Gas array-based method for locating and tracing hazardous chemical storage leaks

Abstract

The technical solution of the present invention relates to the field of gas detection technology, specifically a method for locating and tracing the source of hazardous chemical storage leaks using a gas array. Gas concentrations, components, and environmental physical field parameters are collected through a three-dimensional heterogeneous sensor array to generate a multimodal data set; a transfer learning algorithm is used to match the noise pattern feature library in the knowledge base, dynamically correct the sensor baseline drift, and output calibrated data. The closed-loop adaptive mechanism adjusts sensor parameters and network hyperparameters through meta-reinforcement learning, optimizes the model structure through neural architecture search, and updates the fluid mechanics boundary conditions through the differential evolution algorithm. The present invention solves the problem of signal distortion in complex environments through the collaboration of physical constraints and data-driven, improves the accuracy and robustness of leak source positioning, and is suitable for the field of hazardous chemical storage safety monitoring.

Description

Technical Field

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

[0002] In hazardous chemical storage scenarios, leak location and traceability technology based on gas sensor arrays rapidly identifies leak sources by analyzing multi-node gas concentration distribution characteristics and spatiotemporal correlations. This technology relies on a densely deployed gas sensor network to collect gas diffusion parameters in real time. Combining fluid dynamics models with leak source location algorithms, it determines the propagation path of leak signals through time-difference analysis or concentration gradient analysis, and employs an adaptive weighted fusion model to eliminate environmental interference noise. Further incorporating Bayesian reasoning or neural network algorithms, it can reversely deduce the time series and spatial coordinates of leak events. Dynamic corrections are then made based on the three-dimensional topology of the storage area and environmental parameters (such as wind speed, temperature, and humidity), enabling precise reconstruction of the leak source probability distribution.

[0003] When gas sensor arrays are used to locate leaks in hazardous chemical storage, the gas diffusion path is susceptible to the nonlinear superposition of the turbulent effects of the storage structure, the cross-interference of multi-component gases, and the hysteresis characteristics of the sensor response due to the coupling of complex environmental multi-physical fields. This leads to the distortion of the spatial correlation between the concentration gradient and the time series signal, thereby reducing the analytical accuracy of the fluid mechanics inverse inversion model for the spatial coordinates of the leak source. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a method for locating and tracing the source of hazardous chemical storage leaks using a gas array, which is used to solve the problem in the existing technology that the gas diffusion signal is distorted due to the coupling of multiple physical fields in a complex environment, affecting the accuracy of the inverse inversion of the hazardous chemical leakage source based on the gas sensor array.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The present invention provides a method for locating and tracing the source of a hazardous chemical storage leak in a gas array, comprising:

[0007] Step 1: A heterogeneous gas sensor array deployed in three-dimensional space collects gas concentration, components, and environmental physical field parameters in the storage environment in real time, generating a multimodal raw data set including gas concentration time series signals, multi-component spectral characteristics, and three-dimensional wind speed fields.

[0008] Step 2: Input the multimodal raw data set into a preset 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 of the real-time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and output a calibrated multi-physics field data set that has been subjected to environmental noise suppression and sensor zero drift compensation;

[0009] Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physics field dataset, assign noise suppression weights using a bidirectional gated attention mechanism based on the turbulence and diffusion correlation map in the knowledge base, and simultaneously construct a transfer learning-driven chromatographic response migration model by calling the gas adsorption kinetic parameter library to output the de-noised concentration gradient signal of the target gas and the separated component feature vectors;

[0010] Step 4: Based on the turbulent vortex distribution map generated by computational fluid dynamics dynamic simulation, the response lag effect between sensor nodes is quantified. A particle swarm optimization algorithm is used to iteratively solve the time delay compensation coefficient and align the multi-node timing signals. The compensated concentration gradient data and physical field constraints are combined with a generative adversarial network to synthesize a baseline diffusion feature vector, generating a gas diffusion path map under multi-physics field coupling conditions.

[0011] Step 5: The computational fluid dynamics mass conservation equation is embedded as a hard constraint in the Bayesian and LSTM hybrid neural network. The network parameters are jointly optimized using the projected gradient descent algorithm to generate a posterior probability distribution of the leakage source coordinates that satisfies the fluid dynamics laws and the measured data characteristics.

[0012] Step 6: Input the posterior probability distribution output by the Bayesian and LSTM hybrid neural network, the computational fluid dynamics simulation results, and the LSTM prediction data into the integrated 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 with residual distribution-positioning error mapping network verification.

[0013] Furthermore, in the method for locating and tracing the source of hazardous chemical storage leaks 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 pressure pulsation time series signal;

[0014] Performing spectrum similarity comparison on the noise pattern feature library inputting the three-dimensional wind speed field and the temperature gradient distribution into the knowledge base;

[0015] An adaptive filtering signal based on transfer learning domain adaptation is generated according to the comparison result, temperature drift compensation is performed on the air pressure pulsation time series signal, and a corrected three-dimensional physical field parameter set is output.

[0016] Furthermore, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0017] performing Daubechies wavelet packet decomposition on the gas concentration time series signal in the corrected three-dimensional physical field parameter set to extract high-frequency noise components and low-frequency diffusion components;

[0018] The turbulence and diffusion association map in the knowledge base is used to allocate attention weights to the high-frequency noise components, and the target gas concentration gradient field is reconstructed by combining the chromatographic response migration model driven by transfer learning.

[0019] The adsorption coefficient matrix in the gas adsorption kinetic parameter library is dynamically optimized through a meta-learning framework to generate the spectral characteristics of pure components after removing cross-interference.

[0020] Furthermore, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0021] Based on the turbulent vortex distribution map output by computational fluid dynamics simulation, a vortex propagation delay model between sensor nodes is constructed;

[0022] An improved particle swarm optimization algorithm is used to iteratively solve the phase offset parameter in the vortex propagation delay model, and the reconstructed target gas concentration gradient field is aligned in the time domain;

[0023] Generate a spatiotemporally synchronized multi-node concentration gradient time series signal set.

[0024] Furthermore, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0025] Embed the discretized constraints of the computational fluid dynamics mass conservation equation in the hidden layers of the Bayesian and LSTM networks;

[0026] Extracting vortex propagation direction features from the multi-node concentration gradient time series signal set through a spatiotemporal attention mechanism;

[0027] The projected gradient descent algorithm is used to constrain the flow field divergence of the output tensor of the LSTM unit, forcing the prediction results to satisfy the law of conservation of mass.

[0028] Furthermore, in the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention, step 6 includes:

[0029] 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 quantitative indicator;

[0030] Integrate the posterior probability data of Bayesian inference, the time series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation into an integrated learning framework;

[0031] The NSGA-II multi-objective optimization algorithm is used to perform multi-dimensional weight allocation on the posterior probability data, time series diffusion trajectory and physical field constraint results to generate a Pareto front solution set.

[0032] Furthermore, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0033] Based on the deviation distribution between the credibility rating positioning results and the real-time monitoring residual signal, the sensor sensitivity parameters and the Bayesian and LSTM hybrid neural network hyperparameters are dynamically adjusted through a meta-reinforcement learning framework. At the same time, the network depth is reconstructed based on neural architecture search technology to form a closed-loop adaptive learning mechanism.

[0034] 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;

[0035] Adaptively adjusting the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network using a neural architecture search technique;

[0036] Optimization of boundary condition parameters of turbulent vortex intensity in computational fluid dynamics model based on differential evolution algorithm.

[0037] Furthermore, in the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention, step 4 includes:

[0038] The spatiotemporal trajectory data of historical leakage events stored in the knowledge base is used to construct a training sample set for the adversarial generative network;

[0039] The diffusion characteristic tensor is synthesized by combining the time-domain aligned multi-node concentration gradient time series signal set with the mass conservation equation constraint;

[0040] The discriminator module of the generative adversarial network outputs a diffusion path heat map marked with the vortex propagation direction.

