A combined reasoning method for initial state elements of gas explosion

Abstract

The present application relates to a kind of gas explosion initial state element combined reasoning method, comprising: obtaining the multimodal evidence related to gas explosion scene, and input pre-trained completed deep learning network, the deep learning network includes multimodal perception coding layer, structured bayesian latent space layer and uncertainty reasoning decision layer;Based on multimodal perception coding layer, the feature extraction and fusion of multimodal evidence are carried out, and the encoding feature vector of output is obtained;Based on structured bayesian latent space layer, the encoding feature vector of the output of multimodal perception coding layer is mapped into the probability distribution space of pre-defined multiple initial state elements of gas explosion;Based on uncertainty reasoning decision layer, according to the probability distribution space of multiple initial state elements output by structured bayesian latent space layer, sampling and calculation are carried out, and the combination result of multiple initial state elements according to confidence is output.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and public safety, specifically to the deep integration of multiple technical directions such as computer vision, deep learning, probabilistic graphical models, accident causation investigation and computational fluid dynamics, and particularly to a joint reasoning method for initial state elements of gas explosion that integrates physical semantic constraints and variational inference. Background Technology

[0002] Gas explosions are a common type of accident, with complex causes and high destructive power. Post-accident investigations are of vital strategic importance.

[0003] However, traditional accident investigation methods face numerous serious challenges and inherent limitations when dealing with complex, multi-source, and even partially missing on-site evidence, mainly manifested in the following problems:

[0004] (1) Limitations of evidence processing: Traditional investigations rely heavily on the human experience and intuitive judgment of experts. Faced with multimodal heterogeneous evidence such as three-dimensional spatial damage (e.g., wall displacement), microscopic physical and chemical traces (e.g., molten metal beads), and time-series data (e.g., alarm records), there is a lack of a unified mathematical framework to effectively integrate and quantify it. This makes the reasoning process difficult to reproduce and makes it impossible to scientifically assess the uncertainty of the evidence; (2) The "black box" problem of the reasoning process: With the development of artificial intelligence technology, some studies have tried to use simple machine learning or neural network models for accident analysis. However, these models are often based on pure data-driven reasoning and lack basic physical logic constraints, which may lead to results that do not conform to physical logic, and the scientific validity and reliability of their conclusions are difficult to guarantee; (3) Certainty bias in conclusions: Traditional investigation conclusions are usually single, "black and white" judgments, such as directly identifying a certain electrical appliance as the ignition source. Such certain conclusions cannot reflect the multiple possibilities that exist in the real world, nor can they provide investigators with a probabilistic ranking of possible combinations when evidence is insufficient. More importantly, it cannot proactively suggest the direction of the investigation and lacks the decision support capability to provide intelligent guidance for subsequent supplementary investigations. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to provide a joint reasoning method for initial state elements of a gas explosion, which can solve the difficulties faced by traditional technologies, address the problem of unified modeling of multimodal heterogeneous evidence, and embed physical laws into the reasoning framework of the model, making the reasoning scientific, while also providing support for quantifying the uncertainty of the reasoning conclusions and decision-making.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, this application provides a method for joint reasoning of initial state elements in a gas explosion, including: Acquire multimodal evidence related to the gas explosion scene and input it into a pre-trained deep learning network, which includes a multimodal perception encoding layer, a structured Bayesian latent space layer, and an uncertainty reasoning decision layer. Based on the multimodal sensing coding layer, feature extraction and fusion are performed on multimodal evidence to obtain the output coded feature vector; Based on the structured Bayesian latent space layer, the encoded feature vector output by the multimodal sensing coding layer is mapped to the probability distribution space of multiple predefined initial state elements of gas explosion. Based on the uncertainty reasoning decision layer, sampling and calculation are performed on the probability distribution space of multiple initial state elements output by the structured Bayesian latent space layer, and the combined result of multiple initial state elements arranged according to confidence level is output.

[0007] In one implementation, the multimodal evidence includes: spatial damage evidence of a gas explosion site, wherein the spatial damage evidence includes a three-dimensional scan point cloud of the site or its textual description; The multimodal evidence also includes: microscopic images of combustion traces at the scene of a gas explosion or a textual description thereof; Alternatively, the multimodal evidence may also include: time-series data of gas concentration detected by sensors.

[0008] In one implementation, the structured Bayesian latent space layer encodes feature vectors, performs feature decoupling and mapping, performs heterogeneous distribution modeling based on the different attributes of multiple predefined initial state elements of gas explosions, and samples different mathematical representations for initial state elements with different attributes. In one implementation, the attributes of the initial state elements include discrete attributes and continuous attributes; Initial state elements of discrete attributes include: gas type ( ) and stimulation method ( ); Initial state elements of continuous attributes include: leakage location ( ), Detonation location ( ) and leakage methods ( ).

