Indoor fire detection method based on multi-source heterogeneous data fusion

Abstract

This invention discloses an indoor fire detection method based on multi-source heterogeneous data fusion, comprising the following steps: collecting multi-source heterogeneous fire data to construct a visible light-infrared light image dataset and a temperature-smoke concentration-CO concentration sensor parameter dataset; extracting fire features from the datasets and selecting core features; inputting the core features into a pre-constructed FA-Bi-LSTM thermal smoke detection model and a PSO-Adaboost flame image detection model to obtain the corresponding fire probability distribution results; dividing the detection scene based on image entropy and sensor data variance, dynamically allocating the decision weights of the thermal smoke detection model and the flame image detection model according to the detection scene, and weighted fusing the probability output results of the two models to obtain the final fire detection result. This indoor fire detection method based on multi-source heterogeneous data fusion can significantly improve the accuracy of fire detection in complex and variable indoor environments.

Description

Technical Field

[0001] This invention relates to the field of fire detection technology, and in particular to an indoor fire detection method based on multi-source heterogeneous data fusion. Background Technology

[0002] Fire detection is a core component of indoor safety protection. Existing fire detection methods are mainly divided into two categories: detection methods based on sensor signals and detection methods based on image signals. However, both of them have obvious technical defects.

[0003] In the field of fire detection based on sensor signals, traditional sensors such as temperature, smoke concentration, and CO concentration rely on threshold judgments of single physical parameters. In the early stages of a fire, the temperature rises slowly and the smoke concentration is low, resulting in a lag in sensor response and a tendency to miss the best time to extinguish the fire. At the same time, interference factors such as cooking steam, cigarette smoke, and dust in scenarios such as kitchens, smoking rooms, and dusty workshops can easily cause false alarms. In some places, the false alarm rate of sensors can even exceed 40%. Moreover, a single sensor cannot capture the multi-dimensional characteristics of a fire, resulting in insufficient accuracy in identifying complex fires.

[0004] In the field of fire detection based on image signals, visible light image detection relies on visual features such as flame color and shape. In scenarios such as dense smoke, low illumination, and the presence of fire-like interfering objects, visual features are easily obscured or confused, resulting in a false alarm rate as high as 20%-30%. Although infrared image detection can penetrate some smoke and capture thermal radiation features, it is not good at depicting smoke texture and dynamic changes in flames. It is also difficult to distinguish between high-temperature areas of fire and static high-temperature sources such as industrial equipment and direct sunlight, leading to prominent false alarm problems in industrial scenarios.

[0005] In existing technologies, fire sample data research often uses sensor features or image features alone, rarely fusing the two for analysis. Detection methods based on single-modal features ignore the information complementarity between different modal features, resulting in limited information and low discriminative power, making it difficult to significantly improve the accuracy of fire detection. With the development of IoT and AI technologies, how to effectively fuse sensor features and image features to improve the accuracy and robustness of fire detection in complex indoor scenarios has become an urgent problem to be solved. Summary of the Invention

[0006] In view of this, the present invention addresses the problems of high false alarm and missed alarm rates, poor complementarity of single-modal feature information, and difficulty in adapting to complex indoor scenarios in existing indoor fire detection methods. It provides an indoor fire detection method based on multi-source heterogeneous data fusion, which realizes the complementary advantages of sensor features and image features, and improves the accuracy, robustness and real-time performance of fire detection.

[0007] To achieve the above objectives, this invention provides an indoor fire detection method based on multi-source heterogeneous data fusion, specifically including the following steps:

[0008] Collect multi-source heterogeneous fire data to construct a visible light-infrared light image dataset and a temperature-smoke concentration-CO concentration sensor parameter dataset;

[0009] Fire features were extracted from the visible-infrared image dataset and the temperature-smoke-CO concentration sensor parameter dataset, and core features were selected to form a core feature set;

[0010] The core features are input into the pre-constructed FA-Bi-LSTM hot smoke detection model and PSO-Adaboost flame image detection model to obtain the fire probability distribution results of the two models. The input of the FA-Bi-LSTM hot smoke detection model is the core features of fire sensing parameters, and the input of the PSO-Adaboost flame image detection model is the core features of image type.

[0011] The detection scenarios are divided based on image entropy and sensor data variance. The decision weights of the FA-Bi-LSTM hot smoke detection model and the PSO-Adaboost flame image detection model are dynamically allocated according to the detection scenarios, and the probability output results of the two models are weighted and fused to obtain the final fire detection result.

[0012] Preferably, the extracted fire features include five categories: static features of visible light flame images, dynamic features of visible light flame images, visible light smoke image features, infrared light image features, and fire sensing parameter features.

