Train fire cloud early warning and prediction system and method

Abstract

The invention discloses a train fire cloud early warning and prediction system and method, and the method comprises the steps: constructing a multi-modal fire early warning model and a multi-modal fire situation prediction model based on deep learning, and carrying out the training; the vehicle-mounted terminal obtains multi-source heterogeneous real-time data through a sensor, performs feature extraction and fusion on the multi-source heterogeneous real-time data, inputs the fused features into the trained multi-mode fire early warning model, and outputs the highest probability state as a real-time fire early warning result; transmitting the obtained multi-source heterogeneous real-time data to a cloud end, and receiving and storing the real-time data by the cloud end; combining stored historical data with real-time data, and performing feature extraction and fusion study and judgment on all multi-source heterogeneous data by using the trained multi-mode fire situation prediction model to obtain fire situation development prediction; according to the invention, independent and accurate fire early warning under the condition that the computing power of the vehicle-mounted terminal is limited and accurate fire situation development prediction under the condition that the computing power of the cloud terminal is high are realized.

Description

Technical Field

[0001] The present invention relates to the field of train fire early warning and prediction, and specifically discloses a train fire vehicle-cloud early warning and prediction system and method. Background Art

[0002] The background of train fires is complex, and its ignition sources include various factors such as electrical failures, illegal use of open flames, and spontaneous combustion of flammable materials. Due to the relatively enclosed space of the train, the combustibles present result in a large fire load. The interaction of factors such as air convection inside the train and surface conduction of combustibles causes the fire to spread rapidly. Generally speaking, train fires have many characteristics such as strong initial concealment, fast development speed, rapid spread of smoke, great difficulty in rescue, and large harmful impacts.

[0003] Currently, in engineering, a traditional train fire automatic alarm system is generally adopted. For physical quantities such as smoke and heat in the initial stage of a fire, they are collected through fire detectors such as smoke-temperature composite detectors and temperature-sensing cables arranged in positions such as carriages, and transmitted to the train fire alarm controller. When a single judgment data source such as the collected smoke concentration or temperature exceeds a certain threshold, an alarm signal will be generated and informed to the driver in the form of sound and light or a human-machine interface. However, due to the complex train environment, the detectors are easily interfered by the outside world and cause missed alarms and false alarms, and the reliability of the system cannot meet the requirements of real-time and accurate fire detection, let alone provide the ability of early fire perception and early warning and prediction of the future development trend of the fire situation. In recent years, fire recognition methods based on image information have been widely used in engineering. Fire image processing algorithms based on machine learning distinguish fire pixels and non-fire pixels by extracting features such as color, shape, texture, and motion features, so as to judge whether a fire has occurred. However, this method is still a single judgment data source, and the open-source data set of machine learning is for non-train scenarios, and the effectiveness of fire recognition in the train scenario cannot be guaranteed. Summary of the Invention

[0004] The present invention aims to provide a train fire vehicle-cloud early warning and prediction system and method. Based on the vehicle-cloud collaborative architecture, it realizes independent and accurate fire early warning and rapid linkage disposal under the condition of limited on-vehicle computing power, as well as accurate prediction of the development trend of the fire situation and comprehensive emergency disposal under the condition of high computing power in the cloud; based on the multi-source heterogeneous data fusion algorithm, it comprehensively represents the fusion judgment of multi-type sensor information of disaster-causing factors, and improves the accuracy and reliability of train fire early warning and prediction; based on the vehicle-ground dual communication mechanism, it ensures the reliability of information transmission and the robustness of the system in case of emergency.

[0005] A train fire vehicle-cloud early warning and prediction method provided by the present invention includes the following steps: Construct a multi-modal fire warning model and a multi-modal fire situation prediction model based on deep learning, and train the multi-modal fire warning model and the multi-modal fire situation prediction model to obtain the trained multi-modal fire warning model and multi-modal fire situation prediction model; Obtain multi-source heterogeneous real-time data through sensors installed on the train, including temperature data, smoke data, wind speed data, and fire image data; extract and fuse the features of the multi-source heterogeneous real-time data, and then input the fused features into the trained multi-modal fire warning model to give the real-time fire state probabilities of no fire, smoldering, and open fire. The state with the highest probability output is the real-time fire warning result, and an audible and visual warning or alarm signal for the fire state is output, providing a method for independent and accurate fire warning and rapid linkage disposal under the condition of limited computing power on the vehicle-mounted side; Transmit the multi-source heterogeneous real-time data obtained through sensors to the cloud, and the cloud receives and stores the real-time data; combine the stored historical data with the real-time data, and use the trained multi-modal fire situation prediction model to extract and fuse and analyze the features of all multi-source heterogeneous data to obtain the prediction of the development of the fire situation, providing a method for accurate prediction of the development of the fire situation and comprehensive emergency disposal under the condition of high computing power in the cloud.

