Hazardous chemical storage leakage positioning and tracing method based on gas array

Abstract

The technical scheme of the invention relates to the technical field of gas detection, in particular to a hazardous chemical storage leakage positioning and tracing method based on a gas array. The method comprises the following steps: collecting gas concentration, components and environmental physical field parameters through a three-dimensional heterogeneous sensor array, and generating a multi-modal data set; and matching the noise mode feature library in the knowledge base by using a transfer learning algorithm, dynamically correcting the baseline drift of the sensor, and outputting calibrated data. A closed-loop adaptive mechanism adjusts sensor parameters and network hyper-parameters through meta reinforcement learning, neural architecture searches and optimizes a model structure, and a differential evolution algorithm updates fluid mechanics boundary conditions. Through cooperation of physical constraint and data driving, the problem of signal distortion in a complex environment is solved, the leakage source positioning precision and robustness are improved, and the method is suitable for the field of hazardous chemical substance storage safety monitoring.

Description

Technical Field

[0001] The present invention relates to the cross - application technical field of the dynamic diffusion characteristics of substances and the traceability algorithm in the field of safety monitoring of hazardous chemical storage, and particularly to a method for leak location and traceability of hazardous chemicals in a gas array storage. Background Art

[0002] In the scenario of hazardous chemical storage, the leak location and traceability technology based on a gas sensor array realizes the rapid identification of the leak source through the analysis of the multi - node gas concentration distribution characteristics and the spatio - temporal correlation. This technology relies on a high - density gas - sensitive sensor network to collect gas diffusion parameters in real time, combines the fluid mechanics model and the leak source location algorithm, determines the leak signal propagation path through the time - difference method or concentration gradient analysis, and uses an adaptive weighted fusion model to eliminate environmental interference noise. Further introducing Bayesian inference or neural network algorithms can inversely deduce the time series and spatial coordinates of the leak event, and at the same time, coupling the three - dimensional topological structure and environmental parameters (such as wind speed, temperature and humidity) of the storage area for dynamic correction to achieve the precise reconstruction of the probability distribution of the leak source.

[0003] When the gas sensor array is applied to the leak location of hazardous chemical storage, affected by the coupling of multiple physical fields in a complex environment, the gas diffusion path is easily affected by the non - linear superposition of the turbulence effect of the storage structure, the cross - interference of multi - component gases and the sensor response lag characteristics, resulting in the distortion of the spatial correlation between the concentration gradient and the time - series signal, and further reducing the analytical accuracy of the fluid mechanics inverse inversion model for the spatial coordinates of the leak source. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for leak location and traceability of hazardous chemicals in a gas array storage, which is used to solve the problem that the distortion of the gas diffusion signal caused by the coupling of multiple physical fields in the complex environment in the prior art affects the inverse inversion accuracy of the hazardous chemical leak source based on the gas sensor array.

[0005] To solve the above - mentioned technical problems, the specific technical solutions of the present invention are as follows: The method for leak location and traceability of hazardous chemicals in a gas array storage provided by the present invention includes: Step 1, collecting the gas concentration, components and environmental physical field parameters of the storage environment in real time through a heterogeneous gas sensor array deployed in three - dimensional space, and generating a multi - modal original data set including the gas concentration time - series signal, multi - component spectral characteristics and three - dimensional wind speed field; Step 2, inputting the multi - modal original data set into a preset knowledge base, performing cross - scenario domain adaptation pre - training on the sensor array through a transfer learning algorithm, generating an adaptive baseline correction signal based on the dynamic matching of the real - time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and outputting a calibrated multi - physical field data set after environmental noise suppression and sensor zero - point drift compensation; Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physical field dataset, allocate noise suppression weights using a bidirectional gated attention mechanism based on the turbulence-diffusion correlation map in the knowledge base, and simultaneously call the gas adsorption kinetic parameter library to construct a migration learning-driven chromatographic response migration model, and output the denoised concentration gradient signal of the target gas and the separated component feature vectors; Step 4: Based on the turbulence vortex distribution map generated by computational fluid dynamics dynamic simulation, quantify the response lag effect between sensor nodes, use the particle swarm optimization algorithm to iteratively solve the time delay compensation coefficient and align the multi-node time series signals, and combine the generative adversarial network to synthesize the compensated concentration gradient data with the physical field constraints to generate the reference diffusion feature vectors, and generate the gas diffusion path map under the condition of multi-physical field coupling; Step 5: Embed the computational fluid dynamics mass conservation equation as a hard constraint into the hybrid neural network of Bayesian and LSTM, jointly optimize the network parameters through the projected gradient descent algorithm, and generate the posterior probability distribution of the leakage source coordinates that satisfies the hydrodynamic law and the measured data characteristics; Step 6: Input the posterior probability distribution output by the hybrid neural network of Bayesian and LSTM, the computational fluid dynamics simulation results, and the LSTM prediction data into the ensemble learning framework, use the NSGA-II multi-objective optimization algorithm to dynamically allocate model weights and screen the Pareto optimal solution set, and generate a credibility rating localization result verified by the residual distribution-localization error mapping network.

[0006] Furthermore, for the method for leakage location and traceability of hazardous chemical storage using the gas array of the present invention, the environmental physical field parameters include three-dimensional wind speed field, temperature gradient distribution, and air pressure pulsation time series signal; Input the three-dimensional wind speed field and temperature gradient distribution into the noise pattern feature library in the knowledge base for spectral similarity comparison; Generate an adaptive filtering signal based on transfer learning domain adaptation according to the comparison result, perform temperature drift compensation on the air pressure pulsation time series signal, and output the corrected three-dimensional physical field parameter set.

[0007] Furthermore, the method for leakage location and traceability of hazardous chemical storage using the gas array of the present invention further includes: Perform Daubechies wavelet packet decomposition on the gas concentration time series signal in the corrected three-dimensional physical field parameter set, and extract high-frequency noise components and low-frequency diffusion components; Call the turbulence-diffusion correlation map in the knowledge base to allocate attention weights to the high-frequency noise components, and reconstruct the target gas concentration gradient field in combination with the migration learning-driven chromatographic response migration model; Dynamically optimize the adsorption coefficient matrix in the gas adsorption kinetic parameter library through the meta-learning framework, and generate pure component spectral features after stripping cross-interference.

[0008] Further, the method for locating and tracing the leakage of hazardous chemicals in the gas array of the present invention further includes: Based on the turbulent vortex distribution map output by computational fluid dynamics simulation, a vortex propagation delay model between sensor nodes is constructed; The improved particle swarm optimization algorithm is used to iteratively solve the phase shift parameter in the vortex propagation delay model, and time domain alignment is performed on the reconstructed target gas concentration gradient field; A multi-node concentration gradient time series signal set synchronized in space and time is generated.

[0009] Further, the method for locating and tracing the leakage of hazardous chemicals in the gas array of the present invention further includes: The discretization constraint conditions of the computational fluid dynamics mass conservation equation are embedded in the hidden layer of the Bayesian and LSTM networks; The vortex propagation direction features in the multi-node concentration gradient time series signal set are extracted through a spatio-temporal attention mechanism; The projection gradient descent algorithm is used to perform a divergence constraint on the output tensor of the LSTM unit, forcing the prediction result to satisfy the law of mass conservation.

[0010] Further, in the method for locating and tracing the leakage of hazardous chemicals in the gas array of the present invention, step 6 includes: The posterior probability distribution residual output by the Bayesian and LSTM hybrid neural network is input into the positioning error mapping network to generate a positioning uncertainty quantification index; In the integrated learning framework, the posterior probability data inferred by Bayesian inference, the time series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation are fused; The NSGA-II multi-objective optimization algorithm is used to perform multi-dimensional weight allocation on the posterior probability data, the time series diffusion trajectory, and the physical field constraint results to generate a Pareto front solution set.

[0011] Further, the method for locating and tracing the leakage of hazardous chemicals in the gas array of the present invention further includes: According to the deviation distribution of the credibility rating positioning result and the real-time monitoring residual signal, the sensor sensitivity parameters and the hyperparameters of the Bayesian and LSTM hybrid neural network are dynamically adjusted through the meta-reinforcement learning framework. At the same time, based on the neural architecture search technology, the network depth is reconstructed to form a closed-loop adaptive learning mechanism; According to the residual distribution characteristics of the Pareto front solution set, the sensor sensitivity adjustment strategy and the hyperparameter optimization scheme of the Bayesian and LSTM networks are matched from the knowledge base; The number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network is adaptively adjusted through the neural architecture search technology; Optimizing the boundary condition parameters of the turbulent vortex intensity in the computational fluid dynamics model based on the differential evolution algorithm.

