Fire early warning method and system based on smoke image joint feature analysis

Abstract

The invention provides a fire early warning method and system based on smoke image joint feature analysis, which are applied to fire early warning in an indoor scene, and the method comprises the steps: collecting smoke concentration distribution data and an infrared thermogram of an indoor environment based on a smoke sensor array and an infrared sensor array, and collecting a visible light image of the indoor environment based on a camera; extracting spatio-temporal characteristics of smoke concentration distribution by using a 3D convolutional neural network, and extracting texture and shape characteristics of the visible light image and temperature distribution characteristics of the infrared thermal image by using a 2D convolutional neural network; performing weighted fusion on the three features to generate joint feature representation; expressing the fused joint features as a feature sequence based on a time sequence, and obtaining a fire risk level of a current time window based on a trained multi-layer perceptron classification model; according to the method, the accuracy, timeliness and comprehensiveness of early warning are improved, and fire risks in various complex scenes can be better dealt with.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire warning, and in particular relates to a fire warning method and system based on smoke image joint feature analysis. Background Art

[0002] Existing fire warning technologies mainly rely on single or limited sensor data for judgment. The most common methods include using smoke detectors, temperature sensors or flame detectors; however, these warning methods based on single data sources have obvious defects.

[0003] Although smoke detectors are widely used, normal cooking activities or production processes in kitchens or industrial environments may produce harmless smoke, leading to frequent false alarms. In the early stages of a fire, if the smoke is not enough to trigger the alarm threshold, it may cause missed alarms and delay the best time for rescue. Temperature sensors rely solely on temperature changes to determine the fire condition, which has a lag. In large open spaces or high-ceiling buildings, it may take a long time for heat to reach the sensor location when a fire occurs, resulting in delayed warnings. In addition, some types of fires (such as electrical fires) may not produce significant temperature changes in the early stages, making it difficult for systems that rely solely on temperature sensors to detect potential dangers in a timely manner. Although flame detectors can quickly detect open flames, their effectiveness is greatly reduced in the presence of obstructions. In complex indoor environments, such as offices with partitions or libraries with multi-layer bookshelves, the line of sight of the flame detector may be blocked, resulting in the fire not being discovered in time. Although carbon monoxide detectors are effective in detecting smokeless fires, they cannot cope with all types of fires. For example, in the case of electrical fires or the combustion of certain chemicals, not enough carbon monoxide may be produced to trigger an alarm.

[0004] In addition, existing technologies lack data fusion capabilities, and most systems are unable to effectively integrate multi-source data, resulting in one-sided or inaccurate judgments on fire conditions. Traditional systems are mostly based on simple threshold judgments and lack the ability to comprehensively analyze complex environmental factors. Fixed alarm thresholds are difficult to adapt to changes in different scenarios and environmental conditions, and can easily cause oversensitivity or slow response. Existing systems mainly focus on detecting fires that have already occurred and lack the ability to predict potential fire hazards. Due to the inability to comprehensively analyze environmental factors, false alarms occur frequently, reducing users' trust in the system.

[0005] In view of the above problems, a new fire warning method is needed that can comprehensively analyze multi-source data, has intelligent feature extraction and fusion capabilities, and can adapt to complex environments. Summary of the invention

[0006] The present invention provides a fire warning method and system based on smoke image joint feature analysis, aiming to overcome the limitations of the prior art, improve the accuracy, timeliness and comprehensiveness of the warning, and better cope with fire risks in various complex scenarios.

[0007] In a first aspect, the present invention provides a fire warning method based on joint feature analysis of smoke images, which is applied to fire warning in indoor scenes. The method comprises the following steps: Step S1, setting the data collection frequency of each sensor array and camera to be the same, collecting smoke concentration distribution data and infrared heat map of the indoor environment based on the smoke sensor array and the infrared sensor array, and collecting visible light images of the indoor environment based on the camera.

[0008] Step S2, using a 3D convolutional neural network to extract the spatiotemporal features of smoke concentration distribution, using a 2D convolutional neural network to extract the texture and shape features of the visible light image, and the temperature distribution features of the infrared thermal image; performing weighted fusion on the three features to generate a joint feature representation.

[0009] In step S3, the fused joint features are represented as a feature sequence based on time order, and the fire risk level of the current time window is obtained based on the trained multi-layer perceptron classification model.