[0013] Further optimization, the core feature selection method is as follows:

[0014] For each fire feature, the importance score is calculated independently using the random forest algorithm and the gradient boosting tree algorithm. The scores from the two algorithms for a single feature are then combined and averaged to obtain the final importance score for each fire feature.

[0015] All features are ranked by importance, and low-importance, redundant, and irrelevant features are removed based on the ranking results. The core features with the highest discriminative value among all types of features are retained to form a core feature set.

[0016] Further optimization reveals that the static core features of visible light flame images are the mean of the grayscale histogram, local binary mode, and color saturation; the dynamic core features of visible light flame images are the flame movement speed, temporal changes in flame color, and edge straightness; the core features of visible light smoke images are color features, texture features, and dynamic motion features; the infrared light image features retain penetration characteristics, temperature features, and thermal diffusion features; and the core features of fire sensing parameters are carbon monoxide slope, temperature slope, and smoke slope.

[0017] Further optimization reveals that the FA-Bi-LSTM thermal smoke detection model includes an FA feature processing layer and an optimized Bi-LSTM layer. The FA feature processing layer dynamically weights and enhances the core features of the fire sensing parameters, and the optimized Bi-LSTM layer bidirectionally captures the temporal feature dependencies, ultimately outputting the probability of fire / no fire from the FA-Bi-LSTM thermal smoke detection model.

[0018] Further preferably, the data processing procedure of the FA feature processing layer is as follows:

[0019] The input sensing time-series data is compressed globally, and the feature global representation vector is extracted by global average pooling.

[0020] The feature global representation vector is nonlinearly transformed by a single fully connected layer to obtain the feature similarity score. The feature similarity score is then normalized by the Softmax function to obtain the FA weight distribution.

[0021] The original feature vector is weighted and adjusted using the FA weights to obtain a weighted feature time series.

[0022] Further preferred, the data processing procedure for the optimized Bi-LSTM layer is as follows:

[0023] The bidirectional contextual dependencies of the weighted feature time series output by the FA feature processing layer are captured by two independent hidden state flows, forward and backward, and then fused to form a comprehensive representation.

[0024] The comprehensive representation is linearly transformed and mapped with activation functions through a fully connected layer to output the fire probability distribution of the thermal smoke detection model.

[0025] Further preferred, the training and usage methods of the PSO-Adaboost flame image detection model are as follows:

[0026] The core features of the input image are standardized and preprocessed, and then the training set and test set are randomly divided.

[0027] Configure the core parameters of the PSO algorithm and define the hyperparameter search space of the Adaboost algorithm;

[0028] Construct a fitness function to evaluate the individual particle performance, iteratively update the individual historical best solution and the group global best solution, and dynamically adjust the particle search position and movement speed.

[0029] Iterate until the convergence condition is met, output the optimized Adaboost parameter combination, and train the PSO-Adaboost flame image detection model.

[0030] Input the core features of the image to be predicted, complete the inference, and output the fire probability distribution of the flame image detection model.

[0031] Further optimization involves dividing the detection scenarios based on image entropy and sensor data variance, and dynamically allocating decision weights for the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model according to the detection scenarios, including the following steps:

[0032] Real-time calculation of image entropy based on image grayscale distribution; real-time calculation of sensor data variance based on short sliding window for sensor data;

[0033] The detection scenarios are divided based on the image entropy and sensor data variance, wherein the detection scenarios include scenarios with clear image features, scenarios with disturbed image features, and scenarios with drastic fluctuations in sensor data.

[0034] Differentiated weights were assigned to the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model for different detection scenarios. Specifically, for scenarios with clear image features, the weights of the FA-Bi-LSTM hot flue gas detection model were adjusted accordingly. <Weights of the PSO-Adaboost flame image detection model For scenarios where image features are disturbed, the weights of the FA-Bi-LSTM thermal smoke detection model... > Weights of the PSO-Adaboost flame image detection model For scenarios with drastic fluctuations in sensor data, the weights of the FA-Bi-LSTM hot flue gas detection model... =Weights of the PSO-Adaboost flame image detection model .

[0035] Further optimization yields the following probability outputs from the weighted fusion of the two models:

[0036] The probability outputs of the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model are integrated by weighted summation to obtain the fused comprehensive probability of fire and comprehensive probability of no fire.

[0037] The combined probability of fire and the combined probability of no fire are compared with a preset confidence threshold to obtain the final detection result.

