Multi-sensor dynamic fusion fire monitoring and identification method and system

Abstract

The invention relates to a multi-sensor dynamic fusion fire monitoring and identification method and system, and the method comprises the steps: S1, collecting fire sample data under different working conditions, and obtaining a sample set; s2, selecting video data with the maximum information amount as input of a multi-sensor dynamic fusion learning model, preprocessing a sample set, and calculating a probability value representing whether flames exist or not by using a # imgabs0 # function; s3, constructing a correction function # imgabs 1 #, fitting a curve graph based on physical data to calculate a correction value # imgabs 2 #, and finely adjusting the preliminary prediction result without changing the gradient of the original tensor; and S4, adjusting model output by using a correction value # imgabs3 # and a self-defined loss function # imgabs4 #, controlling the adjustment amplitude of a predicted value to pass through a probability coefficient # imgabs5 #, calculating a loss value by using a negative logarithm likelihood loss function, and performing back propagation to optimize model parameters. And the response speed and accuracy of the fire monitoring and identification system are improved.

Description

Technical Field

[0001] The present invention relates to the fields of fire and smoke detection, multi-source data analysis and image-type fire detectors, and in particular to a multi-sensor dynamic fusion fire monitoring and identification method and system. Background Art

[0002] Fire is one of the most common and most harmful disasters in real life, which is directly related to people's life safety and property safety. In order to reduce the serious consequences of fire, it is crucial to develop efficient fire monitoring technology. Existing fire detection technologies, such as smoke and heat fire detectors, although they show certain efficiency in their respective application scenarios, still have obvious limitations. For example, smoke fire detectors cannot effectively alarm when there is no smoke or the smoke concentration is low; heat fire detectors require the ambient temperature to reach a certain threshold to trigger an alarm, and the early detection effect of smoldering fires is not good; image-based fire detectors cannot detect fires when there are obstructions that block the fire source and smoke.

[0003] With the rapid development of technologies such as the Internet of Things, artificial intelligence, and big data analysis, multi-source data fire detection technology has emerged, providing a new solution. However, the multimodal fire detection system in the existing technology usually relies on a single data processing process or complex threshold settings, which often makes it difficult to achieve optimal performance in dynamic and complex actual environments, and fails to effectively utilize the complementarity of various sensor data. The impact of environmental factors on sensor performance is usually not considered, resulting in unstable detection performance in complex environments and failure to achieve ideal fire warning effects.

[0004] The patent application published in the patent application "A comprehensive flame detection method based on ultraviolet, infrared and vision (CN112069975A)" uses ultraviolet and infrared detectors to detect the current environment respectively, and its trigger signal is connected to the counter through the AND gate; in addition, the current image is obtained through the camera, and the image is analyzed to determine whether there is a fire in the current environment. If there is a fire, the current fire level is judged by the area of ​​the flame and the brightness of the surrounding environment. This invention relies entirely on thresholds for fire judgment. Although the setting is relatively simple, it is difficult to set a suitable threshold in complex scenes. The patent application published in the patent application "A flame detection system based on multimodal information fusion technology (CN115620102A)" uses multiple sensors such as infrared, binocular cameras and TOF sensors. The system mainly relies on static fusion strategies and preset thresholds, which may cause the system to respond in a timely or accurate manner under dynamically changing environmental conditions. In addition, the description of the system in terms of real-time data processing and feedback adjustment is not clear enough, and it may not be efficient when processing large-scale real-time changing scenes.

[0005] Therefore, there is an urgent need to develop a fire monitoring and identification method and system that can dynamically fuse multi-sensor data under changing environmental conditions to improve the accuracy and robustness of fire monitoring. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a multi-sensor dynamic fusion fire monitoring and identification method, which includes four steps: analyzing the actual significance of data, selecting a main model, setting a correction function, and correcting a loss function.