[0012] Furthermore, for the method for locating and tracing the leakage of hazardous chemicals in the gas array storage of the present invention, step 4 includes: Invoking the spatio-temporal trajectory data of historical leakage events stored in the knowledge base to construct a training sample set for the generative adversarial network; Synthesizing a diffusion feature tensor by aligning the multi-node concentration gradient time series signals in the time domain with the mass conservation equation constraint; Outputting a thermal map of the diffusion path with vortex propagation direction markers through the discriminator module of the generative adversarial network.

[0013] Furthermore, for the method for locating and tracing the leakage of hazardous chemicals in the gas array storage of the present invention, it further includes: Training the residual distribution model of the positioning error mapping network based on the three-dimensional reconstruction data of historical leakage scenarios; Performing Monte Carlo sampling on the posterior probability distribution output by the hybrid Bayesian and LSTM neural network to extract the boundary features of the spatial confidence interval; Overlaying the thermal map of the diffusion path and the spatial confidence interval features and visualizing the output as a probability cloud map.

[0014] Furthermore, for the method for locating and tracing the leakage of hazardous chemicals in the gas array storage of the present invention, step 1 includes: A distributed array of metal oxide semiconductor sensors and photoionization detectors, suppressing the cross-sensitivity effect through a transfer learning model; An integrated temperature and humidity sensor and a synchronous acquisition module of a three-dimensional ultrasonic anemometer, real-time outputting multi-physical field parameters with time stamp alignment; An anti-interference coating provided in the electrochemical sensor array and the chromatographic response transfer model cooperate to strip the interference component spectra.

[0015] Advantages of the present invention; Through the multi-modal data acquisition of the heterogeneous gas sensor array and the dynamic noise suppression of transfer learning, the present invention eliminates the temperature drift and cross-sensitivity interference in complex environments and restores the original features of the gas concentration signal. By combining computational fluid dynamics simulation and improved particle swarm optimization algorithm, the turbulent lag effect between sensor nodes is quantified, and the diffusion path map constrained by physical laws is reconstructed. Based on the Bayesian and LSTM hybrid neural network embedded with the hard constraint of the mass conservation equation, the projection gradient descent is used to force the model output to conform to the hydrodynamic laws, solving the problem that the data-driven model violates the physical field evolution in the turbulent environment. The NSGA-II multi-objective optimization algorithm is used to fuse multi-source heterogeneous data, and combined with the residual distribution-location error mapping network verification, a Pareto optimal solution set with credibility ratings is generated, forming a dual verification mechanism of data-driven and physical laws. The closed-loop adaptive learning framework dynamically adjusts the model structure and boundary conditions through neural architecture search and differential evolution algorithm, realizing the continuous optimization of system parameters during long-term operation. Finally, millimeter-level leakage source localization accuracy is achieved in the hazardous chemical storage scenario, significantly improving the robustness and reliability of inverse inversion in complex multi-physical field coupling environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0017] Figure 1 It is a flowchart of the method for leakage location and traceability of hazardous chemical storage of the gas array provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. To better understand the objectives of the present invention, the present invention is further described in detail below.

[0019] Please refer to Figure 1 , the method for leakage location and traceability of hazardous chemical storage of the gas array provided by the present invention includes: Step 1, real-time collect the gas concentration, components, and environmental physical field parameters of the storage environment through a heterogeneous gas sensor array deployed in three-dimensional space, and generate a multi-modal original data set including gas concentration time series signals, multi-component spectral features, and three-dimensional wind speed fields. Step 2: Input the multi-modal original dataset into a pre-established knowledge base, perform cross-scenario domain adaptation pre-training on the sensor array through a transfer learning algorithm, generate an adaptive baseline correction signal based on the dynamic matching between the real-time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and output a calibrated multi-physical field dataset after environmental noise suppression and sensor zero-drift compensation; Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physical field dataset, allocate noise suppression weights using a bidirectional gated attention mechanism based on the turbulence-diffusion correlation map in the knowledge base, and simultaneously call the gas adsorption kinetic parameter library to construct a transfer learning-driven chromatographic response transfer model, and output the denoised concentration gradient signal of the target gas and the separated component feature vectors; Step 4: Quantify the response lag effect between sensor nodes based on the turbulence vortex distribution map generated by computational fluid dynamics dynamic simulation, use the particle swarm optimization algorithm to iteratively solve the time delay compensation coefficient and align the multi-node time series signals, and combine the generative adversarial network to synthesize the compensated concentration gradient data with the physical field constraints into a reference diffusion feature vector to generate a gas diffusion path map under multi-physical field coupling conditions; Step 5: Embed the computational fluid dynamics mass conservation equation as a hard constraint into a hybrid neural network of Bayesian and LSTM, jointly optimize the network parameters through the projected gradient descent algorithm, and generate a posterior probability distribution of the leakage source coordinates that satisfies the hydrodynamic laws and the measured data characteristics; Step 6: Input the posterior probability distribution output by the hybrid neural network of Bayesian and LSTM, the computational fluid dynamics simulation results, and the LSTM prediction data into an ensemble learning framework, use the NSGA-II multi-objective optimization algorithm to dynamically allocate model weights and screen the Pareto optimal solution set, and generate a credibility rating localization result verified by a residual distribution-localization error mapping network.

[0020] The method for locating and tracing the leakage of hazardous chemicals in a gas array provided by the present invention realizes the precise localization of the leakage source in a complex environment through multi-modal data acquisition, signal processing, and physical model fusion. The following is a detailed description of each step of the technical solution: In Step 1, the gas concentration, components, and environmental physical field parameters of the storage environment are collected in real time through a heterogeneous gas sensor array deployed in a three-dimensional space. The heterogeneous gas sensor array consists of a metal oxide semiconductor sensor, a photoionization detector, and an electrochemical sensor with an anti-interference coating, which are distributed at key monitoring nodes in the storage space to synchronously collect gas concentration time series signals, multi-component spectral characteristics, and three-dimensional wind speed field data. The temperature and humidity sensor and the three-dimensional ultrasonic anemometer are integrated into the same acquisition module to ensure the time stamp alignment of multi-physical field parameters, forming a multi-modal original dataset with spatio-temporal correlation.

[0021] In Step 2, the multi-modal original data set is input into a preset knowledge base, and the sensor array is pre-trained for cross-scenario domain adaptation through a transfer learning algorithm. The knowledge base stores the noise pattern feature library of historical leakage scenarios. Through the dynamic matching of the real-time environmental parameter spectrum and the feature library, an adaptive baseline correction signal based on domain adaptation is generated. This signal compensates for the temperature drift component in the air pressure pulsation time series signal, eliminates the sensor zero drift and environmental mutation noise, and outputs a calibrated multi-physical field data set.

[0022] In Step 3, multi-scale wavelet packet decomposition and gas group decoupling processing are performed on the calibrated data set output in Step 2. The Daubechies wavelet packet decomposition algorithm is used to separate the high-frequency noise component and the low-frequency diffusion component of the gas concentration time series signal. Combining the turbulence and diffusion correlation maps in the knowledge base, the noise suppression weights of each frequency band are dynamically allocated through a bidirectional gated attention mechanism. At the same time, the gas adsorption kinetic parameter library is called to construct a transfer learning-driven chromatographic response transfer model, stripping the spectral features of multi-component gas cross-interference, and reconstructing the noise-reduced concentration gradient field and pure component feature vector of the target gas.

[0023] In Step 4, based on the turbulence vortex distribution map generated by computational fluid dynamics dynamic simulation, the response lag effect between sensor nodes is quantified. A vortex propagation delay model between sensor nodes is constructed, and the improved particle swarm optimization algorithm is used to iteratively solve the phase shift parameters to perform time domain alignment on the reconstructed target gas concentration gradient field. The compensated concentration gradient data and physical field constraints are used to synthesize a benchmark diffusion feature vector through a generative adversarial network, generating a spatio-temporal correlation map representing the gas diffusion path under multi-physical field coupling conditions.