[0009] In one implementation, the deep learning network further includes a PINN physical consistency constraint layer, which trains the network parameters based on a pre-designed loss function that incorporates physical laws during the pre-training phase of the deep learning network.

[0010] In one implementation, the loss function includes a gas diffusion consistency term and an energy conservation constraint term.

[0011] In one implementation, during the pre-training phase, different differential reparameter sampling methods are used for discrete and continuous attributes, wherein: For discrete properties, the Gumbel-Softmax operator is used to transform the non-differentiable discrete sampling process into a continuously differentiable function mapping, thereby realizing the backpropagation of the physical residual gradient. For continuous properties, an anisotropic Gaussian reparameterization method is adopted. By introducing auxiliary noise, the sampling process is separated from the distribution parameters, so as to achieve complete differentiability of the sampling points with respect to the mean and variance.

[0012] In one implementation, the method achieves adaptive evolution of the inference results through a gradient collaborative optimization mechanism, including: Dynamic annealing and shrinkage coupling: During training, the dynamic decay of the annealing temperature parameter is used to drive the discrete branches to tend towards deterministic classification. At the same time, combined with the dynamic shrinkage of the anisotropic standard deviation, the probability weights that do not conform to the physical logic region are automatically suppressed. State evolution: Through the coupling optimization of the parameters, the reasoning conclusion is driven to evolve from the fuzzy probability space to the high-confidence physical initial state conclusion.

[0013] In one implementation, the method further includes: a supplementary exploration decision guidance mechanism based on cognitive uncertainty, which includes uncertainty measurement, decision triggering logic, and intelligent guidance generation. Uncertainty Measurement: The cognitive uncertainty of the reasoning conclusion is quantified by using the variance of each initial state element output from the structured Bayesian latent space layer. Decision triggering logic: A preset uncertainty threshold or distribution flatness evaluation criterion is used. When the prediction variance of a certain initial state element exceeds the threshold, the initial state element is identified as the target. Intelligent guidance generation: Based on the physical semantics of the target, it automatically associates it with the corresponding physical space area or physical evidence type, and generates targeted supplementary inspection direction guidance to assist investigators in conducting accurate supplementary inspections.

[0014] In one implementation, the logic for generating supplementary survey direction guidance includes: If the variance of the ignition method exceeds the threshold, based on the coordinate distribution of the detonation location, the system will automatically extract potential sources of electrical wire melting marks, switch sparks, or static electricity risks in the surrounding area as supplementary exploration prompts. If the variance of the leak location exceeds the threshold, the investigators will be prompted to focus on reviewing the concealed pipelines or locations where the seals have failed, based on the areas where the gas diffusion consistency constraint has failed.

[0015] Compared with the prior art, the technical solution proposed in this invention brings the following three significant advantages and beneficial effects: 1. Significantly Enhanced Scientific Rigor and Interpretability of Reasoning: By creatively introducing a physical information loss function, this invention completely liberates deep learning models from the limitations of traditional "black boxes." The reasoning process is no longer a simple data fitting process, but follows inherent physical logic. This makes the final conclusions not only more scientific and reliable, but also possess a certain degree of interpretability. For example, the system can explicitly point out that a seemingly possible combination of conclusions is assigned an extremely low probability because it violates the law of conservation of energy (e.g., the total leakage amount is insufficient to cause on-site damage).

[0016] 2. Probabilistic Conclusions and Dynamic Decision Support Capabilities: The greatest technological leap achieved by this invention lies in its departure from 'black-and-white' judgments to 'probabilistic guidance with physical logic'. The system outputs not only the most probable accident scenario but also other possibilities and their probabilities, and quantifies the confidence level (variance) of each key element. When the confidence level of a certain element is low, the system can automatically identify and generate prompts, such as "The confidence level of the triggering method is low; it is recommended to supplement the investigation by examining the wire melting marks at location XX," thereby providing dynamic and intelligent decision support for accident investigation work, greatly improving investigation efficiency and accuracy.

[0017] 3. Robustness and Efficient Evidence Fusion: The multimodal heterogeneous encoder designed in this invention can process various information in parallel, including 3D point clouds from the scene, wreckage photographs, sensor data, and even textual descriptions from professionals. This enables comprehensive utilization of on-site evidence and adaptability to real-world applications. More importantly, based on a Bayesian probabilistic framework, the system can effectively address common situations in real-world scenarios such as conflicting evidence or missing key evidence. It will not collapse due to the lack of a small amount of evidence, but rather reflects the insufficient evidence by increasing the uncertainty (variance) of the corresponding elements. This fundamentally solves the problem of "insufficient evidence leading to wild guesses" in traditional models, significantly improving the system's robustness.