[0006] Further, the multi-source heterogeneous real-time data is specifically: The temperature data is collected by an optical fiber temperature sensor, and the temperature along the entire length of the train car is measured, which is one-dimensional spatial data; The smoke data is collected by a smoke sensor, and the wind speed data is collected by a wind speed sensor. The smoke sensor and the wind speed sensor respectively collect the smoke and wind speed in the train car, both of which are one-dimensional time data measured at spatial points; The fire image data is collected by a visible light image sensor or / and an infrared image sensor, and the plane image in the train car is measured, which is 2D spatial data.

[0007] Further, the method of extracting and fusing the features of the multi-source heterogeneous real-time data, and then inputting the fused features into the trained multi-modal fire warning model to give the real-time fire state probabilities of no fire, smoldering, and open fire, and the state with the highest probability output is the real-time fire warning result: (1) Extract the features of the multi-source heterogeneous real-time data through a feature extraction network: The temperature data is a one-dimensional spatial sequence. First, use the method of discrete wavelet transform to transform the temperature data into two-dimensional image-like data, and then use a convolutional neural network (CNN) to extract features from it. The convolutional neural network (CNN) includes a convolutional layer, an attention module, a pooling layer, and a normalization layer; The described flue gas data and wind speed data are one-dimensional time series data of spatial points, and the LSTM model with self-attention mechanism is used to extract features from them respectively. The LSTM model includes an LSTM layer, an attention layer and a fully connected layer; The described fire image data is two-dimensional spatial data, and an improved VGG16 model is used to extract features from it. The VGG16 model includes a convolutional layer, an attention module, a pooling layer and a normalization layer; (2) Take the extracted temperature features, flue gas features, wind speed features and fire image features as input elements, and input them into the feature fusion module of the MS-FRFM network built to perform feature fusion to obtain the fusion features of multi-source heterogeneous real-time data; (3) Input the fusion features into a multi-layer perceptron MLP network model with a double hidden layer for decision-making. The output neurons are the real-time fire state probabilities of no fire, smoldering and open fire, and the state with the highest output probability is the real-time fire warning result.

[0008] Furthermore, the method for obtaining the prediction of the development of the fire situation by using the trained multi-modal fire situation prediction model to extract features and perform fusion research and judgment on all multi-source heterogeneous data is as follows: First, perform feature extraction based on the Transformer architecture, then use the cross-modal cross-attention mechanism for modal interaction to achieve feature fusion, and then use the fusion features as input elements, and use two layers of long short-term memory network LSTM and multi-layer perceptron MLP network model to obtain the prediction of the development of the fire situation.

[0009] A train fire vehicle-cloud warning and prediction system, the system includes: An on-vehicle controller, which realizes real-time train fire warning based on a multi-modal fire warning model through the on-vehicle controller and links the train fire linkage control system; A transmission network, which transmits the on-vehicle real-time data to the cloud; The cloud, the cloud receives and stores real-time data through a storage platform, and realizes the prediction of the development of the train fire situation based on a multi-modal fire situation prediction model through a computing platform.

[0010] Furthermore, the on-vehicle controller specifically includes: A data acquisition module, which acquires multi-source heterogeneous real-time data collected by sensors installed on the train, including temperature data, flue gas data, wind speed data, and fire image data; A data fusion processing module, which extracts and fuses features from multi-source heterogeneous real-time data, and then inputs the fusion features into a decision model to give the real-time fire state probabilities of no fire, smoldering and open fire, and the state with the highest output probability is the real-time fire warning result; A display module that displays the real-time data collected by the data acquisition module, the real-time processing process of the data fusion processing module, the fire warning result, and the acoustic and optical warning or alarm signals of the fire status; A linkage control interface module for connecting to the train fire linkage control system to achieve precise and rapid train fire disposal at the on-vehicle end.

