On-site monitoring device of railway tunnel intelligent fire-fighting monitoring system

Abstract

The invention discloses an on-site monitoring device of a railway tunnel intelligent fire-fighting monitoring system. The on-site monitoring device comprises a sensing module, a control module and a fire-fighting module, the sensing module and the fire-fighting module are respectively in communication connection with the control module; the sensing module is used for detecting a tunnel environment and acquiring environment information; the environment information comprises fire detector data and fire source images formed at different positions in the tunnel environment; the control module is used for extracting fire source features according to the fire source image and predicting a fire source spatial position in combination with fire detector data; the fire-fighting module is adjusted according to the fire source space position to execute fixed-point fire fighting; according to the invention, accurate fire source positioning can be realized in combination with image data of a fire scene and detector data, and fire-fighting measures aiming at the position of the fire source are taken.

Description

Technical Field

[0001] The present invention relates to the field of on-site fire fighting technology, and more particularly to an on-site monitoring device for an intelligent fire fighting monitoring system in a railway tunnel. Background Art

[0002] With the rapid development of railway transportation, railway tunnels, as an important part of railway lines, are of vital importance for safe operation. Railway tunnels have complex environments and relatively closed spaces. Once a fire occurs, the fire spreads rapidly, smoke is difficult to discharge, and rescue is extremely difficult, which can easily cause serious casualties and property losses.

[0003] At present, the traditional railway tunnel fire monitoring system mainly relies on fire detectors, such as smoke detectors, temperature detectors, etc. Although these detectors can detect the occurrence of fire to a certain extent, they have many limitations. On the one hand, the detectors can only provide simple fire alarm signals, and it is difficult to accurately determine the specific location of the fire source, and cannot provide accurate information support for fire rescue. As a result, firefighters may not be able to quickly find the fire source when extinguishing the fire, delaying the fire extinguishing opportunity. On the other hand, the traditional system lacks a comprehensive and intuitive understanding of the fire scene, and cannot flexibly adjust firefighting measures according to the actual situation of the fire. It often adopts large-scale firefighting operations, which not only wastes firefighting resources, but also may cause unnecessary damage to other facilities in the tunnel.

[0004] Therefore, it is of great practical significance to develop an on-site monitoring device for railway tunnel intelligent fire monitoring system that can accurately determine the location of the fire source and take targeted fire fighting measures according to the location of the fire source. Your invention came into being based on this background, which can effectively solve the shortcomings of the existing technology and provide more reliable protection for the fire safety of railway tunnels. Summary of the invention

[0005] In view of this, the present invention provides an intelligent fire monitoring system for railway tunnels, which can accurately locate the fire source by combining the image data and detector data of the fire scene, and take fire fighting measures targeting the fire source location.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A field monitoring device for an intelligent fire monitoring system of a railway tunnel comprises a sensing module, a control module and a fire module; the sensing module and the fire module are respectively connected to the control module for communication.

[0008] The perception module is used to detect the tunnel environment and obtain environmental information; the environmental information includes fire detector data and fire source images at different positions in the tunnel environment; the control module is used to extract fire source features based on the fire source image, and predict the spatial position of the fire source in combination with the fire detector data; and adjust the fire fighting module to execute fixed-point fire fighting according to the spatial position of the fire source.

[0009] Preferably, the control module includes a fire source locating unit and an execution control unit.

[0010] The fire source positioning unit is used to issue a fire alarm based on the fire detector data and confirm the alarm area; it is used to obtain the fire source image of the alarm area and predict the spatial position of the fire source using a trained neural network model in combination with the fire detector data; the execution control unit is used to control the fire protection module in the alarm area based on the spatial position of the fire source.

[0011] Preferably, the fire source locating unit confirms the potential alarm area by comparing the fire detector data with a preset threshold, and uses the first neural network model in combination with the spatiotemporal changes of the fire detector data to predict the fire spread trend, and dynamically adjusts the boundary of the potential alarm area to obtain the final alarm area.

[0012] Preferably, it is characterized in that the fire source locating unit obtains the corresponding fire source image according to the alarm area, extracts image features, combines the fire detector data and the image features, and uses a second neural network model to predict the spatial position of the fire source.

[0013] Preferably, the first neural network model includes a first feature extraction unit, a first feature fusion unit and a first prediction unit.