[0018] In summary, this invention achieves a paradigm shift from traditional experience-driven to a dual-driven "data + physics" approach to accident investigation by deeply integrating data-driven deep learning with knowledge-driven physical models, providing a new and effective technical path for the intelligent development of this field. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture and data flow of the system provided in the embodiments of this application; Figure 2 This is a schematic diagram of the data flow of inference, loss calculation, and sampling feedback in the system of this application embodiment; Figure 3 This is a detailed network diagram of the system based on the five-element branch in this embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0021] Existing solutions based on machine learning networks to infer the initial state of a gas explosion are prone to producing inferences that do not conform to scientific logic, or that only provide a definitive result without providing multiple possible results with uncertainty, thus failing to provide supplementary investigations for users based on multiple possibilities.

[0022] To address the above problems, embodiments of the present invention provide a method for joint reasoning of initial state elements in a gas explosion, comprising: Acquire multimodal evidence related to the gas explosion scene and input it into a pre-trained deep learning network, which includes a multimodal perception encoding layer, a structured Bayesian latent space layer, and an uncertainty reasoning decision layer. Based on the multimodal sensing coding layer, feature extraction and fusion are performed on multimodal evidence to obtain the output coded feature vector; Based on the structured Bayesian latent space layer, the encoded feature vector output by the multimodal sensing coding layer is mapped to the probability distribution space of multiple predefined initial state elements of gas explosion. Based on the uncertainty reasoning decision layer, sampling and calculation are performed on the probability distribution space of multiple initial state elements output by the structured Bayesian latent space layer, and the combined result of multiple initial state elements arranged according to confidence level is output.

[0023] This will solve the following three core technical problems: (1) The problem of unified modeling of multimodal heterogeneous evidence: How to construct a unified deep learning model that can effectively integrate and understand various types and structures of evidence obtained from on-site investigation, such as spatial point clouds, microscopic images, time series data, and text descriptions, and extract key information for reasoning.

[0024] (2) The problem of integrating deep learning with physical laws: How to embed physical laws (such as gas diffusion equation, energy conservation law, etc.) in the reasoning framework of deep learning models in a computable and differentiable way to ensure that the reasoning results of the models have the self-consistency and scientific nature of physical logic, thereby avoiding the absurd conclusions that violate basic physical common sense.

[0025] (3) Quantification of uncertainty in reasoning conclusions and decision support issues: How to upgrade traditional deterministic "single-point" conclusions into probabilistic "distribution" conclusions with confidence assessment, and accurately quantify the "cognitive uncertainty" of the model for each reasoning element when evidence is insufficient, so as to provide intelligent directional guidance for subsequent supplementary investigation work.

[0026] The following description, based on the accompanying drawings, further details the network model (system) of the method provided in this application, including its structure, training, application examples, and technical effects.

[0027] This invention proposes a system (i.e., a machine learning network model for joint inference of the five initial elements of a gas explosion, integrating physical semantic constraints and variational inference). The core idea of ​​this system is based on the concepts of a Structured Conditional Variational Autoencoder (S-CVAE) and a Physics-Informed Neural Network (PINN), constructing a probabilistic model capable of inferring the initial physical conditions from the final state damage evidence. By introducing a residual term conforming to physical laws into the loss function of the S-CVAE, the model's representation in the latent space not only satisfies statistical distribution but also follows physical logic constraints. This residual term, calculated by PINN, covers fundamental physical laws such as mass conservation, energy conservation, and state equations, ensuring that the variable evolution during the inference process conforms to the actual physical scenario. The model achieves collaborative inference of the five elements—gas type, leak location, detonation location, activation mode, and leak mode—by jointly optimizing the evidence likelihood and physical residual loss. The inference results are output in the form of probability distribution, which supports uncertainty quantification and multiple solution analysis, and significantly improves the robustness and interpretability of accident reconstruction.

[0028] Figure 1 The overall architecture of the system is shown.

[0029] The system is logically divided into four tightly coupled core modules, forming a complete closed loop from initial evidence input to probabilistic conclusion output. These four modules are: Multimodal sensing coding layer (Input & Encoding Layer): Responsible for receiving and processing various heterogeneous evidence collected on-site, and transforming them into a unified high-dimensional feature vector.

[0030] Structured Bayesian Latent Space Layer: The core inference engine of the system, responsible for mapping feature vectors to a probability distribution space with explicit physical meaning.

[0031] PINN Physics-Informed Constraint Layer: Through a specially designed loss function, physical laws are applied as strong constraints to the model training process to ensure the scientific nature of the inference.

[0032] Uncertainty-Quantified Inference & Decision Layer: Responsible for sampling and analyzing the probability distribution of the latent space, outputting the most likely conclusion and the confidence assessment of each element.