[0024] In Step 5, the computational fluid dynamics mass conservation equation is embedded as a hard constraint into the hybrid neural network of Bayesian and LSTM. Prior data on the spatial distribution of the leakage source is introduced into the Bayesian framework, and the LSTM branch learns the diffusion time series correlation rules through a spatio-temporal attention mechanism. The projection gradient descent algorithm jointly optimizes the network parameters, forcing the model output to satisfy the mass conservation law and the measured data characteristics, and generating the posterior probability distribution of the leakage source coordinates.

[0025] In Step 6, the posterior probability distributions of the Bayesian and LSTM networks, the computational fluid dynamics simulation results, and the time series diffusion trajectories predicted by LSTM are fused through an ensemble learning framework. The NSGA-II multi-objective optimization algorithm is used to perform multi-dimensional weight allocation on heterogeneous data sources, screening the Pareto optimal solution set. The residual distribution-location error mapping network performs uncertainty quantification verification on the solution set, and outputs a positioning result with a credibility rating, providing a benchmark for closed-loop adaptive learning.

[0026] A one-way progressive and reversely amendable data processing chain is formed among the steps. The multi-modal data of Step 1 provides input for the noise suppression of Step 2, and the noise-reducing component features of Step 3 lay the foundation for the spatio-temporal alignment of Step 4. The physical constraint inversion of Step 5 and the multi-model fusion of Step 6 cooperate to improve the positioning accuracy, and finally, the system parameters are dynamically optimized through a closed-loop mechanism. The technical solution solves the problem of signal distortion caused by complex multi-physical field coupling through the dual constraints of data-driven and physical laws, and realizes the accurate traceability of the source of hazardous chemical leaks.

[0027] Specifically, in the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention, the environmental physical field parameters include a three-dimensional wind speed field, a temperature gradient distribution, and a barometric pressure pulsation time series signal; The three-dimensional wind speed field and the temperature gradient distribution are input into the noise pattern feature library in the knowledge base for spectral similarity comparison; An adaptive filtering signal based on transfer learning domain adaptation is generated according to the comparison result, and temperature drift compensation is performed on the barometric pressure pulsation time series signal, and a corrected three-dimensional physical field parameter set is output.

[0028] In the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention, the acquisition and baseline drift correction of environmental physical field parameters constitute the core link of data preprocessing. The measurement of the three-dimensional wind speed field is realized by a distributed three-dimensional ultrasonic anemometer. Based on the Doppler effect, the air flow vector distribution in the storage space is captured, and a wind speed field matrix with spatio-temporal correlation is generated. The temperature gradient distribution data is synchronously collected by an embedded high-precision temperature sensor array. The sensor nodes are deployed in layers along the vertical and horizontal directions, and the spatial gradient change characteristics of the temperature field are output in real time. The barometric pressure pulsation time series signal is obtained by a micro differential pressure sensor with anti-interference design, and the sampling frequency is synchronized with the wind speed and temperature data to form a multi-physical field original data set with time stamp alignment.

[0029] During the baseline drift correction process, the historical noise pattern feature library stored in the knowledge base contains the spectral features of sensor baseline drift under different environmental conditions. After the three-dimensional wind speed field and the temperature gradient distribution data are input into the feature library, scene domain invariant features are extracted through a transfer learning algorithm, and a spectral similarity comparison model is constructed. The model identifies the steady-state and transient components in the real-time environmental parameters and generates an adaptive filtering signal based on domain adaptation. The filtering signal acts on the barometric pressure pulsation time series signal, performs phase compensation on the sensor zero drift component caused by sudden temperature changes, eliminates the non-linear error in the barometric pressure data caused by the thermodynamic effect, and outputs a three-dimensional physical field parameter set corrected in space and time.

[0030] The calibrated three-dimensional physical field parameter set provides high-confidence input for subsequent signal processing. The coupling relationship between the temperature gradient data and the wind speed field matrix is verified through the turbulence evolution model in the knowledge base to ensure the dynamic consistency of the physical field parameters. The compensation result of the air pressure pulsation signal is associated with the residual distribution characteristics of the transfer learning model, and the noise pattern feature library in the knowledge base is dynamically updated to form a closed-loop optimization mechanism for environmental parameter calibration. The technical solution lays a data foundation for gas diffusion feature reconstruction and leakage source inversion through multi-dimensional collaborative calibration of physical field parameters.

[0031] Specifically, the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention further includes: Perform Daubechies wavelet packet decomposition on the gas concentration time series signal in the calibrated three-dimensional physical field parameter set, and extract the high-frequency noise component and the low-frequency diffusion component; Call the turbulence-diffusion correlation map in the knowledge base to assign attention weights to the high-frequency noise components, and reconstruct the target gas concentration gradient field in combination with the chromatographic response transfer model driven by transfer learning; Dynamically optimize the adsorption coefficient matrix in the gas adsorption kinetics parameter library through the meta-learning framework to generate pure component spectral features after stripping cross-interference.

[0032] In the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention, step 3 realizes gas concentration field reconstruction and component separation through multi-scale signal decomposition and dynamic parameter optimization. After the baseline drift correction in step 2, the gas concentration time series signal is input into the multi-scale wavelet packet decomposition module. The signal is decomposed into five layers using the Daubechies wavelet basis function to extract the high-frequency noise component in the frequency band of 0.1 - 10 Hz and the low-frequency diffusion component in the frequency band of 0.01 - 0.1 Hz. During the decomposition process, the effective frequency band boundary is determined through the energy entropy threshold, and the low-frequency signal components associated with the turbulence diffusion characteristics are retained.

[0033] The suppression of the high-frequency noise component is achieved through the turbulence-diffusion correlation map in the knowledge base. The correlation map stores the mapping relationship between the turbulence intensity and the noise spectrum in historical leakage scenarios, and matches the similar noise patterns based on the three-dimensional wind speed field data of the current environment. The bidirectional gated attention mechanism dynamically assigns the noise suppression weights for each frequency band according to the matching result, and adaptively filters the high-frequency components. The filtered low-frequency diffusion component is coupled with the chromatographic response transfer model driven by transfer learning to reconstruct the target gas concentration gradient field.

[0034] The training of the chromatographic response migration model relies on the gas adsorption kinetic parameter library in the knowledge base. The parameter library contains the adsorption isotherms of multi-component gases on the sensor surface and the reaction activation energy data. The model migrates the adsorption characteristics of historical scenarios to the current environment through a transfer learning strategy to construct a component separation matrix. The meta-learning framework optimizes the cross-sensitivity parameters in the adsorption coefficient matrix in real time, and combines the spatio-temporal modulation data of the sensor array to strip the spectral response characteristics of interfering gases. The optimized parameter matrix is fused with the multi-scale decomposition results to generate a pure component spectral feature vector with spatial continuity.

[0035] Each technical link forms a progressive processing chain. Wavelet packet decomposition provides a frequency-domain analysis basis for noise suppression, and the attention mechanism allocates weights to enhance the ability to retain turbulent correlation features. The synergistic effect of the transfer learning model and the meta-learning framework realizes the generalization of the component separation ability across scenarios. The reconstructed concentration gradient field and the pure spectral features provide high signal-to-noise ratio inputs for the spatio-temporal lag compensation in Step 4, supporting the accurate modeling of subsequent diffusion paths. The technical solution solves the problem of concentration field distortion caused by multi-component cross-interference through multi-dimensional coordination of signal decomposition and parameter optimization.

[0036] Specifically, the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention further includes: Based on the turbulent vortex distribution map output by computational fluid dynamics simulation, a vortex propagation delay model between sensor nodes is constructed; The improved particle swarm optimization algorithm is used to iteratively solve the phase shift parameter in the vortex propagation delay model to perform time-domain alignment on the reconstructed target gas concentration gradient field; A spatio-temporally synchronized multi-node concentration gradient time series signal set is generated.