[0011] Furthermore, the transmission network specifically includes: A 5G network for data transmission between the on-vehicle end and the cloud; A LoRa network for data transmission between the on-vehicle end and the cloud in the case of the failure of vehicle-to-ground communication in a self-organizing network mode.

[0012] Furthermore, the cloud specifically includes: A storage platform for storing the real-time data output by the transmission network; A computing platform that provides the computing power conditions for combining the stored historical data with the real-time data to achieve fire prediction, is used for predicting the development of the fire situation, reduces the probability of train fires, provides effective guarantee for train emergency disposal, and improves the safety operation level of trains.

[0013] A non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements a train fire vehicle-cloud warning and prediction method as described above.

[0014] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a train fire vehicle-cloud warning and prediction method as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention proposes a train fire vehicle-cloud warning and prediction system and method. Through the collaborative work of the on-vehicle controller, the transmission network, and the cloud, the system realizes the independent and precise fire warning detection and linkage rapid disposal capabilities under the condition of limited on-vehicle computing power, as well as the precise fire situation development prediction and comprehensive emergency disposal capabilities under the condition of high computing power in the cloud, providing effective guarantee for the safe operation of trains; (2) The present invention proposes a train fire vehicle-cloud warning and prediction system and method. Based on the multi-source heterogeneous data fusion algorithm, it comprehensively fuses and judges various types of sensor information, effectively reducing the interference of single data source noise on the judgment result, and significantly improving the accuracy and reliability of train fire warning and prediction; (3) The present invention proposes a train fire vehicle-cloud warning and prediction system and method. The system performs vehicle-to-ground transmission through the dual communication mechanism of 5G and LoRa self-organizing network, providing double guarantee for information transmission in case of emergencies such as fires on trains, and effectively improving the robustness of the system. Brief Description of the Drawings

[0016] Figure 1 The present invention provides a composition diagram of a train fire vehicle-cloud early warning and prediction system. Among them, 1 is a vehicle-mounted controller, 2 is a transmission network, and 3 is the cloud; Figure 2 The main modules of the vehicle-mounted controller in a train fire vehicle-cloud early warning and prediction system proposed by the present invention; Figure 3 The image data in the open fire state of a set of train fire multi-source heterogeneous data selected for simulation in an example of the present invention; Figure 4 The accuracy evaluation result of the vehicle-mounted end real-time fire state early warning method in an example of the present invention; Figure 5 The accuracy evaluation result of the cloud fire situation development prediction method in an example of the present invention. Detailed Embodiments

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the following further describes the embodiments of the present invention in detail.

[0018] Embodiment 1

[0019] This embodiment provides a train fire vehicle-cloud early warning and prediction method, including the following steps: Construct a multi-modal fire early warning model and a multi-modal fire situation prediction model based on deep learning, and train the multi-modal fire early warning model and the multi-modal fire situation prediction model to obtain the trained multi-modal fire early warning model and multi-modal fire situation prediction model; Obtain multi-source heterogeneous real-time data through sensors installed on the train, including temperature data, smoke data, wind speed data, and fire image data; extract and fuse the features of the multi-source heterogeneous real-time data, and then input the fused features into the trained multi-modal fire early warning model to give the real-time fire state probabilities of no fire, smoldering, and open fire. The state with the highest probability output is the real-time fire early warning result, and an audible and visual early warning or alarm signal for the fire state is output, providing a method for independent and accurate fire early warning and linkage rapid disposal under the condition of limited computing power at the vehicle-mounted end; Transmit the obtained multi-source heterogeneous real-time data to the cloud, and the cloud receives and stores the real-time data; combine the stored historical data with the real-time data, and use the trained multi-modal fire situation prediction model to extract and fuse and analyze all the multi-source heterogeneous data to obtain the prediction of the fire situation development, providing a method for accurate fire situation development prediction and comprehensive emergency disposal under the condition of high computing power at the cloud.

[0020] The multi-source heterogeneous real-time data is specifically: The temperature data is collected by an optical fiber temperature sensor, measuring the temperature along the entire length of the train carriage, which is one-dimensional spatial data; The flue gas data is collected by a flue gas sensor, and the wind speed data is collected by a wind speed sensor. The flue gas sensor and the wind speed sensor respectively collect the flue gas and wind speed in the train carriage, both of which are one-dimensional time data measured at spatial points; The fire image data is collected by a visible light image sensor or / and an infrared image sensor, measuring the planar image in the train carriage, which is 2D spatial data.