[0033] Based on a multimodal sensing coding layer and a structured Bayesian latent space layer, the core of the system is a designed structured conditional variational autoencoder (S-CVAE). Its most critical innovation lies in the fact that we forcibly decompose the unstructured and uninterpretable latent space vector z in traditional CVAEs into five probabilistic random variables with independent physical semantics, which respectively represent the five initial elements of a gas explosion: gas type, leakage location, detonation location, activation mode, and leakage mode.

[0034] The PINN Physical Consistency Constraint Layer employs a Physics-Informed Loss function. This function simplifies complex nonlinear physical equations (such as the gas diffusion equation and the law of energy conservation) into differentiable regularization terms and embeds them into the neural network's training process. This is equivalent to putting "physical shackles" on the neural network's parameter optimization, forcing the network to adhere to fundamental physical laws while learning data patterns.

[0035] The system fully leverages the inherent advantages of variational inference. For each initial state element, the model output is not a fixed value, but a probability distribution (usually derived from the mean). and variance (Parameterization). During reasoning, by performing Monte Carlo sampling on this distribution, not only can we obtain the most probable combination of conclusions, but more importantly, we can quantify the confidence level of the conclusions by analyzing the magnitude of the variance, thereby accurately identifying cognitive uncertainty caused by insufficient evidence.

[0036] Figure 2 The data flow is displayed in detail across four levels.

[0037] Detailed descriptions are provided for each layer.

[0038] Multimodal sensing coding layer The multimodal sensing coding layer is the system's "sensory organ," responsible for transforming unstructured evidence of different forms collected on-site into a unified high-dimensional feature vector that can be processed by subsequent networks.

[0039] The table below shows the various types of multimodal evidence that the multimodal sensing coding layer can process.

[0040] Structured Bayesian Latent Space Layer This layer is responsible for mapping the fused features output by the multimodal sensing coding layer into five probability distributions with clear physical meanings that characterize the "five elements" of the initial state of the explosion.

[0041] Structured variable definition: These five probabilistic random variables together constitute the structured latent space: These represent the leak location, ignition location, activation method, gas type, and leak method, respectively.

[0042] Probability distribution representation: Each variable is represented by its mean. and variance Representation, i.e.

[0043] Uncertainty output: via variance The size of the evidence indicates a conflict, guiding further on-site investigation.

[0044] In this detailed embodiment, the core of achieving joint reasoning of the "five elements" based on the Structured Bayesian Latent Space lies in decoupling the high-dimensional fusion features and mapping them to probability distribution branches with clear physical meaning.

[0045] Summary of the hierarchical implementation method of the "five elements" To ensure the structured and scientific nature of the reasoning process, the system employs a four-step method: "feature extraction—decoupling mapping—distributed reconstruction—physical verification." 1) Feature Decoupling and Mapping: The fused feature vector output by the multimodal coding layer After entering the structured latent space layer, the distribution parameters (such as mean) of the five elements are calculated through five independent parallel fully connected sub-networks (Heads). and variance ).

[0046] 2) Heterogeneous Distribution Modeling: This approach uses different mathematical representations for the physical attributes of different elements. Among them, for discrete attributes ( The key elements are: using Categorical or Dirichlet distribution modeling, combined with the Gumbel-Softmax technique to solve the problem of non-differentiability of discrete variables.

[0047] For spatial / continuous properties ( ): Models using a Gaussian distribution, outputting a probability density plot of coordinates or physical parameters.

[0048] 3) Reparameterization: All elements are sampled from the distribution through reparameterization techniques to ensure that the gradient can propagate back through the random variables to the perceptual coding layer.

[0049] 4) Physical Semantic Alignment: The output of each branch is not only driven by data, but must also be forced to meet the corresponding physical constraints (such as ignition energy matching, diffusion law, etc.) through the loss function.

[0050] A detailed network diagram of the "five elements" is shown below. Figure 3 .

[0051] The detailed technical specifications of each element branch are shown in the table below.

[0052] In this embodiment, the PINN physical consistency constraint layer provides strong scientific prior knowledge for the deep network's learning process by designing a special physical information loss function, forcing it to obey physical laws. The total loss function consists of three parts:

[0053] in, The standard reconstruction loss is used to ensure that the latent space can effectively encode the input information; KL divergence is used to constrain the closeness between the posterior and prior distributions. and These are adjustable hyperparameters used to balance the weights of different loss terms. Additionally, in the early stages of model training, to prevent posterior collapse due to a rapid decrease in KL divergence, the weight coefficients in the loss function are adjusted. The preferred approach is dynamic annealing (Warm-up) strategy, which allows the temperature to smoothly increase from a small value to a preset value with the number of iterations. This strategy works in conjunction with the dynamic decay logic of annealing temperature in the discrete attribute inference branch to ensure that the model can explore the distribution space of evidence features more fully in the early stages of learning, thereby guiding the inference conclusion to evolve from fuzzy probability to high-confidence physical reality.

