Intelligent fire detection method and device based on dynamic space-time temperature analysis

Abstract

The invention relates to the field of fire detection, and provides an intelligent fire detection method and device based on dynamic space-time temperature analysis, and the method employs an image processing analysis technology based on artificial intelligence and deep learning to carry out the analysis of a thermal infrared image obtained through the real-time monitoring of an infrared thermal imager. According to the method, temperature distribution space features and time fusion information of a target area are captured, then a fire risk score value is obtained through analysis based on the captured temperature distribution information, and finally, the fire risk score value is compared with a preset threshold value to determine whether fire alarm is triggered or not. In this way, more effective and intelligent fire detection can be realized.

Description

Technical Field

[0001] The present application relates to the field of fire detection, and more specifically, to an intelligent fire detection method and device based on dynamic spatiotemporal temperature analysis. Background Art

[0002] With the acceleration of industrial IoT, early fire warning is becoming increasingly important in key scenarios such as power facilities and chemical parks. Traditional smoke and flame recognition methods based on visible light cameras suffer from limitations such as low sensitivity at night and susceptibility to failure under occlusion. However, temperature detection technology based on infrared thermal imaging cameras, due to its all-weather capability and temperature sensitivity, has gradually become a research hotspot in the field of intelligent fire prevention.

[0003] However, existing infrared thermal imaging fire warning systems face two core challenges: first, the high false alarm rate caused by background interference such as heat generated by equipment operation in industrial scenarios (such as normal temperature rise of transformers) and ambient temperature fluctuations (such as changes in sunlight); second, the traditional single-frame temperature threshold method has difficulty capturing the unique thermodynamic evolution laws of early fires, resulting in delayed response, which in turn affects fire detection and prevention.

[0004] Therefore, an intelligent fire detection solution based on dynamic spatiotemporal temperature analysis is desired. Summary of the Invention

[0005] In response to the shortcomings of the prior art, the present application provides an intelligent fire detection method and device based on dynamic spatiotemporal temperature analysis.

[0006] According to one aspect of the present application, there is provided an intelligent fire detection method based on dynamic spatiotemporal temperature analysis, which includes: A time sequence of thermal infrared images of a target area acquired by an infrared thermal imager is obtained; Preprocessing and data formatting are performed on each thermal infrared image in the time queue of thermal infrared images to obtain a time queue of thermal infrared temperature distribution matrix; Extracting temperature distribution spatial features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic spatial feature encoding vectors of the target area; Performing thermodynamic temporal dynamic coding on the time queue of the target area thermodynamic spatial feature coding vector to obtain the spatiotemporal fusion coding feature of the target area thermal distribution, including: performing causal correlation feature analysis on every two target area thermodynamic spatial feature coding vectors in the time queue of the target area thermodynamic spatial feature coding vector, and then performing temporal context dynamic fusion coding based on the thermodynamic causal correlation feature to obtain the spatiotemporal fusion coding feature of the target area thermal distribution; Fire detection is performed based on the spatiotemporal fusion coding features of the thermal distribution of the target area to determine a fire risk score value and determine whether to trigger a fire alarm.

[0007] According to another aspect of the present application, there is provided an intelligent fire detection device based on dynamic spatiotemporal temperature analysis, comprising: a thermal infrared image acquisition module for acquiring a time queue of thermal infrared images of a target area acquired by an infrared thermal imager; a data preliminary processing module, configured to pre-process and format the data of each thermal infrared image in the time queue of the thermal infrared images to obtain a time queue of the thermal infrared temperature distribution matrix; a temperature distribution space feature extraction module, configured to extract temperature distribution space features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic space feature encoding vectors of a target area; a thermodynamic temporal dynamic coding module, configured to perform thermodynamic temporal dynamic coding on the time queue of the target area thermodynamic spatial feature coding vectors to obtain a temporal and spatial fusion coding feature of the target area thermal distribution, wherein the thermodynamic temporal dynamic coding module is configured to: perform causal correlation feature analysis on every two target area thermodynamic spatial feature coding vectors in the time queue of the target area thermodynamic spatial feature coding vectors, and then perform temporal context dynamic fusion coding based on the thermodynamic causal correlation feature to obtain a temporal and spatial fusion coding feature of the target area thermal distribution; The fire risk assessment alarm module is used to perform fire detection based on the spatiotemporal fusion coding features of the thermal distribution of the target area to determine the fire risk score value and determine whether to trigger a fire alarm.

[0008] This application has significant technical effects due to the adoption of the above technical solutions: This application provides an intelligent fire detection method and device based on dynamic spatiotemporal temperature analysis. It uses image processing and analysis techniques based on artificial intelligence and deep learning to analyze thermal infrared images captured by a thermal imager in real time. This method captures the spatial characteristics of the temperature distribution in the target area and combines this temporal information with the captured temperature distribution information. This data is then used to generate a fire risk score, which is then compared with a preset threshold to determine whether to trigger a fire alarm. This allows for more efficient and intelligent fire detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 Flowchart of an intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application.

[0011] Figure 2 Flowchart of step S120 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application.

[0012] Figure 3 Flowchart of step S141 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application.

[0013] Figure 4 Flowchart of step S150 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application.

[0014] Figure 5 4 is a block diagram of an intelligent fire detection device based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0016] With the accelerating adoption of the Industrial Internet of Things (IIoT), early fire warning is becoming increasingly important in key applications such as power facilities and chemical parks. Traditional smoke and flame detection technology, which relies on visible light cameras, has limitations due to its lack of sensitivity at night and susceptibility to failure under obstruction. In contrast, temperature detection technology based on infrared thermal imaging cameras, with its 24 / 7 operation capability and high sensitivity to temperature changes, is becoming a research focus in the field of intelligent fire prevention.

[0017] However, current infrared thermal imaging fire warning systems still face two core challenges: on the one hand, background interference such as heat generated by normal equipment operation in industrial environments (such as the temperature rise of transformers) and fluctuations in ambient temperature (such as changes caused by direct sunlight) can easily lead to a high false alarm rate; on the other hand, traditional detection methods based on single-frame temperature thresholds have difficulty effectively capturing the unique thermodynamic evolution characteristics of early fires, resulting in delayed response, which has a certain impact on the timely detection and prevention of fires.