[0007] The technical solution of the present invention is: a multi-sensor dynamic fusion fire monitoring and identification method, comprising the following steps:

[0008] Step S1: using multiple sensors including electric smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras to collect fire sample data under different working conditions, and obtain a sample set, wherein the sample set includes physical and video data: smoke data, temperature data, flame radiation data, visible light video data and infrared video data, and each sample is labeled with a label of flame or no flame;

[0009] Step S2: Based on the analysis results of the physical meaning of multi-source sensor data, the video data mode carrying key information is selected as mode one and input into the multi-sensor dynamic fusion learning model to obtain two prediction values, which represent the flame situation of the current scene. The two prediction values ​​are used to The function is normalized to obtain the probability values ​​of two opposing events: fire and no fire , represent the probability of flame and the probability of no flame respectively;

[0010] Step S3: Use the unselected physical data of mode 1 as the input of mode 2, and construct a correction function based on the physical data feature fitting and the prediction value of the current scene obtained by training the data through a neural network. , for two probability values Fine-tune the data without changing the gradient of the original tensor. Select the physical data sampling in the middle of a video sample as a representative, and use its physical meaning to fit the data-fire occurrence curve. Find a suitable function expression based on the curve to calculate the correction value. , adjust the probability value;

[0011] Step S4: Use the correction function Calculated correction value And a custom loss function The probability value obtained from the multi-sensor dynamic fusion learning model Perform normalization adjustment to obtain the adjusted , by giving the correction value Add a probability coefficient in the range (0,1) , regulate the adjustment range of the predicted value to ensure the advantage of the multi-sensor dynamic fusion learning model output, use the negative log-likelihood loss function to calculate the loss value, and perform back propagation to the model parameters.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] (1) A multi-source data fire detection dataset that is lacking in the field of fire detection was constructed; a new model for predicting fire conditions was constructed, the multi-sensor dynamic fusion learning model (DFM). This model uses multiple sensors including electrical smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras, and infrared video cameras to dynamically obtain flame feature data, and then trains the physical data and video data through neural networks to obtain the predicted value of the current fire situation, which can monitor the development of fire and potential risks in real time.

[0014] (2) The prediction results of the model constructed by single modal data have limitations, with varying degrees of false positives or omissions, and cannot achieve an accuracy of "1". The multi-source sensor data fusion model integrates multiple modal data, with rich data types and high dimensions, and the prediction results can all achieve an accuracy of "1". The multi-sensor dynamic fusion learning model proposed in the present invention has a higher confidence level among all the proposed fusion models, with a confidence level 0.05 to 0.1 higher than that of the traditional fusion model, and in the case of data anomalies, it can obtain better prediction capabilities than other fusion models.

[0015] (3) The present invention discloses a multi-sensor dynamic fusion fire monitoring and identification system, which can monitor flame and smoke signals simultaneously, avoiding the shortcomings of relying on a single detection method and significantly improving the reliability of fire monitoring. Compared with traditional fire detectors using the threshold method, the present system uses deep learning technology to analyze multi-source data, which can more fully extract fire characteristics and effectively distinguish fire targets from other interferences, thereby improving the accuracy of identification. The present system uses a design with multiple interfaces and independent coupling of functional modules, which not only improves practicality, but also facilitates the subsequent transformation and upgrading of equipment, thereby promoting the intelligence of fire-fighting equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a multi-sensor dynamic fusion fire monitoring and identification method in an embodiment of the present invention;

[0017] Figure 2 Schematic diagram of video data acquisition under different working conditions in an embodiment of the present invention;

[0018] Figure 3Schematic diagram of a multi-sensor dynamic fusion learning model in an embodiment of the present invention;

[0019] Figure 4 This is an overall structural diagram of a multi-sensor dynamic fusion learning model in an embodiment of the present invention;

[0020] Figure 5 This is a structural block diagram of a multi-sensor dynamic fusion fire monitoring and identification system based on deep learning in an embodiment of the present invention.

[0021] In the figure, the reference numerals are:

[0022] Multimodal data acquisition unit 1, signal input unit 2, data processing unit 3, signal output unit 4, alarm unit 5, power supply module 6. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0024] The present invention provides a multi-sensor dynamic fusion fire monitoring and identification method, which obtains fire monitoring data through multiple channels, better integrates different types of data, and performs fire alarm more quickly.

[0025] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below through specific implementations and in conjunction with the accompanying drawings.