[0038] This invention provides an indoor fire detection method based on multi-source heterogeneous data fusion. It collects sensor parameter data and visible-infrared image data to construct a multi-source heterogeneous dataset, achieving complementary advantages between sensor features and image features. This method characterizes fire features from multiple dimensions, including visual, thermodynamic, and chemical perspectives, covering the entire fire stage from smoldering to open flame. It solves the problems of limited information and low discriminative power of single-modal features. By selecting core features, redundant features can be effectively removed, avoiding the curse of dimensionality, reducing computational complexity, and improving the efficiency and discriminative power of fire detection. The constructed FA-Bi-LSTM model introduces a feature attention mechanism, dynamically weighting core features at different fire stages. The combined bidirectional temporal modeling capability of Bi-LSTM can accurately capture the dynamic evolution of sensor data, improving the robustness of hot smoke detection. The constructed PSO-Adaboost model optimizes the hyperparameters of Adaboost through the PSO algorithm, avoiding local optima and enhancing the visual recognition capability of flames and smoke in complex scenarios, balancing recognition accuracy and real-time performance. The designed dynamic weight decision-level fusion strategy adaptively adjusts the model weights based on image entropy and sensor data variance, making the fusion results accurately adaptable to different indoor detection scenarios, significantly reducing the false alarm rate and false alarm rate of fire detection, and improving the accuracy and reliability of fire detection in complex and variable indoor environments.

[0039] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the disclosure of the present invention. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart illustrating the fire detection method based on multi-source heterogeneous information fusion provided by this invention;

[0043] Figure 2 This is a schematic diagram of the FA-Bi-LSTM hot flue gas detection model;

[0044] Figure 3 This is a flowchart of the PSO-AdaBoost algorithm. Detailed Implementation

[0045] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of methods consistent with some aspects of the invention as detailed in the appended claims.

[0046] To further explain the present invention, specific experimental examples are provided below.

[0047] like Figure 1 As shown, this invention provides an indoor fire detection method based on multi-source heterogeneous data fusion. Indoor fire detection is achieved through a series of steps: multi-source heterogeneous fire data acquisition, core feature screening, dual-detection model construction and parallel inference, and dynamic weighted decision-level fusion judgment. The specific steps are as follows:

[0048] S1: Collect multi-source heterogeneous fire data and construct a visible light-infrared light image dataset and a temperature-smoke concentration-CO concentration sensor parameter dataset;

[0049] Existing detection methods based on single-modal features neglect the complementarity of information between different modal features, resulting in limited information and low discriminative power, making it difficult to significantly improve the accuracy of fire detection. Therefore, this invention collects multi-source heterogeneous fire data and constructs a visible light-infrared light image dataset and a temperature-smoke concentration-CO concentration sensor parameter dataset. This comprehensively characterizes fire features from multiple dimensions, including visual, thermodynamic, and chemical perspectives, covering the entire stage of a fire from smoldering to open flame. The sensor data reflects the temporal changes of physicochemical parameters such as temperature, smoke concentration, and CO concentration. The visible light image data accurately records the shape, color, and flickering characteristics of flames, as well as visual features such as smoke diffusion dynamics and transparency, providing data support for detection methods based on traditional image processing. The infrared light image data effectively captures the temperature distribution and thermal radiation characteristics during combustion, can penetrate some smoke, and can still provide effective heat source information in low-light environments, compensating for the limitations of visible light images in specific scenarios.

[0050] S2: Extract fire features from the visible light-infrared light image dataset and the temperature-smoke concentration-CO concentration sensing parameter dataset, and filter core features to form a low-dimensional, high-discrimination core feature set;

[0051] Among them, fire features include five categories: static features of visible light images of flames, dynamic features of visible light images of flames, visible light image features of smoke, infrared light image features, and fire sensing parameter features, which are integrated to form a fire multimodal feature set.

[0052] Specifically, static features of visible light images of flames: from color, texture, and grayscale histograms. Figure 3 Eight features are extracted from each dimension: color histogram peak value, color mean, color saturation, grayscale histogram mean, grayscale histogram kurtosis, grayscale co-occurrence matrix contrast, wavelet texture energy, and local binary pattern features.

[0053] Dynamic features of visible light images of flames: Six representative dynamic features are extracted from three dimensions: motion characteristics, temporal changes, and edge morphology. Specifically, these include: the motion speed and direction of the flame region, the trend of brightness change over time, the trend of color change over time, the edge length of the flame outline, and the straightness of its edge.

[0054] Visible light image features of smoke: extracting smoke color features, smoke texture features, and smoke motion dynamic features from smoke texture features and diffusion patterns;

[0055] Infrared image features: Extraction of penetration characteristics, temperature characteristics, and thermal diffusion characteristics;

[0056] Fire sensing parameter characteristics: Nine typical early fire parameters were extracted, namely: smoke slope, smoke amplitude, smoke kurtosis, carbon monoxide slope, carbon monoxide amplitude, carbon monoxide kurtosis, temperature slope, temperature amplitude, and temperature kurtosis.

