An intelligent image recognition method, device, equipment and medium for fire fighting fire warning

Abstract

The application relates to the technical field of event recognition of monitoring video. An intelligent image recognition method, device, equipment and medium for fire disaster early warning are provided, wherein the method comprises the following steps: constructing a physical agent model library based on the physical generation mechanism of non-fire disturbance sources; performing multi-scale time sequence feature extraction on a monitoring video stream to obtain dynamic evolution features; performing physical constraint matching on the dynamic evolution features based on environmental parameters collected by an environmental sensor and the physical agent model library to generate a physical constraint matching result; decoupling features of an input image through a causal decoupling network to separate fire causal features and scene interference features; and performing physical regularization joint training based on the physical constraint matching result, the fire causal features and the scene interference features to construct a fire recognition model, so that the technical effects of improving industrial scene recognition accuracy, reducing environmental interference false alarm rate and enhancing complex environment adaptive capacity are achieved.

Description

Technical Field

[0001] This invention relates to the field of event recognition technology in surveillance video, and in particular to an intelligent image recognition method, device, equipment, and medium for fire early warning. Background Technology

[0002] With the rapid development of artificial intelligence and the Internet of Things (IoT) technologies, intelligent fire protection systems are playing an increasingly crucial role in industrial safety, urban security, and other fields. Early fire identification is a core element in preventing major safety accidents and also a research challenge in the field of computer vision. Effective fire early warning systems require high-precision, low-false-alarm real-time identification in complex industrial environments.

[0003] Traditional technologies rely on deep learning for video fire detection and conventional visual detection algorithms for fire analysis. However, these technologies lack the ability to physically model interference sources and are susceptible to misjudgments due to various physical factors in industrial environments. Mainstream detection models directly extract visual features from the original image for identification, but they cannot remove redundant interference features or effectively decouple the core fire features. Feature confusion significantly reduces the model's discrimination performance. Furthermore, traditional detection schemes operate on a purely data-driven model, lacking the collaborative optimization design of physical constraints and intelligent algorithms. This makes it difficult to dynamically correct the recognition logic according to complex environmental changes, resulting in significant deficiencies in scene adaptation and environmental adaptability. Summary of the Invention

[0004] Based on this, it is necessary to provide an intelligent image recognition method, device, equipment, and medium for fire early warning, in order to address the above-mentioned technical problems and achieve the technical effects of improving the recognition accuracy in industrial scenes, reducing the false alarm rate due to environmental interference, and enhancing the adaptability to complex environments.

[0005] In a first aspect, this application provides an intelligent image recognition method for fire early warning, the method comprising:

[0006] A physical proxy model library is constructed based on the physical generation mechanism of non-fire interference sources;

[0007] Multi-scale temporal feature extraction is performed on the surveillance video stream to obtain dynamic evolution features;

[0008] Based on environmental parameters collected by environmental sensors and a physical proxy model library, physical constraint matching is performed on dynamic evolution characteristics to generate physical constraint matching results.

[0009] The input image is decoupled by a causal decoupling network to separate fire causal features and scene interference features. The scene interference features are used to block the visual correlation of non-fire physical interference sources.

[0010] A fire identification model is constructed by jointly training physical regularization based on physical constraint matching results, fire causal features, and scene interference features.

[0011] In one embodiment, based on environmental parameters collected by environmental sensors and a physical surrogate model library, physical constraint matching is performed on dynamic evolution characteristics to generate physical constraint matching results, including:

[0012] The multidimensional environmental parameters collected by environmental sensors are analyzed to obtain an environmental parameter vector;

[0013] By processing environmental parameter vectors through a physical proxy model library, the evolutionary behavior of interference sources is predicted, and expected dynamic patterns are generated.

[0014] Quantitatively compare the spatial distribution and temporal evolution consistency of dynamic evolution characteristics and expected dynamic patterns to generate similarity metrics;

[0015] Based on the environmental adaptive threshold similarity metric, physical constraint matching results are generated.

[0016] In one embodiment, environmental parameter vectors are processed through a physical proxy model library to predict the evolutionary behavior of interference sources and generate expected dynamic patterns, including:

[0017] Multiphysics coupling analysis is performed on the environmental parameter vector to obtain the initial evolution state;

[0018] The initial evolution state is processed based on the physical proxy model library, the morphological evolution of the interference source is simulated, and a dynamic evolution sequence is generated.

[0019] Extract the spatiotemporal invariant features of the dynamic evolution sequence to generate the expected dynamic pattern.

[0020] In one embodiment, a physical proxy model library is constructed based on a physical generation mechanism for non-fire interference sources, including:

[0021] Define a set of non-fire interference sources, including water vapor, welding sparks, and industrial dust.

[0022] Multiphysics simulations were performed on each category in the set of non-fire interference source categories to obtain the corresponding dynamic morphological sequence data;

[0023] Based on the environmental parameter vectors collected by environmental monitoring equipment, a differentiable surrogate model is performed on the dynamic morphological sequence data to obtain a differentiable physical surrogate model;

[0024] Differentiable physical proxy models are mapped to corresponding interference source categories to establish a physical proxy model library.

[0025] In one embodiment, based on environmental parameter vectors collected by environmental monitoring equipment, a differentiable surrogate model is performed on dynamic morphological sequence data to obtain a differentiable physical surrogate model, including:

[0026] Physical sensing encoding is performed on the environmental parameter vector to obtain a joint environmental-physical representation;

[0027] By processing the environment-physical joint characterization and dynamic morphological sequence data through a differentiable solver, surrogate model parameters are generated.

[0028] Based on variational inference, optimize the surrogate model parameters to generate a differentiable physical surrogate model.

[0029] In one embodiment, a causal decoupling network is used to decouple features from the input image, separating fire causal features and scene interference features, including:

[0030] Thermodynamic causal analysis is performed on the input image to extract heat release rate features and carbon particle diffusion features, and initial fire features are generated.

[0031] Environmental interference analysis is performed on the input image to extract optical reflection features and non-fire motion features, and initial interference features are generated.

[0032] By eliminating the statistical correlation between initial fire features and initial disturbance features through causal intervention units, decoupled fire features are obtained.

[0033] By blocking the fire semantic association of the initial interference features through the adversarial discriminator, purification interference features are generated;

[0034] It integrates and decouples fire characteristics and purification interference characteristics to output fire causal characteristics and scene interference characteristics.

[0035] In one embodiment, a fire identification model is constructed by performing joint training with physical regularization based on physical constraint matching results, fire causal features, and scene interference features, including:

[0036] By integrating fire causal characteristics, physical constraint matching results, and scene interference characteristics, a physical guidance feature vector is generated.

[0037] The optimization direction of the physically guided feature vector is adjusted by the gradient reweighting mechanism to obtain the regularized gradient signal;

[0038] The robustness of the regularized gradient signal is verified by adversarial perturbation testing, and the cleaned gradient parameters are generated.

[0039] A fire identification model is constructed by iteratively optimizing the neural network weights based on the purification gradient parameters.

[0040] Secondly, this application also provides an intelligent image recognition device for fire early warning, the device comprising:

[0041] The interference source modeling module is used to build a physical proxy model library based on the physical generation mechanism of non-fire interference sources;

[0042] The spatiotemporal feature extraction module is used to extract multi-scale temporal features from surveillance video streams to obtain dynamic evolution features;

[0043] The physical matching module is used to perform physical constraint matching on dynamic evolution characteristics based on environmental parameters collected by environmental sensors and a physical proxy model library, and generate physical constraint matching results.