[0041] Furthermore, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0042] The residual distribution model of the positioning error mapping network is trained based on 3D reconstruction data of historical leakage scenes;

[0043] Performing Monte Carlo sampling on the posterior probability distribution output by the Bayesian and LSTM hybrid neural network to extract boundary features of the spatial confidence interval;

[0044] The diffusion path thermal map is superimposed with the spatial confidence interval feature to generate a probability cloud map visualization output.

[0045] Furthermore, in the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention, step 1 comprises:

[0046] Distributed arrays of metal oxide semiconductor sensors and photoionization detectors, suppressing cross-sensitivity effects through transfer learning models;

[0047] A synchronous acquisition module integrating temperature and humidity sensors and a three-dimensional ultrasonic anemometer outputs multi-physics field parameters with time stamp alignment in real time;

[0048] The anti-interference coating provided in the electrochemical sensor array cooperates with the chromatographic response migration model to strip off the interfering component spectrum.

[0049] Beneficial effects of the present invention:

[0050] The present invention eliminates temperature drift and cross-sensitivity interference in complex environments and restores the original characteristics of gas concentration signals through dynamic noise suppression of multimodal data acquisition and transfer learning of heterogeneous gas sensor arrays; combines computational fluid dynamics simulation with an improved particle swarm optimization algorithm to quantify the turbulence hysteresis effect between sensor nodes and reconstruct the diffusion path map constrained by physical laws; embeds hard constraints of the mass conservation equation based on a Bayesian and LSTM hybrid neural network, and forces the model output to conform to the laws of fluid mechanics through projected gradient descent, solving the problem of data-driven models violating the evolution of physical fields in turbulent environments; adopts the NSGA-II multi-objective optimization algorithm to fuse multi-source heterogeneous data, combined with residual distribution-positioning error mapping network verification, to generate a Pareto optimal solution set with credibility rating, 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, realizes continuous optimization of system parameters during long-term operation, and ultimately achieves millimeter-level positioning accuracy of the leakage source in hazardous chemical storage scenarios, significantly improving the robustness and reliability of inverse inversion in complex multi-physical field coupling environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0052] Figure 1 A flow chart of a method for locating and tracing a leak in a hazardous chemical storage facility of a gas array provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, 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 of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are 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. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0054] See also Figure 1 The present invention provides a method for locating and tracing the source of a hazardous chemical storage leak in a gas array, comprising:

[0055] Step 1: A heterogeneous gas sensor array deployed in three-dimensional space collects gas concentration, components, and environmental physical field parameters in the storage environment in real time, generating a multimodal raw data set including gas concentration time series signals, multi-component spectral characteristics, and three-dimensional wind speed fields.

[0056] Step 2: Input the multimodal raw data set into a preset 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 of the real-time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and output a calibrated multi-physics field data set that has been subjected to environmental noise suppression and sensor zero drift compensation;

[0057] Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physics field dataset, assign noise suppression weights using a bidirectional gated attention mechanism based on the turbulence and diffusion correlation map in the knowledge base, and simultaneously construct a transfer learning-driven chromatographic response migration model by calling the gas adsorption kinetic parameter library to output the de-noised concentration gradient signal of the target gas and the separated component feature vectors;

[0058] Step 4: Based on the turbulent vortex distribution map generated by computational fluid dynamics dynamic simulation, the response lag effect between sensor nodes is quantified. A particle swarm optimization algorithm is used to iteratively solve the time delay compensation coefficient and align the multi-node timing signals. The compensated concentration gradient data and physical field constraints are combined with a generative adversarial network to synthesize a baseline diffusion feature vector, generating a gas diffusion path map under multi-physics field coupling conditions.

[0059] Step 5: The computational fluid dynamics mass conservation equation is embedded as a hard constraint in the Bayesian and LSTM hybrid neural network. The network parameters are jointly optimized using the projected gradient descent algorithm to generate a posterior probability distribution of the leakage source coordinates that satisfies the fluid dynamics laws and the measured data characteristics.

[0060] Step 6: Input the posterior probability distribution output by the Bayesian and LSTM hybrid neural network, the computational fluid dynamics simulation results, and the LSTM prediction data into the integrated 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 with residual distribution-positioning error mapping network verification.

[0061] The gas array hazardous chemical storage leak location and source tracing method provided by this invention achieves accurate location of the leak source in complex environments through multimodal data acquisition, signal processing, and physical model integration. The following is a detailed description of each step of the technical solution:

[0062] In step 1, a heterogeneous gas sensor array deployed in three dimensions collects gas concentrations, components, and physical field parameters of the storage environment in real time. This heterogeneous gas sensor array, comprised of metal oxide semiconductor sensors, photoionization detectors, and electrochemical sensors with anti-interference coatings, is distributed across key monitoring nodes within the storage space and simultaneously collects gas concentration time series signals, multi-component spectral signatures, and three-dimensional wind velocity field data. Temperature and humidity sensors and a three-dimensional ultrasonic anemometer are integrated into the same acquisition module, ensuring the timestamp alignment of multiple physical field parameters and forming a multimodal raw data set that incorporates temporal and spatial correlations.

[0063] In step 2, the multimodal raw dataset is input into a pre-built knowledge base, and a transfer learning algorithm is used to pre-train the sensor array for cross-scenario domain adaptation. The knowledge base stores a library of noise pattern signatures from historical leak scenarios. By dynamically matching the real-time environmental parameter spectrum with the signature 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, eliminating sensor zero-point drift and sudden environmental noise, and outputting a calibrated multi-physics field dataset.

[0064] Step 3 performs multiscale wavelet packet decomposition and gas component decoupling on the calibrated dataset output from Step 2. The Daubechies wavelet packet decomposition algorithm is used to separate the high-frequency noise and low-frequency diffusion components of the gas concentration time series signal. Combined with the turbulence and diffusion correlation maps in the knowledge base, a bidirectional gated attention mechanism dynamically assigns noise suppression weights to each frequency band. Simultaneously, a transfer learning-driven chromatographic response migration model is constructed using the gas adsorption kinetics parameter library. This model removes spectral features of multi-component gas cross-interference and reconstructs the de-noised concentration gradient field and pure component feature vectors of the target gas.

[0065] Step 4 quantifies the response lag between sensor nodes based on the turbulent eddy distribution map generated by computational fluid dynamics (CFD) dynamic simulation. A model for eddy propagation delay between sensor nodes is constructed, and an improved particle swarm optimization algorithm is used to iteratively solve for phase offset parameters. The reconstructed target gas concentration gradient field is then aligned in the time domain. The compensated concentration gradient data and physical field constraints are combined through an adversarial generative network to synthesize a baseline diffusion feature vector, generating a spatiotemporal correlation map representing the gas diffusion path under multi-physics coupling conditions.

[0066] In step 5, the computational fluid dynamics mass conservation equation is embedded as a hard constraint in a hybrid Bayesian and LSTM neural network. Prior data on the spatial distribution of leak sources is introduced into the Bayesian framework, and the LSTM branch learns diffusion temporal association rules through a spatiotemporal attention mechanism. A projected gradient descent algorithm jointly optimizes the network parameters, forcing the model output to satisfy the mass conservation law and the characteristics of the measured data, thereby generating a posterior probability distribution of the leak source coordinates.

[0067] Step 6 uses an integrated learning framework to fuse the posterior probability distributions of the Bayesian and LSTM networks, computational fluid dynamics simulation results, and the time-series diffusion trajectory predicted by the LSTM. The NSGA-II multi-objective optimization algorithm is used to assign multi-dimensional weights to heterogeneous data sources and screen for Pareto-optimal solutions. A residual distribution-positioning error mapping network verifies the uncertainty of the solution set and outputs positioning results with confidence ratings, providing a benchmark for closed-loop adaptive learning.