[0018] In response to the above technical problems, the technical solution of this application proposes an intelligent fire detection method based on dynamic spatiotemporal temperature analysis, which can use an infrared thermal imager to monitor and collect thermal infrared images of the target area in real time, and adopt image processing and analysis algorithms based on artificial intelligence and deep learning to analyze these thermal infrared images, so as to capture the spatial characteristics and temporal fusion information of the temperature distribution of the target area. In this way, while retaining the spatial heterogeneity of the temperature field, it can capture the exponential temperature rise characteristics that conform to the physical laws of fire development, avoid the interference of background temperature, and thus help to achieve more effective and intelligent fire detection to determine whether to trigger a fire alarm.

[0019] Figure 1 FIG is a flow chart of an intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. Figure 1 As shown, according to an embodiment of the present application, an intelligent fire detection method based on dynamic spatiotemporal temperature analysis includes: S110, obtaining a time queue of thermal infrared images of a target area collected by an infrared thermal imager; S120, preprocessing and data formatting each thermal infrared image in the time queue of the thermal infrared images to obtain a time queue of a thermal infrared temperature distribution matrix; S130, extracting temperature distribution spatial features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic spatial feature coding vectors of the target area; S140, performing thermodynamic time series dynamic coding on the time queue of the thermodynamic spatial feature coding vectors of the target area to obtain a spatiotemporal fusion coding feature of the thermal distribution of the target area; S150, performing fire detection based on the spatiotemporal fusion coding feature of the thermal distribution of the target area to determine a fire risk score value, and determine whether to trigger a fire alarm.

[0020] In step S110, a time series of thermal infrared images of the target area captured by the infrared thermal imager is acquired. It should be understood that the thermal infrared images of the target area primarily contain temperature information about the target area. Specifically, objects of different temperatures appear in different colors in a thermal infrared image. High-temperature objects, such as operating heat-generating electrical equipment and potential fire sources, appear as brighter areas or specific warm tones; while low-temperature objects, such as surrounding cool air and building structures at normal temperatures, appear as darker areas or cool tones. The distribution of this temperature information provides an intuitive understanding of the temperature differences within the target area. Accordingly, given that the occurrence of a fire is a dynamic process, early temperature changes are often gradual and regular. By acquiring a time series of thermal infrared images, the temperature changes of the target area over time can be observed, capturing dynamic information such as the temperature rise trend and rate, which helps to more accurately determine whether a fire has occurred and its current stage. It is important to note that in industrial scenarios, heat generation from normal equipment operation and natural fluctuations in ambient temperature may be difficult to distinguish from temperature changes caused by a fire in a single image frame. However, from a time series perspective, normal temperature fluctuations typically exhibit a certain periodicity or stability, while temperature changes caused by fires can exhibit unusual rising or fluctuating patterns. By analyzing the temporal sequence of thermal infrared images, we can exploit this temporal difference to better eliminate the interference of normal temperature fluctuations and accurately identify fire-related temperature changes.

[0021] In step S120, each thermal infrared image in the time queue of the thermal infrared images is preprocessed and data formatted to obtain a time queue of the thermal infrared temperature distribution matrix. Specifically, Figure 2 FIG. 1 is a flow chart of step S120 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. Figure 2 As shown, the step S120 includes: S121, preprocessing each thermal infrared image in the time queue of the thermal infrared image to obtain a time queue of preprocessed thermal infrared images, wherein the preprocessing includes Gaussian filtering and histogram equalization; S122, formatting data of each preprocessed thermal infrared image in the time queue of the preprocessed thermal infrared image to obtain a time queue of the thermal infrared temperature distribution matrix.

[0022] In step S121, each thermal infrared image in the time queue of thermal infrared images is preprocessed to obtain a time queue of preprocessed thermal infrared images. The preprocessing includes Gaussian filtering and histogram equalization. It should be understood that in practical applications, infrared thermal imagers are subject to various noise interferences, such as electronic noise and thermal noise from the infrared imager itself, interference with infrared radiation from air turbulence, dust, water vapor, and other factors, as well as vibration and electromagnetic interference from equipment operation in industrial scenarios. This noise can cause random temperature fluctuations or anomalies in thermal images, which may be mistakenly identified as potential fire signals. Gaussian filtering, as a linear smoothing filter, can effectively suppress high-frequency noise while preserving the primary temperature distribution features in the image. Its mathematical principle is to reduce the impact of random noise in thermal infrared images by weightedly averaging neighboring pixel values ​​using a convolution kernel. Furthermore, thermal infrared images captured by infrared imagers typically suffer from low contrast and limited dynamic range, making it difficult to distinguish temperature distribution details. Histogram equalization can enhance the global contrast and brightness of thermal infrared images by redistributing image grayscale levels, making temperature gradients more pronounced. Its mathematical principle is a nonlinear mapping of image grayscale levels. In other words, early fires often manifest as localized temperature increases. Histogram equalization can enhance the contrast in these areas, facilitating subsequent feature extraction. Furthermore, given the significant variations in ambient temperature across seasons and time periods, histogram equalization can adaptively adjust the image brightness distribution, improving the algorithm's robustness.

[0023] The following is a detailed description of a specific implementation process of “preprocessing each thermal infrared image in the time queue of thermal infrared images to obtain a time queue of preprocessed thermal infrared images”: Acquiring thermal infrared images is the starting point of the entire implementation process. When reading the target area's thermal infrared image timeline from the infrared camera's data storage, it's important to pay attention to the image format. Common formats such as TIFF and JPEG are acceptable. During the reading process, ensuring the integrity and accuracy of the image data is crucial; this is essential for all subsequent processing. The read images are then loaded into memory for further processing.

[0024] Next, the read image needs to be Gaussian filtered. Gaussian filtering is a linear smoothing filter whose core lies in the Gaussian kernel. In practical applications, the size and standard deviation of the Gaussian kernel must be carefully selected based on the specific characteristics of the thermal infrared image. Kernel sizes are generally odd numbers, such as 3×3, 5×5, and 7×7. Smaller kernels require less computation, but their smoothing effect is limited. Larger kernels, while able to smooth a wider area, also require significantly more computation. The standard deviation determines the shape of the Gaussian function's distribution. A larger standard deviation results in a smoother filtered image, but may lose some image detail. Therefore, determining the most appropriate standard deviation typically requires multiple trials, taking into account the image's noise level and the desired smoothing effect. Once the parameters are determined, Gaussian filtering is performed pixel by pixel on each thermal infrared image loaded into memory. The Gaussian kernel slides across the image. For each pixel, the pixel values ​​within the kernel's coverage area are weighted averaged according to the Gaussian function's weights to produce the new filtered pixel value. For example, in a 5×5 filter kernel, the center pixel has the highest weight, while pixels farther from the center have smaller weights. By performing this operation on every pixel in the image, the entire image is Gaussian filtered, effectively suppressing noise and making the image smoother, providing a clearer image foundation for subsequent analysis.