[0026] Embodiment 1

[0027] like Figure 1 As shown, a multi-sensor dynamic fusion fire monitoring and identification method provided by an embodiment of the present invention includes the following steps:

[0028] Step S1: using multiple sensors including electric smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras to collect fire sample data under different working conditions, and obtain a sample set, wherein the sample set includes physical and video data: smoke data, temperature data, flame radiation data, visible light video data and infrared video data, and each sample is labeled with a label of flame or no flame;

[0029] Step S2: Based on the analysis results of the physical meaning of multi-source sensor data, the video data mode carrying key information is selected as mode one and input into the multi-sensor dynamic fusion learning model to obtain two prediction values, which represent the flame situation of the current scene. The two prediction values ​​are used to The function is normalized to obtain the probability values ​​of two opposing events: fire and no fire , represent the probability of flame and the probability of no flame respectively;

[0030] Step S3: Use the unselected physical data of mode 1 as the input of mode 2, and construct a correction function based on the physical data feature fitting and the prediction value of the current scene obtained by training the data through a neural network. , for two probability values Fine-tune the data without changing the gradient of the original tensor. Select the physical data sampling in the middle of a video sample as a representative, and use its physical meaning to fit the data-fire occurrence curve. Find a suitable function expression based on the curve to calculate the correction value. , adjust the probability value;

[0031] Step S4: Use the correction function Calculated correction value And a custom loss function The probability value obtained from the multi-sensor dynamic fusion learning model Perform normalization adjustment to obtain the adjusted , by giving the correction value Add a probability coefficient in the range (0,1) , regulate the adjustment range of the predicted value to ensure the advantage of the multi-sensor dynamic fusion learning model output, use the negative log-likelihood loss function to calculate the loss value, and perform back propagation to optimize the model parameters.

[0032] In one embodiment, the above step S1: according to the existing national standards, several combustibles such as cotton rope, wood, n-heptane, polyurethane, n-heptane, etc. are selected for ignition experiments, and multiple sensors including electric smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras are used to dynamically obtain flame feature data, obtain fire sample data under different working conditions, and obtain a sample set, which includes five types of data: smoke data, temperature data, flame radiation data, visible light video data and infrared video data, and each sample is marked with a label with flame or no flame, and a total of 36 groups of experimental data are obtained, and the video duration exceeds 350 minutes; these sample data from different dimensions can fully play a complementary role, and they are comprehensively analyzed to determine whether they show a strong correlation, specifically including:

[0033] like Figure 2As shown, several combustibles such as cotton rope, wood, n-heptane, polyurethane, and n-heptane were selected for ignition experiments. Multiple sensors including electric smoke sensors, temperature-sensitive fire detectors, flame radiation detectors, visible light video cameras, and infrared video cameras were used to collect fire flame and smoke data under interference-free conditions, as well as flame and smoke data under different wind speeds, light intensities, experimental positions, and experimental dosages. The data without flames were labeled as 0, and the data with flames were labeled as 1. According to the scenes that are easily misreported in reality, two negative samples, stage smoke and fine water mist, were designed for interference experiments to obtain video data in laboratory scenes and corresponding physical data such as smoke values, temperature values, and flame radiation values. Among them, Figure 2 (a) is the wood burning video data, Figure 2 (b) is the video data of cotton rope burning. Figure 2 (c) is the video data of polyurethane foam combustion. Figure 2 (d) is the video data of n-heptane combustion. Figure 2 (e) is the stage smoke video data, Figure 2 (f) is the water mist video data, one of the interference items is whether the visible light video is blocked. When the visible light video data is completely black, that is, there is an obstruction, but there is flame smoke at this time, the video sample should also be marked as having flame. ,in For smoke data, is the temperature data, is the flame radiation data, is the visible light video data, Infrared video data.

[0034] In one embodiment, the above step S2: based on the multi-source sensor data in step S1 The results of the physical analysis reveal the response patterns of the sensor data under different fire conditions. The video data mode carrying the key information is selected as mode one and input into the multi-sensor dynamic fusion learning model. The model uses multiple sensors including electric smoke sensors, temperature-sensitive fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras to dynamically obtain flame feature data, and the physical data and video data are trained by the neural network to obtain the predicted value of the current fire situation. It can monitor the development of the fire and potential risks in real time. The key information is the flame feature in the video. After being processed by the neural network, two predicted values ​​are output to represent the flame situation of the current scene. The results are used The function is normalized to obtain the probability values ​​of these two opposing events, which represent the probability of flame and the probability of no flame, respectively. Specifically, Figure 3 As shown, based on multi-source sensor data The analysis results of physical meaning,Considering that video data usually provides rich information about the dynamic changes of fire, the video data carrying the most critical information is selected as the input of the multi-sensor dynamic fusion learning model, such as Figure 4 As shown in the figure, the video data first passes through four 3D convolution blocks. Each convolution block consists of four parts, namely 3D convolution layer, BN (Batch Normalization) layer, activation function (select ReLU) and 3D pooling layer, to obtain the feature vector, and then pass it through two fully connected layer networks to obtain the predicted value of the current scene, indicating the fire situation of the current scene. The result is used After normalization, the function outputs the average prediction value, and obtains the probability values ​​of two opposing events: fire and no fire , representing the probability of fire and the probability of no fire respectively, and the sum of the two probabilities is 1.