[0068] Each step forms a unidirectional, progressive, and reversibly correctable data processing chain. The multimodal data from step 1 provides input for noise suppression in step 2, and the noise reduction component characteristics from step 3 lay the foundation for spatiotemporal alignment in step 4. The physical constraint inversion in step 5 and the multi-model fusion in step 6 synergistically improve positioning accuracy, ultimately dynamically optimizing system parameters through a closed-loop mechanism. Through the dual constraints of data-driven and physical laws, this technical solution addresses the signal distortion caused by the coupling of complex multi-physical fields, enabling precise tracing of hazardous chemical leaks.

[0069] Specifically, in the method for locating and tracing the source of hazardous chemical storage leaks 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 pressure pulsation time series signal;

[0070] Performing spectrum similarity comparison on the noise pattern feature library inputting the three-dimensional wind speed field and the temperature gradient distribution into the knowledge base;

[0071] An adaptive filtering signal based on transfer learning domain adaptation is generated according to the comparison result, temperature drift compensation is performed on the air pressure pulsation time series signal, and a corrected three-dimensional physical field parameter set is output.

[0072] In the method for locating and tracing leaks in hazardous chemical storage using a gas array described in the present invention, the collection of environmental physical field parameters and baseline drift correction constitute the core links of data preprocessing. The measurement of the three-dimensional wind speed field is achieved through a distributed three-dimensional ultrasonic anemometer, which captures the airflow vector distribution in the storage space based on the Doppler effect and generates a wind speed field matrix with time-space correlation. The temperature gradient distribution data is synchronously collected by an embedded high-precision temperature sensor array. The sensor nodes are layered in the vertical and horizontal directions, and the spatial gradient change characteristics of the temperature field are output in real time. The air pressure pulsation timing signal is obtained through a micro-pressure difference sensor with anti-interference design. 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.

[0073] During baseline drift correction, the knowledge base stores a library of historical noise pattern features containing spectral signatures of sensor baseline drift under different environmental conditions. Three-dimensional wind velocity field and temperature gradient distribution data are input into the feature library. A transfer learning algorithm is then used to extract scene-domain invariant features and construct a spectral similarity comparison model. This model identifies the steady-state and transient components of real-time environmental parameters and generates an adaptive filtering signal based on domain adaptation. The filtered signal is applied to the pressure pulsation time series signal to phase compensate for the sensor zero-point drift component caused by sudden temperature changes, eliminating nonlinear errors in the pressure data caused by thermodynamic effects. The resulting output is a spatiotemporally corrected set of three-dimensional physical field parameters.

[0074] The corrected three-dimensional physical field parameter set provides high-confidence input for subsequent signal processing. The coupling relationship between temperature gradient data and the wind velocity field matrix is ​​verified using the turbulence evolution model in the knowledge base, ensuring the dynamic consistency of the physical field parameters. The compensation results of the pressure pulsation signal are correlated with the residual distribution characteristics of the transfer learning model, and the noise pattern feature library in the knowledge base is dynamically updated, forming a closed-loop optimization mechanism for environmental parameter calibration. Through multi-dimensional coordinated correction of physical field parameters, this technical solution lays the data foundation for gas diffusion feature reconstruction and leakage source inversion.

[0075] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0076] performing Daubechies wavelet packet decomposition on the gas concentration time series signal in the corrected three-dimensional physical field parameter set to extract high-frequency noise components and low-frequency diffusion components;

[0077] The turbulence and diffusion association map in the knowledge base is used to allocate attention weights to the high-frequency noise components, and the target gas concentration gradient field is reconstructed by combining the chromatographic response migration model driven by transfer learning.

[0078] The adsorption coefficient matrix in the gas adsorption kinetic parameter library is dynamically optimized through a meta-learning framework to generate the spectral characteristics of pure components after removing cross-interference.

[0079] In the method for locating and tracing leaks in hazardous chemical storage using a gas array described in the present invention, step 3 achieves gas concentration field reconstruction and component separation through multi-scale signal decomposition and dynamic parameter optimization. After the baseline drift correction of 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 high-frequency noise components in the 0.1-10 Hz frequency band and low-frequency diffusion components in the 0.01-0.1 Hz frequency band. During the decomposition process, the effective frequency band boundary is determined by the energy entropy threshold, and the low-frequency signal components associated with the turbulent diffusion characteristics are retained.

[0080] High-frequency noise components are suppressed using a turbulence-diffusion correlation map stored in a knowledge base. This correlation map stores the mapping between turbulence intensity and noise spectra in historical leak scenarios. Similar noise patterns are matched based on the current three-dimensional wind velocity field data. A bidirectional gated attention mechanism dynamically assigns noise suppression weights to each frequency band based on the matching results, adaptively filtering high-frequency components. The filtered low-frequency diffusion components are coupled with a chromatographic response migration model driven by transfer learning to reconstruct the target gas concentration gradient field.

[0081] The chromatographic response transfer model is trained using a knowledge base containing gas adsorption kinetic parameter libraries, including adsorption isotherms and reaction activation energy data for multiple gas components on the sensor surface. The model uses a transfer learning strategy to transfer adsorption characteristics from historical scenarios to the current environment, constructing a component separation matrix. A meta-learning framework optimizes cross-sensitivity parameters in the adsorption coefficient matrix in real time, combining temporal and spatial modulation data from the sensor array to remove spectral response characteristics of interfering gases. The optimized parameter matrix is ​​then integrated with the multiscale decomposition results to generate spatially continuous spectral feature vectors for pure components.

[0082] Each technical link forms a progressive processing chain. Wavelet packet decomposition provides the frequency domain analysis foundation for noise suppression, and the attention mechanism assigns weights to enhance the ability to retain turbulence-related features. The synergy between the transfer learning model and the meta-learning framework enables generalized component separation capabilities across scenarios. The reconstructed concentration gradient field and pure spectral features provide high signal-to-noise ratio input for spatiotemporal lag compensation in step 4, supporting the precise modeling of subsequent diffusion paths. The technical solution addresses the concentration field distortion caused by multi-component cross-interference through the multi-dimensional synergy of signal decomposition and parameter optimization.

[0083] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0084] Based on the turbulent vortex distribution map output by computational fluid dynamics simulation, a vortex propagation delay model between sensor nodes is constructed;

[0085] An improved particle swarm optimization algorithm is used to iteratively solve the phase offset parameter in the vortex propagation delay model, and the reconstructed target gas concentration gradient field is aligned in the time domain;

[0086] Generate a spatiotemporally synchronized multi-node concentration gradient time series signal set.

[0087] In the gas array method for locating and tracing hazardous chemical storage leaks described in this invention, step 4 achieves spatiotemporal alignment of gas diffusion characteristics through physical field modeling and signal timing correction. Based on the target gas concentration gradient field output in step 3, a computational fluid dynamics (CFD) dynamic simulation engine is invoked to generate a turbulent vortex distribution map, combining the three-dimensional topological structure of the storage space and environmental parameters. This map, characterized by vortex intensity, propagation direction, and energy dissipation rate, quantifies the effects of nonlinear perturbations on the gas diffusion path between sensor nodes.

[0088] The construction of the vortex propagation delay model relies on the spatiotemporal evolution characteristics of the turbulent vortex distribution map. The model maps the spatial coordinates of sensor nodes and the vortex propagation direction into a time delay function, establishing a phase offset relationship for signal propagation between nodes. An improved particle swarm optimization algorithm introduces an adaptive inertia weight adjustment strategy, using the vortex energy attenuation coefficient as a constraint to iteratively solve for the optimal solution set of phase offset parameters. During the parameter optimization process, the time domain alignment of the target gas concentration gradient field is achieved through dynamic sliding window matching to eliminate signal timing misalignment caused by turbulence in the storage structure.