[0010] Furthermore, in step S2, a 3D convolutional neural network is used to extract the spatiotemporal characteristics of the smoke concentration distribution. The structure of the 3D convolutional neural network includes: 3 3D convolutional layers, each layer is followed by a BatchNormalization batch normalization function and a ReLU activation function, 2 3D maximum pooling layers and 1 fully connected layer; the input is T-frame smoke concentration distribution data, and the output is a 256-dimensional feature vector.

[0011] The mathematical expression of the 3D convolution operation is: ; in, Indicates that the output feature map is at position Place The value of the output channel; Indicates that the input data is at position Place The value of the input channel; represents the weight of the 3D convolution kernel, represents the spatial position of the convolution kernel, Indicates The bias term for each output channel, Respectively represent the height, width and depth in the convolution kernel space dimension, Represents the total number of channels of input data.

[0012] Furthermore, in step S2, the input of the 3D convolutional neural network is continuous p frames of smoke concentration distribution data, where the value of p ranges from 10 to 30, and the input of the 2D convolutional neural network is a single frame of visible light image or infrared thermal image.

[0013] Furthermore, in step S2, the 2D convolutional neural network adopts the ResNet-50 structure, including: 1 7x7 convolutional layer, 16 residual blocks, each residual block contains 3 convolutional layers, 1 global average pooling layer and 1 fully connected layer; the input is a visible light image or an infrared thermal map, and the output is a 512-dimensional feature vector.

[0014] The mathematical expression of the residual block is: ,in Represents the residual map, which consists of 3 convolutional layers. represents the output of the residual block, represents the input of the residual block, Represents the set of weight parameters in the residual map.

[0015] Furthermore, in step S2, the feature fusion adopts the attention mechanism, and the feature fusion method includes: performing linear transformation on the three features, calculating the attention weights, and performing weighted summation to obtain the fused features.

[0016] The mathematical expression of the linear transformation is: ; The mathematical expression of the attention weight is: ; The fusion feature obtained by weighted summation is expressed as: ; in, Indicates characteristics, They correspond to smoke concentration, visible light image and infrared thermal image features respectively; Indicates the query to be generated ,key Sum The weight matrix of Indicates The query, key and value of each feature; Indicates that from the characteristics To Features The attention weight of represents the transpose of the key vector, which is used to perform dot product operations with the query vector. Represents feature dimension; Represents the final fused features.

[0017] Furthermore, in step S3, the multilayer perceptron classification model includes: an input layer, two hidden layers and an output layer; the input layer contains 1024 neurons, the first hidden layer contains 512 neurons, the second hidden layer contains 256 neurons, both hidden layers use ReLU activation function, the output layer contains 4 neurons, corresponding to 4 risk levels, and uses Softmax activation; the fire risk level is divided into four levels: no risk, low risk, medium risk and high risk.

[0018] ; Input feature vector; Output of the hidden layer; The final output is the probability distribution of risk levels; The weight matrix of each layer; Bias vector for each layer.

[0019] Furthermore, in step S1, the data acquisition frequency of the smoke sensor array, the infrared sensor array and the camera is set to 1 Hz; the resolution of the smoke concentration distribution data collected by the smoke sensor array is 8x8 pixels, the resolution of the infrared heat map collected by the infrared sensor array is 32x24 pixels, and the resolution of the visible light image collected by the camera is 640x480 pixels.

[0020] Furthermore, in step S3, the length of the feature sequence based on time order is L, and the value range of L is 5 to 15.

[0021] In a second aspect, the present invention provides a fire warning system based on joint feature analysis of smoke images, which is used to execute the fire warning method described in the first aspect. The system includes: a data acquisition module, a feature extraction module, a feature fusion module and a risk assessment module connected in sequence.

[0022] The data acquisition module is used to collect smoke concentration distribution data, infrared thermal images and visible light images of the indoor environment. The data acquisition module includes a smoke sensor array, an infrared sensor array and a visible light camera.

[0023] The feature extraction module is used to extract spatiotemporal features, texture and shape features, and temperature distribution features from the collected data. The feature extraction module includes a 3D convolutional neural network and a 2D convolutional neural network.

[0024] The feature fusion module is used to perform weighted fusion on the extracted features to generate a joint feature representation. The feature fusion module adopts a feature fusion algorithm based on an attention mechanism.