[0057] The method for selecting core features is as follows:

[0058] Fire feature importance is ranked based on Random Forest (RF) and Gradient Boosting Tree (XGBoost) algorithms. Low-importance, redundant and irrelevant features are removed based on the ranking results, and the core features with the highest discriminative value among all types of features are retained to form a low-dimensional, high-discriminative core feature set.

[0059] The method for ranking fire feature importance based on the Random Forest (RF) algorithm and the Gradient Boosting Tree (XGBoost) algorithm is as follows:

[0060] For each independent feature in the fire multimodal feature set, the importance score is calculated independently using the Random Forest (RF) algorithm and the Gradient Boosting Tree (XGBoost) algorithm. The scores of the two algorithms for a single feature are then combined and averaged to obtain the final importance score for each feature.

[0061] After all features have completed their final importance score calculations, all features are ranked by importance in a uniform manner.

[0062] Using multimodal features can more accurately and comprehensively reflect the development of a fire, but the feature vector dimension increases significantly, which can easily generate noise and redundancy, and even lead to the "curse of dimensionality", seriously affecting computational efficiency and accuracy. Therefore, this invention sorts the fire features by importance based on RF and XGBoost. Then, based on the sorting results, irrelevant and redundant features can be removed from the initial feature set to extract the optimal multimodal feature set that is necessary and sufficient to solve the problem.

[0063] Among them, the two feature selection algorithms, RF and XGBoost, have certain differences in importance scores. This is mainly due to the different calculation principles and evaluation perspectives they are based on. Compared with the ranking results of relying on only a single algorithm, by combining the scores of the two algorithms and taking the average, the feature preferences and evaluation tendencies of different models can be taken into account to a greater extent, so that the final feature importance ranking is more comprehensive and objective.

[0064] Among the eight static features of visible light flame images, after ranking the importance of the extracted fire features using the Random Forest (RF) algorithm and the Gradient Boosting Tree (XGBoost) algorithm, the scores of the three features—mean grayscale histogram (0.157), local binary pattern (0.146), and color saturation (0.132)—were significantly higher than those of the other features. This indicates that these features have high discriminative power and representational importance in static flame recognition. To ensure that feature information is not lost in large quantities, this embodiment retains the top three static flame features as input to the subsequent flame image detection model to improve the accuracy of fire recognition.

[0065] In the dynamic features of visible light flame images, the flame's motion speed (0.188), flame color temporal change (0.181), and edge straightness (0.168) have high importance scores, while the scores of other features are relatively low. In order to reduce the loss of effective feature information, this embodiment selects the top three dynamic features in terms of importance as the input of the flame image detection model.

[0066] In the visible light image features of smoke, color features (mean 0.347), texture features (mean 0.343), and motion dynamic features (mean 0.311) are the most significant.

[0067] Among the infrared image features, the penetration characteristics (mean 0.379), temperature characteristics (mean 0.311), and thermal diffusion characteristics (mean 0.310) are:

[0068] Among the fire sensing parameter features, the importance scores of the nine features showed significant differences. Among them, the three rate-of-change features, namely carbon monoxide slope (0.215), temperature slope (0.198), and smoke slope (0.186), had significantly higher importance than other parameters. This ranking result indicates that in early fire identification, the trend of sensor readings over time has a stronger discriminative ability than their instantaneous amplitude or distribution shape. This phenomenon is consistent with the typical process of continuous evolution of physical parameters in the initial stage of a fire, further confirming the key role of temporal change features in fire early warning. In order to construct an efficient and highly interpretable feature set, this embodiment retains the top three temporal change features based on importance, namely carbon monoxide slope, temperature slope, and smoke slope, and discards the other six features with lower importance. This screening strategy aims to highlight the most sensitive dynamic change signals in the early stage of a fire, providing more compact and discriminative parameter inputs for subsequent thermal smoke detection models.

[0069] In summary, the static core features of visible light flame images are the mean of the grayscale histogram, local binary mode, and color saturation.

[0070] The core dynamic features of visible light images of flames are flame movement speed, temporal changes in flame color, and edge straightness.

[0071] The core features of visible light images of smoke are color features, texture features, and motion dynamic features;

[0072] Infrared light images retain characteristics such as penetration properties, temperature characteristics, and thermal diffusion characteristics.

[0073] The core characteristics of fire sensing parameters are carbon monoxide slope, temperature slope, and smoke slope.

[0074] S3: Input the core features into the pre-built FA-Bi-LSTM hot smoke detection model and PSO-Adaboost flame image detection model to obtain the fire probability distribution results of the two models;

[0075] Specifically, the core features of fire sensing parameters (carbon monoxide slope, temperature slope, and smoke slope) from the core feature set are input into the FA-Bi-LSTM thermal smoke detection model, and the core features of image types (flame visible light, smoke visible light, and infrared light image core features) are input into the PSO-Adaboost flame image detection model to achieve parallel inference of the two models.