[0014] The first feature extraction unit uses multi-scale convolution and time convolution to extract multi-scale spatial features and multi-scale time series features respectively; the first feature fusion unit is used to splice the multi-scale spatial features and the multi-scale time series features and reduce the dimension through a fully connected layer to generate multi-scale spatiotemporal features; the first prediction unit uses a direction prediction subnetwork and a speed prediction subnetwork to predict the direction and speed of fire spread respectively; the direction prediction subnetwork maps the input large-scale spatial features and long time series features to a preset direction dimension through a fully connected layer, and generates a direction probability distribution through a softmax function; the speed prediction subnetwork maps the input small-scale spatial features and short time series features to a one-dimensional space through a fully connected layer and outputs a speed estimate through a linear regression layer.

[0015] Preferably, the first extraction unit adopts a hybrid neural network composed of MS-CNN and TCN; the MS-CNN is composed of parallel convolution branches, which respectively use convolution kernels of different sizes to extract spatial features of different scales; the TCN captures dependencies of different time scales by multi-level dilated convolution.

[0016] Preferably, it is characterized in that the second neural network includes a multimodal data preprocessing unit, a feature interaction fusion unit and a positioning prediction unit; the multimodal data preprocessing unit is used to preprocess the fire source image and the fire detector data; the fire source image is divided into multiple image blocks, and mapped into feature vectors through a linear projection matrix, and the attention weight is calculated through a Transformer encoder to obtain global features; the fire detector data is subjected to Fourier transformation and then convolution processing is performed to obtain local features;

[0017] The feature interaction fusion unit is used to perform global average pooling of the global features and the local features in the channel dimension respectively, and confirm the fusion weight through two independent fully connected layers; it is used to fuse the global features and the local features according to the fusion weight to obtain the fusion feature; the positioning prediction unit outputs the spatial position coordinates through a multi-layer fully connected network according to the fusion feature.

[0018] Preferably, the positioning prediction unit includes a first fully connected layer, a second fully connected layer and a third fully connected layer;

[0019] The first fully connected layer performs preliminary aggregation based on the fused features and the image block coordinates; the second fully connected layer is used to secondary aggregate the features of the preliminary aggregation and output the aggregated features; the third fully connected layer is used to perform final aggregation based on the aggregated features and the normalized global plane coordinates of the image block pixels to obtain the depth value, and obtain the spatial position coordinates in combination with the rotation matrix and the translation vector.

[0020] Preferably, when dividing the image blocks, the multimodal data preprocessing unit is also used to mark the positions of the image blocks to obtain the image block coordinates; obtain normalized local plane coordinates by normalizing the image block coordinates; and confirm the normalized global plane coordinates based on the normalized local plane coordinates.

[0021] Preferably, the normalized global plane coordinates are specifically:

[0022]

[0023] Among them, (x n ,y n ) is the normalized global plane coordinate, is the normalized local plane coordinate corresponding to image block i, ωi is the weight of image block i, N is the number of image blocks, f i is the feature vector corresponding to image block i, ||f i ||2 is the L2 norm, which is used to represent the intensity weight of the image block, d is the dimension of the feature vector, and f ij is the jth element in the feature vector corresponding to image i; p i is the position weight of the image block; W and H represent the width and height of the image respectively, and are the image patch coordinates, α and β are hyperparameters used to balance the influence of intensity weight and position weight.

[0024] Preferably, it also includes an emergency drone module, which includes a drone body and an image acquisition unit and a detection unit carried by the drone body; the drone is used to receive control instructions from the control module to reach a specified position; the image acquisition unit is used to capture fire source images.

[0025] Preferably, the image acquisition component further includes a display screen, and the display screen is used to receive instructions from the control module.

[0026] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a railway tunnel intelligent fire monitoring and on-site monitoring device, which realizes efficient linkage from information acquisition, fire source location to fire execution through the coordinated cooperation of perception, control and fire modules. The system combines fire detector data with fire source images, changes the disadvantage of fuzzy positioning of traditional fire monitoring systems, accurately locks the spatial position of the fire source, and directs the fire module to implement fixed-point fire fighting, significantly improving the efficiency of fire fighting and avoiding waste of resources and additional damage to tunnel facilities caused by blind large-scale operations.