[0044] The feature decoupling module is used to decouple the features of the input image through a causal decoupling network, separating the fire causal features and scene interference features, where the scene interference features are used to block the visual correlation of non-fire physical interference sources.

[0045] The joint training module is used to perform physical regularization joint training based on physical constraint matching results, fire causal features, and scene interference features to build a fire identification model.

[0046] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0047] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0048] This application provides an intelligent image recognition method, device, equipment, and medium for fire early warning. By constructing a physical proxy model library based on the physical generation mechanism of non-fire interference sources, and combining environmental parameters collected by environmental sensors, it performs physical constraint matching on the multi-scale temporal dynamic evolution features extracted from the monitoring video stream. This effectively correlates environmental dynamic changes with the evolution patterns of visual features, reducing the misleading influence of environmental interference factors on feature judgment. Simultaneously, it utilizes a causal decoupling network to separate fire causal features from scene interference features and blocks the visual correlation of non-fire physical interference sources, helping the model filter redundant interference information and more accurately pinpoint the core fire features.

[0049] This technical approach, through a deep integration of physical mechanisms and visual recognition, enables the model to avoid judgment biases caused by interference sources in industrial scenarios, thereby reducing the false alarm rate caused by environmental interference and improving the accuracy of fire feature identification. Furthermore, the synergistic effect of real-time environmental parameters and the physical proxy model helps the model better adapt to the dynamic changes in complex environments, further enhancing its adaptive capabilities in complex scenarios. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of an intelligent image recognition method for fire early warning, as shown in one embodiment of the present invention.

[0052] Figure 2 This is a flowchart illustrating how, in one embodiment of the present invention, environmental parameter vectors are processed through a physical proxy model library to predict the evolutionary behavior of interference sources and generate expected dynamic patterns.

[0053] Figure 3 This is a structural diagram of an intelligent image recognition device for fire early warning, according to one embodiment of the present invention. Detailed Implementation

[0054] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0055] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, an intelligent image recognition method, device, equipment, and medium for fire early warning are provided, applicable to, but not limited to, industrial fire safety monitoring scenarios such as industrial workshops, mining areas, and warehouses.

[0056] In illustrative purposes, the intelligent image recognition method, device, equipment and medium for fire early warning provided in the embodiments of this application can also be applied to conventional fire safety monitoring application scenarios such as urban parks, commercial buildings, transportation hubs, and outdoor stations. This is only an example and does not limit the specific application scenarios.

[0057] like Figure 1 As shown, this application provides an intelligent image recognition method for fire early warning, the method comprising:

[0058] S101: Construct a physical proxy model library based on the physical generation mechanism of non-fire interference sources.

[0059] For example, the fire early warning identification terminal defines the scope of typical non-fire interference sources in industrial scenarios, deeply analyzes the physical generation mechanism of various non-fire interference sources, sorts out their formation principles and morphological evolution laws, and then collects data characterizing the physical attributes and dynamic features of interference sources through physical simulation or real-world data collection. The collected data is then filtered and organized, redundant and invalid content is removed, and key data required for modeling is retained.

[0060] Based on the physical generation mechanism of non-fire interference sources, the fire early warning and identification terminal formulates an appropriate modeling logic, builds initial models of various interference sources based on the standardized key data, verifies the physical adaptability of the initial models, corrects model deviations and optimizes performance, integrates all optimized model resources according to the interference source category specifications, and constructs a physical proxy model library based on the physical generation mechanism of non-fire interference sources.

[0061] S102: Perform multi-scale temporal feature extraction on the surveillance video stream to obtain dynamic evolution features.

[0062] For example, the fire early warning identification terminal receives the monitoring video stream transmitted by the monitoring equipment, performs frame-level preprocessing on the monitoring video stream to remove noise interference and invalid frame data, and completes image calibration to ensure the spatial consistency of the frame data. The fire early warning identification terminal performs feature extraction operations on the preprocessed continuous video frames according to the preset scale levels, and simultaneously captures spatial detail features and global correlation features at each scale to obtain a multi-scale spatial feature set.

[0063] Based on the temporal sequence relationships of continuous video frames, the fire early warning and identification terminal performs temporal correlation analysis on features in a multi-scale spatial feature set, tracks the changing trajectories and evolution trends of various features at different time points, and establishes a feature temporal correlation mapping. Subsequently, the fire early warning and identification terminal fuses and regularizes the multi-scale temporal correlation features, eliminates redundant feature information, and extracts core features that reflect the dynamic changes of the target, thus obtaining dynamic evolution features.

[0064] S103: Based on environmental parameters collected by environmental sensors and a physical proxy model library, perform physical constraint matching on dynamic evolution characteristics and generate physical constraint matching results.

[0065] For example, by integrating environmental parameters uploaded by environmental sensors, the fire early warning and identification terminal standardizes and organizes the environmental parameters, removes abnormal data, synchronously retrieves the built-in physical proxy model library, and combines the organized environmental parameters to deduce the evolution of interference sources and output a standard evolution paradigm adapted to the on-site environment.

[0066] Based on the extracted dynamic evolution features, the fire early warning identification terminal performs multi-dimensional content comparison, checks the consistency of the temporal changes and spatial performance of the features against the evolutionary paradigm formed by the deduction, completes constraint verification based on the standard judgment logic, and completes comprehensive judgment by integrating various comparison information to generate physical constraint matching results.

[0067] S104: The input image is decoupled by a causal decoupling network to separate fire causal features and scene interference features. The scene interference features are used to block the visual correlation of non-fire physical interference sources.

[0068] For example, the fire early warning recognition terminal receives the input image to be processed, performs standardization preprocessing on the input image, completes spatial calibration and noise suppression, and provides a qualified data foundation for feature decoupling. Then, it calls the causal decoupling network to perform a full-domain scan of the preprocessed input image and extracts initial comprehensive features covering various types of visual information.

[0069] Based on the core physical mechanism and causal logic of fire, the fire early warning and identification terminal uses a dedicated branch of the causal decoupling network to screen fire-related feature components from the initial comprehensive features and preliminarily separate candidate fire causal features; at the same time, for the visual attributes of non-fire interference sources, it extracts interference visual information through another feature branch to obtain the initial scene interference features.

[0070] The fire early warning and identification terminal uses the causal intervention unit of the causal decoupling network to eliminate the statistical correlation between the two types of features and purify the fire causal features. It strengthens the exclusive representation of the initial scene interference features through the adversarial constraint mechanism, blocks their association with fire semantics, and forms scene interference features that can block the visual correlation of non-fire physical interference sources.

[0071] S105: A fire identification model is constructed by jointly training physical regularization based on physical constraint matching results, fire causal features, and scene interference features.

[0072] For example, the fire early warning and identification terminal integrates physical constraint matching results, fire causal features, and scene interference features, and performs feature-level synergistic fusion of the three types of information to generate a physical guidance feature vector that takes into account both physical constraint requirements and visual representation integrity. Then, a physical regularization mechanism is introduced, and the optimization direction of the physical guidance feature vector is adjusted through a gradient reweighting strategy, so that the model training process conforms to the actual physical evolution law and a regularized gradient signal is obtained.

[0073] The fire early warning and identification terminal performs robustness verification on the regularized gradient signal. By simulating various environmental disturbance scenarios, it verifies the stability of the signal, removes abnormal interference components, obtains purification gradient parameters, and then uses the purification gradient parameters as the core optimization basis to continuously iterate and adjust the weight parameters of the neural network, thereby continuously improving the model's accuracy in identifying fire features and its anti-interference ability, and finally constructing a fire identification model.