[0089] After time-domain alignment, the multi-node concentration gradient time series signal set is spatially reconstructed using a spatiotemporal interpolation algorithm. This interpolation algorithm integrates the physical constraints of the CFD simulation with the measured concentration gradient data to generate a spatiotemporally synchronized signal set. This synchronized signal set preserves the time-varying characteristics and spatial correlations of gas diffusion, providing high-precision input for leak source inversion in Step 5. This technical solution, through the combined optimization of physical field modeling and data-driven optimization, addresses the problem of time series signal distortion caused by sensor response lag and improves the reliability of diffusion path reconstruction.

[0090] The synergistic effect of these technical links is reflected in the following: CFD simulation provides physical field evolution rules for the time-delay model, improved optimization algorithms achieve signal alignment through dynamic parameter adjustment, and spatiotemporal interpolation enhances the spatial consistency of the data. This processing chain effectively suppresses the interference of complex turbulent environments on the spatiotemporal correlation of sensor signals, laying the data foundation for subsequent inverse inversion.

[0091] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0092] Embed the discretized constraints of the computational fluid dynamics mass conservation equation in the hidden layers of the Bayesian and LSTM networks;

[0093] Extracting vortex propagation direction features from the multi-node concentration gradient time series signal set through a spatiotemporal attention mechanism;

[0094] The projected gradient descent algorithm is used to constrain the flow field divergence of the output tensor of the LSTM unit, forcing the prediction results to satisfy the law of conservation of mass.

[0095] In the method for locating and tracing leaks in hazardous chemical storage systems using gas arrays, the hybrid neural network architecture in step 4 improves leak source inversion accuracy by deeply integrating physical constraints with data-driven models. In the hidden layers of the Bayesian and LSTM networks, the computational fluid dynamics mass conservation equation is discretized into partial differential constraints and embedded as tensors in the network's forward propagation process. The discretized constraints are generated based on the gridding of the storage space. The residual of the mass conservation equation for 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.

[0096] The spatiotemporal attention mechanism extracts features from the multi-node concentration gradient time series signal set generated in step 4. The attention weight calculation module dynamically allocates the contribution of the time series signals from different sensor nodes by combining the vortex propagation direction characteristics with the concentration gradient change rate. The directional features are extracted using a convolutional neural network to extract local vortex vectors from the diffusion path heat map. This directional feature is then combined with the attention weight matrix for a Hadamard product, enhancing the model's ability to detect turbulence-dominated diffusion paths.

[0097] The projected gradient descent algorithm operates on the output tensor of the LSTM unit, forcing the prediction results to conform to the mass conservation law through flow field divergence constraints. This divergence constraint, based on the discretized mass conservation equation, maps the divergence residual of the network output tensor to an orthogonal basis in the projected space. During the gradient descent process, the network parameter update direction is constrained to lie within the orthogonal complement space of the residuals, ensuring that the predicted leakage source distribution conforms to both physical laws and the characteristics of the measured data.

[0098] The logical connections between the technical solutions are as follows: physical constraint embedding provides the fundamental laws of fluid motion for the attention mechanism, attention weights optimize the spatiotemporal correlations of signal feature extraction, and projected gradient descent enforces model output consistency with physical field evolution through mathematical optimization. Embedding discretized constraints avoids direct modifications to the network structure, ensuring the model's trainability. The synergy between the attention mechanism and gradient projection addresses the issue of data-driven models violating physical laws in complex turbulent environments.

[0099] The above processing chain forms a leakage source location mechanism enhanced by physical information. By constraining the consistency between the mathematical characteristics of the network output and the laws of fluid mechanics, it significantly improves the robustness of the inverse inversion model under noise interference, providing technical support for generating a high-confidence posterior probability distribution in step 5.

[0100] Specifically, in the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention, step 6 includes:

[0101] 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 quantitative indicator;

[0102] Integrate the posterior probability data of Bayesian inference, the time series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation into an integrated learning framework;

[0103] The NSGA-II multi-objective optimization algorithm is used to perform multi-dimensional weight allocation on the posterior probability data, time series diffusion trajectory and physical field constraint results to generate a Pareto front solution set.

[0104] In the gas array hazardous chemical storage leak location and traceability method described in this invention, step 6 verifies the credibility of the leak source location results through multi-source data fusion and multi-objective optimization. The posterior probability distribution residuals output by the Bayesian and LSTM hybrid neural network represent the deviation between the model prediction and the measured data. After being input into the positioning error mapping network, the network generates a positioning uncertainty quantification index based on the residual distribution pattern of historical leak scenarios. This index extracts the spatiotemporal correlation characteristics of the residuals through a convolutional neural network and, combined with the error-environmental parameter mapping relationship in the knowledge base, outputs a confidence weight matrix for each sensor node.

[0105] The ensemble learning framework receives three heterogeneous data sources: posterior probability data from Bayesian inference, time-series diffusion trajectories from LSTM predictions, and physical field constraint results from CFD simulations. The framework employs a feature-level fusion strategy to uniformly map the spatial confidence of the posterior probability distribution, the temporal correlation of the LSTM trajectories, and the directional constraints of the CFD physical field into a high-dimensional feature space. A self-attention mechanism dynamically assigns contribution weights to each data source in the feature space, generating a fused joint decision vector that represents cross-model consistency for leak source location.

[0106] The NSGA-II multi-objective optimization algorithm assigns multi-dimensional weights to the joint decision vector, optimizing positioning accuracy, model consistency, and computational efficiency. The algorithm uses a non-dominated sorting strategy to select the Pareto front solution set. Each candidate point in the solution set must satisfy a posterior probability confidence threshold, physical field divergence constraints, and temporal trajectory continuity conditions. A residual distribution-positioning error mapping network performs secondary validation on the solution set, eliminating anomalous solutions that violate historical leakage patterns or environmental parameter constraints, and generating a final, trustworthy positioning result.

[0107] The logical chain of technical solutions is as follows: residual analysis quantifies the uncertainty of model predictions, providing a credibility benchmark for data fusion; an integrated framework coordinates the complementarity of multi-source data through feature space mapping; and multi-objective optimization balances positioning accuracy and model robustness under constraints. The verification mechanism of the positioning error mapping network and the screening strategy of NSGA-II form a dual verification, ensuring that the output results conform to both data-driven characteristics and physical laws. Through multi-level verification and optimization, this process solves the problem of insufficient credibility of single-model positioning results in complex environments and provides highly reliable input for closed-loop adaptive learning.

[0108] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0109] Based on the deviation distribution between the credibility rating positioning results and the real-time monitoring residual signal, the sensor sensitivity parameters and the Bayesian and LSTM hybrid neural network hyperparameters are dynamically adjusted through a meta-reinforcement learning framework. At the same time, the network depth is reconstructed based on neural architecture search technology to form a closed-loop adaptive learning mechanism.

[0110] 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;

[0111] Adaptively adjusting the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network using a neural architecture search technique;

[0112] Optimization of boundary condition parameters of turbulent vortex intensity in computational fluid dynamics model based on differential evolution algorithm.

[0113] In the gas array hazardous chemical storage leak location and traceability method described in the present invention, a closed-loop adaptive learning mechanism achieves continuous system optimization through dynamic parameter adjustment and model reconstruction. The credibility rating positioning results and the deviation distribution of the real-time monitoring residual signal are input into the meta-reinforcement learning framework, which matches the current environmental state based on the historical optimization strategy library in the knowledge base. The meta-reinforcement learning policy network dynamically adjusts the sensor sensitivity parameters, including the response gain and sampling frequency of the gas sensor, through a multi-objective reward function, while optimizing the regularization coefficient and learning rate hyperparameters in the Bayesian and LSTM hybrid neural network to balance the model's generalization ability and fitting accuracy.