[0025] Next, the Gaussian filtered image needs to be subjected to histogram equalization. The first task of histogram equalization is to calculate the image's histogram. By counting the frequency of each grayscale level in the image, a grayscale histogram is generated. This intuitively displays the distribution of different grayscale values ​​within the image, providing insights into the image's contrast and brightness. If the image's grayscale values ​​are concentrated within a small range, the image's contrast is low, resulting in a blurry appearance. Next, the cumulative distribution function (CDF) is calculated based on the generated grayscale histogram. The CDF reflects the proportion of pixels in the image with grayscale values ​​less than or equal to the current grayscale level and plays a key mapping role in the histogram equalization process. Finally, the CDF is used to map the grayscale values ​​of each pixel in the original image. For example, if the grayscale value of a pixel in the original image is x, its corresponding CDF value is CDF(x). For an 8-bit grayscale image (with a total number of grayscale levels L = 256), the grayscale value of that pixel will be updated to CDF(x) × (L - 1).

[0026] After Gaussian filtering and histogram equalization, the resulting preprocessed thermal infrared image time queue is stored in its original order. The storage format can be selected based on the specific needs of subsequent data processing and analysis, but the original image format is generally used. These preprocessed images provide high-quality data support for subsequent data analysis.

[0027] In step S122, data formatting is performed on each preprocessed thermal infrared image in the time queue of preprocessed thermal infrared images to obtain a time queue of the thermal infrared temperature distribution matrix. It should be understood that thermal images captured by infrared cameras are essentially two-dimensional temperature distribution maps, with each pixel corresponding to a temperature value. However, raw thermal image data is typically stored in image formats (such as RGB or grayscale), making it unsuitable for direct use in temperature analysis and fire detection. Therefore, in the technical solution of the present application, data formatting is further performed on each preprocessed thermal infrared image in the time queue of preprocessed thermal infrared images to obtain a time queue of the thermal infrared temperature distribution matrix. Specifically, the purpose of data formatting is to convert the image data into a structured temperature distribution matrix, in which each element directly represents the temperature value at a corresponding location in the target area. This resulting thermal infrared temperature distribution matrix directly reflects the temperature field distribution of the target area at a specific point in time, facilitating the extraction of spatial features such as local high-temperature areas and temperature gradients. After converting multiple frames of thermal images into a time queue of the thermal infrared temperature distribution matrix, temperature variation trends at the same location at different points in time can be analyzed to capture the temperature rise patterns unique to fires, providing a basis for subsequent fire detection and alarming.

[0028] In step S130, temperature distribution spatial features are extracted from each thermal infrared temperature distribution matrix in the temporal queue of the thermal infrared temperature distribution matrices to obtain a temporal queue of target region thermodynamic spatial feature encoding vectors. Specifically, in an embodiment of the present application, step S130 includes: passing each thermal infrared temperature distribution matrix in the temporal queue of the thermal infrared temperature distribution matrices through a temperature distribution spatial feature extractor based on a dilated convolutional neural network model to obtain a temporal queue of target region thermodynamic spatial feature encoding vectors. It should be understood that the early stages of a fire typically manifest as a localized temperature increase, but the affected area may gradually expand. Traditional convolution operations (such as pooling) expand the receptive field while reducing spatial resolution, resulting in loss of temperature distribution details. However, dilated convolution, through staggered sampling, can expand the receptive field while maintaining spatial resolution, thereby preserving the fine structure of the temperature field. This is particularly important for fire detection, as early fires may only manifest as temperature anomalies within a small area. In other words, by increasing the receptive field, dilated convolution can capture temperature distribution features over a wider range without increasing computational complexity. In this way, not only the local high-temperature area in the target area can be captured, but also the temperature gradient changes in a larger range in the area can be captured.

[0029] In step S140, the time queue of the target area thermodynamic space feature coding vector is subjected to thermodynamic time series dynamic coding to obtain the target area thermal distribution spatiotemporal fusion coding feature. Specifically, in the embodiment of the present application, the step S140 includes: S141, after performing causal correlation feature analysis on every two target area thermodynamic space feature coding vectors in the time queue of the target area thermodynamic space feature coding vector, the time series context dynamic fusion coding is performed based on the thermodynamic causal correlation feature to obtain the target area thermal distribution spatiotemporal fusion coding feature. More specifically, Figure 3 FIG. 1 is a flow chart of step S141 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. Figure 3 As shown, the step S141 includes: S1411, performing deep implicit feature extraction on each target area thermodynamic space feature coding vector in the time queue of the target area thermodynamic space feature coding vector to obtain a set of target area temperature distribution space deep implicit feature coding vectors; S1412, constructing the causal association topological features of the set of target area temperature distribution space deep implicit feature coding vectors to obtain the target area thermal field semantic causal topological feature matrix as the thermodynamic causal association feature; S1413, using the target area thermal field semantic causal topological feature matrix as the semantic causal association structure information, performing context dynamic propagation fusion on the time queue of the target area thermodynamic space feature coding vector and the set of the target area temperature distribution space deep implicit feature coding vector to obtain the target area thermal distribution spatiotemporal fusion coding vector as the target area thermal distribution spatiotemporal fusion coding feature.

[0030] It should be understood that since fire development is a dynamic process, its temperature changes exhibit distinct temporal characteristics. Specifically, in the early stages, local temperatures rise slowly, potentially manifesting as a subtle upward trend; in the development phase, temperatures rise rapidly, and the heat diffusion range gradually expands; and in the outbreak phase, temperatures rise dramatically, resulting in significant thermal radiation and convection. Relying solely on spatial features (such as single-frame temperature distribution) is difficult to capture these dynamic changes. Therefore, to capture the dynamic patterns of temperature changes in the target area from a temporal perspective and combine spatial features to achieve space-time fusion coding, the technical solution of this application further performs thermodynamic temporal dynamic coding on the temporal sequence of the target area's thermodynamic spatial feature encoding vectors to obtain a spatiotemporal fusion encoding vector for the target area's thermal distribution. In other words, temperature changes in a fire are driven not only by spatial distribution but also by temporal evolution. For example, heat diffusion in a localized high-temperature area will affect surrounding areas over time. Therefore, thermodynamic temporal dynamic coding can explicitly model the causal relationships and temporal dependencies of temperature spatial distribution changes, enabling more accurate identification of fire precursors and providing a more comprehensive feature representation for fire detection. Specifically, the method of thermodynamic temporal dynamic encoding of the time sequence of the target area's thermodynamic spatial feature encoding vectors can explicitly model the causal relationship between the target area's temperature spatial distribution characteristics in the time dimension through causal inference theory. At the same time, it can combine surface and latent semantic encoding to capture the explicit patterns and implicit laws of temperature changes, thereby avoiding misjudgments caused by relying solely on correlation. This process can extract a multi-level, dynamic semantic representation of the target area's temperature distribution space in the time dimension from the time sequence of the target area's thermodynamic spatial feature encoding vectors, not only improving the accuracy of fire detection, but also providing reliable technical support for fire early warning in complex industrial scenarios.