[0074] One embodiment of this application provides an intelligent image recognition method for fire early warning. By constructing a physical proxy model library based on the physical generation mechanism of non-fire interference sources, and combining environmental parameters collected by environmental sensors, the method performs physical constraint matching on the multi-scale temporal dynamic evolution features extracted from the monitoring video stream. This effectively correlates dynamic environmental changes with the evolution patterns of visual features, reducing the misleading influence of environmental interference factors on feature judgment. Simultaneously, a causal decoupling network is used to separate fire causal features from scene interference features and block the visual correlation of non-fire physical interference sources, helping the model filter redundant interference information and more accurately pinpoint the core fire features.

[0075] This technical approach, through a deep integration of physical mechanisms and visual recognition, enables the model to avoid judgment biases caused by interference sources in industrial scenarios, thereby reducing the false alarm rate caused by environmental interference and improving the accuracy of fire feature identification. Furthermore, the synergistic effect of real-time environmental parameters and the physical proxy model helps the model better adapt to the dynamic changes in complex environments, further enhancing its adaptive capabilities in complex scenarios.

[0076] In one embodiment, based on environmental parameters collected by environmental sensors and a physical surrogate model library, physical constraint matching is performed on dynamic evolution characteristics to generate physical constraint matching results, including:

[0077] (1) Analyze the multidimensional environmental parameters collected by the environmental sensors to obtain the environmental parameter vector.

[0078] For example, after receiving multi-dimensional environmental parameters collected by environmental sensors, the fire early warning identification terminal checks the completeness and validity of each type of environmental parameter. It marks any missing or abnormal data, suppresses abnormal fluctuations in the data through smoothing, and arranges the verified and normalized multi-dimensional environmental parameters in an orderly manner according to preset parameter classification standards and dimensional priority. This integrates the scattered multi-dimensional environmental parameters into a structured and formatted environmental parameter vector. The structured construction formula for the environmental parameter vector is as follows:

[0079]

[0080] In the formula, Represents an environmental parameter vector; This represents the first dimension of environmental parameters; This represents the second dimension of environmental parameters; Indicates the first Dimensional environment parameters; This represents the matrix transpose symbol.

[0081] Among them, multidimensional environmental parameters refer to various physical quantity parameters that reflect the current environmental state collected by environmental sensors, including but not limited to temperature, humidity, air pressure, airflow speed, etc.; the environmental parameter vector is a one-dimensional numerical vector formed by arranging multidimensional environmental parameters in a fixed dimensional order, which is used to provide standardized input for subsequent model processing.

[0082] (2) By processing environmental parameter vectors through the physical proxy model library, the evolution behavior of interference sources is predicted and the expected dynamic pattern is generated.

[0083] For example, after retrieving the constructed physical proxy model library, the fire early warning identification terminal selects non-fire interference source models that match the current environmental conditions and interference source types from the physical proxy model library based on the type classification, numerical range, and environmental scene adaptability of multi-dimensional environmental parameters. The standardized environmental parameter vector is input into the selected non-fire interference source model. According to the preset non-fire interference source physical generation mechanism and evolution rules, the model performs phased deduction and prediction of the evolution behavior of non-fire interference sources under the current environmental conditions, such as initial form, development rate, spatial diffusion range, and temporal change nodes, and generates an expected dynamic pattern that can reflect the dynamic change law of non-fire interference sources.

[0084] Among them, the physical proxy model library is a collection of physical generation mechanism models of various non-fire interference sources, used to simulate the evolution process of non-fire interference sources under different environmental conditions; the expected dynamic model refers to the collection of spatial distribution and temporal change characteristics of non-fire interference sources that may be exhibited in the current environment, which are predicted by the physical proxy model library.

[0085] (3) Quantitatively compare the spatial distribution and temporal evolution consistency of dynamic evolution characteristics and expected dynamic patterns to generate similarity metrics.

[0086] For example, by using feature parsing technology and time series analysis methods, the fire early warning identification terminal extracts key attributes representing spatial distribution from dynamic evolution features, including spatial location coordinates, coverage boundaries, morphological and structural features, etc., and at the same time extracts key spatial distribution attributes corresponding to the expected dynamic pattern; by using time series analysis methods, the change trajectory, stage transition features and trend of the two types of features over time are extracted to obtain a complete time series evolution trajectory.

[0087] The fire early warning identification terminal employs a dual-dimensional quantification comparison method to calculate the consistency of two types of features. A spatial feature matching algorithm calculates the percentage of overlapping areas, morphological contour fit, and key point correspondence between the two types of features, yielding spatial distribution similarity. A dynamic time warping algorithm calculates the distance metric between the temporal trajectories of the two types of features, and the distance metric result is normalized and converted into temporal evolution similarity. A weighted fusion strategy integrates the spatial distribution similarity and temporal evolution similarity to obtain a comprehensive similarity metric. The specific calculation formula is as follows:

[0088]

[0089] In the formula, This represents the similarity metric, with a value range of [0,1]. The closer the value is to 1, the higher the consistency between the dynamic evolution characteristics and the expected dynamic pattern. The weighting coefficient represents the spatial distribution similarity, with a value range of [0,1], and is used to balance the importance of the spatial dimension and the temporal dimension; It represents the spatial distribution similarity, with a value range of [0,1]. It is obtained by calculating the proportion of overlapping areas and the morphological matching degree of the spatial distributions of the two types of features. Represents temporal evolution similarity, with a value range of [0,1]. It is obtained by calculating the dynamic time warping distance of the two types of feature temporal trajectories.

[0090] Among them, spatial distribution consistency refers to the degree of fit between dynamic evolution characteristics and expected dynamic patterns in terms of spatial attributes such as spatial location, coverage, and morphological structure; temporal evolution consistency refers to the degree of fit between the two in terms of temporal attributes such as rhythm, trend, and stage characteristics that change over time; similarity metric is a comprehensive index that quantifies the two-dimensional consistency of the two types of features and is used to intuitively reflect the degree of matching between the two.

[0091] (4) Generate physical constraint matching results based on the environmental adaptive threshold judgment similarity metric.

[0092] For example, combining the environmental factor combination state of the current environmental parameter vector and the characteristics of the adaptation model, the fire early warning identification terminal refers to the evolution sensitivity, environmental adaptability and other characteristic parameters of the adaptation non-fire interference source model in the physical proxy model library, and determines the environmental adaptive threshold adapted to the current scenario through a threshold dynamic adjustment algorithm. This threshold can be dynamically adjusted in real time according to changes in environmental conditions to meet the matching judgment requirements under different scenarios.

[0093] The fire early warning identification terminal strictly follows the preset decision logic, comparing the calculated similarity metric with an environmental adaptive threshold. If the similarity metric is greater than or equal to the environmental adaptive threshold, the dynamic evolution feature is determined to match the expected dynamic pattern; if the similarity metric is less than the environmental adaptive threshold, the two are determined not to match. Based on this clear decision result, a physical constraint matching result containing a matching status identifier and quantitative information on the degree of matching is generated. The decision formula for the physical constraint matching result is as follows:

[0094]

[0095] In the formula, This indicates the result of physical constraint matching; This represents the similarity metric. This indicates the environmental adaptive threshold.