[0025] The risk assessment module is used to obtain the fire risk level of the current time window based on the fused feature sequence using a multi-layer perceptron classification model, and includes a trained multi-layer perceptron classification model.

[0026] Furthermore, the system also includes an alarm module for issuing an alarm when a medium risk or a high risk is detected.

[0027] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides comprehensive environmental information by combining smoke concentration distribution, infrared heat map and visible light image, thus significantly improving the accuracy and reliability of fire warning. It adopts 3D and 2D convolutional neural networks to effectively extract spatiotemporal features, texture shape features and temperature distribution features, thus enhancing the system's ability to understand complex environments. It uses a weighted fusion method to adaptively adjust the importance of different features, thus improving the system's adaptability to different fire scenarios. The accuracy, real-time and adaptability of the fire warning of the present invention effectively reduce false alarms and missed alarms, thus providing strong technical support for protecting personal and property safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of a fire warning method based on smoke image joint feature analysis of the present invention; Figure 2 The figure is a schematic diagram of the composition of a fire warning system based on smoke image joint feature analysis of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are 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. Example 1

[0030] like Figure 1 FIG. 1 is a flow chart of a fire warning method based on smoke image joint feature analysis according to the present invention, wherein the method comprises the following steps: Step S1, setting the data collection frequency of each sensor array and camera to be the same, collecting smoke concentration distribution data and infrared heat map of the indoor environment based on the smoke sensor array and the infrared sensor array, and collecting visible light images of the indoor environment based on the camera.

[0031] All sensors and cameras collect data at the same frequency, which ensures that different types of data are synchronized in time, facilitating subsequent feature extraction and fusion. The typical collection frequency can be set to 1 Hz (once per second), which is sufficient to capture dynamic changes in the environment without generating excessive data burden.

[0032] Using an array of multiple smoke sensors, spatially distributed smoke concentration information can be obtained. This is more advantageous than a single-point smoke sensor and can reflect the diffusion pattern and concentration gradient of smoke. The infrared sensor array collects indoor temperature distribution information and can detect possible heat sources and temperature anomaly areas. Infrared heat maps can penetrate smoke and provide information that visible light cameras may not be able to capture. Visible light cameras collect real-time images of the room and can capture the visual characteristics of flames and smoke, as well as personnel activities.

[0033] The data acquisition frequency of the smoke sensor array, infrared sensor array and camera is set to 1 Hz; the resolution of the smoke concentration distribution data collected by the smoke sensor array is 8x8 pixels, the resolution of the infrared heat map collected by the infrared sensor array is 32x24 pixels, and the resolution of the visible light image collected by the camera is 640x480 pixels.

[0034] Step S2, using a 3D convolutional neural network to extract the spatiotemporal features of smoke concentration distribution, using a 2D convolutional neural network to extract the texture and shape features of the visible light image, and the temperature distribution features of the infrared thermal image; performing weighted fusion on the three features to generate a joint feature representation.

[0035] 3D convolutional neural networks are used to process smoke concentration distribution data. 3D convolution can consider both spatial and temporal dimensions, making it very suitable for capturing the dynamic characteristics of smoke diffusion. It can learn the pattern of smoke concentration changes over time and space, which is crucial for early fire detection. 2D convolutional neural networks are used to process visible light images and extract texture and shape features. This can help identify the visual features of flames and smoke, as well as detect abnormal environmental changes. It is also used to process infrared thermal images, extract temperature distribution features, and help identify information such as the location of heat sources and heat diffusion patterns.

[0036] The three features are combined using a weighted fusion method. Attention mechanisms or other advanced fusion algorithms can be used to dynamically adjust the weights of various features according to the current environment conditions to generate the most representative joint feature representation.

[0037] When processing time series data (such as the distribution of smoke concentration in consecutive frames), 3D convolutional neural networks can capture the spatiotemporal features in the data. In the smoke concentration detection task, these features help to accurately determine the existence, concentration and changing trend of smoke. For example, a surveillance video records the process of smoke generation and diffusion in a certain area. By extracting consecutive frames from this video as input, the 3D convolutional neural network can capture the shape changes of smoke in space and its diffusion pattern over time. These features are extracted and converted into 256-dimensional feature vectors, providing strong support for subsequent classification or regression tasks.