[0021] A method for extracting and fusing features from multi-source heterogeneous real-time data, then inputting the fused features into a trained multi-modal fire warning model to give the real-time fire state probabilities of no fire, smoldering and open fire, and the state with the highest probability output is the real-time fire warning result: (1) Extract features from multi-source heterogeneous real-time data through a feature extraction network; The temperature data is a one-dimensional spatial sequence. First, use the method of discrete wavelet transform to transform the temperature data into two-dimensional image-like data, and then use a convolutional neural network CNN to extract features from it. The convolutional neural network CNN includes a convolutional layer, an attention module, a pooling layer and a normalization layer; The flue gas data and wind speed data are one-dimensional time series data of spatial points, and use an LSTM model with a self-attention mechanism to extract features from them respectively. The LSTM model includes an LSTM layer, an attention layer and a fully connected layer; The fire image data is two-dimensional spatial data, and use an improved VGG16 model to extract features from it. This model includes a convolutional layer, an attention module, a pooling layer and a normalization layer; (2) Take the extracted temperature features, flue gas features, wind speed features and fire image features as input elements, and input them into the MS-FRFM feature fusion module for feature fusion to obtain the fused features of multi-source heterogeneous real-time data; (3) Input the fused features into a multi-layer perceptron MLP network model with a double hidden layer for decision-making. The output neurons are the real-time fire state probabilities of no fire, smoldering and open fire, and the state with the highest probability output is the real-time fire warning result.

[0022] A method for extracting features and conducting fusion research and judgment on all multi-source heterogeneous data by using a trained multi-modal fire situation prediction model: First, perform feature extraction based on the Transformer architecture, then use a cross-modal cross-attention mechanism for modal interaction to achieve feature fusion, and then take the fused features as input elements, and use a two-layer long short-term memory network LSTM and a multi-layer perceptron MLP network model to obtain the prediction of the development of the fire situation.

[0023] Specifically, in this embodiment, the trained multi-modal fire warning model is used to extract features and fuse and judge multi-source heterogeneous real-time data, and the specific processing process of real-time fire status warning on the vehicle-mounted side is as follows: Step 1: Build a full-scale train model. Select a certain train model for 1:1 full-scale simulation. In this embodiment, the selected train is 197 m in length and includes 8 carriages. Step 2: Construct a train fire simulation data set. Set the fire ignition point position, fire heat release rate, sensor position and parameters, etc., and conduct a simulation of the train fire combustion process to simulate different fire situations, obtaining 1,800 sets of multi-source heterogeneous data sets. Each set of data includes 4 types: temperature, smoke, wind speed, and fire images. And the data set is labeled by manually dividing it into three judgment results: no fire, smoldering, and open fire. The labeled data set is divided into a training set and a validation set according to a ratio of 4:1.

[0024] Step 3: Preprocess multi-source heterogeneous data. In this embodiment, a set of train fire multi-source heterogeneous data is selected as an example for illustration; Collect temperature data at 508 points along the entire length of the carriage. The data format is 1×508, which is 1D spatial data. Table 1 shows part of the temperature data in this carriage at this moment; Tables 2 and 3 show the smoke and wind speed data in this carriage at this moment, which are 1D time data, with a sampling rate of 11 times per second, and the data format is 1×11 format; It should be noted that since the data in Tables 1, 2, and 3 cannot be displayed in one line, they are processed by carriage return into multiple lines; Figure 3 It is an image in a certain carriage at a certain moment, which is 2D spatial data, and the data format is 1920×1080; Table 1