[0096] Among them, the environment adaptive threshold is a decision benchmark value that is dynamically adjusted according to the current environmental state and the characteristics of the adaptive model. It is used to distinguish whether the dynamic evolution features match the expected dynamic pattern. The physical constraint matching result is the result information that clearly represents the matching status and degree of the dynamic evolution features and the expected dynamic pattern, and provides a physical constraint basis for subsequent model training.

[0097] like Figure 2 As shown, by processing environmental parameter vectors through a physical proxy model library, the evolutionary behavior of interference sources is predicted, and expected dynamic patterns are generated, including:

[0098] S201: Perform multiphysics coupling analysis on the environmental parameter vector to obtain the initial evolution state.

[0099] For example, after receiving the environmental parameter vector, the fire early warning identification terminal analyzes the various physical field parameters contained in the environmental parameter vector, clarifies the types and numerical distributions of multiple physical fields such as temperature field, humidity field, and airflow field, establishes a multi-physics coupling correlation model, and quantifies the coupling strength and influence weight between fields by analyzing the energy transfer, material exchange, and interaction laws between the various physical fields. Based on the coupling dynamics equation, it calculates the state parameters of the interference source at the initial moment, including the initial shape, distribution range, energy level, etc., to obtain the initial evolution state. The core calculation formula of the multi-physics coupling analysis is as follows:

[0100]

[0101] In the formula, This represents the total action of multiphysics coupling; Indicates the number of physical fields involved in the coupling; Indicates the first Independent action weighting coefficients for quasi-physical fields; Indicates the first Field strength parameters of quasi-physical fields; Indicates the first Physics Fields and the First Cross-coupling weighting coefficients for physics-like fields.

[0102] Among them, multiphysics coupling analysis refers to the comprehensive analysis of multiple interconnected physical fields in the environment, taking into account the common impact of inter-field interactions on the evolution of the interference source; the initial evolution state refers to the set of core state parameters such as morphology, distribution, and energy of the interference source under the current environmental conditions after the multiphysics coupling effect.

[0103] S202: Based on the physical proxy model library, the initial evolution state is processed, the morphological evolution of the interference source is simulated, and a dynamic evolution sequence is generated.

[0104] For example, after retrieving the adapted non-fire interference source model from the physical proxy model library, the fire early warning and identification terminal inputs the obtained initial evolution state into the model, sets the time step and evolution constraints for the model simulation, and the constraints include environmental boundary restrictions, the range of changes in physical field parameters, etc. Based on the preset physical evolution law of non-fire interference sources, the model iteratively calculates and deduces the morphological changes, spatial location migration, range expansion or contraction, etc. of the interference source at each time step, and records the complete state parameters of the interference source at each time node in sequence, forming a continuous and orderly dynamic evolution sequence.

[0105] Among them, the dynamic evolution sequence refers to a continuous set of state parameters of the interference source at different time points, arranged in chronological order, which can fully reflect the evolution process of the interference source over time; the physical proxy model library is a collection of physical generation mechanism models of various non-fire interference sources, used to simulate the evolution process of non-fire interference sources under different environmental conditions.

[0106] S203: Extract the spatiotemporal invariant features of the dynamic evolution sequence to generate the expected dynamic pattern.

[0107] For example, for the generated dynamic evolution sequence, the fire early warning identification terminal extracts features of the interference source state corresponding to each time node, obtaining multi-dimensional basic features such as morphological contour features, spatial distribution features, and density distribution features. Through spatiotemporal correlation analysis, features that remain stable across all time nodes, do not change over time, or exhibit strong regularity in their change trends are selected; these are spatiotemporally invariant features, which reflect the core essential laws of interference source evolution. Subsequently, the selected spatiotemporally invariant features are integrated and standardized to obtain the expected dynamic pattern that can characterize the evolution laws of non-fire interference sources. The extraction and selection formulas for the spatiotemporally invariant features are as follows:

[0108]

[0109] In the formula, Indicates the spacetime invariant characteristic; This represents the total number of time points in the dynamic evolution sequence; Indicates the first A certain type of basic characteristic value corresponding to the state of the interference source at each time point; This represents the mean of this type of basic feature across all time points.

[0110] Among them, spatiotemporally invariant characteristics refer to the characteristics in the dynamic evolution sequence that do not change significantly over time or whose change patterns are stable and can reflect the core evolutionary essence of the interference source; expected dynamic patterns refer to the set of spatial distribution and temporal change characteristics that non-fire interference sources may exhibit in the current environment, as predicted by the physical proxy model library.

[0111] In one embodiment, a physical proxy model library is constructed based on a physical generation mechanism for non-fire interference sources, including:

[0112] (1) Define a set of non-fire interference sources, including water vapor, welding sparks and industrial dust.

[0113] For example, combining the actual interference characteristics of industrial fire monitoring scenarios, the fire early warning and identification terminal investigates various interference targets in industrial operation scenarios that are prone to causing misjudgment of fire identification, establishes standardized criteria for classifying and judging interference sources, strictly screens non-fire targets that appear frequently and have visual interference characteristics, delineates the scope of core interference objects, and successively includes water vapor interference substances commonly found in industrial scenarios, welding sparks generated by welding operations, and industrial dust suspended and diffused in the production process into the control scope, clarifies the definition standards and characteristic boundaries of each type of interference source, and completes the standardized and complete definition of the set of non-fire interference source categories.

[0114] Among them, the non-fire interference source category set refers to the collection of industrial interference objects classified according to the interference cause and visual appearance, which is used to uniformly constrain the target range of subsequent simulation modeling.

[0115] (2) Perform multiphysics simulation on each category in the set of non-fire interference sources to obtain the corresponding dynamic morphological sequence data.

[0116] For example, for each independent interference source object in the set of non-fire interference source categories, the fire early warning identification terminal individually matches and adapts the physical boundary conditions and environmental action conditions, completely replicates the real physical environment of the natural generation, diffusion and evolution of various non-fire interference sources, integrates multiple physical fields such as thermal field, flow field, and particle motion field for joint and collaborative simulation calculation, continuously captures the contour changes, spatial diffusion, state transition and other full-dimensional evolution information of the interference source in the time-series process, retains the simulation sampling data in an orderly manner according to continuous time nodes, filters, organizes and sorts the original simulation data in a time sequence, removes invalid segments and abnormal deviation data, and outputs dynamic morphological sequence data matching the characteristics of the corresponding interference source for each category.

[0117] Among them, multiphysics simulation refers to the numerical simulation process that reproduces the law of action of the real environment based on physical boundary conditions, while dynamic morphological sequence data refers to the continuous sample data set that records the morphological changes of the interference source throughout the time sequence.

[0118] (3) Based on the environmental parameter vectors collected by the environmental monitoring equipment, the dynamic morphological sequence data is modeled by a differentiable proxy to obtain a differentiable physical proxy model.

[0119] For example, relying on the environmental monitoring equipment deployed throughout the site, the fire early warning identification terminal continuously collects real-time environmental perception raw data. The collected raw data is then processed by anomaly removal, noise suppression, dimensional calibration, and standardization to eliminate data discrepancies and generate environmental parameter vectors with uniform format and regular dimensions.