[0027] The first and second neural network models played a key role. The former extracted and integrated the spatiotemporal features of fire data at multiple scales to accurately predict the direction and speed of fire spread; the latter integrated multimodal data and combined image block coordinates to achieve precise fire source location. The combination of the two improved the accuracy and comprehensiveness of feature extraction, provided the system with rich and accurate information, and significantly improved the accuracy and reliability of fire source location. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0029] Figure 1 A schematic diagram of the structure of an on-site monitoring device of an intelligent fire monitoring system for a railway tunnel provided by the present invention;

[0030] Figure 2 Schematic diagram of the working process of the fire source locating unit in the embodiment of the present invention;

[0031] Figure 3 The figure is a schematic diagram of the flow of locating the spatial position of a fire source in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] like Figure 1 , Figure 2 and Figure 3 The embodiment of the present invention discloses an on-site monitoring device of an intelligent fire monitoring system for a railway tunnel, including a sensing module, a control module and a fire module; the sensing module and the fire module are respectively connected to the control module for communication.

[0034] The perception module is used to detect the tunnel environment and obtain environmental information; the environmental information includes fire detector data and fire source images at different locations in the tunnel environment; the control module is used to extract fire source features based on the fire source image and predict the spatial location of the fire source in combination with the fire detector data; and the fire fighting module is adjusted according to the spatial location of the fire source to execute fixed-point fire fighting.

[0035] Among them, the perception module includes a variety of sensors, such as smoke sensors, temperature sensors, and flame sensors, which are distributed in different locations in the tunnel through sensor arrays, continuously collect tunnel environment data, and use data fusion algorithms to integrate them into fire detector data; and obtain fire source images through high-definition cameras.

[0036] In one embodiment, the control module includes a fire source locating unit and an execution control unit; the fire source locating unit is used to issue a fire alarm based on fire detector data and confirm the alarm area; it is used to obtain a fire source image in the alarm area and predict the spatial position of the fire source using a trained neural network model in combination with the fire detector data; the execution control unit is used to control the fire fighting module in the alarm area according to the spatial position of the fire source.

[0037] In order to further implement the above technical solution, the fire source positioning unit confirms the potential alarm area by comparing the fire detector data with the preset threshold, and uses the first neural network model combined with the spatiotemporal changes of the fire detector data to predict the fire spread trend, and dynamically adjusts the boundary of the potential alarm area to obtain the final alarm area.

[0038] Further, the first neural network model includes a first feature extraction unit, a first feature fusion unit and a first prediction unit;

[0039] The first feature extraction unit uses multi-scale convolution and time convolution to extract multi-scale spatial features and multi-scale time series features respectively;

[0040] The first feature fusion unit is used to concatenate multi-scale spatial features and multi-scale time series features and perform dimension reduction through a fully connected layer to generate multi-scale spatiotemporal features;

[0041] The first prediction unit uses a direction prediction subnetwork and a speed prediction subnetwork to predict the direction and speed of fire spread respectively; the direction prediction subnetwork maps the input large-scale spatial features and long time series features to the preset direction dimension through a fully connected layer, and generates a direction probability distribution through a softmax function; the speed prediction subnetwork maps the input small-scale spatial features and short time series features to a one-dimensional space through a fully connected layer and outputs a speed estimate through a linear regression layer.

[0042] In this embodiment, the first extraction unit adopts a hybrid neural network consisting of a multi-scale convolutional neural network (MS-CNN) and a temporal convolutional network (TCN); the MS-CNN is composed of parallel convolution branches, which use convolution kernels of different sizes to extract spatial features of different scales; the TCN captures dependencies of different time scales by multi-level dilated convolutions.

[0043] Exemplarily, a multi-scale convolutional neural network (MS-CNN) is used to extract the spatial features of fire detector data; MS-CNN includes three parallel convolution branches, which use convolution kernels of different sizes (such as 3×3, 5×5, and 7×7) to perform convolution operations on the input data: large-scale spatial features: global spatial features are extracted through a 7×7 convolution kernel to capture the overall trend of fire spread; medium-scale spatial features: regional spatial features are extracted through a 5×5 convolution kernel to capture the mesoscopic changes in fire spread; small-scale spatial features: local spatial features are extracted through a 3×3 convolution kernel to capture the detailed changes in fire spread; the outputs of the three branches are feature fused to generate multi-scale spatial features.

[0044] Time series features are extracted using a temporal convolutional network (TCN); TCN captures dependencies of different time scales through multi-level dilated convolutions: long time series features: a larger dilation coefficient (such as d=4) is used to capture dependencies of long time spans (such as data trends in the past 10 minutes); short time series features: a smaller dilation coefficient (such as d=1) is used to capture dependencies of short time spans (such as data fluctuations in the past 1 minute); features of different time scales are fused to generate multi-scale time series features.