[0038] A 3D convolutional neural network is used to extract the spatiotemporal characteristics of smoke concentration distribution. The structure of the 3D convolutional neural network includes: 3 3D convolutional layers, each layer is followed by a BatchNormalization function and a ReLU activation function, 2 3D maximum pooling layers and 1 fully connected layer; the input is T frames of smoke concentration distribution data, and the output is a 256-dimensional feature vector.

[0039] The mathematical expression of the 3D convolution operation is: ; in, Indicates that the output feature map is at position Place The value of the output channel; Indicates that the input data is at position Place The value of the input channel; represents the weight of the 3D convolution kernel, represents the spatial position of the convolution kernel, Indicates The bias term for each output channel, Respectively represent the height, width and depth in the convolution kernel space dimension, Represents the total number of channels of input data.

[0040] The input of the 3D convolutional neural network is continuous p frames of smoke concentration distribution data, where p ranges from 10 to 30. The input of the 2D convolutional neural network is a single frame of visible light image or infrared thermal image.

[0041] 2D convolutional neural networks (such as ResNet-50) can extract detailed features in images, such as texture, edges, and shapes. Taking visible light images as an example, there is a picture of indoor smoke. Through the ResNet-50 network, we can extract the smoke features in the picture, such as the color, shape, and texture of the smoke. These features are converted into 512-dimensional feature vectors for subsequent analysis and judgment.

[0042] The 2D convolutional neural network adopts the ResNet-50 structure, including: 1 7x7 convolutional layer, 16 residual blocks, each residual block contains 3 convolutional layers, 1 global average pooling layer and 1 fully connected layer; the input is a visible light image or an infrared thermal map, and the output is a 512-dimensional feature vector.

[0043] The mathematical expression of the residual block is: ,in Represents the residual map, which consists of 3 convolutional layers. represents the output of the residual block, represents the input of the residual block, Represents the set of weight parameters in the residual map.

[0044] Feature fusion adopts the attention mechanism. The feature fusion method includes: linear transformation of the three features, calculation of attention weights, and weighted summation to obtain fused features.

[0045] The mathematical expression of the linear transformation is: ; The mathematical expression of the attention weight is: ; The fusion feature obtained by weighted summation is expressed as: ; in, Indicates characteristics, They correspond to smoke concentration, visible light image and infrared thermal image features respectively; Indicates the query to be generated ,key Sum The weight matrix of Indicates The query, key and value of each feature; Indicates that from the characteristics To Features The attention weight of represents the transpose of the key vector, which is used to perform dot product operations with the query vector. Represents feature dimension; Represents the final fused features.

[0046] In the smoke detection task, when the smoke density is high, the smoke density feature may be more important; when the smoke is light or difficult to identify, the visible light image and infrared thermal image features may provide more effective supplementary information. Through the attention mechanism, we can dynamically adjust the weights of different features to more accurately judge the presence and density of smoke.

[0047] The specific implementation steps include: data preprocessing, frame extraction of monitoring video to obtain smoke concentration distribution data of continuous frames; at the same time, collecting visible light images and infrared thermal images as supplementary information; feature extraction, using 3D convolutional neural network to extract the spatiotemporal features of smoke concentration distribution data to obtain 256-dimensional feature vectors; using ResNet-50 network to extract image features of visible light images and infrared thermal images to obtain 512-dimensional feature vectors; feature fusion, using attention mechanism to perform linear transformation on three features, calculate attention weights and weighted summation to obtain fused features; classification or regression, inputting the fused features into the classifier or regressor for smoke detection or concentration prediction.

[0048] In step S3, the fused joint features are represented as a feature sequence based on time order, and the fire risk level of the current time window is obtained based on the trained multi-layer perceptron classification model.

[0049] The multilayer perceptron classification model includes: an input layer, two hidden layers and an output layer; the input layer contains 1024 neurons, the first hidden layer contains 512 neurons, the second hidden layer contains 256 neurons, and both hidden layers use ReLU activation function. The output layer contains 4 neurons, corresponding to 4 risk levels, and uses Softmax activation; the fire risk level is divided into four levels: no risk, low risk, medium risk and high risk.

[0050] ; Input feature vector; Output of the hidden layer; The final output is the probability distribution of risk levels; The weight matrix of each layer; Bias vector for each layer.