[0037] In the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention, Step 4 realizes the spatio-temporal alignment of gas diffusion characteristics through physical field modeling and signal time series correction. Based on the target gas concentration gradient field output in Step 3, a computational fluid dynamics (CFD) dynamic simulation engine is called, and a turbulent vortex distribution map is generated in combination with the three-dimensional topological structure and environmental parameters of the storage space. The map quantifies the non-linear perturbation effect of the gas diffusion path between sensor nodes with vortex intensity, propagation direction, and energy dissipation rate as characteristic dimensions.

[0038] The construction of the vortex propagation delay model relies on the spatio-temporal evolution characteristics of the turbulent vortex distribution atlas. The model maps the spatial coordinates of sensor nodes and the vortex propagation direction into a delay function, and establishes the phase shift relationship of signal propagation between nodes. The improved particle swarm optimization algorithm introduces an inertial weight adaptive adjustment strategy, and takes the vortex energy attenuation coefficient as a constraint condition to iteratively solve the optimal solution set of the phase shift parameters. During the parameter optimization process, the time-domain alignment of the target gas concentration gradient field is realized by dynamic matching of sliding windows, eliminating the signal time-series misalignment caused by the turbulence of the storage structure.

[0039] The spatially continuous reconstruction of the multi-node concentration gradient time-series signal set after time-domain alignment is carried out by a spatio-temporal interpolation algorithm. The interpolation algorithm fuses the physical field constraints of CFD simulation and the measured concentration gradient data to generate a spatio-temporally synchronized signal set. The synchronized signal set retains the time-varying characteristics and spatial correlation of gas diffusion, providing high-precision input for the leakage source inversion in step 5. The technical solution solves the problem of time-series signal distortion caused by sensor response lag through the joint optimization of physical field modeling and data-driven, and improves the reliability of diffusion path reconstruction.

[0040] The synergistic effect of each technical link is reflected in that: CFD simulation provides the physical field evolution rules for the delay model, the improved optimization algorithm realizes signal alignment through dynamic parameter adjustment, and spatio-temporal interpolation enhances the spatial consistency of data. This processing chain effectively suppresses the interference of complex turbulent environments on the spatio-temporal correlation of sensor signals, laying a data foundation for subsequent inverse inversion.

[0041] Specifically, the method for locating and tracing the leakage of hazardous chemicals in a storage using a gas array according to the present invention further includes: Embedding the discretized constraint conditions of the computational fluid dynamics mass conservation equation into the hidden layer of the Bayesian and LSTM networks; Extracting the vortex propagation direction features in the multi-node concentration gradient time-series signal set through a spatio-temporal attention mechanism; Using the projection gradient descent algorithm to perform a divergence constraint on the output tensor of the LSTM unit, forcing the prediction result to satisfy the law of mass conservation.

[0042] In the method for locating and tracing the leakage of hazardous chemicals in a storage using a gas array according to the present invention, the hybrid neural network architecture in step 4 improves the leakage source inversion accuracy through the deep integration of physical constraints and data-driven models. In the hidden layer of the Bayesian and LSTM networks, the computational fluid dynamics mass conservation equation is discretized into partial differential constraint conditions and embedded in the forward propagation process of the network in the form of tensors. The discretized constraint conditions are generated based on the grid division of the storage space, and the residual of the mass conservation equation at each grid node is added to the loss function as a regularization term, enabling the network to implicitly learn the physical laws of fluid motion.

[0043] The spatio-temporal attention mechanism extracts features from the multi-node concentration gradient time-series signal set generated in step 4. The attention weight calculation module combines the vortex propagation direction feature and the concentration gradient change rate to dynamically allocate the contribution degrees of the time-series signals of different sensor nodes. The direction feature extracts the local vortex vectors in the diffusion path heat map through a convolutional neural network and performs a Hadamard product operation with the attention weight matrix to enhance the model's perception ability of the turbulence-dominated diffusion path.

[0044] The projected gradient descent algorithm acts on the output tensor of the LSTM unit, and forces the prediction result to satisfy the law of mass conservation through the divergence constraint of the flow field. The divergence constraint is constructed based on the discretized mass conservation equation, and maps the divergence residual of the network output tensor to the orthogonal basis of the projection space. During the gradient descent process, the update direction of the network parameters is constrained within the orthogonal complement space of the residual, so that the predicted leakage source distribution conforms to both physical laws and the characteristics of the measured data.

[0045] The logical association of the technical solution is reflected as follows: the physical constraint is embedded to provide the basic law of fluid motion for the attention mechanism, the attention weight optimizes the spatio-temporal correlation of signal feature extraction, and the projected gradient descent forces the model output to be consistent with the physical field evolution through mathematical optimization. The embedding method of the discretized constraint conditions avoids directly modifying the network structure and ensures the trainability of the model; the synergistic effect of the attention mechanism and the gradient projection solves the problem that the data-driven model violates physical laws in a complex turbulent environment.

[0046] The above processing chain forms a leakage source localization mechanism with enhanced physical information. By constraining the consistency between the mathematical characteristics of the network output and the laws of fluid mechanics, the robustness of the inverse inversion model under noise interference is significantly improved, providing technical support for generating a high-confidence posterior probability distribution in step 5.

[0047] Specifically, for the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention, step 6 includes: Input the posterior probability distribution residual output by the Bayesian and LSTM hybrid neural network into the localization error mapping network to generate a localization uncertainty quantification index; Fuse the posterior probability data inferred by Bayesian inference, the time-series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation in an ensemble learning framework; Perform multi-dimensional weight allocation on the posterior probability data, the time-series diffusion trajectory, and the physical field constraint results through the NSGA-II multi-objective optimization algorithm to generate a Pareto front solution set.

[0048] In the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention, in step 6, the credibility verification of the leakage source location result is achieved through multi-source data fusion and multi-objective optimization. The residual of the posterior probability distribution output by the Bayesian and LSTM hybrid neural network characterizes the deviation between the model prediction and the measured data. After being input into the location error mapping network, the network generates a location uncertainty quantification index based on the residual distribution pattern of historical leakage scenarios. This index extracts the spatio-temporal correlation features of the residual through a convolutional neural network, and combines the error-environment parameter mapping relationship in the knowledge base to output the confidence weight matrix of each sensor node.

[0049] The integrated learning framework receives three types of heterogeneous data sources: the posterior probability data inferred by Bayesian inference, the temporal diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation. The framework adopts a feature-level fusion strategy to uniformly map the spatial confidence of the posterior probability distribution, the temporal correlation of the LSTM trajectory, and the direction constraint of the CFD physical field to a high-dimensional feature space. The feature space dynamically assigns the contribution weights of each data source through a self-attention mechanism to generate a fused joint decision vector, which characterizes the cross-model consistency features of the leakage source location.

[0050] The NSGA-II multi-objective optimization algorithm performs multi-dimensional weight allocation on the joint decision vector, and the optimization objectives cover location accuracy, model consistency, and computational efficiency. The algorithm screens the Pareto front solution set based on the non-dominated sorting strategy, and each candidate point in the solution set needs to meet the posterior probability confidence threshold, the physical field divergence constraint, and the temporal trajectory continuity condition. The residual distribution-location error mapping network performs a secondary verification on the solution set, eliminates abnormal solutions that violate the historical leakage pattern or environmental parameter constraints, and generates the final credibility-rated location result.

[0051] The logical chain of the technical solution is as follows: Residual analysis quantifies the uncertainty of model prediction and provides a credibility benchmark for data fusion; the integrated framework coordinates the complementarity of multi-source data through feature space mapping; multi-objective optimization balances location accuracy and model robustness under constraint conditions. The verification mechanism of the location error mapping network and the screening strategy of NSGA-II form a double verification to ensure that the output results conform to both data-driven features and physical laws at the same time. The above processing flow solves the problem of insufficient credibility of the single-model location result in a complex environment through multi-level verification and optimization, and provides highly reliable input for closed-loop adaptive learning.

[0052] Specifically, the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention further includes: According to the deviation distribution between the credibility-rated location result and the real-time monitoring residual signal, the sensor sensitivity parameters and the hyperparameters of the Bayesian and LSTM hybrid neural network are dynamically adjusted through a meta-reinforcement learning framework. At the same time, the network depth is reconstructed based on the neural architecture search technology to form a closed-loop adaptive learning mechanism; According to the residual distribution characteristics of the Pareto front solution set, match the sensor sensitivity adjustment strategy and the hyperparameter optimization scheme of the Bayesian and LSTM networks from the knowledge base; Adopt neural architecture search technology to adaptively adjust the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network; Optimize the boundary condition parameters of the turbulent vortex intensity in the computational fluid dynamics model based on the differential evolution algorithm.