[0035] In one embodiment, the above step S3: using the physical data not selected in mode 1 in step S2 as the input of mode 2, the physical data set includes smoke value, temperature value and flame radiation value including 5.0 Electromagnetic wave wavelength data, 4.4 Electromagnetic wave wavelength data, 3.8 There are 6 sensor data, including electromagnetic wave wavelength data and ultraviolet count value. According to the physical data characteristics, the data is fitted and trained through neural network, such as Figure 4 As shown in the figure, these 6 physical data values ​​are used as input. The input data has 6 columns and 60 rows of physical values, forming a 60×6 matrix. This matrix is ​​used as a feature map. It first passes through 3 2D convolution blocks. Each convolution block consists of 4 parts, namely 2D convolution layer, BN (Batch Normalization) layer, activation function (select ReLU) and 2D pooling layer to obtain the feature vector. Then, it passes through two fully connected layer networks to obtain the predicted value of the current scene. Use the obtained predicted value to construct the correction function , the probability values ​​of the two results are fine-tuned, but the adjustment process does not change the gradient of the original tensor, and by selecting the middle physical data sampling of a video sample as a representative, the data-fire occurrence curve is fitted using its physical meaning, and the correction value is calculated based on the curve to find a suitable function expression , adjust the results, including:

[0036] like Figure 3 As shown in the figure, the physical data sampling in the middle of a video sample is selected as a representative, and the smoke value of this sampling is taken as the , temperature value , UV count value As an independent variable get , where the UV count value is the wavelength data collected by the flame radiation detector, and the main factor is 4.4 The wavelength and UV count data are strongly correlated and have the same ability to express fire scenes, so the UV count is used. participate Function construction is shown in formula (1)(2).

[0037] design The function is used to improve the single mode problem of using video data alone, and to add data of other modes to reflect the current flame situation. When there is fire, the function value should be a number less than 1, and when there is no fire, it should be equal to or close to 0. Therefore, the value range of this function should be within According to the video data used, the two predicted values ​​and data output after the neural network training are Launch of prediction results with better and As shown in formulas (1) and (2), we obtain the models DFM-NET1 and DFM-NET2 respectively. and In and The minimum value of the two terms is 0 to ensure the non-negativity of the function value.

[0038] (1)

[0039] (2)

[0040] In one embodiment, the above step S4: using the correction function Calculated correction value And a custom loss function The probability values ​​obtained for the main model Perform normalization adjustment to obtain the new adjusted , by giving the correction value Add a probability coefficient in the range (0,1) , adjust the adjustment range of the predicted value to ensure the advantage of the output of the multi-sensor dynamic fusion learning main model, use the negative log-likelihood loss function to calculate the loss value, and perform back propagation to optimize the parameters of the multi-sensor dynamic fusion learning model, including:

[0041] like Figure 3 As shown, using the correction function Calculated correction value The prediction results obtained by the multi-sensor dynamic fusion learning model are used with the function Normalize and adjust the probability value The calculation method of is shown in formula (3), and then the new adjusted In the calculation process, Adding probability coefficients It is used to affect the adjustment range of the predicted value. The fusion result is mapped to the range of 0-1 according to the proportion, so that it represents the probability of having a flame after adjustment. The value range of the probability coefficient is (0,1). Then the negative log-likelihood loss is used to calculate its loss function value to obtain the loss function As shown in formula (4), the final The values ​​are back-propagated to optimize the parameters of the multi-sensor dynamic fusion learning model.