[0045] In another embodiment, the fire source locating unit obtains the corresponding fire source image according to the alarm area, extracts image features, combines the fire detector data and the image features, and uses a second neural network model to predict the spatial position of the fire source.

[0046] In order to further implement the above technical solution, the second neural network includes a multimodal data preprocessing unit, a feature interaction fusion unit and a positioning prediction unit.

[0047] The multimodal data preprocessing unit is used to preprocess the fire source image and detector data; the fire source image is divided into multiple image blocks, and mapped into feature vectors through a linear projection matrix, and the attention weight is calculated through the Transformer encoder to obtain the global feature; the fire detector data is convolved after Fourier transformation to obtain the local feature; the feature interaction fusion unit is used to perform global average pooling of the global features and local features in the channel dimension respectively, and confirm the fusion weight through two independent fully connected layers; it is used to fuse the global features and local features according to the fusion weight to obtain the fusion feature; the positioning prediction unit outputs the spatial position coordinates according to the fusion feature through a multi-layer fully connected network.

[0048] Furthermore, the positioning prediction unit includes a first fully connected layer, a second fully connected layer and a third fully connected layer; the first fully connected layer performs preliminary aggregation based on the fused features and the image block coordinates; the second fully connected layer is used to secondary aggregate the preliminary aggregated features and output the aggregated features; the third fully connected layer is used to perform final aggregation based on the aggregated features and the normalized global plane coordinates of the image block pixels to obtain the depth value, and obtain the spatial position coordinates in combination with the rotation matrix and the translation vector.

[0049] In this embodiment, when dividing the image block, the multimodal data preprocessing unit is also used to mark the position of the image block to obtain the image block coordinates; obtain the normalized local plane coordinates by normalizing the image block coordinates; and confirm the normalized global plane coordinates based on the normalized local plane coordinates.

[0050] Exemplarily, in the multimodal data preprocessing unit, for the fire source image, its size is uniformly adjusted to 224×224 pixels using a bilinear interpolation algorithm, and divided into image blocks of 16×16 pixels. During the division process, the pixel coordinates of the upper left corner pixel of each image block in the original image are recorded. These coordinates will be used to determine the position information of the fire source in the image later and participate in the calculation of the spatial position of the fire source in the camera coordinate system positioning module. Each image block is mapped to a feature vector with a dimension of 768 through a linear projection matrix P, which is a 768×256 matrix. Its element values ​​are randomly initialized and optimized during training, and then a 12-layer Transformer encoder is used, with 12 attention heads in each layer and an attention head dimension of 64. The attention weights are calculated through the Q, K, and V matrices to extract global features.

[0051] For the fire detector data, the fast Fourier transform (FFT) algorithm is first used to convert the collected time series data from the time domain to the frequency domain, and then it is processed by a convolutional neural network with three convolutional layers. The convolution kernel sizes are (3,1), (3,1), and (3,1), respectively, with a step size of 1 and the padding method is the same. Each convolution layer is followed by a ReLU activation function to extract local features.

[0052] In the feature interaction fusion module, a dynamic weight allocation mechanism is used to weightedly fuse global features and local features. The specific process includes: performing global averaging of the channel dimension of global features and local features to obtain the corresponding statistics; then mapping the statistics to scalar values ​​through two independent fully connected layers, and the proportion of the scalar value is the fusion weight.

[0053] In addition, a multi-scale fusion strategy is used to fuse and obtain fused features. Specifically, the global features and local features are downsampled respectively, and the maximum pooling operation is used to obtain feature maps of different resolutions, such as 1 / 2 resolution, 1 / 4 resolution, etc. At each resolution, fusion is performed according to the above-mentioned dynamic weight allocation mechanism, and then the bilinear interpolation upsampling is used to restore the original resolution and splice, and finally a convolution layer (convolution kernel size is (1,1), step size is 1, and padding method is the same) is used to integrate the features to obtain fused features.

[0054] Finally, the final positioning is achieved through the positioning prediction unit. The location of the fire source in the camera coordinate system is confirmed by the positioning prediction unit, so that the specific location of the fire source can be determined according to the conversion relationship between the camera coordinate system and the world coordinate system.