[0120] The fire early warning and identification terminal uses the processed environmental parameter vector as external physical constraints. Combined with dynamic morphological sequence data specific to various non-fire interference sources, it constructs a differentiable surrogate modeling framework with gradient backpropagation capabilities. This framework deeply explores the intrinsic relationship between the physical generation mechanism and morphological evolution of non-fire interference sources, continuously iteratively adjusting the model's internal operational parameters, constantly correcting fitting errors, and strengthening the model's ability to represent physical evolution characteristics. After convergent training and performance verification, stable and reliable differentiable physical surrogate models are obtained for each category. The core expression for the differentiable surrogate modeling is as follows:

[0121]

[0122] In the formula, Indicates the first Differentiable physical proxy model corresponding to non-fire interference sources This represents a differentiable surrogate modeling mapping function. This represents the dynamic morphological sequence data corresponding to a single type of interference source. This represents a vector of environmental parameters collected and constructed by environmental monitoring equipment.

[0123] Among them, the differentiable physical surrogate model refers to a lightweight alternative model that has the ability to backpropagate gradients and can fit the physical evolution law of the interference source, while the environmental parameter vector refers to a combination of multi-dimensional environmental monitoring parameters after standardization.

[0124] (4) Map the differentiable physical proxy model to the corresponding interference source category and establish a physical proxy model library.

[0125] For example, by referring to the previously defined classification rules for non-fire interference source categories, the fire early warning and identification terminal will associate and bind each trained differentiable physical proxy model with the corresponding interference source category, establish a one-to-one exclusive mapping relationship, and simultaneously improve the basic attribute annotation, working condition adaptation range, physical constraint rules and calling permission configuration of various models.

[0126] The fire early warning and identification terminal standardizes the data interaction format, operation call interface and hierarchical storage rules of all models. It classifies, collects, partitions and centrally manages different types of differentiable physical proxy models, integrates all model resources that have completed mapping and matching, realizes the systematic integration of multiple categories of physical proxy models, and builds a physical proxy model library with a complete structure and can be called quickly.

[0127] Among them, the physical surrogate model library refers to a systematic resource library that centrally collects various differentiable physical surrogate models, which can provide physical constraint deduction support for subsequent fire identification.

[0128] In one embodiment, based on environmental parameter vectors collected by environmental monitoring equipment, a differentiable surrogate model is performed on dynamic morphological sequence data to obtain a differentiable physical surrogate model, including:

[0129] (1) Physical sensing encoding is performed on the environmental parameter vector to obtain the environmental-physical joint representation.

[0130] For example, after obtaining the standardized environmental parameter vector, the fire early warning identification terminal analyzes the physical attributes and evolutionary correlation features corresponding to each dimension parameter in the environmental parameter vector one by one, builds a physical perception coding network to perform deep feature mapping on the environmental parameter vector, transforms discrete environmental monitoring values ​​into continuous physical mechanism feature vectors, integrates environmental spatial distribution features and physical constraint features to complete feature aggregation, and integrates environmental semantic information and physical evolution prior knowledge through nonlinear coding operations to obtain the environmental-physical joint representation.

[0131] Among them, physical perception coding refers to a coding method that combines prior physical knowledge to map the features of environmental parameters; environment-physical joint representation refers to a high-dimensional feature vector that integrates environmental monitoring information and physical evolution mechanisms.

[0132] (2) Generate surrogate model parameters by processing the environment-physical joint representation and dynamic morphological sequence data through a differentiable solver.

[0133] For example, after loading the environmental-physical joint representation and dynamic morphological sequence data, the fire early warning and identification terminal synchronously inputs the two types of data into the built-in differentiable solver. The environmental-physical joint representation is used as the physical constraint input condition, and the dynamic morphological sequence data is used as the evolution supervision target. The terminal constructs a differentiable mapping relationship between the morphological evolution of the interference source and the physical conditions. The terminal fits the physical evolution law through continuous numerical iterative calculation, solves the core parameters such as the weight coefficient and bias term in the mapping relationship, and outputs the surrogate model parameters that have completed the initial fitting.

[0134] Among them, the differentiable solver refers to the numerical solution tool that supports gradient backpropagation; the surrogate model parameters refer to the set of weights and biases that describe the physical evolution mapping relationship of the interference source.

[0135] (3) Optimize the parameters of the surrogate model based on variational inference to generate a differentiable physical surrogate model.

[0136] For example, based on the surrogate model parameters obtained from the initial solution, the fire early warning and identification terminal builds a variational inference optimization framework, constructs the prior probability distribution and approximate posterior probability distribution of the surrogate model parameters, achieves parameter regularization constraints by minimizing the difference between the probability distributions, iteratively optimizes the fitting accuracy and generalization ability of the surrogate model parameters, eliminates abnormal fluctuations and overfitting components in the parameters, and substitutes the optimized stable parameters into the modeling framework to generate a differentiable physical surrogate model. The variational inference parameter optimization formula is as follows:

[0137]

[0138] In the formula, This represents the optimized proxy model parameters; This represents the parameters of the proxy model to be optimized; express Divergence operation function; This represents the approximate posterior probability distribution of the parameters; This represents the true posterior probability distribution of the parameters; Representing a combined environmental and physical characterization; It represents dynamic morphological sequence data.

[0139] Variational inference refers to parameter optimization methods based on probability distribution fitting; differentiable physical surrogate models refer to lightweight models that possess gradient propagation characteristics and fit the laws of physical evolution.

[0140] In one embodiment, a causal decoupling network is used to decouple features from the input image, separating fire causal features and scene interference features, including:

[0141] (1) Perform thermodynamic causal analysis on the input image, extract heat release rate features and carbon particle diffusion features, and generate initial fire features.

[0142] For example, based on the fully acquired input image content, the fire early warning and identification terminal extracts the global pixel information and temperature mapping correlation data of the input image, and performs a comprehensive thermodynamic causal analysis by combining the inherent thermodynamic causal laws of the fire combustion reaction. Relying on energy transfer mechanisms and thermodynamic evolution rules, the fire early warning and identification terminal deeply mines the energy release variation patterns in high-temperature regions, and refines the heat release rate characteristics that can stably characterize changes in combustion reaction intensity.

[0143] The fire early warning and identification terminal simultaneously locks onto areas of carbon particle aggregation within the image, analyzes the spatial diffusion range, movement trajectory, and distribution variation patterns of the carbon particles, and fully extracts their diffusion characteristics. The terminal then performs dimensional calibration, data normalization, and feature recombination operations on the two extracted core features to unify the feature representation, systematically fuse and overlay features, and stably generate initial fire features that are structurally complete and conform to the thermodynamic laws of fire. The formula for generating the initial fire features is as follows:

[0144]

[0145] In the formula, Indicates the initial fire characteristics; The fusion weighting coefficients represent the characteristics of the heat release rate; It represents the heat release rate characteristic, used to characterize the intensity and rhythm of energy release during fire combustion; Fusion weighting coefficients representing the diffusion characteristics of carbon particles; It represents the diffusion characteristics of carbon particles and is used to characterize the spatial diffusion properties of carbon-based particles generated by fire.

[0146] Among them, thermodynamic causal analysis refers to the analysis method that mines the feature associations directly related to the cause of fire in the input image based on the thermodynamic laws of fire combustion; heat release rate features are features that reflect the intensity and trend of fire energy release; carbon particle diffusion features are features that describe the spatial propagation state of carbon-based particles generated by fire; initial fire features are a set of features that are initially extracted and contain the core thermodynamic properties of fire but may be mixed with interfering associations.

[0147] (2) Perform environmental interference analysis on the input image, extract optical reflection features and non-fire motion features, and generate initial interference features.