[0042] (3)

[0043] (4)

[0044] Embodiment 2

[0045] like Figure 5 As shown, the embodiment of the present invention provides a multi-sensor dynamic fusion fire monitoring and identification system, including the following modules:

[0046] The multimodal data acquisition unit 1 is used to collect multi-channel data including: electric smoke sensor, temperature fire detector, flame radiation detector, visible light video camera and infrared video camera;

[0047] The multimodal data acquisition unit contains a variety of sensors for collecting data from the fire monitoring site. The sampling frequency can be set according to actual needs, and the collected data is sent to the signal input unit through the interface;

[0048] The signal input unit 2 includes a plurality of serial interfaces and a plurality of video signal interfaces, and is used to receive the multi-channel data collected by the multi-modal data collection unit;

[0049] The data processing unit 3 processes the multi-channel data input by the signal input unit in real time, performs real-time data preprocessing, uses a multi-sensor dynamic fusion learning model to extract features, and obtains flame detection results;

[0050] The module contains a processor that can process multiple input data in real time, including one-dimensional sensor signals and two-dimensional video image signals, and can perform real-time signal preprocessing, signal analysis based on deep learning algorithms, and has multiple serial port data outputs;

[0051] A signal output unit 4, used to output flame detection results;

[0052] The output result is 1 or 0. 1 indicates that a fire has occurred and the signal is transmitted to the alarm unit for alarm; 0 indicates that no fire has occurred and fire monitoring will continue.

[0053] Alarm unit 5, including light emitting diode and buzzer, used for fire alarm prompt;

[0054] If the signal output unit receives a signal of 1, it indicates that a fire has occurred, and the alarm model is driven to emit a sound and a flashing light emitting diode alarm.

[0055] The power module 6 is used to provide power to the multimodal data acquisition unit, the data processing unit and the alarm unit.

[0056] The data processing unit 3 uses the above-mentioned multi-sensor dynamic fusion fire monitoring and identification method to process the multi-channel data input by the signal input interface unit in real time.

[0057] The present invention discloses a fire monitoring and identification method and system with dynamic fusion of multiple sensors. Through the design of multiple interfaces and independent coupling of functional modules, the practicability of the system is improved, which is conducive to the subsequent equipment modification. In the data processing part, deep learning methods will be used for information analysis to realize fire detection and identification and promote the intelligence of fire fighting equipment.

[0058] The above embodiments are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present invention should all be included within the scope of the present invention.

Claims

1. A fire monitoring and identification method based on dynamic fusion of multiple sensors, characterized in that: The steps include: Step S1: using multiple sensors including electric smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras to collect fire sample data under different working conditions, and obtain a sample set, wherein the sample set includes physical and video data: smoke data, temperature data, flame radiation data, visible light video data and infrared video data, and each sample is labeled with a label of flame or no flame; Step S2: Based on the analysis results of the physical meaning of multi-source sensor data, the video data mode carrying key information is selected as mode 1 and input into the multi-sensor dynamic fusion learning model to obtain two prediction values, which represent the flame situation of the current scene. The two prediction values ​​are used The function is normalized to obtain the probability values ​​of two opposing events: fire and no fire , represent the probability of flame and the probability of no flame respectively; Step S3: Use the unselected physical data of mode 1 as the input of mode 2, and construct a correction function based on the physical data feature fitting and the current scene prediction value obtained through neural network training , for two probability values Fine-tune the data without changing the gradient of the original tensor. Select the physical data sampling in the middle of a video sample as a representative, and use its physical meaning to fit the data-fire occurrence curve. Find a suitable function expression based on the curve to calculate the correction value. , adjust the probability value; Step S4: Use the correction function Calculated correction value And a custom loss function The probability value obtained from the multi-sensor dynamic fusion learning model Perform normalization adjustment to obtain the adjusted , by giving the correction value Add a probability coefficient in the range (0,1) , regulate the adjustment range of the predicted value to ensure the advantage of the multi-sensor dynamic fusion learning model output, use the negative log-likelihood loss function to calculate the loss value, and perform back propagation to the model parameters.

2. A multi-sensor dynamic fusion fire monitoring and identification method according to claim 1, characterized in that: The step S1 specifically includes: Several combustible materials including cotton rope, wood, n-heptane, polyurethane, and n-heptane were selected for ignition experiments. Fire flame and smoke data under interference-free conditions were collected based on multiple sensors including electric smoke sensors, temperature-sensitive fire detectors, flame radiation detectors, visible light video cameras, and infrared video cameras. Flame and smoke data were also collected under different wind speeds, light intensities, experimental positions, and experimental dosages. The data without flames were labeled as 0, and the data with flames were labeled as 1. Two negative samples, stage smoke and fine water mist, were designed for interference experiments to obtain video data in the experimental scene and the corresponding physical data of smoke values, temperature values, and flame radiation values. One of the interference items is whether the visible light video is blocked. When the visible light video data is completely black, that is, there is an obstruction but there is flame smoke at this time, the video sample should also be labeled as having flames, thereby constructing sample data. ,in For smoke data, is the temperature data, is the flame radiation data, is the visible light video data, Infrared video data.