[0114] Neural architecture search technology adaptively adjusts the number of layers in the LSTM branch of a hybrid Bayesian and LSTM neural network. The search space is defined as a combination of network depth and number of hidden units, and a set of candidate network structures is generated based on real-time residual distribution characteristics. The search process uses a weight-sharing strategy to evaluate the validation loss of different structures. A gradient-directed optimizer is then used to select lightweight network architectures that meet computational efficiency constraints, dynamically expanding or compressing the time series modeling capabilities of the LSTM branch.

[0115] The sensor sensitivity adjustment strategies stored in the knowledge base are matched with Bayesian and LSTM hyperparameter optimization schemes based on the residual distribution characteristics of the Pareto front solution set. The matching algorithm quantifies the correlation between residual patterns and historical scenarios using cosine similarity to extract the optimal parameter adjustment template. The sensitivity parameters and hyperparameter combinations in the template are transferred to the current system configuration using a transfer learning framework, preventing parameter optimization from falling into local optima.

[0116] The turbulent eddy intensity boundary condition parameters for the computational fluid dynamics model were optimized using a differential evolution algorithm. The algorithm uses the residual between the diffusion path map generated in step 4 and the measured concentration gradient field as a fitness function and iteratively updates the population distribution of the eddy intensity parameters. Physical field divergence constraints are introduced during the population update process to enforce that the optimized boundary parameters satisfy the mass conservation equation, ensuring consistency between the CFD simulation results and the measured data.

[0117] The logical connections between the technical solutions are as follows: a meta-reinforcement learning framework dynamically adjusts front-end sensor and back-end model parameters based on real-time deviations; a neural architecture search optimizes the network structure to adapt to the dynamic environment; a knowledge base matching strategy improves the efficiency of parameter migration; and a differential evolution algorithm ensures the accuracy of the physical model's boundary conditions. These steps form a closed-loop mechanism for collaborative optimization of inner and outer loops. The inner loop optimizes sensor and model parameters, while the outer loop reconstructs the network structure and physical field boundaries. This two-way feedback mechanism continuously improves system performance.

[0118] The closed-loop adaptive mechanism addresses model degradation caused by dynamic environmental changes through a multi-level optimization strategy, ensuring the leak location system maintains high accuracy and robustness over the long term. The synergistic effect of parameter adjustment and model reconstruction provides dynamic adaptability for the entire data processing process from Steps 1 to 6, enhancing the practicality of the technical solution in complex multi-physics coupling scenarios.

[0119] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention, step 4 includes:

[0120] The spatiotemporal trajectory data of historical leakage events stored in the knowledge base is used to construct a training sample set for the adversarial generative network;

[0121] The diffusion characteristic tensor is synthesized by combining the time-domain aligned multi-node concentration gradient time series signal set with the mass conservation equation constraint;

[0122] The discriminator module of the generative adversarial network outputs a diffusion path heat map marked with the vortex propagation direction.

[0123] In the gas array method for locating and tracing hazardous chemical storage leaks described in this invention, step 4 achieves accurate modeling of gas diffusion paths by integrating historical data with physical constraints. The spatiotemporal trajectory data of historical leak events stored in a knowledge base undergoes structured processing to extract the leak source coordinates, diffusion time series, and combined features of environmental parameters, forming a multidimensional spatiotemporal trajectory matrix. This matrix is ​​interpolated to supplement missing data points, and Gaussian noise is injected to enhance sample diversity, constructing a training sample set for the generative adversarial network. This sample set covers diffusion scenarios under different storage structures, leak intensities, and environmental disturbance patterns, providing the generative adversarial network with cross-scenario generalization capabilities.

[0124] The synthesis of the time-domain aligned multi-node concentration gradient time series signal set and the mass conservation equation constraints is achieved using tensor fusion technology. The concentration gradient time series signal set is encoded into a spatiotemporal feature map through three-dimensional convolution, while the mass conservation equation is discretized into divergence constraints for the spatial grid nodes. The feature map and the divergence constraint matrix are fused using a Hadamard product operation to generate a diffusion feature tensor enhanced with physical information. The physical constraints embedded in the tensor force the diffusion path to conform to the laws of fluid mechanics, suppressing path distortion caused by data noise.

[0125] The discriminator module of the generative adversarial network receives the diffusion feature tensor as input and extracts local and global features of the diffusion path through a multi-level convolutional network. The discriminator's output layer combines a vortex propagation direction classifier with a thermal value regressor to generate a directional thermal map of the diffusion path. The directional labels are based on the angular distribution of the vortex vector field, and the thermal value represents the cumulative intensity of the concentration gradient at the spatial grid nodes. The generated map is fed back to the generator module via a residual connection, iteratively optimizing the path generation for realism and physical consistency.

[0126] The logical connections between the technical solutions are as follows: a training set constructed from historical data enhances the scenario adaptability of the generative adversarial network; the integration of physical constraints and real-time data ensures the consistency of the diffusion feature tensor; and the multi-task output of the discriminator enables joint modeling of path visualization and direction labeling. The resulting diffusion path heat map provides high-fidelity input for leak source inversion in step 5 and supports credibility rating verification in step 6. Driven by both historical experience and physical laws, the technical solution effectively addresses the distortion problem of diffusion path reconstruction in complex turbulent environments.

[0127] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention further includes:

[0128] The residual distribution model of the positioning error mapping network is trained based on 3D reconstruction data of historical leakage scenes;

[0129] Performing Monte Carlo sampling on the posterior probability distribution output by the Bayesian and LSTM hybrid neural network to extract boundary features of the spatial confidence interval;

[0130] The diffusion path thermal map is superimposed with the spatial confidence interval feature to generate a probability cloud map visualization output.

[0131] In the gas array method for locating and tracing hazardous chemical storage leaks described in this invention, step 4 achieves accurate modeling of gas diffusion paths by integrating historical data with physical constraints. The spatiotemporal trajectory data of historical leak events stored in a knowledge base undergoes structured processing to extract the leak source coordinates, diffusion time series, and combined features of environmental parameters, forming a multidimensional spatiotemporal trajectory matrix. This matrix is ​​interpolated to supplement missing data points, and Gaussian noise is injected to enhance sample diversity, constructing a training sample set for the generative adversarial network. This sample set covers diffusion scenarios under different storage structures, leak intensities, and environmental disturbance patterns, providing the generative adversarial network with cross-scenario generalization capabilities.

[0132] The synthesis of the time-domain aligned multi-node concentration gradient time series signal set and the mass conservation equation constraints is achieved using tensor fusion technology. The concentration gradient time series signal set is encoded into a spatiotemporal feature map through three-dimensional convolution, while the mass conservation equation is discretized into divergence constraints for the spatial grid nodes. The feature map and the divergence constraint matrix are fused using a Hadamard product operation to generate a diffusion feature tensor enhanced with physical information. The physical constraints embedded in the tensor force the diffusion path to conform to the laws of fluid mechanics, suppressing path distortion caused by data noise.

[0133] The discriminator module of the generative adversarial network receives the diffusion feature tensor as input and extracts local and global features of the diffusion path through a multi-level convolutional network. The discriminator's output layer combines a vortex propagation direction classifier with a thermal value regressor to generate a directional thermal map of the diffusion path. The directional labels are based on the angular distribution of the vortex vector field, and the thermal value represents the cumulative intensity of the concentration gradient at the spatial grid nodes. The generated map is fed back to the generator module via a residual connection, iteratively optimizing the path generation for realism and physical consistency.

[0134] The logical connections between the technical solutions are as follows: a training set constructed from historical data enhances the scenario adaptability of the generative adversarial network; the integration of physical constraints and real-time data ensures the consistency of the diffusion feature tensor; and the multi-task output of the discriminator enables joint modeling of path visualization and direction labeling. The resulting diffusion path heat map provides high-fidelity input for leak source inversion in step 5 and supports credibility rating verification in step 6. Driven by both historical experience and physical laws, the technical solution effectively addresses the distortion problem of diffusion path reconstruction in complex turbulent environments.