[0031] Specifically, deep implicit feature extraction is first performed on each target region thermodynamic spatial feature encoding vector in the time queue of the target region thermodynamic spatial feature encoding vector to obtain a set of target region temperature distribution spatial deep implicit feature encoding vectors. The above process can be expressed as: ; ; ;

[0032] in, A time sequence of thermodynamic spatial feature encoding vectors for the target region, , , and are the first, second, and third time series of the target region thermodynamic space feature encoding vector. and Thermodynamic spatial feature encoding vector of the target region, is point convolutional coding, is the convolution activation function, , , , and They are the first, second, and third in the set of spatial depth implicit feature encoding vectors of the temperature distribution of the target area. , and The spatial depth implicit feature encoding vector of the temperature distribution of the target area, It is a set of spatial depth implicit feature encoding vectors of the temperature distribution in the target area.

[0033] It should be understood that deep latent feature extraction can extract deep semantic features from the temporal sequence of the target region's thermodynamic spatial feature encoding vectors. These deep semantic features contain richer and more accurate information about the target region's temperature distribution. For example, in a fire detection scenario, the original feature vector may simply reflect the surface distribution of temperature, but after deep latent feature extraction, it can deeply reveal the potential connection between temperature changes and the likelihood of fire. For example, abnormal patterns of temperature changes in specific areas may not be obvious in the original features, but they are crucial for fire detection. Moreover, the data processing method of deep latent feature extraction can also identify and eliminate redundant information caused by noise, preventing it from misleading fire judgments.

[0034] Specifically, in an embodiment of the present application, the step S1412 includes: calculating the semantic association measurement factor between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of the target area temperature distribution spatial depth implicit feature coding vectors to obtain a target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors; inputting the target area thermal field semantic causal topology matrix into a gated causal triggering network based on a gated activation function to obtain the target area thermal field semantic causal topology feature matrix.

[0035] More specifically, in an embodiment of the present application, a semantic association measurement factor between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of target area temperature distribution spatial depth implicit feature coding vectors is calculated to obtain a target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors, including: calculating an association matrix between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of target area temperature distribution spatial depth implicit feature coding vectors to obtain a set of target area temperature distribution association matrices; calculating the semantic association measurement factor of each target area temperature distribution association matrix in the set of target area temperature distribution association matrices to obtain a target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors. The target area thermal field semantic causal topology matrix is ​​composed of the temperature distribution semantic association measurement factor, and the target area temperature distribution semantic association measurement factor is calculated by the mean, variance, maximum value and association compensation coefficient of the target area temperature distribution association matrix; wherein, in response to the variance of the target area temperature distribution association matrix being greater than or equal to a predetermined threshold, the weighted mean of the distance between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of the target area temperature distribution spatial depth implicit feature coding vectors is used as the association compensation coefficient; in response to the variance of the target area temperature distribution association matrix being less than the predetermined threshold, the weighted value of the variance of the target area temperature distribution association matrix is ​​used as the association compensation coefficient. The above process can be expressed by the formula: ; ; ; in, , are the first and second implicit feature encoding vectors of the target area temperature distribution space depth. , The spatial depth implicit feature encoding vector of the temperature distribution of the target area, is matrix multiplication, yes The transposed vector of yes and The target area temperature distribution correlation matrix between yes The variance of To obtain The maximum value in yes The mean of is the correlation compensation coefficient, yes The corresponding semantic association measurement factor of the target area temperature distribution, for and The distance between yes The number of vectors in is a predetermined threshold, and is the weighted optimization coefficient, , , and are the semantic correlation measurement factors of the target area temperature distribution at each position in the target area thermal field semantic causal topology matrix, is the semantic causal topology matrix of the thermal field of the target area.

[0036] It is understandable that changes in temperature distribution in fire scenarios involve complex causal relationships. By calculating semantic association metrics and constructing a semantic causal topology matrix for the target region's thermal field, we can deeply explore the hidden causal relationships between the deep implicit feature encoding vectors of the target region's temperature distribution space. For example, in a large industrial warehouse, a temperature rise in one area may be due to heat conduction from equipment failure in an adjacent area. This potential causal relationship is not easily apparent in explicit features, but can be clearly presented through causal association analysis of implicit features, helping the model accurately grasp the underlying mechanisms of fire occurrence and development. Specifically, by using a causal association energy metric function to explicitly quantify and encode causal relationships, we obtain a specific semantic association metric for the target region's temperature distribution. This eliminates the fuzzy concept of causal relationships between the deep implicit feature encoding vectors of the temperature distribution space of different target regions and provides a precise numerical representation. This provides a unified standard for causal assessment, enabling the use of this metric to evaluate causal relationships between temperature distribution features across different industrial scenarios or target regions. This enhances the comparability and credibility of assessment results, facilitating unified fire risk management and analysis across diverse scenarios.

[0037] Here, by treating the low-level semantic causal associations of target region temperature distribution in complex systems as molecular-level relationships inferred from statistical correlations, we can further predict the intervention energy of the target region temperature distribution semantic causal association based on a global, fine-grained statistical association representation. This allows us to study the fine-grained structure of causal associations and their dynamic regulation based on the high-dimensional and heterogeneous omics-based representation of the target region temperature distribution semantic causal associations. When the aggregate distribution representation of the causal graph exceeds a predetermined threshold, source data integration is biased based on a matrixed graph node effect representation of the target region temperature distribution semantic association metric. When the aggregate distribution representation of the causal graph is below the predetermined threshold, agglomerative structural modeling can be performed directly through feature pattern integration and compression. This approach not only encodes the semantic causal energy of the target region temperature distribution in the system but also condenses the implicit target region temperature distribution semantic causal intervention prediction results, thereby more efficiently revealing key causal associations.