[0054] The core lies in the physical loss term. It consists of the following two core components: (1) Gas diffusion consistency term

[0055] This project utilizes simplified physical operators (based on the advection-diffusion equations) to constrain the logical consistency between the leak point, gas type, and observational evidence. Its core is to transform complex fluid dynamics laws into a "physical constraint chain" understandable by neural networks. This loss term is defined by the following formula:

[0056] in, The specific physical expansion of the function is:

[0057] in, This indicates the gas concentration field inferred by the system based on on-site evidence; It is determined by latent space variables The diffusion coefficient is determined by the type of gas. It is by (Location leaked) and The source item defined by (leakage method).

[0058] This approach enforces the laws of physics through a closed-loop mechanism of "prediction-comparison-penalty-correction": Physics-guided forward propagation: When S-CVAE prediction It is a basement and When the gas is lighter than air, a buoyancy-driven diffusion model is automatically applied to the physical consistency layer. Logically, the gas will rapidly accumulate upwards.

[0059] Multimodal evidence verification: The system extracts information from the multimodal sensing coding layer (such as photographs of the ceiling or descriptions of smoke marks). If the on-site evidence shows that the ceiling remains clean, i.e., the observation value... .

[0060] Residual Conflicts and Penalties: The "theoretical high concentration" calculated by physical formulas and the "actual low concentration" from field evidence can produce huge mathematical discrepancies (i.e., physical residuals). (Significantly increased). At this point, the system will generate a huge penalty value (High Loss).

[0061] Backward gradient correction: This penalty value, through differential operators such as Gumbel-Softmax, backpropagates the gradient to the front-end encoder, forcing the network to abandon the erroneous assumption that "natural gas is leaking in the basement" and instead search for combinations that better conform to physical laws. For example, the leak point is above, or the gas is liquefied petroleum gas, which is heavier than air.

[0062] (2) Energy conservation constraint terms

[0063] This study uses an explosion overpressure formula or energy balance equation based on TNT equivalent to verify whether the inferred energy release matches the on-site structural damage on order of magnitude. Its form can be expressed as:

[0064] Its physical logic is: the system will verify the "gas type + leakage method / total amount" deduced from the data. The theoretical chemical energy released (calculated) Whether it is related to the mechanical work required for destruction calculated based on the on-site structural damage (X). These are on the same order of magnitude. If the theoretical energy is much greater or less than the energy required for actual destruction, this term will also incur a significant penalty. This is a preset smoothing term for tiny positive real numbers, used to ensure the stability of numerical calculations. Its preferred range is... to In actual calculations, it is often taken as .

[0065] Specifically, chemical energy release calculations The system calculates the theoretical total energy based on the inferred initial state elements:

[0066] in, For the reason The determined mass flow rate; It is the lower heating value of the gas; It is the explosion efficiency factor (usually taken as 0.03-0.15).

[0067] On-site damage calculation The displacement and failure state of components were extracted using the on-site 3D point cloud. The mechanical energy absorbed by the structure is calculated using an equivalent single degree of freedom (SDOF) model:

[0068] in, It is the first Measured residual displacement of a damaged component (such as a wall or window); It is the resistance function of the component, representing the work required to destroy the component.

[0069] In this embodiment, in order to perform model training and variational inference, differential reparameter sampling needs to be implemented for the structured latent space variables.

[0070] In the "Gas Explosion Initial State Joint Inference System," the inference results include discrete semantic variables (such as gas type) and continuous physical variables (such as leak location and detonation energy). To address the gradient truncation problem caused by the non-differentiability of sampling operations, this invention designs differentiating and reparameterizing operators for initial state elements with different characteristics, ensuring that the loss feedback of the Physical Information Neural Network (PINN) can guide feature extraction across layers.

[0071] Gas types (such as methane, liquefied petroleum gas, manufactured gas, etc.) are typical discrete semantic variables. Due to traditional discrete sampling operations (such as...) The sampling method (or random category sampling) is mathematically non-differentiable, causing the physical residual gradient generated by the backend Physical Information Neural Network (PINN) to be unable to penetrate the sampling layer and be fed back to the frontend encoder. To solve this "gradient discontinuity" problem, this embodiment employs the Gumbel-Softmax (Concrete Distribution) technique to transform the discrete category selection process into a continuously differentiable function mapping. The specific implementation scheme is as follows: (1) Continuous modeling of category distribution Suppose the raw log-odds vector (Logits) output by the system inference head (Head_Gas) is: , ,in The total number of preset gas categories. To simulate the random sampling process while maintaining the differentiability of the computation graph, independent and identically distributed standard Gumbel noise is introduced. The calculation method is as follows:

[0072] By superimposing noise terms onto the Logits space, a continuous classification feature vector containing randomness is constructed. The formulas for calculating each component are as follows:

[0073] in, For annealing temperature parameters, when Time: Output vector It approximates one-hot encoding (discrete sampling effect). When Time: The output is a continuously differentiable probability vector, allowing gradient penetration. The data is then fed back to the encoder network. During training, adjustments can be made to... Employ a dynamic decay strategy (such as exponential decay).