[0053] In the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention, the closed-loop adaptive learning mechanism realizes the continuous optimization of the system through dynamic parameter adjustment and model reconstruction. The deviation distribution between the confidence rating positioning result and the real-time monitoring residual signal is input into the meta-reinforcement learning framework, and the framework matches the current environmental state based on the historical optimization strategy library in the knowledge base. The policy network of the meta-reinforcement learning dynamically adjusts the sensor sensitivity parameters through a multi-objective reward function, including the response gain and sampling frequency of the gas sensor, and simultaneously optimizes the regularization coefficient and learning rate hyperparameters in the Bayesian and LSTM hybrid neural network to balance the model generalization ability and fitting accuracy.

[0054] Neural architecture search technology adaptively adjusts the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network. The search space is defined as a combination of network depth and the number of hidden units, and a network structure candidate set is generated based on the real-time residual distribution characteristics. The search process uses a weight sharing strategy to evaluate the validation loss of different structures, and combines a gradient direction optimizer to screen for lightweight network architectures that meet the computational efficiency constraints, dynamically expanding or compressing the temporal modeling ability of the LSTM branch.

[0055] The sensor sensitivity adjustment strategy and the Bayesian and LSTM hyperparameter optimization scheme stored in the knowledge base are matched through the residual distribution characteristics of the Pareto front solution set. The matching algorithm quantifies the relevance between the residual pattern and the historical scenario based on cosine similarity, and extracts the optimal parameter adjustment template. The combination of sensitivity parameters and hyperparameters in the template is migrated to the current system configuration through a transfer learning framework to avoid the parameter optimization falling into a local optimum.

[0056] The boundary condition parameters of the turbulent vortex intensity in the computational fluid dynamics model are optimized by the differential evolution algorithm. The algorithm uses the residual between the diffusion path map generated in step 4 and the measured concentration gradient field as the fitness function to iteratively update the population distribution of the vortex intensity parameters. A physical field divergence constraint condition is introduced during the population update process to force the optimized boundary parameters to satisfy the mass conservation equation, ensuring the consistency between the CFD simulation results and the measured data.

[0057] The logical association of the technical solution is reflected in that the meta-reinforcement learning framework dynamically adjusts the front-end sensor and the back-end model parameters according to the real-time deviation, the neural architecture search optimizes the network structure to adapt to the dynamic environment, the knowledge base matching strategy improves the parameter migration efficiency, and the differential evolution algorithm ensures the accuracy of the boundary conditions of the physical model. The above steps form a closed-loop mechanism for collaborative optimization of the inner and outer loops. The inner loop optimizes the sensor and model parameters, and the outer loop reconstructs the network structure and the physical field boundary, and continuously improves the system performance through two-way feedback.

[0058] The closed-loop adaptive mechanism solves the problem of model degradation caused by the dynamic change of the environment through a multi-level optimization strategy, ensuring that the leakage location system maintains high accuracy and robustness during long-term operation. The synergistic effect of parameter adjustment and model reconstruction provides dynamic adaptation capabilities for the full-process data processing from step 1 to step 6, strengthening the practicality of the technical solution in complex multi-physical field coupling scenarios.

[0059] Specifically, in the method for leakage location and traceability of hazardous chemical storage of the gas array described in the present invention, step 4 includes: Calling the spatio-temporal trajectory data of historical leakage events stored in the knowledge base to construct a training sample set for the generative adversarial network; Synthesizing a diffusion feature tensor by aligning the multi-node concentration gradient time series signal set in the time domain with the mass conservation equation constraint; Outputting a diffusion path heat map with vortex propagation direction markers through the discriminator module of the generative adversarial network.

[0060] In the method for leakage location and traceability of hazardous chemical storage of the gas array described in the present invention, step 4 realizes accurate modeling of the gas diffusion path through the fusion of historical data-driven and physical constraints. The spatio-temporal trajectory data of historical leakage events stored in the knowledge base is structurally processed, and the combined features of the leakage source coordinates, diffusion time series, and environmental parameters are extracted to form a multi-dimensional spatio-temporal trajectory matrix. The matrix supplements missing data points through an interpolation algorithm and injects Gaussian noise to enhance sample diversity, constructing a training sample set for the generative adversarial network. The sample set covers diffusion scenarios under different storage structures, leakage intensities, and environmental disturbance patterns, providing cross-scenario generalization capabilities for the generative adversarial network.

[0061] The synthesis process of the multi-node concentration gradient time series signal set aligned in the time domain and the mass conservation equation constraint is realized by using tensor fusion technology. The concentration gradient time series signal set is encoded into a spatio-temporal feature map through three-dimensional convolution, and the mass conservation equation is discretized into divergence constraint conditions for spatial grid nodes. The feature map and the divergence constraint matrix are fused through Hadamard product operation to generate a diffusion feature tensor enhanced with physical information. The physical constraint conditions embedded in the tensor force the diffusion path to conform to the laws of fluid mechanics and suppress path distortion caused by data noise.

[0062] The discriminator module of the generative adversarial network receives the diffusion feature tensor as input and extracts the local and global features of the diffusion path through a multi-level convolutional network. The output layer of the discriminator combines the vortex propagation direction classifier and the heat value regressor to generate a heat map of the diffusion path with direction markers. The direction markers are divided based on the angular distribution of the vortex vector field, and the heat value represents the cumulative intensity of the concentration gradient at the spatial grid nodes. The generated map is fed back to the generator module through residual connections to iteratively optimize the authenticity and physical consistency of the path generation.

[0063] The logical relationship of the technical solution is manifested as follows: The training sample set constructed from historical data enhances the scenario adaptation ability of the generative adversarial network. The fusion of physical constraints and real-time data ensures the regular consistency of the diffusion feature tensor. The multi-task output of the discriminator realizes the joint modeling of path visualization and direction marking. The generated heat map of the diffusion path provides a high-fidelity input for the leakage source inversion in step 5 and simultaneously supports the credibility rating verification in step 6. Driven by both historical experience and physical laws, the technical solution effectively solves the problem of distortion in the reconstruction of diffusion paths in complex turbulent environments.

[0064] Specifically, the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention further includes: Training the residual distribution model of the positioning error mapping network based on the three-dimensional reconstruction data of historical leakage scenarios; Performing Monte Carlo sampling on the posterior probability distribution output by the Bayesian and LSTM hybrid neural network to extract the boundary features of the spatial confidence interval; Overlaying the heat map of the diffusion path and the spatial confidence interval features and visualizing the output as a probability cloud map.

[0065] In the method for locating and tracing the leakage of hazardous chemicals in a gas array according to the present invention, step 4 realizes the accurate modeling of the gas diffusion path through the fusion of historical data-driven and physical constraints. The spatio-temporal trajectory data of historical leakage events stored in the knowledge base is structurally processed to extract the combined features of the leakage source coordinates, diffusion time series, and environmental parameters, forming a multi-dimensional spatio-temporal trajectory matrix. The matrix supplements the missing data points through an interpolation algorithm and injects Gaussian noise to enhance the sample diversity, constructing a training sample set for the generative adversarial network. The sample set covers diffusion scenarios under different storage structures, leakage intensities, and environmental perturbation patterns, providing cross-scenario generalization ability for the generative adversarial network.

[0066] The synthesis process of the multi-node concentration gradient time-series signal set after time-domain alignment and the mass conservation equation constraint is realized by tensor fusion technology. The concentration gradient time-series signal set is encoded into a spatio-temporal feature map through three-dimensional convolution, and the mass conservation equation is discretized into divergence constraint conditions for spatial grid nodes. The feature map and the divergence constraint matrix are fused through Hadamard product operation to generate a diffusion feature tensor with enhanced physical information. The physical constraint conditions embedded in the tensor force the diffusion path to conform to the laws of hydrodynamics and suppress the path distortion caused by data noise.