[0076] The FA-Bi-LSTM hot flue gas detection model is explained below:

[0077] Traditional LSTM models apply equal attention to all feature dimensions of the input, failing to adjust feature importance weights according to the dynamic changes in the fire scene. This results in the dilution of the discriminative information of the selected core features and the amplification of the influence of interfering features, making it difficult to fully utilize the value of high-discriminative features. To address this issue, the proposed FA-Bi-LSTM thermal smoke detection model, considering the temporal continuity and stage-specific differences in sensor parameter data, introduces a common feature attention mechanism (FA) into the Bi-LSTM model. Through dynamic weighting and enhancement of core features, it achieves adaptive adjustment of feature importance for different fire stages, enabling accurate identification of fire thermal smoke sensing signals and outputting the probability of fire / no fire from the sensor model.

[0078] like Figure 2 As shown, the FA-Bi-LSTM hot smoke detection model includes an FA feature processing layer and an optimized Bi-LSTM layer. The core function of the FA feature processing layer is to dynamically adjust the model's focus on different feature dimensions based on the correlation between the core features of fire sensing parameters and the fire identification task. The FA feature processing layer is used to perform the following operations:

[0079] 1) Global information compression is performed on the input sensing time-series data, and the global feature representation vector is extracted through global average pooling:

[0080] Let the input sensing timing data be... ,in, Represents the set of real numbers. For batch size, For timing length, For the feature dimensions, corresponding to CO slope, temperature slope, and smoke slope; to adapt the weight calculation for the feature dimensions, the input shape is first adjusted to... Then, global average pooling (GAP) is used to extract the global representation vector of the features. , among which, the Global representation of each feature The formula is as follows:

[0081] ;

[0082] in, For the first The feature in the first The value at each time step, The timing length;

[0083] 2) The feature global representation vector is nonlinearly transformed by a single fully connected layer to obtain the feature similarity score. Then, the feature similarity score is normalized by the Softmax function to obtain the FA weight distribution.

[0084] The global feature representation vector is represented by a single fully connected layer. Perform a nonlinear transformation to obtain the feature similarity score. The formula is as follows:

[0085] ;

[0086] in, For the learnable weight matrix of the fully connected layer, For bias terms, This is the feature similarity score vector;

[0087] The feature similarity scores are normalized using the Softmax function to obtain the FA weight distribution that satisfies the probability distribution characteristics. , among which, the FA weight formula for each feature The formula is as follows:

[0088] ;

[0089] in, The range of values ​​is The sum of the FA weights of all features is 1;

[0090] 3) The original feature vector is weighted and adjusted using the FA weights to obtain a weighted feature vector (weighted feature time series).

[0091] In this step, the time series of each feature dimension will be strengthened or weakened according to the corresponding FA weight, so that the discriminative information of key features is highlighted (e.g., increasing the CO slope weight during smoldering and increasing the temperature slope weight during open flame), and the influence of interfering features is suppressed. The weighting process formula is as follows:

[0092] ;

[0093] in, For the first The weighted feature vector at each time step For the first The feature in the first The original values ​​for each time step. For the first The FA weights of each feature, ;

[0094] This weighting method helps the model focus on features that are crucial for fire identification: for example, in smoldering fires, the average CO slope weight can reach 0.62, amplifying the discriminative information of its time sequence; in kitchen interference scenarios, the smoke slope weight is reduced to 0.07, effectively suppressing noise caused by steam interference; in open flame scenarios, the temperature slope weight is increased to 0.58, accurately capturing the feature of rapid fire spread.

[0095] The weighted feature time sequence will be used as the input to the optimized Bi-LSTM layer. Temporal feature extraction is performed in the optimized Bi-LSTM layer. This weighted sequence not only preserves the temporal continuity of the original sensor data, but also forms an optimized input adapted to the fire identification task. Through this process, the FA feature extraction layer further mines the contextual discrimination value of the features from the core features, dynamically adapts to the feature requirements of different fire stages and interference scenarios, lays the foundation for the LSTM layer to accurately capture temporal dependencies, and ultimately improves the robustness and recognition accuracy of the model.