[0038] Then, the target area thermal field semantic causal topology matrix is ​​input into the gated causal trigger network based on the gated activation function to obtain the target area thermal field semantic causal topology feature matrix. The above process can be expressed as follows: ; in, yes The corresponding semantic association measurement factor of the target area temperature distribution, is a nonlinear activation function, is the normalized threshold, For Perform gated activation processing, It is the semantic causal topological feature matrix of the thermal field of the target area.

[0039] It should be understood that the gated causal triggering network, based on a gated activation function, can dynamically model the semantic causal topological matrix of the target region's thermal field, enabling more detailed feature extraction of the semantic causal topological relationships of the temperature distribution. Specifically, the causal relationships between temperature changes in different regions are complex, and a simple topological matrix only presents basic causal information. The gated causal triggering network, however, can delve deeper into more subtle and critical features, identifying potential indirect and subtle causal connections, thereby providing a more comprehensive and in-depth basis for fire prediction and prevention. It should be noted that the core of the gated causal triggering network lies in the dynamic gating mechanism and nonlinear activation function. Specifically, the dynamic gating mechanism enables the model to identify key causal paths in a dynamic context. During the development of a fire, not all causal relationships play an equally important role in its evolution. The gating mechanism can screen and strengthen those causal paths that are crucial to fire development. For example, when an abnormally high temperature in a particular area is a key starting point for a fire and has close causal connections with multiple other areas, the gating mechanism highlights this critical path, allowing the model to focus more closely on these important connections, significantly improving its ability to detect early signs of a fire. Nonlinear activation functions in the network can significantly enhance the model's expressive power. Complex fire causal structures contain many high-order, nonlinear regularities that are difficult to describe using simple linear models. The introduction of activation functions allows the model to capture these high-order features hidden within these complex causal structures. For example, during the development of a fire, the rate of temperature change and the coordination of temperature changes across different areas may exhibit nonlinear trends. Activation functions can help the model better fit these trends, leading to more accurate predictions of fire development and providing reliable support for subsequent fire warning systems to issue alerts more promptly and accurately.

[0040] Specifically, in the embodiment of the present application, step S1413 includes: inputting the target area thermal field semantic causal topological feature matrix and the target area thermodynamic space feature encoding vector into the time queue of the feature sequence dynamic walk encoder based on the graph convolutional neural network model to obtain the target area temperature distribution surface context dynamic propagation semantic encoding vector. This process can be expressed by the formula: ;

[0041] in, is the first in the time queue of the target area thermodynamic space feature encoding vector Thermodynamic spatial feature encoding vector of the target region, is the semantic causal topological feature matrix of the thermal field of the target area, It is graph convolution processing, It is the semantic encoding vector of the dynamic propagation of the surface context of the temperature distribution in the target area; The target area thermal field semantic causal topological feature matrix and the target area temperature distribution spatial depth implicit feature encoding vector are input into the feature sequence dynamic walk encoder based on the graph convolutional neural network model to obtain the target area temperature distribution hidden layer context dynamic propagation semantic encoding vector. This process can be expressed as follows: ;

[0042] in, is the first in the set of spatial depth implicit feature encoding vectors of the temperature distribution of the target area. The spatial depth implicit feature encoding vector of the temperature distribution of the target area, is the semantic causal topological feature matrix of the thermal field of the target area, It is graph convolution processing, is the target area temperature distribution hidden layer context dynamic propagation semantic encoding vector; The target area temperature distribution hidden layer context dynamic propagation semantic coding vector and the target area temperature distribution surface layer context dynamic propagation semantic coding vector are fused to obtain the target area thermal distribution spatiotemporal fusion coding vector. This process can be expressed as follows: ;

[0043] in, is the target area temperature distribution surface context dynamic propagation semantic encoding vector, is the target area temperature distribution hidden layer context dynamic propagation semantic encoding vector, is the fusion weight parameter, It is the spatiotemporal fusion coding vector of the thermal distribution of the target area.

[0044] It should be understood that the target region's thermal field semantic causal topological feature matrix contains causal information about the temperature distribution, while the temporal sequence of the target region's thermodynamic spatial feature encoding vector reflects the spatial and temporal variations in temperature. Inputting these two into the feature sequence dynamic walk encoder strengthens the causal and spatial connections between features. For example, during a fire, a temperature rise in one area may causally affect the temperatures of adjacent areas, while also exhibiting a certain temporal trend. The dynamic walk encoder can integrate these causal relationships with spatial and temporal variation information, enabling the model to more accurately understand the underlying mechanisms of temperature variation, thereby improving fire detection accuracy.

[0045] Accordingly, considering that the latent semantics of the target area temperature distribution contain richer and deeper information, surface information alone cannot fully capture the complex patterns and inherent laws of temperature changes. In order to mine deep semantic information, in this application, it is necessary to input the target area thermal field semantic causal topological feature matrix and the set of deep implicit feature encoding vectors of the target area temperature distribution space into the feature sequence dynamic walk encoder based on the graph convolutional neural network model for processing. Specifically, the dynamic walk mechanism can explore the implicit features of each node along the topological structure of the graph in the latent space, and discover deep semantic information that is not easy to be directly observed, which helps the model to understand the fire development more accurately. Since the latent semantics enable the potential embedding of features, it is necessary to focus on the multi-hop propagation of high-order information and the distributed decoupling of deep features during the walk process. This means that the model can capture the complex changes and interactions of features after multi-step propagation, separate and integrate deep features at different levels and aspects, so that the obtained target area temperature distribution hidden context dynamic propagation semantic encoding vector can more comprehensively and accurately reflect the intrinsic characteristics of the temperature distribution. This latent representation captures profound temporal dependencies and complex semantic patterns, providing greater generalization for feature expression. This allows the model to better adapt to a wide range of fire scenarios and temperature variations, beyond the specific conditions in the training data. It can also more accurately analyze and predict temperature distributions in new, unseen fire scenarios, providing more reliable support for fire detection and early warning.

[0046] It should be understood that the dynamic propagation semantic encoding vector of the target area's surface context temperature distribution primarily captures the explicit semantic information of the temperature distribution, such as direct temperature changes at various locations in the target area and obvious causal relationships. In contrast, the dynamic propagation semantic encoding vector of the target area's hidden context temperature distribution further explores the implicit semantic information of the temperature distribution, such as underlying temperature trends and complex causal transmission mechanisms. By fusing these two encoding vectors, explicit and implicit semantic information can be integrated to comprehensively capture the various characteristics of the target area's temperature distribution, thereby helping the model more accurately predict fire development. Overall, through appropriate fusion and weight adjustment, the spatiotemporal fusion encoding vector of the target area's thermal distribution can more accurately represent the target area's temperature distribution. It can more precisely reflect the spatial and temporal variations in temperature and the complex causal relationships underlying these variations. For fire detection, such an accurate representation facilitates more precise fire risk assessment and reduces false alarms and missed alerts. For example, in a complex chemical park environment, the fused encoding vector can more accurately distinguish between normal process temperature fluctuations and abnormal temperature changes in the early stages of a fire, improving the reliability of fire warnings.