[0074] (2) Physical gradient cross-layer backpropagation mechanism During the joint training phase, the total physical losses (from the residual of the diffusion equation) and energy conservation residual Composition) Probability distribution parameters of gas types The feedback gradient is achieved using the following chain rule:

[0075] in, This is the Kronecker delta symbol.

[0076] Advantages of differential reparameterization of discrete semantic variables: 1) Eliminate gradient truncation: Through the above-mentioned differential mapping, the error of "the gas type does not match the on-site thermal trace" found by the physical verification layer can be transmitted back to the feature fusion layer and the multimodal encoder in the form of gradient without loss.

[0077] 2) Semantic correction guided by physical properties: Different types of gas correspond to different diffusion coefficients. and chemical reaction energy Within the framework of this invention, the physical loss term... This will "force" the classification branches. For example, if the current point cloud evidence supports "large-scale leakage," but physical calculations indicate that only high-calorific-value liquefied petroleum gas can cause the structural damage observed on-site, the system will automatically lower the prediction probability of low-calorific-value gas (such as blast furnace gas). .

[0078] 3) End-to-end co-evolution: This implementation ensures that discrete semantic recognition ("what kind of gas") and continuous physical field simulation ("how badly it exploded") are co-optimized in the same manifold space, which significantly improves the logical consistency of reasoning results in complex field environments.

[0079] For continuous variables such as leak location, initial concentration, and detonation scale, this invention employs anisotropic Gaussian reparameterization to achieve precise shaping of the spatial probability distribution: (1) Parametric characterization: the mean of the output position distribution of the system inference head and anisotropic standard deviation This characterizes the most likely location and its cognitive uncertainty.

[0080] (2) Sampling formula: By introducing independent and identically distributed auxiliary noise , sampling points Represented as:

[0081] in This is an element-wise multiplication. This mapping strips away randomness. , making For parameters and Completely differentiable.

[0082] (3) Spatial Focusing Logic: Physical loss is achieved through the following: Differentiation: It dynamically compresses the probability weights of non-physical logic regions to achieve automatic focusing on the source of leakage on site.

[0083] In the total loss function Driven by this, the system establishes the following gradient feedback flow: Semantic-physical cooperation: when As the number of discrete branches gradually decreases, the discrete branches tend towards deterministic classification. It will automatically shrink based on the physical verification results, so that continuous variables and discrete categories can achieve logical self-consistency within the physical boundaries.

[0084] End-to-end correction: The gradient eventually penetrates the sampling layer and is fed back to the front-end multimodal encoder, guiding the model to prioritize the extraction of features most relevant to the physical evolution logic when processing 3D point cloud or sensor data.

[0085] In this embodiment, the system's functionality is implemented in two stages: offline pre-training and on-site inference.

[0086] The goal of offline pre-training is to enable the model to learn, from massive amounts of simulated data, the universal, non-linear mapping relationship between the "five elements of the initial state of the explosion" and the "final state of the damage at the scene," as well as the inherent weak causal constraints among the five elements.

[0087] (1) Generation of synthetic datasets: Using professional computational fluid dynamics and explosion dynamics numerical simulation software (such as FLACS or PyroSim), large-scale and diverse "initial-final state" paired synthetic datasets are generated by randomly combining different initial state five elements. Historical case or experimental data are collected and organized, and necessary data are supplemented to form synthetic datasets from real cases. The two types of data are merged to form a training dataset, and each data point contains a set of initial state parameters and the resulting final damage results (such as 3D point clouds, pressure cloud maps, etc.).

[0088] (2) Model training: On the generated large-scale synthetic dataset, the entire S-CVAE model is trained end-to-end using the total loss function L, which includes the physical information loss function as defined above, until the model converges.

[0089] The on-site reasoning phase involves deploying the pre-trained model to the real accident scene to perform efficient and accurate causal reasoning.

[0090] (1) Evidence softening: All kinds of evidence collected at the scene, including definitive evidence (such as photos of wreckage), ambiguous evidence (such as eyewitness descriptions of "hearing a muffled sound"), and historical sensor data, are uniformly input into the corresponding multimodal encoder and transformed into a unified input tensor with probability weights that the model can understand.