[0055] In the positioning prediction unit, the first fully connected layer concatenates the fusion features and the coordinates of each image block into a vector V1 as input. The number of neurons in this layer is 1024, the size of the weight matrix W1 of the fully connected layer is 1024×(768+2), and the size of the bias vector b1 is 1024×1. Both are randomly initialized and optimized during training. Feature extraction is performed through the ReLU activation function to obtain the preliminary fusion feature F1, F1=ReLU(W1V1+B1). The second fully connected layer further extracts the preliminary fusion features. The number of neurons in this layer is 512, the size of the weight matrix W2 of the fully connected layer is 512×1024, and the size of the bias vector b2 is 512×1. Both are randomly initialized and optimized during training. The preliminary fusion features are further refined and the output feature F2 is output, F2=ReLU(W2F1+b2). ​​The third fully connected layer combines the feature F2 with the normalized plane coordinates (x n ,y n ) is concatenated into vector V2 as input. The weight matrix W3 of this layer is 1×(512+2), and the bias vector b3 is 1×1. Both are randomly initialized and optimized during training. The depth value z in the camera coordinate system is output through linear transformation. c , z c =W3V3+b3.

[0056] The global normalized plane coordinates are obtained by enhancing and averaging multiple image blocks.

[0057] The local normalized plane coordinates corresponding to each image block can be obtained from the image block coordinates and the camera intrinsic parameters. The details are as follows:

[0058]

[0059] Among them, f x 、f y is the focal length of the camera in the x and y directions, and u0 and v0 are the coordinates of the principal point of the image.

[0060] The normalized global plane coordinates are:

[0061]

[0062] Among them, (x n ,y n ) is the normalized global plane coordinate, is the normalized local plane coordinate corresponding to image block i, ω i is the weight of image block i, N is the number of image blocks, f i is the feature vector corresponding to image block i, ||f i ||2 is the L2 norm, which is used to represent the intensity weight of the image block, d is the dimension of the feature vector, and f ijis the jth element in the feature vector corresponding to image i; p i is the position weight of the image block; W and H represent the width and height of the image respectively, and are the image patch coordinates, α and β are hyperparameters used to balance the influence of intensity weight and position weight.

[0063] Then, the coordinates of the fire source in the camera coordinate system are obtained according to the coordinate system conversion:

[0064]

[0065] Among them, X c , Y c , Z c is the coordinate of the fire source in the camera coordinate system, R is the rotation matrix of the coordinate system transformation, and t is the translation matrix of the coordinate system transformation.

[0066] During the training process, the back propagation algorithm is used to continuously adjust the weights of the fully connected layer based on the mean square error loss function between the predicted coordinates and the actual fire source coordinates in the camera coordinate system to minimize the loss function.

[0067] In order to further implement the above technical solution, it also includes an emergency drone module, which includes a drone body and an image acquisition unit and a detection unit carried by the drone body; the drone is used to receive control instructions from the control module to reach a designated location; the image acquisition unit is used to acquire images of fire sources; the image acquisition component also includes a display screen, which is used to receive instructions from the control module.

[0068] In this embodiment, emergency data collection is achieved by using a drone equipped with a sensing module. When the built-in sensing module of the railway fails to work, the control module can achieve immediate firefighting based on the data collected by the drone.

[0069] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0070] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A field monitoring device for an intelligent fire monitoring system in a railway tunnel, characterized in that: It includes a sensing module, a control module and a fire fighting module; the sensing module and the fire fighting module are respectively connected to the control module for communication; The perception module is used to detect the tunnel environment and obtain environmental information; the environmental information includes fire detector data and fire source images at different locations in the tunnel environment; The control module is used to extract fire source features according to the fire source image, and predict the spatial position of the fire source in combination with the fire detector data; and adjust the fire fighting module to perform fixed-point fire fighting according to the spatial position of the fire source.

2. The on-site monitoring device of the intelligent fire monitoring system for a railway tunnel according to claim 1 is characterized in that: The control module includes a fire source positioning unit and an execution control unit; The fire source positioning unit is used to make a fire alarm according to the fire detector data and confirm the alarm area; it is used to obtain the fire source image of the alarm area and predict the spatial position of the fire source by using the trained neural network model in combination with the fire detector data; The execution control unit is used to control the fire protection modules in the alarm area according to the spatial location of the fire source.