[0148] For example, for the entire scene area covered by the input image, the fire early warning recognition terminal specifically screens for various non-fire interference targets and conducts a systematic environmental interference analysis based on the surface optical properties of different interference objects. The fire early warning recognition terminal combines the optical principles of natural light reflection and ambient light refraction to capture the differences in light and shadow feedback of different interference media and extract optical reflection features that can distinguish the types of interference.

[0149] The fire early warning and identification terminal continuously monitors dynamic targets without fire association within the screen, tracking their displacement changes, motion rhythms, and morphological changes, and fully extracting non-fire motion features that conform to the laws of environmental disturbance. The terminal removes redundant information and invalid noise data within both types of features, unifies the feature dimensional structure, and achieves orderly fusion processing of multiple types of interference features, stably generating initial interference features that cover scene environmental disturbance information.

[0150] Among them, environmental interference analysis refers to the analysis method of mining the optical, motion and other features of non-fire factors in the input image that may affect fire identification; optical reflection features are the features that reflect the ability of interference objects to reflect light of different wavelengths; non-fire motion features are the features that describe the motion state of non-fire dynamic targets; initial interference features are the set of features that are initially extracted, containing the core attributes of environmental interference but may be related to fire features.

[0151] (3) The statistical correlation between the initial fire features and the initial disturbance features is eliminated by the causal intervention unit to obtain the decoupled fire features.

[0152] For example, after invoking the causal intervention unit built into the causal decoupling network, the fire early warning and identification terminal calculates the data distribution correlation between the initial fire features and the initial interference features, and constructs a statistical covariance matrix that quantifies the degree of correlation between the two types of features. Based on the covariance matrix, the fire early warning and identification terminal divides the independent feature space, establishes orthogonal constraint mapping rules, and reconstructs the spatial dimension of the initial fire features through a high-precision orthogonal projection transformation.

[0153] The fire early warning and identification terminal strictly severs the implicit data association and explicit statistical association between the initial fire features and the initial interference features, fully retains the core effective information in the initial fire features that fits the combustion causal mechanism, completes the feature deassociation optimization process, and obtains independent and uncorrelated decoupled fire features.

[0154] Among them, the causal intervention unit is the core part of the causal decoupling network used to eliminate statistical associations between different features; statistical correlation refers to the linear or nonlinear association between two types of features through data distribution; decoupling fire features is the feature that eliminates the association with interfering features and retains only the core causal attributes of fire.

[0155] (4) By blocking the fire semantic correlation of the initial interference features by the adversarial discriminator, purification interference features are generated.

[0156] For example, the initial interference features are input into the adversarial discriminator mounted on the causal decoupling network. The fire early warning identification terminal relies on the semantic recognition capability pre-trained by the adversarial discriminator to identify the implicit fire semantic information contained in the initial interference features.

[0157] The fire early warning identification terminal continuously outputs semantic discrimination results and forms a closed-loop feedback signal. Using non-fire semantic constraints as the optimization target, it iteratively adjusts the deep feature expression logic of the initial interference features. The fire early warning identification terminal continuously strengthens the expression weight of environmental interference-specific features, gradually weakening and completely severing the association path between the initial interference features and the fire semantic content, avoiding the identification confusion caused by interference features. After multiple rounds of adversarial iterative optimization, it generates purified interference features that are completely decoupled from fire semantic binding.

[0158] Among them, the adversarial discriminator is a model built on deep learning to distinguish the semantic categories of features; fire semantic correlation refers to the semantic information that may exist in the interference features and is easily misjudged as fire; the purification interference feature is a feature that blocks the fire semantic correlation and only retains the environmental interference-specific attributes.

[0159] (5) Integrate and decouple fire characteristics and purification interference characteristics to output fire causal characteristics and scene interference characteristics.

[0160] For example, two core data types—fire features and purification interference features—are collected uniformly. The fire early warning and identification terminal performs standardized processing to unify the dimensional specifications, numerical ranges, and expression formats of the two types of features, eliminating fusion biases caused by differences in feature structures. The fire early warning and identification terminal configures differentiated fusion weights according to the actual needs of the fire identification task, completing the integration and orderly output of feature information while ensuring that the two types of features are independent and do not interfere with each other.

[0161] The fire early warning and identification terminal solidifies the advantages of decoupling fire features from causal representation and the advantages of anti-interference representation of purification interference features, and stably outputs fire causal features with complete physical causal attributes, as well as scene interference features with strong semantic isolation capabilities.

[0162] Among them, fire causal features are features that integrate and decouple the core information of fire features and can reflect the causal laws of fire occurrence and development; scene interference features are features that integrate the exclusive information of purification interference features and are used to block the visual correlation of non-fire physical interference sources.

[0163] In one embodiment, a fire identification model is constructed by performing joint training with physical regularization based on physical constraint matching results, fire causal features, and scene interference features, including:

[0164] (1) Integrate fire causal features, physical constraint matching results and scene interference features to generate a physical guidance feature vector.

[0165] For example, after acquiring fire causal features, physical constraint matching results, and scene interference features, the fire early warning and identification terminal performs independent standardization and normalization processing on the three types of input data, unifying the numerical distribution range and feature dimension specifications of all data, and eliminating fusion interference caused by differences in dimensions and numerical deviations. The fire early warning and identification terminal assigns adaptive fusion weights based on the priority of physical constraints, the contribution of fire causality, and the requirements for scene interference suppression. It then performs deep feature concatenation and nonlinear mapping fusion on the standardized fire causal features, physical constraint matching results, and scene interference features, retaining the core representation information of each type of data, and generating a physical guidance feature vector that integrates prior physical knowledge and visual features.

[0166] Among them, the physical guidance feature vector refers to a high-dimensional joint feature that integrates fire visual causal features, physical constraint matching information and scene interference features, and is used to provide physical prior guidance for model training.

[0167] (2) The optimization direction of the physical guidance feature vector is adjusted by the gradient reweighting mechanism to obtain the regularized gradient signal.

[0168] For example, after loading the physical guidance feature vector, the fire early warning and identification terminal calculates the initial loss gradient of the neural network model, clarifies the original parameter optimization direction of the model, and then activates the built-in gradient reweighting mechanism. This mechanism adjusts the weighting coefficients of the gradient components based on the confidence level of the physical constraint matching results, strengthening gradient contributions that conform to physical laws and suppressing gradient components that violate physical mechanisms. The fire early warning and identification terminal corrects the optimization direction of the model parameters through weighted operations, eliminates gradient shifts without physical meaning, and integrates and normalizes the adjusted gradient components to generate a regularized gradient signal that conforms to physical regularization constraints.

[0169] Among them, the gradient reweighting mechanism refers to the regularization method that adjusts the weights of gradient components and corrects the direction of model optimization based on physical constraint rules; the regularized gradient signal refers to the model optimization gradient that has been corrected by physical constraints and conforms to the laws of physical evolution.

[0170] (3) Based on the adversarial perturbation test, verify the robustness of the regularized gradient signal and generate the purification gradient parameters.

[0171] For example, after acquiring the regularized gradient signal, the fire early warning and identification terminal generates multiple sets of adversarial perturbation data based on a preset interference intensity. This perturbation data is then superimposed onto the regularized gradient signal to simulate noise interference and feature shift conditions in a real-world scenario. The fire early warning and identification terminal performs robustness verification tests on the injected perturbation gradient signal, selecting effective gradient components that can resist interference and maintain stability, eliminating invalid gradient components with abrupt gradient changes or numerical anomalies, and performing smoothing, purification, and numerical calibration on the retained effective gradient components to generate robust purification gradient parameters.