[0091] (2) Variational sampling inference: The processed evidence tensor is input into the trained model. In the structured Bayesian latent space layer, N rounds (e.g., 1000 times) of Monte Carlo sampling are performed. Each sampling will yield a set of possible combinations of five elements. After N samplings, the posterior probability distribution of each element can be obtained, forming a joint probability distribution heatmap.

[0092] (3) Model fine-tuning: In order to further improve the inference accuracy in specific accident scenarios, the pre-trained model can be fine-tuned by a small number of cycles of transfer learning using unique, high-confidence on-site measured data. (4) Result aggregation and output: Mean result: Statistically aggregate the N sampling results (such as taking the mean or mode) to give the most likely combination of the five elements of the initial state. For example: "Liquefied petroleum gas, instantaneous hose detachment, leak location in the kitchen cabinet, and excitation by an electric spark generated by the start of the refrigerator compressor". (5) Variance (Confidence) Assessment: Analyze the variance of the probability distribution of each element's output. If the distribution of a certain element (e.g., "Activation Method") is very flat (i.e., the variance is large), the system will identify it as "high cognitive uncertainty" and automatically generate decision support prompts. For example: "Due to the lack of evidence of wire melting marks in key locations, the confidence of the inference of activation method is low (32%). It is recommended to supplement the investigation of the compressor area at the back of the refrigerator." (6) Cognitive uncertainty-driven supplementary exploration closed loop: The system not only outputs the "most likely scenario" (mean) More importantly, it outputs the "degree of doubt about the scenario" (variance). In practical applications, if the evidence (such as photographs or point clouds) is contradictory or key evidence is missing, the S-CVAE sampling process may exhibit a multi-peaked or flat distribution, leading to variance. A surge. At this point, decision-makers no longer force a single, potentially erroneous answer, but instead use this "uncertainty" to reverse-engineer the weak links in the chain of evidence. For example, when the source of the leak is extremely diffuse in spatial distribution, the system generates supplementary investigation instructions with coordinate suggestions, guiding investigators to look for more subtle physical and chemical traces near specific three-dimensional coordinates, thus forming an intelligent iterative process of "reasoning-supplementary investigation-re-reasoning".

[0093] The following is an application example of this system.

[0094] To more intuitively demonstrate the practical application effect of the present invention, a typical apartment kitchen gas explosion case will be used as an example for explanation below.

[0095] At the scene: An explosion occurred in an apartment building, causing severe damage to the kitchen area and leaving the scene in a mess. The gas meter was blown off, and all the windows were shattered. Preliminary investigation revealed that the apartment used both piped natural gas and bottled liquefied petroleum gas, raising several possible causes of the explosion.

[0096] System reasoning process: 1) Evidence input: Investigators input evidence such as the three-dimensional laser scan data of the scene, close-up photos of obvious tear marks on the metal hose under the stove, and photos of the compressor at the back of the refrigerator into the system.

[0097] 2) Feature Extraction and Mapping: The system's multimodal sensing coding layer begins operation. The spatial damage branch, through analysis of the 3D point cloud, identifies that the energy center of the explosion is located in the lower part of the kitchen, rather than near the ceiling. The microscopic trace branch, on the other hand, identifies typical ductile fracture features from the torn photographs of the metal hose.

[0098] 3) Latent space sampling and physical verification: In the latent space, due to the input of ductile fracture characteristics, The sampling results (for the leakage method) strongly suggest a "high-pressure instantaneous injection," ruling out the possibility of a "slow leakage." Meanwhile, the energy conservation constraint term of the PINN physical consistency constraint layer ( It begins to work. Its verification reveals that only when... When the gas is liquefied petroleum gas, which has a higher energy density, the chemical energy it releases in a short period of time can match the energy logic of the high-intensity damage observed on site (wall displacement, gas meter blown off), thus ruling out the possibility of pipeline natural gas with a very high degree of confidence.

[0099] Final conclusion output: The system finally outputs a comprehensive reasoning report with a logical chain: "Comprehensive reasoning, the most likely initial state of the accident is: the gas type is liquefied petroleum gas, the leakage method is the instantaneous detachment of the gas cylinder hose under high pressure, the leakage location is near the kitchen cabinet, the ignition method is caused by the electric spark of the refrigerator compressor (confidence level 85%), and the explosion location is in the cabinet space near the leakage point." The above system can be implemented in a computer device in hardware or software, so that the computer device can implement the methods in the embodiments of this application. The specific methods can be referred to the description of the foregoing embodiments, and will not be repeated here.

[0100] In this application embodiment, a computer-readable storage medium is also provided, which stores a computer program that implements the method in this application embodiment when the computer device executes the computer program.