[0047] In step S150, fire detection is performed based on the spatiotemporal fusion coding features of the target area thermal distribution to determine the fire risk score value and determine whether to trigger a fire alarm. Specifically, Figure 4 FIG. 1 is a flow chart of step S150 in the intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. Figure 4 As shown, the step S150 includes: S151, inputting the spatiotemporal fusion coding vector of the thermal distribution of the target area into a decoder-based fire detector to obtain a detection result, and the detection result is a fire risk score value; S152, determining whether to trigger a fire alarm based on a comparison between the detection result and a preset threshold.

[0048] In step S151, the spatiotemporal fusion coding vector of the thermal distribution of the target area is input into a fire detector based on a decoder to obtain a detection result, which is a fire risk score value. That is, decoding regression is performed using the temporal fusion feature information of the spatial characteristics of the temperature distribution of the target area in the time dimension to detect the risk score value of the fire, and it is used as a direct basis for judging whether it is necessary to trigger a fire alarm. In particular, in a specific embodiment of the present application, the spatiotemporal fusion coding vector of the thermal distribution of the target area is input into a fire detector based on a decoder to obtain a detection result, which is a fire risk score value, including: multiplying the decoding weight matrix of the decoder with the spatiotemporal fusion coding vector of the thermal distribution of the target area to obtain the decoded spatiotemporal fusion coding vector of the thermal distribution of the target area, and accumulating and summing all the eigenvalues ​​of the decoded spatiotemporal fusion coding vector of the thermal distribution of the target area to obtain the fire risk score value.

[0049] In the technical solution of the present application, each target area thermodynamic space feature coding vector in the time queue of the target area thermodynamic space feature coding vector respectively represents the spatial characteristics of the target area temperature distribution. When performing temporal context coding of the temperature space distribution, the dimensional orthogonality of the temperature space distribution and the temperature time distribution will cause the attention to be sparse during the temporal context coding, thereby resulting in the presence of fine-grained feature expression discretization in the obtained target area thermal distribution spatiotemporal fusion coding vector, which affects the accuracy of the detection results obtained by the decoder-based fire detector input.

[0050] Based on this, in one technical solution of the present application, in the process of inputting the spatiotemporal fusion coding vector of the target area thermal distribution into a decoder-based fire detector to obtain a detection result, the spatiotemporal fusion coding vector of the target area thermal distribution is dynamically quantized and controlled, and the process includes: The target area thermal distribution spatiotemporal fusion coding microscopic self-organization representation matrix of the target area thermal distribution spatiotemporal fusion coding vector is calculated and expressed as: ;

[0051] in, Represents the spatiotemporal fusion coding vector of the thermal distribution of the target area, represents the transpose of a vector, represents matrix multiplication, Represents the spatiotemporal fusion encoding microscopic self-organization representation matrix of the thermal distribution of the target area.

[0052] Based on the attribute association information between each position in the target area thermal distribution spatiotemporal fusion coding vector, the first cross-step association weight matrix of the target area thermal distribution spatiotemporal fusion coding and the second cross-step association weight matrix of the target area thermal distribution spatiotemporal fusion coding are constructed, which are expressed as: ;

[0053] ;

[0054] in, and They represent the first and The eigenvalues ​​at the positions, 、 、 、 Represent the first learnable parameter, the second learnable parameter, the third learnable parameter, and the fourth learnable parameter, respectively. Represents the first-step correlation weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area The value of the position, The second cross-step correlation weight matrix representing the spatiotemporal fusion coding of the thermal distribution of the target area The value of the position; Based on the first cross-step correlation weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area and the second cross-step correlation weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area, a cross-scale topological weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area is constructed; ;

[0055] in, Represents the first cross-step correlation weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area, Represents the second cross-step correlation weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area, and denote the fifth and sixth learnable parameters respectively, Indicates point multiplication by position, It means adding by position. Represents the target area's thermal distribution spatial and temporal fusion encoding cross-scale topological weight matrix.

[0056] Based on the cross-scale topological weight matrix of the spatiotemporal fusion coding of the target area thermal distribution, the spatiotemporal fusion coding vector of the target area thermal distribution is mapped and reconstructed to obtain the multi-scale association configuration vector of the spatiotemporal fusion coding of the target area thermal distribution, which is expressed as: ;

[0057] in, represents the piecewise linear activation function, represents the length of the spatiotemporal fusion coding vector of the thermal distribution of the target area, Represents the spatiotemporal fusion encoding multi-scale correlation configuration vector of the thermal distribution of the target area.

[0058] Based on the cross-scale topological weight matrix of the spatiotemporal fusion coding of the thermal distribution of the target area, the microscopic self-organizing representation matrix of the spatiotemporal fusion coding of the thermal distribution of the target area is mapped and reconstructed to obtain the spatiotemporal fusion coding aggregation cluster topological optimization matrix of the thermal distribution of the target area, which is expressed as: ;

[0059] in, represents the sigmoid activation function, The characteristic scale of the cross-scale topological weight matrix representing the spatiotemporal fusion coding of the thermal distribution of the target area, Represents the spatiotemporal fusion coding cluster topology optimization matrix of the target area thermal distribution.

[0060] Based on the target area thermal distribution spatiotemporal fusion coding aggregation cluster topology optimization matrix, the target area thermal distribution spatiotemporal fusion coding multi-scale association configuration vector is subjected to discrete feature dynamic coupling to obtain the target area thermal distribution spatiotemporal fusion coding compensation coding mechanism vector, which is expressed as: ;

[0061] in, Represents the spatiotemporal fusion coding compensation coding mechanism vector of the thermal distribution of the target area.

[0062] The target area thermal distribution spatiotemporal fusion coding compensation coding mechanism vector and the target area thermal distribution spatiotemporal fusion coding vector are fused to obtain an optimized target area thermal distribution spatiotemporal fusion coding vector, which is expressed as: ;

[0063] in, and denote the seventh and eighth learnable parameters respectively, Represents the spatiotemporal fusion encoding vector of the optimized target region thermal distribution. Those skilled in the art will appreciate that the aforementioned learnable parameters can be set using existing automatic parameter tuning tools, such as AutoML tools or libraries (e.g., Optuna, Hyperopt, etc.).