[0096] The optimized Bi-LSTM layer (Bidirectional Long Short-Term Memory network) is used to perform the following operations:

[0097] 1) The bidirectional contextual dependencies of the weighted feature time series output by the FA feature processing layer are captured by two independent forward and backward hidden state flows, and then fused to form a comprehensive representation;

[0098] 2) The comprehensive representation is linearly transformed and mapped with activation functions through a fully connected layer to output the fire probability distribution of the thermal smoke detection model;

[0099] The optimized Bi-LSTM layer is an extension of the traditional LSTM model, designed to enhance the model's ability to capture contextual information by simultaneously modeling the forward and backward dependencies of sequence data. Its structure can be formally represented by the following two recursive processes:

[0100] ;

[0101] ;

[0102] in, and These represent the forward and backward LSTM at time [time]. The hidden state, and These are the input sequence components in the corresponding directions;

[0103] The states in two directions are usually merged through methods such as splicing to form a comprehensive representation that integrates bidirectional information:

[0104] ;

[0105] This combined output The dimension is It contains bidirectional contextual semantic information before and after the current moment. In regression tasks, Bi-LSTM uses this fused state to predict continuous values. The input is fed into a fully connected layer, and after linear transformation and mapping with an activation function, a predicted output is generated.

[0106] ;

[0107] in; Typically, it is a regression function with weighted parameters, which can be further expanded as:

[0108] ;

[0109] In the formula, For learnable weight matrix, For bias terms, For the model at time The prediction results;

[0110] Compared to unidirectional structures, Bi-LSTM significantly enhances the ability to model the overall dynamic trend of the sequence through bidirectional information flow, forming a feature representation from a full time-series perspective. The slow upward trend of the CO slope in the smoldering stage needs to be combined with the feature changes in subsequent time steps to confirm whether it is a real fire. Bidirectional modeling makes this cross-time step correlation capture more accurate.

[0111] The PSO-Adaboost flame image detection model is explained below:

[0112] The performance of the traditional Adaboost algorithm depends on the selection of weak classifiers and hyperparameter settings. It defaults to a greedy strategy to select weak classifiers, which is prone to getting trapped in local optima, resulting in insufficient generalization ability of the model. To address the above problems, this invention adopts the PSO-Adaboost flame image detection model. To address the problem that image features are easily affected by complex backgrounds and lighting interference, the Particle Swarm Optimization (PSO) algorithm is introduced to optimize the Adaboost ensemble model. By optimizing the weak classifier parameters and ensemble strategy, the visual recognition ability of flames and smoke in complex scenes is enhanced, solving the problem that the traditional Adaboost is prone to getting trapped in local optima, and outputting the probability of fire / no fire of the image detection model.

[0113] Among them, such as Figure 3 As shown, the training and usage methods of the PSO-Adaboost flame image detection model are as follows:

[0114] 1) Standardize the core features of the input image and randomly divide it into training and test sets;

[0115] 2) Configure the core parameters of the PSO algorithm (population size, maximum number of iterations, step size control coefficient, inertia weight, learning rate), and define the hyperparameter search space of the Adaboost algorithm (regularization coefficient C, kernel function parameter γ).

[0116] 3) Construct a fitness function to evaluate the performance of individual particles, iteratively update the historical best solution of the individual and the global best solution of the population, and dynamically adjust the particle search position and movement speed;

[0117] 4) Iterate until the convergence condition is met, output the optimized Adaboost parameter combination, and train to obtain the optimal classification model (the PSO-Adaboost flame image detection model).

[0118] 5) Input the core features of the image to be predicted, complete the inference, and output the fire probability distribution of the flame image detection model;

[0119] S4: Based on image entropy and sensor data variance, the detection scene is divided. The decision weights of the FA-Bi-LSTM hot smoke detection model and the PSO-Adaboost flame image detection model are dynamically allocated according to the detection scene, and the probability output results of the two models are weighted and fused to obtain the final fire detection result.

[0120] This step designs a dynamic weighted decision-level fusion strategy, integrating the advantages of the FA-Bi-LSTM hot smoke detection model and the PSO-Adaboost flame image detection model to further improve the reliability of fire detection in complex scenarios. The core idea is to dynamically adjust the weight ratio of the two models based on scene adaptability, so that the fusion result can accurately adapt to the recognition needs in different environments, ultimately achieving high accuracy, low false alarms, and low false alarms in fire detection. The specific fusion process is as follows:

[0121] S41: Real-time calculation of image entropy based on image grayscale distribution (reflecting image clarity and interference level; the lower the entropy value, the clearer the image and the more prominent the flame features); Real-time calculation of sensor data variance based on short sliding window (reflecting sensor data fluctuation stability and interference level; the larger the variance, the stronger the environmental interference and the lower the data reliability); Both serve as quantitative indicators for scene division.

[0122] S42: Delineate detection scenarios based on two types of indicators (image entropy and sensor data variance);

[0123] The detection scenarios include scenarios with clear image features, scenarios with disturbed image features, and scenarios with drastic fluctuations in sensor data.