[0025] Table 2

[0026] Table 3

[0027] The specific preprocessing process is as follows: perform tensor conversion on the temperature data, perform standardization processing with a mean of 0 and a standard deviation of 1, and apply discrete wavelet transform to convert the 1×508 temperature data into a class image data format of 1×10×60. The 1 in 1×10×60 indicates that the number of channels of the class image data format is 1, 10 indicates that the height of the class image data format is 10, and 60 indicates that the width of the class image data format is 60; perform standardization processing with a mean of 0 and a standard deviation of 1 on the smoke data and wind speed data, and convert the data format to 1×11; perform equal-proportion scaling on the image data, and perform channel-wise standardization processing, and convert the data format to 3×244×244; Step 4: Extract and fuse the features of multi-source heterogeneous real-time data, then input the fused features into the trained multi-modal fire warning model to give the real-time fire state probabilities of no fire, smoldering, and open fire, and the state with the highest probability output is the real-time fire warning result: 1. Feature extraction: Different feature extraction networks are used for feature extraction according to the differences in data nature and structure: (1) Extract the features of the temperature data of the one-dimensional spatial sequence after discrete wavelet transform using the VGG16 structure with 8 convolutional layers; after each convolutional layer, batch normalization is performed using BatchNorm2d, a CBAM attention module is added to enhance the feature representation, and a residual module is added to avoid information loss, and the resulting feature size is 128×1×1; (2) For the smoke and wind speed data of the one-dimensional time sequence, the LSTM_attention network is used for feature extraction respectively; the LSTM_attention network architecture is an LSTM layer, a self-attention layer, an LSTM layer, and a fully connected layer, and the resulting feature size is 128×1×1; (3) For the fire image data with the two-dimensional spatial sequence scaled to 3×244×244 size, an improved version of the VGG16 network is used to extract features; this network contains 10 convolutional layers, a CBAM module attention module is added, batch normalization is performed on the first 8 convolutional layers using BatchNorm2d, and the last 2 convolutional layers are used to enhance the channel features, and the resulting feature size is 128×1×1; 2. Feature fusion: The temperature, smoke, wind speed, and fire image features with a size of 128×1×1 obtained by feature extraction are used as four input elements, and the MS-FRFM network feature fusion module is used to fuse the features to obtain the fused features of multi-source information. The network effectively captures the global and local information between different feature channels through a multi-scale channel attention mechanism; further feature extraction is performed through a one-dimensional convolutional neural network 1D-CNN, and the extraction result is converted into a weight using the Sigmoid operation to highlight the key features in the original information, and the weights of the modal features are assigned based on this to achieve the full and effective fusion of the four modal features.

[0028] 3. Fire decision-making: The fused features are used as input elements and input into an MLP network model with a double hidden layer for decision-making. The output neurons are no fire, smoldering, and open fire, and the resulting output sequence is 1×3, which respectively represent the probabilities of the three real-time fire states. The state with the highest probability is selected as the research and judgment result of the real-time fire state.

[0029] In the training and verification process of the multi-modal fire warning model in this step, Cross-Entropy is selected as the loss function for evaluation.

[0030]

[0031] Among them, is the probability that the true label of sample x is the i-th class, is the probability that the model predicts sample x as the i-th class, and C is the total number of classes. is the cross-entropy loss of the current sample x.

[0032] In this embodiment, the training has a total of 200 epochs. The learning rate for the first 100 epochs is 2.5×10 -5 , and the learning rate for the last 100 epochs is 2.5×10 -6 . After training for 200 epochs, the loss reaches about 1.1×10 -3 , and basically no longer changes. The accuracy of fire state detection can reach 98.8%. As Figure 4 shown, the horizontal axis is the number of training epochs, and the vertical axis is the accuracy.

[0033] In this embodiment, the trained multi-modal fire situation prediction model is used to extract features and conduct fusion research and judgment on all multi-source heterogeneous data. The specific process of obtaining the cloud fire situation development prediction is as follows: Step 1: Transmit the collected multi-source heterogeneous real-time data, namely train temperature, smoke, wind speed, and fire image data, to the cloud storage platform through a 5G or LoRa transmission network; Step 2: The computing platform combines the stored historical data with the real-time data, calculates the fire situation development prediction. The processing process is the same as the path of the in-vehicle detection method. Feature extraction and fusion research and judgment are carried out on the multi-source heterogeneous data. The specific network model selected under high computing power conditions is as follows: First, feature extraction is performed based on the Transformer architecture, then cross-modal cross-attention mechanism is used for modal interaction to achieve feature fusion, and then the fused features are used as input elements. The fire situation development prediction is obtained by using a two-layer LSTM and MLP network model; Taking Figure 5 as an example to show the effectiveness of this network model, the trend of its situation development prediction (the predicted value marked in the figure) is consistent with the simulation result of the simulated fire development (the true value marked in the figure), where the horizontal axis is time and the vertical axis is the heat release rate.