3. The on-site monitoring device of the intelligent fire monitoring system for a railway tunnel according to claim 2 is characterized in that: The fire source locating unit confirms the potential alarm area by comparing the fire detector data with a preset threshold, and uses the first neural network model in combination with the spatiotemporal changes of the fire detector data to predict the fire spread trend, and dynamically adjusts the boundary of the potential alarm area to obtain the final alarm area.

4. A railway tunnel intelligent fire monitoring system on-site monitoring device according to claim 2 or 3, characterized in that: The fire source positioning unit obtains the corresponding fire source image according to the alarm area, extracts image features, combines the fire detector data and the image features, and uses a second neural network model to predict the spatial position of the fire source.

5. The on-site monitoring device of the intelligent fire monitoring system for a railway tunnel according to claim 3 is characterized in that: The first neural network model includes a first feature extraction unit, a first feature fusion unit and a first prediction unit; The first feature extraction unit uses multi-scale convolution and time convolution to extract multi-scale spatial features and multi-scale time series features respectively; The first feature fusion unit is used to splice the multi-scale spatial features and the multi-scale time series features and perform dimension reduction through a fully connected layer to generate multi-scale spatiotemporal features; The first prediction unit uses a direction prediction subnetwork and a speed prediction subnetwork to respectively predict the direction and speed of fire spread; the direction prediction subnetwork maps the input large-scale spatial features and long time series features to a preset direction dimension through a fully connected layer, and generates a direction probability distribution through a softmax function; the speed prediction subnetwork maps the input small-scale spatial features and short time series features to a one-dimensional space through a fully connected layer and outputs a speed estimate through a linear regression layer.

6. The monitoring device of the intelligent fire monitoring system of a railway tunnel according to claim 5 is characterized in that: The first extraction unit adopts a hybrid neural network composed of MS-CNN and TCN; The MS-CNN consists of parallel convolution branches, each using convolution kernels of different sizes to extract spatial features of different scales; The TCN captures dependencies at different time scales by using multi-level dilated convolutions.

7. The on-site monitoring device of the intelligent fire monitoring system for a railway tunnel according to claim 4 is characterized in that: The second neural network includes a multimodal data preprocessing unit, a feature interaction fusion unit and a positioning prediction unit; The multimodal data preprocessing unit is used to preprocess the fire source image and the fire detector data; divide the fire source image into multiple image blocks, and map them into feature vectors through a linear projection matrix, and calculate the attention weights through a Transformer encoder to obtain global features; perform convolution processing on the fire detector data after Fourier transformation to obtain local features; The feature interaction fusion unit is used to perform global average pooling of the global features and the local features in channel dimensions respectively, and confirm the fusion weights through two independent fully connected layers; and is used to fuse the global features and the local features according to the fusion weights to obtain fused features; The positioning prediction unit outputs spatial position coordinates through a multi-layer fully connected network according to the fusion features.

8. The on-site monitoring device of the intelligent fire monitoring system for railway tunnels according to claim 7 is characterized in that: The positioning prediction unit includes a first fully connected layer, a second fully connected layer and a third fully connected layer; The first fully connected layer performs preliminary aggregation based on the fusion features and the image block coordinates; The second fully connected layer is used to secondary aggregate the initially aggregated features and output aggregated features; The third fully connected layer is used to perform final aggregation according to the aggregated features and the normalized global plane coordinates of the image block pixels to obtain a depth value, and to obtain the spatial position coordinates in combination with the rotation matrix and the translation vector.

9. The on-site monitoring device of the intelligent fire monitoring system for railway tunnels according to claim 7 is characterized in that: When dividing the image blocks, the multimodal data preprocessing unit is also used to mark the positions of the image blocks to obtain the image block coordinates; obtain normalized local plane coordinates by normalizing the image block coordinates; and confirm the normalized global plane coordinates based on the normalized local plane coordinates.

10. The on-site monitoring device of the intelligent fire monitoring system for railway tunnels according to claim 9, characterized in that: The normalized global plane coordinates are specifically: Among them, (x n ,y n ) is the normalized global plane coordinate, is the normalized local plane coordinate corresponding to image block i, ω i is the weight of image block i, N is the number of image blocks, f i is the feature vector corresponding to image block i, ||f i ||2 is the L2 norm, which is used to represent the intensity weight of the image block, d is the dimension of the feature vector, and f ij is the jth element in the feature vector corresponding to image i; p i is the position weight of the image block; W and H represent the width and height of the image respectively, and are the image patch coordinates, α and β are hyperparameters used to balance the influence of intensity weight and position weight.

Images