[0135] Specifically, the method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to the present invention comprises:

[0136] Distributed arrays of metal oxide semiconductor sensors and photoionization detectors, suppressing cross-sensitivity effects through transfer learning models;

[0137] A synchronous acquisition module integrating temperature and humidity sensors and a three-dimensional ultrasonic anemometer outputs multi-physics field parameters with time stamp alignment in real time;

[0138] The anti-interference coating provided in the electrochemical sensor array cooperates with the chromatographic response migration model to strip off the interfering component spectrum.

[0139] In the method for locating and tracing leaks in hazardous chemical storage using a gas array according to the present invention, step 1 realizes high-precision data acquisition in a complex environment through the collaborative deployment of multimodal sensors 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 is based on a sensor response feature library for different scenarios in the knowledge base, extracts domain invariant feature parameters, dynamically adjusts the sensitivity weights of the sensor array, and weakens the spectral interference of non-target gases. A spatial interpolation algorithm is used between sensor nodes to compensate for blind spot data and generate a concentration gradient distribution matrix covering the entire storage area.

[0140] The synchronous acquisition module integrates a temperature and humidity sensor and a three-dimensional ultrasonic anemometer, achieving timestamp alignment of multi-physics parameters through a hardware clock synchronization protocol. The temperature and humidity sensor utilizes a multi-point calibration strategy to eliminate the effects of ambient temperature drift on gas concentration inversion. The three-dimensional ultrasonic anemometer uses the Doppler effect to capture the airflow vector distribution in real time and generate a wind velocity field matrix. The output signals of the synchronous acquisition module are time-aligned and fused with the gas concentration data to form a spatiotemporally correlated multi-physics dataset, providing a unified time reference for subsequent signal processing.

[0141] The surface of the electrochemical sensor array is coated with an anti-interference coating. The coating material is designed based on the adsorption characteristics of the target gas, and chemical modification suppresses the adsorption reaction of interfering components. This anti-interference coating works synergistically with the chromatographic response migration model from step 3. The model utilizes adsorption kinetic parameters from a knowledge base to dynamically correct for sensor surface response deviations caused by coating aging or environmental changes. This dual mechanism of physical isolation and algorithmic compensation removes the spectral characteristics of interfering components, improving the accuracy of reconstructing the target gas concentration gradient.

[0142] The logical connection between the technical solutions is as follows: the distributed sensor array optimizes signal quality through transfer learning, the synchronous acquisition module ensures the spatiotemporal consistency of multi-source data, and the synergistic effect of the anti-interference coating and chromatographic model enhances component separation. The physical properties of the sensor hardware complement the dynamic correction of the algorithm model, resolving the challenges of multi-component cross-interference and asynchronous signal acquisition in complex environments, and providing highly reliable input for noise suppression and baseline correction in step 2. This technological chain, through the deep integration of hardware design and algorithm optimization, lays the foundation for data acquisition throughout the entire leak location process.

[0143] The technical features of the present invention are explained as follows:

[0144] Heterogeneous gas sensor arrays: A distributed network composed of multiple sensor types (such as metal oxide semiconductor sensors, photoionization detectors, and electrochemical sensors) is deployed at key monitoring points within the warehouse. Metal oxide sensors detect gas concentration based on changes in surface resistance; photoionization detectors identify gas molecules through ultraviolet light ionization; and electrochemical sensors use redox reactions to determine specific gas concentrations. The array's spatially complementary layout covers blind spots and, combined with transfer learning, mitigates cross-sensitivity effects, improving detection robustness in multi-component mixed environments.

[0145] Transfer learning: A machine learning method that transfers sensor response features from historical scenarios in a knowledge base to the current environment. By extracting domain-invariant features across scenarios (such as temperature-sensitivity mapping and noise spectrum patterns), the algorithm dynamically adjusts sensor parameter weights to eliminate baseline drift caused by sudden environmental changes. During pre-training, the algorithm uses historical data to build a noise pattern feature library. During real-time runtime, the algorithm matches the current environmental parameter spectrum to generate an adaptive filtering signal.

[0146] Computational Fluid Dynamics (CFD) dynamic simulation: Based on the Navier-Stokes equations, numerical simulation of gas flow in the storage space is performed, generating a turbulent vortex distribution map. This map includes parameters such as vortex intensity, propagation direction, and energy dissipation rate, which are used to quantify nonlinear perturbations in the gas diffusion path. The simulation process couples the 3D topology (such as shelf layout and vent location) with environmental parameters (wind speed and temperature) to generate a vortex evolution model under physical constraints.

[0147] Improved Particle Swarm Optimization (PSO): A swarm intelligence optimization algorithm that incorporates an adaptive inertia weight adjustment strategy. This algorithm uses the phase offset of signal propagation between sensor nodes as the optimization variable and the eddy energy attenuation coefficient as the constraint, iteratively solving for the optimal solution for the delay compensation parameter. The optimization process dynamically matches the concentration gradient field using a sliding window to eliminate time-series signal misalignment caused by turbulence in the storage structure.

[0148] A hybrid Bayesian and LSTM neural network: This deep learning architecture combines a Bayesian probabilistic framework with a long short-term memory (LSTM) network. The Bayesian layer incorporates prior data on the spatial distribution of leakage sources to generate a posterior probability distribution. The LSTM branch uses a spatiotemporal attention mechanism to learn the temporal correlation of concentration gradients. The network's hidden layers embed discretized constraints on the mass conservation equation, forcing the output to conform to the laws of fluid dynamics.

[0149] NSGA-II Multi-Objective Optimization Algorithm: A non-dominated sorting genetic algorithm, used to resolve multi-objective conflicting problems. The algorithm optimizes positioning accuracy, model consistency, and computational efficiency, assigning weights to Bayesian inference, LSTM predictions, and CFD simulation results. It uses non-dominated sorting and congestion calculations to filter the Pareto front solution set and output high-confidence candidate solutions.

[0150] Residual Distribution-Location Error Mapping Network: This verification model, built on a convolutional neural network, takes as input the posterior probability residuals output by Bayesian and LSTM models. The network extracts the spatiotemporal correlation features of the residuals through a three-dimensional convolutional layer. Combined with the error distribution patterns of historical leak scenarios, it generates quantitative indicators of positioning uncertainty (such as a 95% confidence interval) to support credibility ratings.

[0151] Neural Architecture Search: An automated machine learning method that dynamically adjusts the number of layers and structure of Bayesian and LSTM networks. The search space is defined as a combination of LSTM branch depth and number of hidden units. A weight-sharing strategy is used to evaluate the validation loss of candidate networks. A gradient optimizer is then used to select lightweight architectures that adapt to dynamic environmental changes.

[0152] Closed-loop adaptive learning mechanism: A dynamic optimization system consisting of meta-reinforcement learning and differential evolution algorithms. The meta-reinforcement learning framework adjusts sensor sensitivity and model hyperparameters based on real-time residual distributions. The differential evolution algorithm uses the residuals from CFD simulations and measured data as a fitness function to optimize the boundary conditions for turbulent eddy intensity, forming a parameter-structure optimization chain that coordinates the inner and outer loops.

[0153] Probability cloud visualization: This visualization outputs a multi-channel fusion of diffusion path thermal maps (concentration gradient intensity) and spatial confidence intervals (Monte Carlo sampling results). Thermal values ​​are encoded as red channel intensities, and confidence intervals are encoded as transparency channels. A 3D interpolation algorithm enhances spatial continuity, generating an intuitive representation of leak source distribution and uncertainty.