[0034] Embodiment 2

[0035] This embodiment describes a train fire vehicle-cloud early warning and prediction system, including an in-vehicle controller 1, a transmission network 2, and a cloud 3, as Figure 1 shown; The in-vehicle controller specifically includes, as Figure 2As shown in the figure: The data acquisition module obtains multi-source heterogeneous real-time data from various types of sensors installed on the train, including temperature, smoke, wind speed, and fire images; the data fusion processing module extracts and fuses the features of the multi-source heterogeneous real-time data, then inputs the fused features into the decision model to give the real-time fire state probabilities of no fire, smoldering fire, and flaming fire, and outputs the state with the highest probability as the real-time fire warning result; the display module displays the real-time data collected by the data acquisition module, the real-time processing process of the data fusion processing module, the fire warning result, and the audible and visual warning or alarm signals of the fire state; the linkage control interface module is used to connect to the train fire linkage control system to achieve accurate and rapid train fire disposal on the vehicle-mounted side; The transmission network specifically includes 5G and LoRa. Among them, 5G is used for data transmission between the vehicle-mounted side and the cloud under normal vehicle-ground communication conditions, and LoRa is used for data transmission between the vehicle-mounted side and the cloud in a self-organizing network mode when vehicle-ground communication fails; The cloud specifically includes a storage platform and a computing platform. Among them, the storage platform is used to store the real-time data output by the transmission network; the computing platform provides the computing power conditions for combining the stored historical data with the real-time data to achieve fire prediction, is used for predicting the development of the fire situation, reducing the probability of train fires, providing effective guarantees for train emergency disposal, and improving the safe operation level of the train.

[0036] This embodiment also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements a train fire vehicle-cloud warning and prediction method as described above.

[0037] Furthermore, the present invention adopts the following technical solutions: An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a train fire vehicle-cloud warning and prediction method as described above.

[0038] Through the description of the above embodiments, those skilled in the art can clearly understand that the facilities of the present invention can be implemented by means of software plus a necessary general hardware platform. The embodiments of the present invention can be implemented using existing processors, or by dedicated processors used for this purpose or other purposes in a suitable system, or by a hardwired system. The embodiments of the present invention also include non-transitory computer-readable storage media, which include machine-readable media for carrying or having machine-executable instructions or data structures stored thereon; such machine-readable media can be any available media accessible by a general or special-purpose computer or other machine with a processor. For example, such machine-readable media can include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disk memories, magnetic disk memories or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of machine-executable instructions or data structures and can be accessed by a general or special-purpose computer or other machine with a processor. When information is transmitted or provided to a machine through a network or other communication connection (hardwired, wireless, or a combination of hardwired and wireless), the connection is also regarded as a machine-readable medium.

[0039] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

Claims

1. A train fire vehicle cloud warning prediction method, characterized in that: The steps include: Constructing a multimodal fire warning model and a multimodal fire situation prediction model based on deep learning, and training the multimodal fire warning model and the multimodal fire situation prediction model to obtain the trained multimodal fire warning model and the multimodal fire situation prediction model; The sensors installed on the train are used to obtain multi-source heterogeneous real-time data, including temperature data, smoke data, wind speed data, and fire image data. The multi-source heterogeneous real-time data are extracted and fused, and the fused features are input into the trained multimodal fire warning model to give the real-time fire state probabilities of no fire, smoldering, and open flames. The highest probability state is output as the real-time fire warning result, and the sound and light warning or alarm signal of the fire state is output. The multi-source heterogeneous real-time data obtained through sensors are transmitted to the cloud, which receives and stores the real-time data. The stored historical data is combined with the real-time data, and the trained multimodal fire situation prediction model is used to extract features and integrate all multi-source heterogeneous data to obtain a fire situation development forecast.

2. A train fire vehicle cloud warning prediction method according to claim 1, characterized in that: The multi-source heterogeneous real-time data is specifically: The temperature data is collected by an optical fiber temperature sensor, which measures the temperature along the entire length of the train compartment and is one-dimensional spatial data; The smoke data is collected by a smoke sensor, and the wind speed data is collected by a wind speed sensor. The smoke sensor and wind speed sensor collect smoke and wind speed in the train compartment respectively, and both are one-dimensional time data measured at a spatial point; The fire image data is collected by a visible light image sensor and / or an infrared image sensor to measure a plane image inside the train compartment, which is 2D spatial data.