[0067] The discriminator module of the generative adversarial network receives the diffusion feature tensor as input and extracts the local and global features of the diffusion path through a multi-level convolutional network. The output layer of the discriminator combines the vortex propagation direction classifier and the heat value regressor to generate a heat map of the diffusion path with direction markings. The direction markings are divided based on the angular distribution of the vortex vector field, and the heat value represents the cumulative intensity of the concentration gradient at the spatial grid nodes. The generated map is fed back to the generator module through residual connections to iteratively optimize the authenticity and physical consistency of path generation.

[0068] The logical association of the technical solution is as follows: The training sample set constructed from historical data enhances the scenario adaptation ability of the generative adversarial network. The fusion of physical constraints and real-time data ensures the law consistency of the diffusion feature tensor. The multi-task output of the discriminator realizes the joint modeling of path visualization and direction marking. The generated heat map of the diffusion path provides a high-fidelity input for the leakage source inversion in step 5 and simultaneously supports the credibility rating verification in step 6. Driven by both historical experience and physical laws, the technical solution effectively solves the problem of distortion in diffusion path reconstruction in complex turbulent environments.

[0069] Specifically, for the method for locating and tracing the leakage of hazardous chemicals in a gas array described in the present invention, step 1 includes: A distributed array of metal oxide semiconductor sensors and photoionization detectors suppresses cross-sensitivity effects through a transfer learning model; A synchronous acquisition module integrating temperature and humidity sensors and three-dimensional ultrasonic anemometers outputs multi-physical field parameters with aligned timestamps in real time; An anti-interference coating provided in the electrochemical sensor array and the chromatographic response transfer model cooperate to strip the spectra of interfering components.

[0070] In the method for locating and tracing the leakage of hazardous chemicals in the gas array described in the present invention, in step 1, high-precision data acquisition in a complex environment is achieved through multi-modal sensor collaborative deployment and signal optimization technology. A distributed array of metal oxide semiconductor sensors and photoionization detectors is arranged at key monitoring nodes in the storage space, and the cross-sensitivity effect of multi-component gases is suppressed through a transfer learning model. The transfer learning model extracts domain-invariant feature parameters based on the sensor response feature library of different scenarios in the knowledge base, dynamically adjusts the sensitivity weights of the sensor array, and weakens the spectral interference of non-target gases. The spatial interpolation algorithm is used between sensor nodes to compensate for blind area data and generate a concentration gradient distribution matrix covering the entire storage area.

[0071] The synchronous acquisition module integrating temperature and humidity sensors and three-dimensional ultrasonic anemometers realizes the timestamp alignment of multi-physical field parameters through the hardware clock synchronization protocol. The temperature and humidity sensors adopt a multi-point calibration strategy to eliminate the influence of environmental temperature drift on gas concentration inversion; the three-dimensional ultrasonic anemometer captures the airflow vector distribution in real time based on the Doppler effect and generates a wind speed field matrix. After the output signals of the synchronous acquisition module are aligned in time sequence, they are fused with gas concentration data into a spatio-temporally correlated multi-physical field data set, providing a unified time reference for subsequent signal processing.

[0072] An anti-interference coating is set on the surface of the electrochemical sensor array. The coating material is designed based on the adsorption characteristics of the target gas, and the adsorption reaction of interfering components is inhibited through chemical modification. The anti-interference coating and the chromatographic response transfer model in step 3 act synergistically. The model calls the adsorption kinetic parameters in the knowledge base to dynamically correct the response deviation of the sensor surface caused by coating aging or environmental mutation. The dual mechanisms of coating physical isolation and algorithm compensation strip the spectral characteristics of interfering components and improve the reconstruction accuracy of the target gas concentration gradient.

[0073] The logical association of the technical solution is reflected as follows: the distributed sensor array optimizes the signal quality through transfer learning, the synchronous acquisition module ensures the spatio-temporal consistency of multi-source data, and the synergistic effect of the anti-interference coating and the chromatographic model strengthens the component separation ability. The physical characteristics of the sensor hardware and the dynamic correction of the algorithm model complement each other, solving the problems of multi-component cross-interference and asynchronous signal acquisition in a complex environment, and providing highly reliable input for noise suppression and baseline correction in step 2. The above technical chain lays the data acquisition foundation for the entire process of leakage location through the deep integration of hardware design and algorithm optimization. The explanations of the technical feature terms of the present invention are as follows: Heterogeneous Gas Sensor Array: A distributed network composed of multiple types of sensors (such as metal oxide semiconductor sensors, photoionization detectors, electrochemical sensors), deployed at key monitoring points in the storage space. Metal oxide sensors detect gas concentration based on surface resistance changes; photoionization detectors identify components by ultraviolet photoionizing gas molecules; electrochemical sensors measure the concentration of specific gases using redox reactions. The array covers monitoring blind spots through spatial complementary layout, combines transfer learning to suppress cross-sensitivity effects, and improves the detection robustness in multi-component mixed environments.

[0074] Transfer Learning Algorithm: A machine learning method that transfers the sensor response characteristics of historical scenarios in the knowledge base to the current environment. By extracting domain-invariant features across scenarios (such as temperature-sensitivity mapping relationships, noise spectrum patterns), it dynamically adjusts the weight of sensor parameters to eliminate baseline drift caused by environmental mutations. The algorithm constructs a noise pattern feature library using historical data in the pre-training stage, and matches the current environmental parameter spectrum during real-time operation to generate an adaptive filtering signal.

[0075] Computational Fluid Dynamics (CFD) Dynamic Simulation: Numerically simulates the gas flow in the storage space based on the Navier-Stokes equations, and outputs a map of turbulent vortex distribution. The map contains parameters such as vortex intensity, propagation direction, and energy dissipation rate, which are used to quantify the non-linear perturbations of the gas diffusion path. The simulation process couples the three-dimensional topological structure (such as shelf layout, vent position) with environmental parameters (wind speed, temperature) to generate a vortex evolution model under physical field constraints.

[0076] Improved Particle Swarm Optimization Algorithm: A swarm intelligence optimization algorithm that introduces an adaptive adjustment strategy for the inertia weight. The algorithm uses the phase shift of signal propagation between sensor nodes as the optimization variable and the vortex energy attenuation coefficient as the constraint condition, and iteratively solves the optimal solution of the time delay compensation parameter. The optimization process dynamically matches the concentration gradient field through a sliding window to eliminate the misalignment of time series signals caused by storage structure turbulence.

[0077] Hybrid Bayesian and LSTM Neural Network: A deep learning architecture that combines the Bayesian probability framework and the long short-term memory network (LSTM). The Bayesian layer introduces prior data on the spatial distribution of the leakage source to generate a posterior probability distribution; the LSTM branch learns the temporal correlation of the concentration gradient through a spatio-temporal attention mechanism. The discretized constraint terms of the mass conservation equation are embedded in the hidden layer of the network to force the output to satisfy the hydrodynamic laws.

[0078] NSGA-II Multi-objective Optimization Algorithm: Non-dominated Sorting Genetic Algorithm, used to solve multi-objective conflict problems. The algorithm takes positioning accuracy, model consistency, and computational efficiency as optimization objectives, and assigns weights to the results of Bayesian inference, LSTM prediction, and CFD simulation. By non-dominated sorting and crowding degree calculation, the Pareto front solution set is screened, and high-confidence candidate solutions are output.

[0079] Residual Distribution - Positioning Error Mapping Network: A verification model based on a convolutional neural network, with the input being the posterior probability residuals output by Bayesian and LSTM. The network extracts the spatio-temporal correlation features of the residuals through 3D convolutional layers, combines the error distribution patterns of historical leakage scenarios, and generates a quantization index of positioning uncertainty (such as 95% confidence interval) to support the credibility rating.

[0080] Neural Architecture Search Technology: An automated machine learning method for dynamically adjusting the number of layers and structure of Bayesian and LSTM networks. The search space is defined as the combination of the depth and the number of hidden units of the LSTM branch. The validation loss of candidate networks is evaluated through a weight sharing strategy, and lightweight architectures are screened in combination with a gradient optimizer to adapt to the dynamic changes of the environment.

[0081] Closed-loop Adaptive Learning Mechanism: A dynamic optimization system composed of meta-reinforcement learning and differential evolution algorithm. The meta-reinforcement learning framework adjusts the sensor sensitivity and model hyperparameters according to the real-time residual distribution; the differential evolution algorithm uses the residuals between CFD simulation and measured data as the fitness function to optimize the boundary conditions of the turbulent vortex intensity, forming a parameter-structure optimization chain with coordinated inner and outer loops.