[0154] Transfer learning model: This model is used to suppress cross-sensitivity effects within sensor arrays. Based on historical scenario data stored in a knowledge base, the model extracts sensor response characteristics (such as temperature-sensitivity mapping and noise spectrum patterns) under different environmental conditions. Domain adaptation training is then used to adjust sensor sensitivity weights. During pre-training, the model constructs a noise pattern feature library. During real-time runtime, the model matches the current environmental parameter spectrum to generate an adaptive filtering signal. This dynamically eliminates baseline drift caused by sudden environmental changes (such as temperature fluctuations), improving signal stability in multi-component mixed environments.

[0155] Generative Adversarial Networks (GANs): Generative adversarial networks are used to reconstruct gas diffusion path maps. The generator module receives time-aligned concentration gradient time series signals and discretized constraints from the mass conservation equation and generates a diffusion feature tensor through three-dimensional convolutional encoding. The discriminator module combines an eddy propagation direction classifier with a thermal value regressor to output a diffusion path thermal map with direction labels. The network iteratively optimizes the generated results through an adversarial training mechanism, ensuring that the diffusion path conforms to the laws of fluid dynamics while faithfully reflecting the spatiotemporal correlation characteristics of the measured data.

[0156] A hybrid Bayesian and LSTM neural network: This network combines a Bayesian probabilistic framework with a long short-term memory (LSTM) model for probabilistic inversion of leak source coordinates. The Bayesian layer incorporates prior data on the spatial distribution of historical leak sources and generates a probabilistic distribution of leak source coordinates through a posterior probability update mechanism. The LSTM branch extracts temporal correlations in concentration gradient signals using a spatiotemporal attention mechanism. The network's hidden layers embed discretized constraints from the computational fluid dynamics mass conservation equation. Using a projected gradient descent algorithm, the output is forced to satisfy flow field divergence conditions, achieving a combined drive of physical laws and data features.

[0157] NSGA-II Multi-Objective Optimization Algorithm: A non-dominated sorting genetic algorithm is used to fuse and optimize the results of multiple models. The algorithm optimizes positioning accuracy, model consistency, and computational efficiency, assigning multi-dimensional weights to the Bayesian posterior probability distribution, LSTM time-series diffusion trajectories, and CFD physics simulation results. A non-dominated sorting strategy is used to screen the Pareto frontier solution set. Combined with the validation results of the residual distribution-positioning error mapping network, anomalous solutions that violate physical constraints or historical error patterns are eliminated, resulting in a high-confidence candidate solution set.

[0158] Residual Distribution-Positioning Error Mapping Network: This verification model, built on a convolutional neural network, quantifies the uncertainty of positioning results. The model inputs the posterior probability residuals output by the Bayesian and LSTM models. A three-dimensional convolutional layer extracts the spatiotemporal correlation features of the residuals. Combined with error distribution data from historical leak scenarios, it generates a quantitative indicator of positioning uncertainty (such as a spatial confidence interval). This indicator provides a statistical basis for confidence ratings and supports the interpretability of positioning results.

[0159] Neural Architecture Search (NAS) technology: An automated machine learning method is used to dynamically adjust Bayesian and LSTM network structures. The search space is defined as the number of layers and hidden units in the LSTM branches. A weight-sharing strategy is used to evaluate the validation loss of candidate networks, combined with a gradient optimizer to select lightweight architectures. The search process dynamically adjusts network depth based on the real-time residual distribution, adapting to the complexity of temporal features caused by environmental changes and improving model generalization.

[0160] Differential Evolution: A swarm optimization algorithm is used to update the boundary condition parameters of the computational fluid dynamics model. Using the residual between the CFD simulation results and the measured concentration gradient field as a fitness function, the algorithm iteratively optimizes the population distribution of turbulent vortex intensity parameters. Divergence constraints from the mass conservation equation are introduced during the population update process to enforce that the optimized boundary parameters conform to the laws of fluid dynamics, ensuring consistency between the physical model and the measured data.

[0161] The specific implementation method of the present invention realizes the precise positioning of the leakage source in the hazardous chemical storage scenario through the collaboration of multi-level technologies. A heterogeneous array composed of metal oxide semiconductor sensors, photoionization detectors and anti-interference coated electrochemical sensors is deployed in the storage space, and the distributed nodes cover the shelf gaps, vents and ground areas. The sensor array dynamically adjusts the sensitivity weight through the transfer learning model. The model training is based on the cross-sensitive feature data set stored in the knowledge base to suppress the spectral interference of multi-component gases such as benzene and hydrogen sulfide. The temperature and humidity sensor and the three-dimensional ultrasonic anemometer are integrated into the same acquisition module, and the hardware clock synchronization protocol is used to achieve millisecond-level timestamp alignment. The multimodal data set containing three-dimensional wind speed field, temperature and humidity gradient and gas concentration time series signals is output, providing a time-space-related physical field input for subsequent processing.

[0162] A multimodal dataset is fed into a pre-built knowledge base for baseline correction. A transfer learning algorithm extracts spectral features of environmental parameters and matches them to temperature drift templates in a noise pattern library. Adaptive filtering signals are used to phase-compensate the pressure pulsation time series data, eliminating sensor zero-point drift caused by diurnal temperature fluctuations. The corrected data undergoes Daubechies wavelet packet decomposition to separate high-frequency noise and low-frequency diffusion components. This data is then combined with an attention weight allocation model from the turbulence-diffusion correlation map to reconstruct the target gas concentration gradient field. A chromatographic response transfer model utilizes activation energy data from an adsorption kinetics parameter library. Using a meta-learning framework, the adsorption coefficient matrix is ​​optimized, removing the spectral features of interfering components such as ethanol and acetone to generate a pure concentration-component feature vector.

[0163] A computational fluid dynamics (CFD) dynamic simulation engine generates a turbulent vortex distribution map based on the three-dimensional warehouse topology and constructs a vortex propagation delay model for sensor nodes. An improved particle swarm optimization algorithm, constrained by the vortex energy attenuation coefficient, iteratively solves for delay compensation parameters and aligns multi-node concentration gradient time series signals. Compensation data and the discretized constraints of the mass conservation equation are combined via a generative adversarial network to synthesize diffusion feature tensors, generating a path heat map with vortex direction markers. A Bayesian-LSTM hybrid neural network embeds a regularization term in the mass conservation equation in the hidden layer. A spatiotemporal attention mechanism extracts vortex propagation direction features, and a projected gradient descent algorithm enforces the network output to meet flow field divergence constraints. The NSGA-II multi-objective optimization algorithm integrates Bayesian posterior probabilities, LSTM time series trajectories, and CFD physics data to screen Pareto front solutions. A residual distribution-to-localization error mapping network generates confidence ratings. A closed-loop adaptive mechanism dynamically adjusts sensor sensitivity and network hyperparameters through meta-reinforcement learning. A neural architecture search optimizes the number of LSTM layers, and a differential evolution algorithm updates CFD boundary conditions, forming an environmentally adaptive leak location system.

[0164] In petrochemical storage scenarios, the above-mentioned implementation method effectively overcomes the diffusion path distortion caused by turbulent interference, multi-component cross-sensitivity and signal lag through hardware-algorithm collaborative optimization, physical-data model fusion and dynamic closed-loop adjustment, achieves millimeter-level positioning accuracy of the leakage source coordinates, and meets the engineering needs of hazardous chemical storage safety monitoring.

[0165] This method solves the problem of gas diffusion signal distortion in complex environments through multi-physics field data fusion and physical constraint modeling, improving the accuracy of leak source inversion. First, a heterogeneous gas sensor array is used to synchronously collect multimodal data such as gas concentration, composition, and three-dimensional wind velocity field. The sensors are pre-trained for cross-domain adaptation using a transfer learning algorithm. The noise pattern feature library in the knowledge base is dynamically matched to generate a baseline-corrected calibration dataset. This calibration process eliminates sensor zero-point drift caused by sudden temperature changes, suppresses environmental noise interference with the original signal, and provides a high signal-to-noise ratio input for subsequent processing.