3. A train fire vehicle cloud warning prediction method according to claim 1, characterized in that: The method extracts and fuses features from multi-source heterogeneous real-time data, and then inputs the fused features into the trained multimodal fire warning model to give the real-time fire state probabilities of no fire, smoldering and open flames. The highest probability state output is the real-time fire warning result: (1) Feature extraction of multi-source heterogeneous real-time data through feature extraction network: The temperature data is a one-dimensional spatial sequence. The temperature data is first transformed into two-dimensional image-like data using a discrete wavelet transform method, and then a convolutional neural network (CNN) is used to extract features. The convolutional neural network (CNN) includes a convolutional layer, an attention module, a pooling layer, and a normalization layer. The smoke data and wind speed data are one-dimensional time series data of spatial points, and features are extracted using an LSTM model with a self-attention mechanism. The LSTM model includes an LSTM layer, an attention layer, and a fully connected layer. The fire image data is two-dimensional spatial data, and features are extracted using an improved VGG16 model, wherein the VGG16 model includes a convolutional layer, an attention module, a pooling layer, and a normalization layer; (2) The extracted temperature features, smoke features, wind speed features, and fire image features are used as input elements and input into the constructed MS-FRFM network feature fusion module for feature fusion to obtain the fusion features of multi-source heterogeneous real-time data; (3) The fused features are input into a multi-layer perceptron (MLP) network model with two hidden layers for decision making. The output neurons are the real-time fire state probabilities of no fire, smoldering, and open flames. The state with the highest probability is the real-time fire warning result.

4. A train fire vehicle cloud warning prediction method according to claim 1, characterized in that: The trained multimodal fire situation prediction model is used to extract features and fuse all multi-source heterogeneous data to obtain the method for predicting the development of the fire situation: First, feature extraction is performed based on the Transformer architecture, and then a cross-modal cross-attention mechanism is used to perform modal interaction to achieve feature fusion. Finally, the fused features are used as input elements, and a two-layer long short-term memory network LSTM and a multi-layer perceptron MLP network model are used to obtain a fire situation development prediction.

5. A train fire vehicle cloud warning prediction system, characterized in that: The system comprises: On-board controller, which realizes real-time train fire warning and linkage train fire linkage control system based on multi-modal fire warning model; Transmission network, which transmits real-time data from the vehicle to the cloud; The cloud receives and stores real-time data through a storage platform, and realizes the prediction of train fire situation development based on a multi-modal fire situation prediction model through a computing platform.

6. A train fire vehicle cloud warning prediction system according to claim 5, characterized in that: The vehicle-mounted controller specifically includes: The data acquisition module collects multi-source heterogeneous real-time data through sensors installed on the train, including temperature data, smoke data, wind speed data, and fire image data; The data fusion processing module extracts and fuses features from multi-source heterogeneous real-time data, and then inputs the fused features into the decision model to give the real-time fire state probabilities of no fire, smoldering and open flames. The highest probability state output is the real-time fire warning result; Display module, which displays the real-time data collected by the data acquisition module, the real-time processing process of the data fusion processing module and the fire warning results, and the sound and light warning or alarm signal of the fire status; The linkage control interface module is used to connect to the train fire linkage control system to achieve accurate and rapid train fire handling on the vehicle side.

7. A train fire vehicle cloud warning prediction system according to claim 5, characterized in that: The transmission network specifically includes: 5G network, used for data transmission between the vehicle and the cloud; The LoRa network is used in a self-organizing manner for data transmission between the vehicle and the cloud when vehicle-to-ground communication fails.

8. A train fire vehicle cloud warning prediction system according to claim 5, characterized in that: The cloud specifically includes: Storage platform, used to store real-time data output by the transmission network; The computing platform provides the computing power to combine stored historical data with real-time data to achieve fire prediction, which is used to predict the development of fire situations.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a train fire vehicle cloud warning prediction method as described in any one of claims 1 to 4 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, a train fire vehicle cloud warning prediction method as described in any one of claims 1 to 4 is implemented.

Images