability; S55. Perform image fusion and annotation processing on the original infrared image, heat map, and leakage point coordinates to obtain a dynamic risk map; The dynamic risk map monitoring interface is displayed in real time, and leakage treatment is carried out according to the displayed leakage point coordinate positions.

8. A gas concentration recognition system based on an AI model, characterized in that, Implement a gas concentration recognition method based on an AI model as described in any one of claims 1-7. The system includes a multi-modal data acquisition module, a dynamic window preprocessing module, a multi-modal feature fusion module, a hybrid intelligent prediction module, and a leakage source location, risk response, and visualization module; The multi-modal data acquisition module uses a multi-band infrared sensor array and a high-precision gas concentration sensor to synchronously collect infrared radiation signals and raw gas concentration data in the environment, and generates a multi-band infrared image sequence and a gas concentration time series signal through spatio-temporal alignment; The dynamic window preprocessing module adaptively determines the sliding window length based on the real-time calculated gas concentration change rate by applying the dynamic window formula; Output a window image sequence and a window concentration sequence associated with the current moment, providing time-aligned and length-adaptive data slices for subsequent analysis; The multi-modal feature fusion module performs DTCWT decomposition on the input window image sequence, extracts image sub-band features including low-frequency approximation and directional details, and generates an image feature vector through directional energy calculation and 1DCNN combined with global average pooling; uses the LSTM network to process the window concentration sequence and extract the temporal feature vector; After splicing and fusing these two heterogeneous feature vectors, a real-time fusion feature vector is obtained; The hybrid intelligent prediction module inputs the real-time fusion feature vector into a two-branch hybrid deep learning model, which includes a regression branch for predicting future gas concentration values and a classification branch for predicting the probability distribution of different hazard levels; the model is trained with historical data and optimized using weighted total loss and Adam optimizer until convergence; The leakage source location, risk response and visualization module is responsible for accurately locating the leakage source, executing hierarchical response instructions and generating a dynamic risk map; based on the predicted gas concentration value, combined with the current concentration field data, uses the Monte Carlo model to simulate a large number of possible leakage diffusion paths and generate a diffusion probability heat map; calculates the spatial gradient of the current concentration field; accurately locates the coordinates of the most likely leakage source by fusing the heat map probability and concentration gradient information; triggers the preset corresponding-level hazard response instructions according to the hazard level probability output by the prediction module; fuses and labels the original infrared image, diffusion probability heat map, and leakage source coordinates to generate a dynamic risk map that intuitively displays the risk situation and is real-time displayed on the monitoring interface to guide the leakage treatment operation.

9. A storage medium, characterized in that, A program is stored thereon, and when the program is executed by a processor, it implements a gas concentration recognition method according to any one of claims 1-7.

Images