[0051] The length of the time-based feature sequence is L, and the value of L ranges from 5 to 15. During the model training phase, we use a labeled dataset to optimize the weight and bias parameters. The labels are known fire risk levels, which are associated with the input feature sequence. Through optimization techniques such as back propagation algorithms and gradient descent, we can minimize the error between the predicted risk level and the actual risk level.

[0052] Once the model is trained, we can use it to predict the risk level of new feature sequences. By inputting the fused feature sequence into the model, we can obtain a probability distribution representing the likelihood of each risk level. We can then select the category with the highest probability as the predicted risk level; this fire risk level assessment method based on a multi-layer perceptron is suitable for various monitoring scenarios, such as industrial plants, warehouses, residential areas, etc. By analyzing the smoke concentration, visible light image, and infrared thermal map features in the surveillance video in real time, we can detect potential fire risks in a timely manner and take corresponding preventive measures to reduce losses. Example 2

[0053] like Figure 2 As shown, it is a schematic diagram of the composition of a fire warning system based on smoke image joint feature analysis of the present invention, and the system includes: a data acquisition module, a feature extraction module, a feature fusion module and a risk assessment module connected in sequence, which are used to execute the fire warning method described in the first aspect.

[0054] The data acquisition module is used to collect smoke concentration distribution data, infrared thermal images and visible light images of the indoor environment. The data acquisition module includes a smoke sensor array, an infrared sensor array and a visible light camera.

[0055] The feature extraction module is used to extract spatiotemporal features, texture and shape features, and temperature distribution features from the collected data. The feature extraction module includes a 3D convolutional neural network and a 2D convolutional neural network.

[0056] The feature fusion module is used to perform weighted fusion on the extracted features to generate a joint feature representation. The feature fusion module adopts a feature fusion algorithm based on an attention mechanism.

[0057] The risk assessment module is used to obtain the fire risk level of the current time window based on the fused feature sequence using a multi-layer perceptron classification model, and includes a trained multi-layer perceptron classification model.

[0058] The system further comprises an alarm module for issuing an alarm when a medium risk or a high risk is detected.

[0059] The system also includes a data preprocessing module that can complete the cleaning and enhancement of the collected data; data cleaning is to add a data cleaning step after the data acquisition module to remove noise, outliers and duplicate data to ensure the quality of data input to the feature extraction module; data enhancement is to enhance the infrared thermal map and visible light image data, such as rotation, scaling, flipping, etc., to increase the generalization ability of the model and reduce the risk of overfitting.

[0060] The fire warning system proposed in the present invention realizes real-time monitoring and accurate warning of indoor smoke through close cooperation of various modules, providing strong protection for the safety of people's lives and property.

[0061] It should be noted that those skilled in the art should understand that various changes and equivalent substitutions may be made to the present invention without departing from the scope of the present invention. In addition, various modifications may be made to the present invention for specific situations or materials without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed, but should include all embodiments falling within the scope of the claims of the present invention.

Claims

1. A fire warning method based on joint feature analysis of smoke images, applied to fire warning in indoor scenes, characterized in that: The method comprises the following steps: Step S1, setting the data collection frequency of each sensor array and camera to be the same, collecting smoke concentration distribution data and infrared heat map of the indoor environment based on the smoke sensor array and the infrared sensor array, and collecting visible light image of the indoor environment based on the camera; Step S2, using a 3D convolutional neural network to extract the spatiotemporal features of smoke concentration distribution, and using a 2D convolutional neural network to extract the texture and shape features of the visible light image, and the temperature distribution features of the infrared thermal image; performing weighted fusion on the three features to generate a joint feature representation; In step S3, the fused joint features are represented as a feature sequence based on time order, and the fire risk level of the current time window is obtained based on the trained multi-layer perceptron classification model.

2. A fire warning method based on smoke image joint feature analysis according to claim 1, characterized in that: In step S2, a 3D convolutional neural network is used to extract the spatiotemporal features of the smoke concentration distribution. The structure of the 3D convolutional neural network includes: 3 3D convolutional layers, each layer is followed by a BatchNormalization function and a ReLU activation function, 2 3D maximum pooling layers and 1 fully connected layer; the input is T frames of smoke concentration distribution data, and the output is a 256-dimensional feature vector; The mathematical expression of the 3D convolution operation is: ； in, Indicates that the output feature map is at position Place The value of the output channel; Indicates that the input data is at position Place The value of the input channel; represents the weight of the 3D convolution kernel, represents the spatial position of the convolution kernel, Indicates The bias term for each output channel, Respectively represent the height, width and depth in the convolution kernel space dimension, Represents the total number of channels of input data.