[0172] Among them, adversarial perturbation testing refers to a test method that verifies the stability of gradient signals by artificially adding perturbations; cleaned gradient parameters refer to stable gradient parameters after robustness verification and removal of outliers.

[0173] (4) Based on the purification gradient parameters, the neural network weights are iteratively optimized to construct a fire identification model.

[0174] For example, after obtaining the purification gradient parameters, the fire early warning and identification terminal inputs these parameters into the backpropagation unit of the neural network, and iteratively updates all network weight parameters, including fully connected layers and convolutional layers, according to the gradient descent algorithm. The fire early warning and identification terminal continuously and iteratively executes the feature fusion, gradient adjustment, robustness verification, and parameter update process until the model loss function converges to a preset threshold and the identification accuracy meets the requirements of industrial applications. The fire early warning and identification terminal integrates the optimized network structure, weight parameters, and physical regularization constraint rules to solidify and encapsulate the model, constructing a fire identification model with physical constraints and anti-interference capabilities.

[0175] Among them, neural network weights refer to the connection parameters between nodes in each layer of the neural network; the fire identification model refers to a deep learning model for industrial fire identification that integrates physical regularization constraints and has high robustness and anti-interference capabilities.

[0176] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0177] In one embodiment, such as Figure 3 As shown, this application also provides an intelligent image recognition device 300 for fire early warning, the device 300 comprising:

[0178] Interference source modeling module 301 is used to build a physical proxy model library based on the physical generation mechanism of non-fire interference sources;

[0179] The spatiotemporal feature extraction module 302 is used to extract multi-scale temporal features from the surveillance video stream to obtain dynamic evolution features;

[0180] The physical matching module 303 is used to perform physical constraint matching on dynamic evolution characteristics based on environmental parameters collected by environmental sensors and a physical proxy model library, and generate physical constraint matching results.

[0181] The feature decoupling module 304 is used to decouple the features of the input image through a causal decoupling network, and separate the fire causal features and scene interference features, wherein the scene interference features are used to block the visual correlation of non-fire physical interference sources.

[0182] The joint training module 305 is used to perform physical regularization joint training based on physical constraint matching results, fire causal features and scene interference features to build a fire identification model.

[0183] Specifically, the fire early warning and identification terminal includes an interference source modeling module 301, a spatiotemporal feature extraction module 302, a physical matching module 303, a feature decoupling module 304, and a joint training module 305.

[0184] The interference source modeling module is based on the survey results of non-fire interference sources that are prone to misjudgment in industrial fire protection scenarios. It categorizes and sorts them according to physical attributes, clarifying the physical laws and mechanisms of action followed by the generation and evolution of various interference sources. A multiphysics simulation environment is built to reproduce the real evolution process of interference sources, collect dynamic morphological sequence data, and fit the physical evolution laws and feature mapping relationships through differentiable modeling to train differentiable physical proxy models corresponding to various interference sources. All models are bound by category labels, standardized in format, and archived to build a physical proxy model library.

[0185] Among them, the physical generation mechanism of non-fire interference sources refers to the physical laws followed in the formation and evolution of various interference sources, including principles such as energy conversion and matter diffusion; the physical proxy model library is a collection of physical evolution models of various interference sources that are systematically stored, providing support for subsequent physical constraint matching.

[0186] The spatiotemporal feature extraction module continuously receives the monitoring video stream, extracts images frame by frame, and performs preprocessing such as noise suppression and distortion correction. Multi-scale convolution operations are used to extract spatial features at different resolution levels. Inter-frame correlation analysis is used to capture the temporal change information of the target. Spatial features and temporal features are aligned in dimension, weighted, and fused to retain core dynamic information and output dynamic evolution features.

[0187] Among them, multi-scale temporal feature extraction is a technique that takes into account both spatial resolution and temporal dimension to capture the static features and dynamic changes of the target; dynamic evolution features are a comprehensive feature set that integrates the spatial morphology and temporal patterns of the target, and fully represents the dynamic development process of the target.

[0188] The physical matching module receives parameters collected by environmental sensors in real time. After validity verification and standardization, it selects interference source models from the physical proxy model library that are suitable for the characteristics of the current environmental parameters. The standardized parameters are input into the model to predict the expected dynamic pattern of the interference source, extract the spatial and temporal information of the dynamic evolution characteristics and the expected pattern, quantify the fit and synchronization between the two, and generate physical constraint matching results.

[0189] Among them, physical constraint matching is a process of verifying the consistency between dynamic evolution characteristics and expected patterns of interference sources based on physical laws and environmental conditions; environmental sensors are used to collect on-site environmental state parameters in real time; and the results of physical constraint matching are clear determination information on the interference source attributes of dynamic evolution characteristics.

[0190] The feature decoupling module receives the input image to be processed and performs preprocessing. It extracts core fire-related features through a built-in thermodynamic causal analysis unit to obtain initial fire features; and mines non-fire interference features through an environmental interference analysis unit to obtain initial interference features. A causal intervention unit is then activated to eliminate the statistical correlation between the two types of features, refining the fire features. An adversarial discriminator is used to block the association between the initial interference features and fire semantics, generating purified interference features. After standardization of the two types of features, they are fused to output fire causal features and scene interference features.

[0191] Among them, the causal decoupling network is the core algorithm framework of the feature decoupling module, which has the ability to separate different semantic associations in mixed features; the fire causal features are a set of features that reflect the core laws of fire occurrence and development; the scene interference features are features that retain only interference attributes and block the semantic associations of fire, and are used to suppress interference effects.

[0192] The joint training module collects physical constraint matching results, fire causal features, and scene interference features. After dimensionality calibration and normalization, adaptive weights are assigned according to physical constraint priority and feature contribution, completing deep fusion and generating a physical-guided feature vector. A physical regularization mechanism is introduced to adjust the optimization direction, generating a regularized gradient signal. Its robustness is verified through adversarial perturbation testing, and outliers are removed to obtain purification gradient parameters. Based on these parameters, the neural network weights are iteratively adjusted to continuously optimize the model's discrimination accuracy and anti-interference capability. Once the model performance meets the standards, it is solidified and encapsulated to construct a fire identification model.

[0193] Among them, physical regularization joint training is a model optimization method that integrates physical constraints and data-driven training, which improves the reliability of the model through physical mechanism constraints; the fire identification model is a deep learning model that can distinguish between fire and interference and achieve rapid early warning.

[0194] The physical matching module 303 is also used for:

[0195] The multidimensional environmental parameters collected by environmental sensors are analyzed to obtain an environmental parameter vector;

[0196] By processing environmental parameter vectors through a physical proxy model library, the evolutionary behavior of interference sources is predicted, and expected dynamic patterns are generated.

[0197] Quantitatively compare the spatial distribution and temporal evolution consistency of dynamic evolution characteristics and expected dynamic patterns to generate similarity metrics;

[0198] Based on the environmental adaptive threshold similarity metric, physical constraint matching results are generated.

[0199] The physical matching module 303 is also used for:

[0200] Multiphysics coupling analysis is performed on the environmental parameter vector to obtain the initial evolution state;

[0201] The initial evolution state is processed based on the physical proxy model library, the morphological evolution of the interference source is simulated, and a dynamic evolution sequence is generated.

[0202] Extract the spatiotemporal invariant features of the dynamic evolution sequence to generate the expected dynamic pattern.

[0203] Interference source modeling module 301 is also used for:

[0204] Define a set of non-fire interference sources, including water vapor, welding sparks, and industrial dust.