[0166] Secondly, a turbulent vortex distribution map is generated based on dynamic computational fluid dynamics simulations, and a model of vortex propagation delay between sensor nodes is constructed. An improved particle swarm optimization algorithm is used to iteratively solve for phase offset parameters, aligning the multi-node concentration gradient time series signals in the time domain to eliminate the time series misalignment caused by turbulence in the storage structure. The compensated concentration gradient data and the mass conservation equation constraints are combined through a generative adversarial network to synthesize the diffusion characteristic tensor. This reconstructs a gas diffusion path map driven by the synergy of physical laws and measured data, restoring the spatiotemporal correlation characteristics distorted by multi-physics field coupling.

[0167] Finally, the mass conservation equation is embedded as a hard constraint in a hybrid Bayesian and LSTM neural network architecture. The network parameters are jointly optimized using a projected gradient descent algorithm to enforce compliance of the model output with the laws of fluid dynamics. By integrating the multi-model results of Bayesian inference, LSTM time series prediction, and CFD simulation, the NSGA-II multi-objective optimization algorithm is employed to dynamically assign model weights and screen the Pareto front solution set. A residual distribution-positioning error mapping network quantifies the uncertainty of the solution set and generates positioning results with confidence ratings. This results in an inversion mechanism that is both data-driven and physically constrained, significantly improving the accuracy of leak source location in complex turbulent environments.

Claims

1. A method for locating and tracing the source of hazardous chemical storage leaks using a gas array, characterized in that: include: Step 1: Collect the gas concentration, composition and environmental physical field parameters of the storage environment to generate a multimodal raw data set; Step 2: Input the multimodal raw data set into a preset 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 of the real-time environmental parameter spectrum and the noise pattern feature library in the knowledge base, and output a calibrated multi-physics field data set; Step 3: Perform multi-scale wavelet packet decomposition on the gas concentration time series signal in the calibrated multi-physics field dataset, assign noise suppression weights using a bidirectional gated attention mechanism based on the turbulence and diffusion correlation map in the knowledge base, and simultaneously construct a transfer learning-driven chromatographic response migration model by calling the gas adsorption kinetic parameter library to output the de-noised concentration gradient signal of the target gas and the separated component feature vectors; Step 4: Based on the turbulent vortex distribution map generated by computational fluid dynamics dynamic simulation, the response lag effect between sensor nodes is quantified. A particle swarm optimization algorithm is used to iteratively solve the time delay compensation coefficient and align the multi-node timing signals. The compensated concentration gradient data and physical field constraints are combined with a generative adversarial network to synthesize a baseline diffusion feature vector to generate a gas diffusion path map. Step 5: The computational fluid dynamics mass conservation equation is embedded as a hard constraint into the Bayesian and LSTM hybrid neural network. The network parameters are jointly optimized using the projected gradient descent algorithm to generate the posterior probability distribution of the leak source coordinates. In step 6, the posterior probability distribution output by the Bayesian and LSTM hybrid neural network, the computational fluid dynamics simulation results, and the LSTM prediction data are input into the integrated learning framework. The NSGA-II multi-objective optimization algorithm is used to dynamically allocate model weights and screen the Pareto optimal solution set to generate the credibility rating positioning result.

2. The method for locating and tracing the source of hazardous chemical storage leaks 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; Performing spectrum similarity comparison on the noise pattern feature library inputting the three-dimensional wind speed field and the temperature gradient distribution into the knowledge base; An adaptive filtering signal based on transfer learning domain adaptation is generated according to the comparison result, temperature drift compensation is performed on the air pressure pulsation time series signal, and a corrected three-dimensional physical field parameter set is output.

3. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 2, characterized in that: Also includes: performing Daubechies wavelet packet decomposition on the gas concentration time series signal in the corrected three-dimensional physical field parameter set to extract high-frequency noise components and low-frequency diffusion components; The turbulence and diffusion association map in the knowledge base is used to allocate attention weights to the high-frequency noise components, and the target gas concentration gradient field is reconstructed by combining the chromatographic response migration model driven by transfer learning. The adsorption coefficient matrix in the gas adsorption kinetic parameter library is dynamically optimized through a meta-learning framework to generate the spectral characteristics of pure components after removing cross-interference.

4. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 3, characterized in that: Also includes: Based on the turbulent vortex distribution map output by computational fluid dynamics simulation, a vortex propagation delay model between sensor nodes is constructed; An improved particle swarm optimization algorithm is used to iteratively solve the phase offset parameter in the vortex propagation delay model, and the reconstructed target gas concentration gradient field is aligned in the time domain; Generate a spatiotemporally synchronized multi-node concentration gradient time series signal set.

5. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 4, characterized in that: Also includes: Embed the discretized constraints of the computational fluid dynamics mass conservation equation in the hidden layers of the Bayesian and LSTM networks; Extracting vortex propagation direction features from the multi-node concentration gradient time series signal set through a spatiotemporal attention mechanism; The projected gradient descent algorithm is used to constrain the flow field divergence of the output tensor of the LSTM unit, forcing the prediction results to satisfy the law of conservation of mass.

6. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 1, characterized in that: The step 6 comprises: 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 quantitative indicator of positioning uncertainty; Integrate the posterior probability data of Bayesian inference, the time series diffusion trajectory predicted by LSTM, and the physical field constraint results of CFD simulation into an integrated learning framework; The NSGA-II multi-objective optimization algorithm is used to perform multi-dimensional weight allocation on the posterior probability data, time series diffusion trajectory and physical field constraint results to generate a Pareto front solution set.

7. The method for locating and tracing the source of a hazardous chemical storage leak in a gas array according to claim 6, characterized in that: Also includes: Based on the deviation distribution between the credibility rating positioning results and the real-time monitoring residual signal, the sensor sensitivity parameters and the Bayesian and LSTM hybrid neural network hyperparameters are dynamically adjusted through a meta-reinforcement learning framework, and the network depth is reconstructed based on neural architecture search technology; 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; Adaptively adjust the number of layers of the LSTM branch in the Bayesian and LSTM hybrid neural network through neural architecture search technology; Optimization of boundary condition parameters of turbulent vortex intensity in computational fluid dynamics model based on differential evolution algorithm.

8. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 1, characterized in that: The step 4 comprises: The spatiotemporal trajectory data of historical leakage events stored in the knowledge base is used to construct a training sample set for the adversarial generative network; The diffusion characteristic tensor is synthesized by combining the time-domain aligned multi-node concentration gradient time series signal set with the mass conservation equation constraint; The discriminator module of the generative adversarial network outputs a diffusion path heat map marked with the vortex propagation direction.

9. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 8, characterized in that: Also includes: The residual distribution model of the positioning error mapping network is trained based on 3D reconstruction data of historical leakage scenes; Monte Carlo sampling is performed on the posterior probability distribution output by the Bayesian and LSTM hybrid neural network to extract the boundary features of the spatial confidence interval; The diffusion path thermal map is superimposed with the spatial confidence interval feature to generate a probability cloud map visualization output.

10. The method for locating and tracing the source of hazardous chemical storage leaks in a gas array according to claim 1, characterized in that: The step 1 comprises: Distributed arrays of metal oxide semiconductor sensors and photoionization detectors, suppressing cross-sensitivity effects through transfer learning models; A synchronous acquisition module integrating temperature and humidity sensors and a three-dimensional ultrasonic anemometer outputs multi-physics field parameters with time stamp alignment in real time; The anti-interference coating provided in the electrochemical sensor array cooperates with the chromatographic response migration model to strip off the interfering component spectrum.

Images