[0082] Probability Cloud Map Visualization: A visual output that performs multi-channel fusion of the diffusion path heat map (concentration gradient intensity) and the spatial confidence interval (Monte Carlo sampling results). The heat value is encoded as the intensity of the red channel, and the confidence boundary is encoded as the transparency channel. The spatial continuity is enhanced through a 3D interpolation algorithm to generate an intuitive expression of the leakage source distribution and uncertainty.

[0083] Transfer Learning Model: The transfer learning model is used to suppress the cross-sensitivity effect of the sensor array. Based on the historical scenario data stored in the knowledge base, the model extracts the sensor response characteristics under different environmental conditions (such as temperature-sensitivity mapping relationship, noise spectrum pattern), and adjusts the sensitivity weights of the sensors through domain adaptation training. The model constructs a noise pattern feature library in the pre-training stage, matches the current environmental parameter spectrum during real-time operation, generates an adaptive filtering signal, and dynamically eliminates the baseline drift caused by sudden environmental changes (such as sudden temperature changes) to improve the signal stability in a multi-component mixed environment.

[0084] Generative Adversarial Network (GAN): The generative adversarial network is used to reconstruct the gas diffusion path map. The generator module receives the time-domain aligned concentration gradient time series signal and the discretization constraint of the mass conservation equation, and generates a diffusion feature tensor through three-dimensional convolutional encoding; the discriminator module combines the vortex propagation direction classifier and the thermal value regressor to output a thermal map of the diffusion path with direction markers. The network iteratively optimizes the generation result through the adversarial training mechanism to ensure that the diffusion path truly reflects the spatio-temporal correlation characteristics of the measured data while conforming to the laws of fluid mechanics.

[0085] Bayesian and LSTM Hybrid Neural Network: This network integrates the Bayesian probability framework and the Long Short-Term Memory (LSTM) model for the probabilistic inversion of the leakage source coordinates. The Bayesian layer introduces the prior data of the historical leakage source spatial distribution and generates the probability distribution of the leakage source coordinates through the posterior probability update mechanism; the LSTM branch extracts the temporal correlation of the concentration gradient signal through the spatio-temporal attention mechanism. The discretization constraint term of the computational fluid dynamics mass conservation equation is embedded in the network hidden layer, and the projection gradient descent algorithm is used to force the output to satisfy the flow field divergence condition, realizing the joint drive of physical laws and data characteristics.

[0086] NSGA-II Multi-Objective Optimization Algorithm: The Non-Dominated Sorting Genetic Algorithm is used for the fusion and optimization of multi-model results. The algorithm takes the positioning accuracy, model consistency, and computational efficiency as the optimization objectives, and performs multi-dimensional weight allocation on the Bayesian posterior probability distribution, the LSTM temporal diffusion trajectory, and the CFD physical field simulation results. The Pareto front solution set is screened through the non-dominated sorting strategy, and combined with the verification results of the residual distribution-location error mapping network, the abnormal solutions that violate the physical constraints or historical error patterns are eliminated, and a high-confidence candidate solution set is output.

[0087] Residual Distribution-Location Error Mapping Network: A verification model based on a convolutional neural network for quantifying the uncertainty of the positioning result. The input of the model is the posterior probability residual output by the Bayesian and LSTM, and the spatio-temporal correlation characteristics of the residual are extracted through the three-dimensional convolutional layer. Combining with the error distribution data of historical leakage scenarios, a quantization index of positioning uncertainty (such as spatial confidence interval) is generated. This index provides a statistical basis for the credibility rating and supports the interpretability of the positioning result.

[0088] Neural Architecture Search (NAS) Technology: An automated machine learning method for dynamically adjusting the structures of the Bayesian and LSTM networks. The search space is defined as the combination of the number of layers and the number of hidden units of the LSTM branch. The validation loss of the candidate network is evaluated through the weight sharing strategy, and the lightweight architecture is selected in combination with the gradient optimizer. The search process dynamically adjusts the network depth according to the real-time residual distribution, adapts to the complexity of the temporal characteristics caused by environmental changes, and improves the generalization ability of the model.

[0089] Differential Evolution Algorithm: A population optimization algorithm is used to update the boundary condition parameters of the computational fluid dynamics model. The algorithm takes the residual between the CFD simulation results and the measured concentration gradient field as the fitness function, and iteratively optimizes the population distribution of the turbulent vortex intensity parameters. During the population update process, the divergence constraint of the mass conservation equation is introduced to force the optimized boundary parameters to conform to the laws of fluid mechanics, ensuring the consistency between the physical model and the measured data.

[0090] In the specific implementation manner of the present invention, precise positioning of the leakage source in the hazardous chemical storage scenario is achieved through multi-level technology collaboration. An heterogeneous array consisting of metal oxide semiconductor sensors, photoionization detectors, and anti-interference coating electrochemical sensors is deployed inside the storage space, and the distributed nodes cover the gaps between shelves, ventilation openings, and the ground area. The sensor array dynamically adjusts the sensitivity weights through a transfer learning model, and the model training is based on the cross-sensitive feature dataset stored in the knowledge base to suppress the spectral interference of multi-component gases such as benzene series and hydrogen sulfide. The temperature and humidity sensor and the three-dimensional ultrasonic anemometer are integrated into the same acquisition module, and the millisecond-level timestamp alignment is achieved using the hardware clock synchronization protocol, outputting a multi-modal dataset containing three-dimensional wind speed field, temperature and humidity gradient, and gas concentration time series signals, providing the physical field input with spatio-temporal correlation for subsequent processing.

[0091] The multi-modal dataset is input into the preset knowledge base for baseline correction, and the transfer learning algorithm extracts the spectral features of the environmental parameters and matches the temperature drift template in the noise pattern feature library. The adaptive filtering signal performs phase compensation on the barometric pulsation time series data to eliminate the zero-point drift of the sensor caused by the diurnal temperature difference. The corrected data is decomposed by the Daubechies wavelet packet to separate the high-frequency noise and the low-frequency diffusion components, and combined with the attention weight assignment model in the turbulence and diffusion correlation map to reconstruct the target gas concentration gradient field. The chromatographic response transfer model calls the activation energy data in the adsorption kinetics parameter library, optimizes the adsorption coefficient matrix through the meta-learning framework, strips the spectral features of interfering components such as ethanol and acetone, and generates a pure concentration-component feature vector.

[0092] The computational fluid dynamics dynamic simulation engine generates a turbulent vortex distribution map based on the three-dimensional topology of the warehouse and constructs a vortex propagation delay model for sensor nodes. The improved particle swarm optimization algorithm uses the vortex energy attenuation coefficient as a constraint condition to iteratively solve the time delay compensation parameters and align the multi-node concentration gradient time series signals. The compensated data and the discretization constraints of the mass conservation equation generate diffusion feature tensors through a generative adversarial network, and generate a path heat map with vortex direction markings. The Bayesian and LSTM hybrid neural network embeds the regularization term of the mass conservation equation in the hidden layer. The spatio-temporal attention mechanism extracts the vortex propagation direction features, and the projected gradient descent algorithm forces the network output to satisfy the flow field divergence constraint. The NSGA-II multi-objective optimization algorithm integrates Bayesian posterior probability, LSTM time series trajectory, and CFD physical field data, screens the Pareto front solution set, and the residual distribution-location error mapping network generates confidence rating results. The closed-loop adaptive mechanism dynamically adjusts the sensor sensitivity and network hyperparameters through meta-reinforcement learning, optimizes the number of LSTM layers through neural architecture search, and updates the CFD boundary conditions through differential evolution algorithm to form an environment-adaptive leakage location system.

[0093] In the petrochemical warehouse scenario, the above-mentioned implementation effectively overcomes the diffusion path distortion caused by turbulent interference, multi-component cross-sensitivity, and signal lag through hardware-algorithm co-optimization, physical-data model fusion, and dynamic closed-loop adjustment, achieves millimeter-level positioning accuracy of the leakage source coordinates, and meets the engineering requirements of hazardous chemical warehouse safety monitoring.