[0205] Multiphysics simulations were performed on each category in the set of non-fire interference source categories to obtain the corresponding dynamic morphological sequence data;

[0206] Based on the environmental parameter vectors collected by environmental monitoring equipment, a differentiable surrogate model is performed on the dynamic morphological sequence data to obtain a differentiable physical surrogate model;

[0207] Differentiable physical proxy models are mapped to corresponding interference source categories to establish a physical proxy model library.

[0208] Interference source modeling module 301 is also used for:

[0209] Physical sensing encoding is performed on the environmental parameter vector to obtain a joint environmental-physical representation;

[0210] By processing the environment-physical joint characterization and dynamic morphological sequence data through a differentiable solver, surrogate model parameters are generated.

[0211] Based on variational inference, optimize the surrogate model parameters to generate a differentiable physical surrogate model.

[0212] Feature decoupling module 304 is also used for:

[0213] Thermodynamic causal analysis is performed on the input image to extract heat release rate features and carbon particle diffusion features, and initial fire features are generated.

[0214] Environmental interference analysis is performed on the input image to extract optical reflection features and non-fire motion features, and initial interference features are generated.

[0215] By eliminating the statistical correlation between initial fire features and initial disturbance features through causal intervention units, decoupled fire features are obtained.

[0216] By blocking the fire semantic association of the initial interference features through the adversarial discriminator, purification interference features are generated;

[0217] It integrates and decouples fire characteristics and purification interference characteristics to output fire causal characteristics and scene interference characteristics.

[0218] The joint training module 305 is also used for:

[0219] By integrating fire causal characteristics, physical constraint matching results, and scene interference characteristics, a physical guidance feature vector is generated.

[0220] The optimization direction of the physically guided feature vector is adjusted by the gradient reweighting mechanism to obtain the regularized gradient signal;

[0221] The robustness of the regularized gradient signal is verified by adversarial perturbation testing, and the cleaned gradient parameters are generated.

[0222] A fire identification model is constructed by iteratively optimizing the neural network weights based on the purification gradient parameters.

[0223] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0224] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0225] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0226] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An intelligent image recognition method for fire early warning, characterized in that, The method includes: A physical proxy model library is constructed based on the physical generation mechanism of non-fire interference sources; Multi-scale temporal feature extraction is performed on the surveillance video stream to obtain dynamic evolution features; Based on environmental parameters collected by environmental sensors and the physical proxy model library, physical constraint matching is performed on the dynamic evolution characteristics to generate physical constraint matching results; The input image is decoupled by a causal decoupling network to separate fire causal features and scene interference features, wherein the scene interference features are used to block the visual correlation of non-fire physical interference sources. A fire identification model is constructed by performing joint training with physical regularization based on the physical constraint matching results, the fire causal features, and the scene interference features.

2. The intelligent image recognition method for fire early warning according to claim 1, characterized in that, The environmental parameters collected by environmental sensors and the physical proxy model library are used to perform physical constraint matching on the dynamic evolution features, generating physical constraint matching results, including: The multidimensional environmental parameters collected by environmental sensors are analyzed to obtain an environmental parameter vector; The environmental parameter vector is processed by the physical proxy model library to predict the evolutionary behavior of interference sources and generate expected dynamic patterns. Quantitatively compare the spatial distribution and temporal evolution consistency of the dynamic evolution features and the expected dynamic patterns to generate a similarity metric. The physical constraint matching result is generated based on the similarity metric value determined by the environment adaptive threshold.

3. The intelligent image recognition method for fire early warning according to claim 2, characterized in that, The process of processing the environmental parameter vector through the physical proxy model library to predict the evolutionary behavior of interference sources and generate expected dynamic patterns includes: Multiphysics coupling analysis is performed on the environmental parameter vector to obtain the initial evolution state; The initial evolutionary state is processed based on the physical proxy model library to simulate the morphological evolution of the interference source and generate a dynamic evolutionary sequence. Extract the spatiotemporal invariant features of the dynamic evolution sequence to generate the expected dynamic pattern.

4. The intelligent image recognition method for fire early warning according to claim 1, characterized in that, The physical proxy model library constructed based on the physical generation mechanism of non-fire interference sources includes: Define a set of non-fire interference source categories, which includes water vapor, welding sparks, and industrial dust. Multiphysics simulation was performed on each category in the set of non-fire interference source categories to obtain the corresponding dynamic morphological sequence data. Based on the environmental parameter vectors collected by environmental monitoring equipment, a differentiable proxy model is performed on the dynamic morphological sequence data to obtain a differentiable physical proxy model. The differentiable physical proxy model is mapped to the corresponding interference source category to establish the physical proxy model library.

5. The intelligent image recognition method for fire early warning according to claim 4, characterized in that, The environmental parameter vectors collected by the environmental monitoring equipment are used to perform differentiable proxy modeling on the dynamic morphological sequence data to obtain a differentiable physical proxy model, including: The environmental parameter vector is physically sensed and encoded to obtain a joint environmental-physical representation; The surrogate model parameters are generated by processing the environment-physical joint characterization and the dynamic morphological sequence data through a differentiable solver. The parameters of the proxy model are optimized based on variational inference to generate the differentiable physical proxy model.

6. The intelligent image recognition method for fire early warning according to claim 1, characterized in that, The step of decoupling features from the input image using a causal decoupling network to separate fire causal features and scene interference features includes: Thermodynamic causal analysis is performed on the input image to extract heat release rate features and carbon particle diffusion features, thereby generating initial fire features; Environmental interference analysis is performed on the input image to extract optical reflection features and non-fire motion features, and initial interference features are generated. By eliminating the statistical correlation between the initial fire features and the initial interference features through a causal intervention unit, decoupled fire features are obtained. By blocking the fire semantic association of the initial interference features through the adversarial discriminator, purification interference features are generated; By fusing the decoupled fire features and the purification interference features, the fire causal features and the scene interference features are output.

7. The intelligent image recognition method for fire early warning according to claim 1, characterized in that, The step of jointly training a fire identification model based on the physical constraint matching results, the fire causal features, and the scene interference features using physical regularization to construct the model includes: By integrating the fire causal features, the physical constraint matching results, and the scene interference features, a physical guidance feature vector is generated. The optimization direction of the physical guidance feature vector is adjusted by a gradient reweighting mechanism to obtain a regularized gradient signal; The robustness of the regularized gradient signal is verified based on the adversarial perturbation test, and the cleanup gradient parameters are generated. The fire identification model is constructed by iteratively optimizing the neural network weights based on the purification gradient parameters.

8. An intelligent image recognition device for fire early warning, characterized in that, The device includes: The interference source modeling module is used to build a physical proxy model library based on the physical generation mechanism of non-fire interference sources; The spatiotemporal feature extraction module is used to extract multi-scale temporal features from surveillance video streams to obtain dynamic evolution features; The physical matching module is used to perform physical constraint matching on the dynamic evolution features based on environmental parameters collected by environmental sensors and the physical proxy model library, and generate physical constraint matching results. The feature decoupling module is used to decouple the features of the input image through a causal decoupling network, and separate the fire causal features and scene interference features, wherein the scene interference features are used to block the visual correlation of non-fire physical interference sources. The joint training module is used to perform physical regularization joint training based on the physical constraint matching results, the fire causal features, and the scene interference features to construct a fire identification model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent image recognition method for fire early warning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent image recognition method for fire early warning as described in any one of claims 1 to 7.

Images