3. The fire warning method based on smoke image joint feature analysis according to claim 2 is characterized in that: In step S2, the 2D convolutional neural network adopts the ResNet-50 structure, including: 1 7x7 convolutional layer, 16 residual blocks, each residual block contains 3 convolutional layers, 1 global average pooling layer and 1 fully connected layer; the input is a visible light image or an infrared thermal map, and the output is a 512-dimensional feature vector; The mathematical expression of the residual block is: ,in Represents the residual map, which consists of 3 convolutional layers. represents the output of the residual block, represents the input of the residual block, Represents the set of weight parameters in the residual map.

4. The fire warning method based on smoke image joint feature analysis according to claim 3 is characterized in that: In step S2, the feature fusion adopts the attention mechanism, and the feature fusion method includes: performing linear transformation on the three features, calculating the attention weights, and obtaining the fusion features by weighted summation; The mathematical expression of the linear transformation is: ； The mathematical expression of the attention weight is: ； The fusion feature obtained by weighted summation is expressed as: ； in, Indicates characteristics, They correspond to smoke concentration, visible light image and infrared thermal image features respectively; Indicates the query to be generated ,key Sum The weight matrix of Indicates The query, key and value of each feature; Indicates that from the characteristics To Features The attention weight of represents the transpose of the key vector, which is used to perform dot product operations with the query vector. Represents feature dimension; Represents the final fused features.

5. The fire warning method based on smoke image joint feature analysis according to claim 4 is characterized in that: In step S3, the multilayer perceptron classification model includes: an input layer, two hidden layers and an output layer; the input layer includes 1024 neurons, the first hidden layer includes 512 neurons, the second hidden layer includes 256 neurons, both hidden layers use ReLU activation function, the output layer includes 4 neurons, corresponding to 4 risk levels, respectively, using Softmax activation; the fire risk level is divided into four levels: no risk, low risk, medium risk and high risk; ； Input feature vector; Output of the hidden layer; The final output is the probability distribution of risk levels; The weight matrix of each layer; Bias vectors for each layer.

6. A fire warning method based on smoke image joint feature analysis according to claim 5, characterized in that: In step S1, the data acquisition frequency of the smoke sensor array, the infrared sensor array and the camera is set to 1 Hz; the resolution of the smoke concentration distribution data collected by the smoke sensor array is 8x8 pixels, the resolution of the infrared heat map collected by the infrared sensor array is 32x24 pixels, and the resolution of the visible light image collected by the camera is 640x480 pixels.

7. The fire warning method based on smoke image joint feature analysis according to claim 6 is characterized in that: In step S2, the input of the 3D convolutional neural network is continuous p frames of smoke concentration distribution data, where the value of p ranges from 10 to 30, and the input of the 2D convolutional neural network is a single frame of visible light image or infrared thermal image.

8. The fire warning method based on smoke image joint feature analysis according to claim 1 is characterized in that: In step S3, the length of the feature sequence based on time order is L, and the value range of L is 5 to 15.

9. A fire warning system based on smoke image joint feature analysis, used to execute the fire warning method according to any one of claims 1 to 8, characterized in that: The system comprises: A data acquisition module, used to collect smoke concentration distribution data, infrared thermal images and visible light images of the indoor environment, the data acquisition module includes a smoke sensor array, an infrared sensor array and a visible light camera; A feature extraction module, used to extract spatiotemporal features, texture and shape features, and temperature distribution features from the collected data, wherein the feature extraction module includes a 3D convolutional neural network and a 2D convolutional neural network; A feature fusion module is used to perform weighted fusion on the extracted features to generate a joint feature representation. The feature fusion module adopts a feature fusion algorithm based on an attention mechanism; The risk assessment module is used to obtain the fire risk level of the current time window based on the fused feature sequence using a multi-layer perceptron classification model, including a trained multi-layer perceptron classification model.

10. The fire warning system based on smoke image joint feature analysis according to claim 9, characterized in that: The system further comprises an alarm module for issuing an alarm when a medium risk or a high risk is detected.

Images