[0094] The present invention solves the problem of gas diffusion signal distortion in complex environments through multi-physical field data fusion and physical constraint modeling, and improves the accuracy of leakage source inverse inversion. First, a heterogeneous gas sensor array is used to synchronously collect multi-modal data such as gas concentration, components, and three-dimensional wind speed field. The sensors are pre-trained for cross-scenario domain adaptation through a transfer learning algorithm, dynamically match the noise pattern feature library in the knowledge base, and generate a calibrated data set after baseline correction. The calibration process eliminates the zero-point drift of the sensor caused by sudden temperature changes, suppresses the interference of environmental noise on the original signal, and provides a high signal-to-noise ratio input for subsequent processing.

[0095] Secondly, based on the computational fluid dynamics dynamic simulation, a turbulent vortex distribution map is generated, and a vortex propagation delay model between sensor nodes is constructed. The improved particle swarm optimization algorithm is used to iteratively solve the phase shift parameters, time-domain align the multi-node concentration gradient time series signals, and eliminate the time series misalignment caused by the warehouse structure turbulence. The compensated concentration gradient data and the mass conservation equation constraints generate diffusion feature tensors through a generative adversarial network, reconstruct the gas diffusion path map jointly driven by physical laws and measured data, and restore the spatio-temporal correlation features distorted by multi-physical field coupling.

[0096] Finally, the mass conservation equation is embedded as a hard constraint into the Bayesian and LSTM hybrid neural network architecture, and the network parameters are jointly optimized by the projected gradient descent algorithm to force the model output to satisfy the hydrodynamic laws. The multi-model results of Bayesian inference, LSTM time series prediction, and CFD simulation are fused, and the NSGA-II multi-objective optimization algorithm is used to dynamically allocate model weights to screen the Pareto front solution set. The residual distribution-location error mapping network performs uncertainty quantification verification on the solution set to generate location results with credibility ratings, forming an inversion mechanism with dual constraints of data-driven and physical laws, which significantly improves the leakage source location accuracy in complex turbulent environments.

Claims

1. A method for locating and tracing the source of hazardous chemical storage leakage in a gas array, characterized in that, Including: Step 1: Collect the gas concentration, components, and environmental physical field parameters of the warehousing environment to generate a multi-modal original data set; Step 2: Input the multi-modal original data set into a pre-set knowledge base, perform cross-scenario domain adaptation pre-training on the sensor array through a transfer learning algorithm, generate an adaptive baseline correction signal based on the dynamic matching between the real-time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and output a calibrated multi-physical field data set; Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physical field data set, allocate noise suppression weights using a bidirectional gated attention mechanism based on the turbulence and diffusion correlation map in the knowledge base, and simultaneously call the gas adsorption kinetics parameter library to construct a transfer learning-driven chromatographic response transfer model, and output the denoised concentration gradient signal of the target gas and the separated component feature vector; Step 4: Based on the turbulence vortex distribution map generated by computational fluid dynamics dynamic simulation, quantify the response lag effect between sensor nodes, use the particle swarm optimization algorithm to iteratively solve the time delay compensation coefficient and align the multi-node time series signals, and combine the generative adversarial network to synthesize the compensated concentration gradient data with the physical field constraint to generate a gas diffusion path map; Step 5: Embed the computational fluid dynamics mass conservation equation as a hard constraint into a hybrid neural network of Bayesian and LSTM, jointly optimize the network parameters through the projected gradient descent algorithm, and generate the posterior probability distribution of the leakage source coordinates; Step 6: Input the posterior probability distribution output by the hybrid neural network of Bayesian and LSTM, the computational fluid dynamics simulation results, and the LSTM prediction data into an ensemble learning framework, use the NSGA-II multi-objective optimization algorithm to dynamically allocate model weights and screen the Pareto optimal solution set, and generate a credibility rating positioning result.

2. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 1, characterized in that, The environmental physical field parameters include three-dimensional wind speed field, temperature gradient distribution, and air pressure pulsation time series signal; Input the three-dimensional wind speed field and temperature gradient distribution into the noise pattern feature library in the knowledge base for spectral similarity comparison; Generate an adaptive filtering signal based on transfer learning domain adaptation according to the comparison result, perform temperature drift compensation on the air pressure pulsation time series signal, and output a corrected three-dimensional physical field parameter set.

3. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 2, characterized in that, Also including: Perform Daubechies wavelet packet decomposition on the gas concentration time series signal in the corrected three-dimensional physical field parameter set, and extract high-frequency noise components and low-frequency diffusion components; Call the turbulence and diffusion correlation map in the knowledge base to allocate attention weights to the high-frequency noise components, and reconstruct the target gas concentration gradient field in combination with the transfer learning-driven chromatographic response transfer model; Dynamically optimize the adsorption coefficient matrix in the gas adsorption kinetics parameter library through a meta-learning framework to generate a pure component spectral feature after stripping cross-interference.

4. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 3, characterized in that, Also including: Based on the turbulence vortex distribution map output by computational fluid dynamics simulation, construct a vortex propagation delay model between sensor nodes; Use an improved particle swarm optimization algorithm to iteratively solve the phase shift parameter in the vortex propagation delay model, and perform time domain alignment on the reconstructed target gas concentration gradient field; Generate a spatio-temporally synchronized multi-node concentration gradient time series signal set.

5. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 4, characterized in that, It also includes: Embedding the discretization constraint conditions of the computational fluid dynamics mass conservation equation in the hidden layer of the Bayesian and LSTM networks; Extracting the vortex propagation direction features in the multi-node concentration gradient time series signal set through the spatio-temporal attention mechanism; Using the projected gradient descent algorithm to perform flow field divergence constraint on the output tensor of the LSTM unit, forcing the prediction result to satisfy the law of mass conservation.

6. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 1, characterized in that The said step 6 includes: Inputting the posterior probability distribution residual output by the Bayesian and LSTM hybrid neural network into the positioning error mapping network to generate a positioning uncertainty quantification index; Fusing the posterior probability data of Bayesian inference, the time series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation in the ensemble learning framework; Performing multi-dimensional weight allocation on the posterior probability data, time series diffusion trajectory, and physical field constraint results through the NSGA-II multi-objective optimization algorithm to generate a Pareto front solution set.

7. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 6, characterized in that, It also includes: According to the deviation distribution between the credibility rating positioning result and the real-time monitoring residual signal, dynamically adjusting the sensor sensitivity parameters and the hyperparameters of the Bayesian and LSTM hybrid neural network through the meta-reinforcement learning framework, and at the same time reconstructing the network depth based on the neural architecture search technology; Matching the sensor sensitivity adjustment strategy and the hyperparameter optimization scheme of the Bayesian and LSTM networks from the knowledge base according to the residual distribution characteristics of the Pareto front solution set; Adapting the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network through the neural architecture search technology; Optimizing the boundary condition parameters of the turbulent vortex intensity in the computational fluid dynamics model based on the differential evolution algorithm.

8. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 1, characterized in that, The said step 4 includes: Invoking the historical leakage event spatio-temporal trajectory data stored in the knowledge base to construct a training sample set for the generative adversarial network; Synthesizing the diffusion feature tensor by combining the multi-node concentration gradient time series signal set after time domain alignment with the mass conservation equation constraint; Outputting a diffusion path heat map with vortex propagation direction markings through the discriminator module of the generative adversarial network.

9. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 8, characterized in that, It also includes: Training the residual distribution model of the positioning error mapping network based on the three-dimensional reconstruction data of the historical leakage scenario; Performing Monte Carlo sampling on the posterior probability distribution output by the Bayesian and LSTM hybrid neural network to extract the boundary features of the spatial confidence interval; Overlaying the diffusion path heat map and the spatial confidence interval features and visualizing the output as a probability cloud map.

10. The method for locating and tracing the leakage of hazardous chemicals in a gas array according to claim 1, wherein, The said step 1 includes: A distributed array of metal oxide semiconductor sensors and photoionization detectors, suppressing the cross-sensitivity effect through a transfer learning model; An integrated temperature and humidity sensor and three-dimensional ultrasonic anemometer synchronous acquisition module, real-time outputting multi-physical field parameters with timestamp alignment; An anti-interference coating set in the electrochemical sensor array and the chromatographic response transfer model cooperate to strip the interference component spectrum.

Images