[0101] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the above-mentioned system (device) and module unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0102] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system device embodiments described above are merely illustrative. For instance, the division of the above-described module units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0103] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for joint reasoning of initial state elements in a gas explosion, characterized in that, include: Acquire multimodal evidence related to the gas explosion scene and input it into a pre-trained deep learning network, which includes a multimodal perception coding layer, a structured Bayesian latent space layer, and an uncertainty reasoning decision layer. Based on the multimodal sensing coding layer, feature extraction and fusion are performed on multimodal evidence to obtain the output coded feature vector; Based on the structured Bayesian latent space layer, the encoded feature vector output by the multimodal sensing coding layer is mapped to the probability distribution space of multiple predefined initial state elements of gas explosion. Based on the uncertainty reasoning decision layer, sampling and calculation are performed on the probability distribution space of multiple initial state elements output by the structured Bayesian latent space layer, and the combined result of multiple initial state elements arranged according to confidence level is output.

2. The method according to claim 1, characterized in that, The multimodal evidence includes: spatial damage evidence at the gas explosion site, which includes a three-dimensional scan point cloud of the site or its textual description; The multimodal evidence also includes: microscopic images of combustion traces at the scene of a gas explosion or a textual description thereof; Alternatively, the multimodal evidence may also include: time-series data of gas concentration detected by sensors.

3. The method according to claim 2, characterized in that, The structured Bayesian latent space layer encodes feature vectors, performs feature decoupling and mapping, and performs heterogeneous distribution modeling based on the different attributes of multiple predefined initial state elements of gas explosions, sampling different mathematical representation forms for initial state elements with different attributes.

4. The method according to claim 3, characterized in that, The attributes of initial state elements include discrete attributes and continuous attributes; Initial state elements of discrete attributes include: gas type ( ) and stimulation method ( ); Initial state elements of continuous attributes include: leakage location ( ), Detonation location ( ) and leakage methods ( ).

5. The method according to claim 4, characterized in that, The deep learning network also includes a PINN physical consistency constraint layer, which trains the network parameters based on a pre-designed loss function that incorporates physical laws during the pre-training phase of the deep learning network.

6. The method according to claim 5, characterized in that, The loss function includes a gas diffusion consistency term and an energy conservation constraint term.

7. The method according to claim 5, characterized in that, During the pre-training phase, different differential reparameter sampling methods are used for discrete and continuous attributes, where: For discrete properties, the Gumbel-Softmax operator is used to transform the non-differentiable discrete sampling process into a continuously differentiable function mapping, thereby realizing the backpropagation of the physical residual gradient. For continuous properties, an anisotropic Gaussian reparameterization method is adopted. By introducing auxiliary noise, the sampling process is separated from the distribution parameters, so as to achieve complete differentiability of the sampling points with respect to the mean and variance.

8. The method according to claim 7, characterized in that, The method achieves adaptive evolution of inference results through a gradient collaborative optimization mechanism, including: Dynamic annealing and shrinkage coupling: During training, the dynamic decay of the annealing temperature parameter is used to drive the discrete branches to tend towards deterministic classification. At the same time, combined with the dynamic shrinkage of the anisotropic standard deviation, the probability weights that do not conform to the physical logic region are automatically suppressed. State evolution: Through the coupling optimization of the parameters, the reasoning conclusion is driven to evolve from the fuzzy probability space to the high-confidence physical initial state conclusion.

9. The method according to claim 8, characterized in that, The method further includes: a supplementary exploration decision guidance mechanism based on cognitive uncertainty, which includes uncertainty measurement, decision triggering logic, and intelligent guidance generation. Uncertainty Measurement: The cognitive uncertainty of the reasoning conclusion is quantified by using the variance of each initial state element output from the structured Bayesian latent space layer. Decision triggering logic: A preset uncertainty threshold or distribution flatness evaluation criterion is used. When the prediction variance of a certain initial state element exceeds the threshold, the initial state element is identified as the target. Intelligent guidance generation: Based on the physical semantics of the target, it automatically associates it with the corresponding physical space area or physical evidence type, and generates targeted supplementary inspection direction guidance to assist investigators in conducting accurate supplementary inspections.

10. The method according to claim 9, characterized in that, The generation logic for supplementary survey direction guidance includes: If the variance of the ignition method exceeds the threshold, based on the coordinate distribution of the detonation location, the system will automatically extract potential sources of electrical wire melting marks, switch sparks, or static electricity risks in the surrounding area as supplementary exploration prompts. If the variance of the leak location exceeds the threshold, the investigators will be prompted to focus on reviewing the concealed pipelines or locations where the seals have failed, based on the areas where the gas diffusion consistency constraint has failed.

Images