3. A multi-sensor dynamic fusion fire monitoring and identification method according to claim 2, characterized in that: The step S2 specifically includes: The video data carrying key information is selected as the modality-input multi-sensor dynamic fusion learning model. The key information is the flame feature in the video. The video data is first passed through four 3D convolution blocks. Each convolution block consists of four parts, namely 3D convolution layer, BN layer, ReLU activation function and 3D pooling layer, to obtain the feature vector. Then, it passes through two fully connected layer networks to obtain the predicted value of the current scene, indicating the fire situation of the current scene. The result is used After normalization, the function outputs the average prediction value, and obtains the probability values ​​of two opposing events: fire and no fire , representing the probability of fire and the probability of no fire respectively, and the sum of the two probabilities is 1.

4. A multi-sensor dynamic fusion fire monitoring and identification method according to claim 3, characterized in that: The step S3 specifically includes: The physical data not selected in mode 1 is used as the input of the multi-sensor dynamic fusion learning model of mode 2. The physical data includes smoke value, temperature value and flame radiation value. The flame radiation value includes 5.0 Electromagnetic wave wavelength data, 4.4 Electromagnetic wave wavelength data, 3.8 Electromagnetic wave wavelength data and ultraviolet count value; take the above physical data values ​​as input, form a matrix of input data, use the matrix as the feature map, first pass through 3 2D convolution blocks, each convolution block consists of 4 parts, namely 2D convolution layer, BN layer, ReLU activation function and 2D pooling layer, get the feature vector, then pass through two fully connected layer networks to get the predicted value of the current scene, and use the predicted value to build the correction function , and select the physical data sampling in the middle of a video data sample as the representative, and use the smoke value of this sampling , temperature value , UV count value As an independent variable through the correction function Get the corrected value .

5. A multi-sensor dynamic fusion fire monitoring and identification method according to claim 4, characterized in that: Design correction function , select the correction function and , respectively, we get the models DFM-NET1 and DFM-NET2, and the correction function and The formula is shown in (1)(2): (1) (2)。 6. A multi-sensor dynamic fusion fire monitoring and identification method according to claim 5, characterized in that: The step S4 specifically includes: Using Correction Functions Calculated correction value The probability value obtained from the model Use function Normalize and adjust the probability value The calculation method of is shown in formula (3), and then the new adjusted , the calculation process is done by giving the correction value Adding probability coefficients It is used to affect the adjustment range of the predicted value. The fusion result is mapped to the range of 0-1 according to the proportion, so that it represents the probability of having a flame after adjustment. The value range of the probability coefficient is (0,1). Then the negative log-likelihood loss is used to calculate its loss function value to obtain the loss function , and finally get The values ​​are back-propagated to optimize the parameters of the multi-sensor dynamic fusion learning model; (3) (4)。 7. A multi-sensor dynamic fusion fire monitoring and identification system as claimed in any one of claims 1 to 6, characterized in that: Includes the following modules: A multi-modal data acquisition unit for collecting multi-channel data including electric smoke sensors, temperature fire detectors, flame radiation detectors, visible light video cameras and infrared video cameras; A signal input unit, comprising a plurality of serial interfaces and a plurality of video signal interfaces, for receiving the multi-channel data collected by the multi-modal data collection unit; The data processing unit processes the multi-channel data input by the signal input unit in real time, performs real-time data preprocessing, and uses the trained multi-sensor dynamic fusion learning model to extract features to obtain flame detection results; A signal output unit, used for outputting flame detection results; Alarm unit, including light emitting diode and buzzer, for fire alarm prompt; A power supply module, used to provide power to the multimodal data acquisition unit, the data processing unit and the alarm unit; The data processing unit processes multiple channels of data in real time using the method described in any one of claims 1 to 6.

8. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, and when the instructions are executed by a processor, the processor implements the method according to any one of claims 1 to 6.

Images