[0124] The specific details for each scenario are as follows:

[0125] Scenes with clear image features: Image entropy value < image entropy threshold (e.g., 3.5) and no dense smoke obstruction (clear image features and high discrimination value);

[0126] Scenarios where image features are interfered with: Image entropy value ≥ image entropy threshold (e.g., 3.5) or there is dense smoke obscuring the image (image features are interfered with and have low effectiveness);

[0127] Scenarios with drastic fluctuations in sensor data: Sensor data variance > Sensor data variance threshold (e.g., 0.1) (Serious fluctuations in sensor data, both features are disturbed);

[0128] S43: For each detection scenario, assign differentiated weights to the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model, as follows:

[0129] Scenes with clear image features: Weights of the FA-Bi-LSTM thermal flue gas detection model <Weights of the PSO-Adaboost flame image detection model This highlights the decision-making dominance of image models, such as: =0.6, =0.4;

[0130] Scenarios where image features are disturbed: Weights of the FA-Bi-LSTM thermal flue gas detection model > Weights of the PSO-Adaboost flame image detection model To enhance the decision contribution of sensor models, such as: =0.3, =0.7;

[0131] Scenarios with drastic fluctuations in sensor data: Weights of the FA-Bi-LSTM hot flue gas detection model =Weights of the PSO-Adaboost flame image detection model Balancing the outputs of the two models reduces the risk of misjudgment. = =0.5;

[0132] S44: The probability outputs of the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model are integrated using a weighted summation method to obtain the fused combined probability of fire and the combined probability of no fire. The calculation formula is as follows:

[0133] ;

[0134] In the formula, The categories are represented by two types: those with fire and those without. The weights of the PSO-Adaboost flame image detection model are... This represents the probability output of the PSO-Adaboost flame image detection model. The weights of the FA-Bi-LSTM hot flue gas detection model are... This represents the probability output of the FA-Bi-LSTM hot flue gas detection model. ;

[0135] S45: Compare the fused combined probability of fire and the combined probability of no fire with the preset confidence threshold to obtain the final detection result;

[0136] If P(fire) ≥ confidence threshold (e.g., 0.65), then it is determined that there is a fire and a fire warning is triggered.

[0137] If P(no fire) ≥ confidence threshold (e.g., 0.65), then it is determined to be no fire.

[0138] If both probabilities are below 0.65, it indicates uncertainty in identification, triggering a secondary detection process. Steps S1-S4 are repeated, and the detection and judgment are re-completed based on the latest collected real-time data, forming a closed-loop verification mechanism.

[0139] The beneficial effects of the indoor fire detection method based on multi-source heterogeneous data fusion provided by this invention are as follows:

[0140] 1) This invention collects sensor parameter data and visible light-infrared light image data to construct a multi-source heterogeneous dataset, realizing the complementary advantages of sensor features and image features, and characterizing fire features from multiple dimensions such as vision, thermodynamics, and chemistry, covering the full stage features of fire from smoldering to open flame, and solving the problems of limited information and low discrimination of single-modal features.

[0141] 2) The RF and XGBoost algorithms are used to fuse scores and select core features, effectively removing redundant features, avoiding the "curse of dimensionality", reducing computational complexity, and improving the efficiency and feature discrimination of fire detection;

[0142] 3) The constructed FA-Bi-LSTM model introduces a feature attention mechanism to dynamically weight the core features of different fire stages. Combined with the bidirectional time series modeling capability of Bi-LSTM, it accurately captures the dynamic evolution of sensor data and improves the robustness of thermal smoke detection.

[0143] 4) The constructed PSO-Adaboost model optimizes the hyperparameters of Adaboost through the PSO algorithm, avoids local optima, enhances the visual recognition capability of flames and smoke in complex scenes, and balances recognition accuracy and real-time performance.

[0144] 5) The designed dynamic weight decision-level fusion strategy adaptively adjusts the model weights based on image entropy and sensor data variance, making the fusion results accurately adapt to different indoor detection scenarios, significantly reducing the false alarm rate and false alarm rate of fire detection, and improving the accuracy and reliability of fire detection in complex and variable indoor backgrounds.

[0145] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0146] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. An indoor fire detection method based on multi-source heterogeneous data fusion, characterized in that, Includes the following steps: Collect multi-source heterogeneous fire data to construct a visible light-infrared light image dataset and a temperature-smoke concentration-CO concentration sensor parameter dataset; Fire features were extracted from the visible-infrared image dataset and the temperature-smoke-CO concentration sensor parameter dataset, and core features were selected to form a core feature set; The core features are input into the pre-constructed FA-Bi-LSTM hot smoke detection model and PSO-Adaboost flame image detection model to obtain the fire probability distribution results of the two models. The input of the FA-Bi-LSTM hot smoke detection model is the core features of fire sensing parameters, and the input of the PSO-Adaboost flame image detection model is the core features of image type. The detection scenarios are divided based on image entropy and sensor data variance. The decision weights of the FA-Bi-LSTM hot smoke detection model and the PSO-Adaboost flame image detection model are dynamically allocated according to the detection scenarios, and the probability output results of the two models are weighted and fused to obtain the final fire detection result.

2. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The extracted fire features include five categories: static features of visible light flame images, dynamic features of visible light flame images, visible light smoke image features, infrared light image features, and fire sensing parameter features.

3. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method for selecting core features is as follows: For each fire feature, the importance score is calculated independently using the random forest algorithm and the gradient boosting tree algorithm. The scores from the two algorithms for a single feature are then combined and averaged to obtain the final importance score for each fire feature. All features are ranked by importance, and low-importance, redundant, and irrelevant features are removed based on the ranking results. The core features with the highest discriminative value among all types of features are retained to form a core feature set.

4. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The static core features of visible light flame images are the mean of the grayscale histogram, local binary mode, and color saturation; the dynamic core features of visible light flame images are the flame movement speed, the temporal variation of flame color, and the straightness of the edges; the core features of visible light smoke images are color features, texture features, and dynamic motion features; and the infrared light image features retain penetration characteristics, temperature features, and thermal diffusion features. The core characteristics of fire sensing parameters are carbon monoxide slope, temperature slope, and smoke slope.

5. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The FA-Bi-LSTM hot smoke detection model includes an FA feature processing layer and an optimized Bi-LSTM layer. The FA feature processing layer dynamically weights and enhances the core features of the fire sensing parameters, and the optimized Bi-LSTM layer captures the temporal feature dependencies in both directions, finally outputting the probability of fire / no fire from the FA-Bi-LSTM hot smoke detection model.

6. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The data processing procedure of the FA feature processing layer is as follows: The input sensing time-series data is compressed globally, and the feature global representation vector is extracted by global average pooling. The feature global representation vector is nonlinearly transformed by a single fully connected layer to obtain the feature similarity score. The feature similarity score is then normalized by the Softmax function to obtain the FA weight distribution. The original feature vector is weighted and adjusted using the FA weights to obtain a weighted feature time series.

7. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The data processing procedure for optimizing the Bi-LSTM layer is as follows: The bidirectional contextual dependencies of the weighted feature time series output by the FA feature processing layer are captured by two independent hidden state flows, forward and backward, and then fused to form a comprehensive representation. The comprehensive representation is linearly transformed and mapped with activation functions through a fully connected layer to output the fire probability distribution of the thermal smoke detection model.

8. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The training and usage methods of the PSO-Adaboost flame image detection model are as follows: The core features of the input image are standardized and preprocessed, and then the training set and test set are randomly divided. Configure the core parameters of the PSO algorithm and define the hyperparameter search space of the Adaboost algorithm; Construct a fitness function to evaluate the individual particle performance, iteratively update the individual historical best solution and the group global best solution, and dynamically adjust the particle search position and movement speed. Iterate until the convergence condition is met, output the optimized Adaboost parameter combination, and train the PSO-Adaboost flame image detection model. Input the core features of the image to be predicted, complete the inference, and output the fire probability distribution of the flame image detection model.

9. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The detection scenarios are divided based on image entropy and sensor data variance. The decision weights of the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model are dynamically allocated according to the detection scenarios, including the following steps: Real-time calculation of image entropy based on image grayscale distribution; real-time calculation of sensor data variance based on short sliding window for sensor data; The detection scenarios are divided based on the image entropy and sensor data variance, wherein the detection scenarios include scenarios with clear image features, scenarios with disturbed image features, and scenarios with drastic fluctuations in sensor data. Differentiated weights were assigned to the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model for different detection scenarios. Specifically, for scenarios with clear image features, the weights of the FA-Bi-LSTM hot flue gas detection model were adjusted accordingly. <Weights of the PSO-Adaboost flame image detection model For scenarios where image features are disturbed, the weights of the FA-Bi-LSTM thermal smoke detection model... > Weights of the PSO-Adaboost flame image detection model For scenarios with drastic fluctuations in sensor data, the weights of the FA-Bi-LSTM hot flue gas detection model... =Weights of the PSO-Adaboost flame image detection model .

10. The indoor fire detection method based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The probability output of the weighted fusion of the two models includes: The probability outputs of the FA-Bi-LSTM hot flue gas detection model and the PSO-Adaboost flame image detection model are integrated by weighted summation to obtain the fused comprehensive probability of fire and comprehensive probability of no fire. The combined probability of fire and the combined probability of no fire are compared with a preset confidence threshold to obtain the final detection result.

Images