[0064] In this way, the spatiotemporal fusion coding vector of the target area's thermal distribution is input into a dynamic coding framework to construct the first-level and second-level association weight matrices, respectively, and further generate a feature expression with a cross-level association topology. On this basis, the original spatiotemporal fusion coding vector of the target area's thermal distribution and its self-organizing representation are subjected to a mapping transformation of heterogeneous node group interaction links to simulate the primitive state excitation paradigm triggered by a discontinuous association feedback mechanism. At the same time, a feature adaptation operator is introduced to dynamically adjust the discretization characteristics of the aggregate cluster association. Utilizing the diverse association mapping features of the cross-level topology, a compensation strategy for the association attenuation gradient at a multi-dimensional scale is established, thereby avoiding the impact of the weakening of association strength caused by the discretization of features on the feature decoding mapping inference performance. In this way, the accuracy of the detection results obtained by the decoder-based fire detector is improved.

[0065] In step S152, based on the comparison between the detection result and the preset threshold, it is determined whether to trigger a fire alarm. It should be understood that in the process of fire detection, the fire risk score obtained through the previous processing is only a quantitative value, which itself cannot directly determine whether to issue a fire alarm signal. Comparing the detection result with the preset threshold can provide the system with a clear and operational decision-making basis. Only when the fire risk score reaches or exceeds the preset threshold is it considered that the possibility of a fire has reached the level that requires an alarm, which can effectively avoid decision-making errors caused by fuzzy standards. Specifically, when the fire risk score exceeds the preset threshold, the system can quickly trigger a fire alarm and promptly notify relevant personnel and fire departments so that they can take measures to extinguish the fire and rescue as soon as possible, thereby minimizing the losses caused by the fire. It should be noted that different areas or equipment can set different preset thresholds according to their fire risk levels, so as to achieve more personalized fire monitoring and alarms.

[0066] In summary, the intelligent fire detection method based on dynamic spatiotemporal temperature analysis, as described in the embodiments of this application, employs image processing and analysis techniques based on artificial intelligence and deep learning to analyze thermal infrared images acquired by a thermal imager in real time. This method captures the spatial characteristics of the temperature distribution in the target area and the temporal fusion information. This captured temperature distribution information is then analyzed to determine a fire risk score. Finally, this fire risk score is compared with a preset threshold to determine whether to trigger a fire alarm. This allows for more efficient and intelligent fire detection.

[0067] Figure 5 FIG is a block diagram of an intelligent fire detection device based on dynamic spatiotemporal temperature analysis according to an embodiment of the present application. Figure 5 As shown, the intelligent fire detection device 100 based on dynamic spatiotemporal temperature analysis according to the embodiment of the present application includes: a thermal infrared image acquisition module 110, which is used to obtain a time queue of thermal infrared images of a target area acquired by an infrared thermal imager; a data preliminary processing module 120, which is used to preprocess and format the data of each thermal infrared image in the time queue of the thermal infrared image to obtain a time queue of a thermal infrared temperature distribution matrix; a temperature distribution spatial feature extraction module 130, which is used to extract temperature distribution spatial features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic spatial feature coding vectors of the target area; a thermodynamic time series dynamic coding module 140, which is used to perform thermodynamic time series dynamic coding on the time queue of the thermodynamic spatial feature coding vectors of the target area to obtain a spatiotemporal fusion coding feature of the thermal distribution of the target area; a fire risk assessment alarm module 150, which is used to perform fire detection based on the spatiotemporal fusion coding feature of the thermal distribution of the target area to determine a fire risk score value, and determine whether to trigger a fire alarm. Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the intelligent fire detection device 100 based on dynamic spatiotemporal temperature analysis have been described above with reference to FIG. Figures 1 to 4 The description of the intelligent fire detection method based on dynamic spatiotemporal temperature analysis has been introduced in detail, and therefore, its repeated description will be omitted.

[0068] In summary, the intelligent fire detection device 100 based on dynamic spatiotemporal temperature analysis according to the embodiment of the present application is described. It uses image processing and analysis techniques based on artificial intelligence and deep learning to analyze thermal infrared images acquired by an infrared thermal imager in real time. This allows the device to capture the spatial characteristics of the temperature distribution and temporal fusion information of the target area. This temperature distribution information is then analyzed to determine a fire risk score. Finally, the fire risk score is compared with a preset threshold to determine whether to trigger a fire alarm. This allows for more efficient and intelligent fire detection.

Claims

1. An intelligent fire detection method based on dynamic spatiotemporal temperature analysis, characterized in that: include: A time sequence of thermal infrared images of a target area acquired by an infrared thermal imager is obtained; Preprocessing and data formatting are performed on each thermal infrared image in the time queue of thermal infrared images to obtain a time queue of thermal infrared temperature distribution matrix; Extracting temperature distribution spatial features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic spatial feature encoding vectors of the target area; Performing thermodynamic temporal dynamic coding on the time queue of the target area thermodynamic spatial feature coding vector to obtain the spatiotemporal fusion coding feature of the target area thermal distribution, including: performing causal correlation feature analysis on every two target area thermodynamic spatial feature coding vectors in the time queue of the target area thermodynamic spatial feature coding vector, and then performing temporal context dynamic fusion coding based on the thermodynamic causal correlation feature to obtain the spatiotemporal fusion coding feature of the target area thermal distribution; Fire detection is performed based on the spatiotemporal fusion coding features of the thermal distribution of the target area to determine a fire risk score value and determine whether to trigger a fire alarm.

2. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 1, characterized in that: Preprocessing and data formatting are performed on each thermal infrared image in the time queue of the thermal infrared images to obtain a time queue of a thermal infrared temperature distribution matrix, including: Preprocessing each thermal infrared image in the time queue of thermal infrared images to obtain a time queue of preprocessed thermal infrared images, wherein the preprocessing includes Gaussian filtering and histogram equalization; Data formatting is performed on each preprocessed thermal infrared image in the time queue of the preprocessed thermal infrared images to obtain the time queue of the thermal infrared temperature distribution matrix.

3. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 2, characterized in that: Extracting temperature distribution spatial features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of the thermodynamic spatial feature coding vector of the target area, including: passing each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix through a temperature distribution spatial feature extractor based on a void convolutional neural network model to obtain a time queue of the thermodynamic spatial feature coding vector of the target area.

4. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 3 is characterized in that: After performing causal correlation feature analysis on every two target area thermodynamic spatial feature coding vectors in the time queue of the target area thermodynamic spatial feature coding vector, dynamic fusion coding of temporal context is performed based on the thermodynamic causal correlation feature to obtain the spatiotemporal fusion coding feature of the target area thermal distribution, including: Performing deep implicit feature extraction on each target region thermodynamic space feature coding vector in the time queue of the target region thermodynamic space feature coding vector to obtain a set of target region temperature distribution space deep implicit feature coding vectors; Constructing a causal association topological feature of a set of spatial depth implicit feature encoding vectors of the target area temperature distribution to obtain a target area thermal field semantic causal topological feature matrix as the thermodynamic causal association feature; Taking the semantic causal topological feature matrix of the thermal field of the target area as the semantic causal association structure information, the time queue of the target area thermodynamic space feature coding vector and the set of the target area temperature distribution space depth implicit feature coding vector are subjected to context dynamic propagation fusion to obtain the target area thermal distribution spatiotemporal fusion coding vector as the target area thermal distribution spatiotemporal fusion coding feature.

5. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 4 is characterized in that: Constructing a causal association topological feature of a set of spatial deep implicit feature encoding vectors of the target area temperature distribution to obtain a target area thermal field semantic causal topological feature matrix as the thermodynamic causal association feature, including: Calculating the semantic association measurement factor between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of the target area temperature distribution spatial depth implicit feature coding vectors to obtain a target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors; The target area thermal field semantic causal topology matrix is ​​input into a gated causal triggering network based on a gated activation function to obtain the target area thermal field semantic causal topology feature matrix.

6. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 5, characterized in that: Calculating the semantic association measurement factor between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of the target area temperature distribution spatial depth implicit feature coding vectors to obtain a target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors, including: Calculating the correlation matrix between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of target area temperature distribution spatial depth implicit feature coding vectors to obtain a set of target area temperature distribution correlation matrices; Calculating the semantic association measurement factor of each target area temperature distribution association matrix in the set of the target area temperature distribution association matrices to obtain the target area thermal field semantic causal topology matrix composed of multiple target area temperature distribution semantic association measurement factors, wherein the target area temperature distribution semantic association measurement factor is calculated by the mean, variance, maximum value and association compensation coefficient of the target area temperature distribution association matrix; wherein, in response to the variance of the target area temperature distribution association matrix being greater than or equal to a predetermined threshold, a weighted mean of the distances between any two target area temperature distribution spatial depth implicit feature coding vectors in the set of the target area temperature distribution spatial depth implicit feature coding vectors is used as the association compensation coefficient; In response to the variance of the target area temperature distribution correlation matrix being smaller than the predetermined threshold, a weighted value of the variance of the target area temperature distribution correlation matrix is ​​used as the correlation compensation coefficient.

7. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 6, characterized in that: The target area thermal field semantic causal topological feature matrix is ​​used as the semantic causal association structure information, and the time queue of the target area thermodynamic space feature coding vector and the set of the target area temperature distribution space depth implicit feature coding vector are subjected to context dynamic propagation fusion to obtain the target area thermal distribution spatiotemporal fusion coding vector as the target area thermal distribution spatiotemporal fusion coding feature, including: Input the target area thermal field semantic causal topological feature matrix and the target area thermodynamic space feature encoding vector into the time queue of the feature sequence dynamic walk encoder based on the graph convolutional neural network model to obtain the target area temperature distribution surface context dynamic propagation semantic encoding vector; Inputting the target area thermal field semantic causal topological feature matrix and the target area temperature distribution spatial depth implicit feature encoding vector set into the feature sequence dynamic walk encoder based on the graph convolutional neural network model to obtain the target area temperature distribution hidden layer context dynamic propagation semantic encoding vector; The target area temperature distribution hidden layer context dynamic propagation semantic coding vector and the target area temperature distribution surface layer context dynamic propagation semantic coding vector are fused to obtain the target area thermal distribution spatiotemporal fusion coding vector.

8. The intelligent fire detection method based on dynamic spatiotemporal temperature analysis according to claim 7, characterized in that: Fire detection is performed based on the spatiotemporal fusion coding features of the thermal distribution of the target area to determine a fire risk score value and determine whether to trigger a fire alarm, including: Inputting the spatiotemporal fusion coding vector of the target area thermal distribution into a decoder-based fire detector to obtain a detection result, which is a fire risk score value; Based on the comparison between the detection result and a preset threshold, it is determined whether to trigger a fire alarm.

9. An intelligent fire detection device based on dynamic spatiotemporal temperature analysis, characterized in that: include: a thermal infrared image acquisition module for acquiring a time queue of thermal infrared images of a target area acquired by an infrared thermal imager; a data preliminary processing module, configured to pre-process and format the data of each thermal infrared image in the time queue of the thermal infrared images to obtain a time queue of the thermal infrared temperature distribution matrix; a temperature distribution space feature extraction module, configured to extract temperature distribution space features from each thermal infrared temperature distribution matrix in the time queue of the thermal infrared temperature distribution matrix to obtain a time queue of thermodynamic space feature encoding vectors of a target area; a thermodynamic temporal dynamic coding module, configured to perform thermodynamic temporal dynamic coding on the time queue of the target area thermodynamic spatial feature coding vectors to obtain a temporal and spatial fusion coding feature of the target area thermal distribution, wherein the thermodynamic temporal dynamic coding module is configured to: perform causal correlation feature analysis on every two target area thermodynamic spatial feature coding vectors in the time queue of the target area thermodynamic spatial feature coding vectors, and then perform temporal context dynamic fusion coding based on the thermodynamic causal correlation feature to obtain a temporal and spatial fusion coding feature of the target area thermal distribution; The fire risk assessment alarm module is used to perform fire detection based on the spatiotemporal fusion coding features of the thermal distribution of the target area to determine the fire risk score value and determine whether to trigger a fire alarm.

10. The intelligent fire detection device based on dynamic spatiotemporal temperature analysis according to claim 9, characterized in that: The fire risk assessment alarm module includes: a fire risk score value generating unit, configured to input the spatiotemporal fusion coding vector of the target area thermal distribution into a decoder-based fire detector to obtain a detection result, wherein the detection result is a fire risk score value; The fire alarm triggering unit is used to determine whether to trigger a fire alarm based on the comparison between the detection result and a preset threshold.

Images