Smoke and fire detection method and system based on deep reinforcement learning and feature fusion

Abstract

The invention discloses a smoke and fire detection method and system based on deep reinforcement learning and feature fusion. According to the method, an unmanned aerial vehicle is used for carrying visible light, infrared and smoke sensors to collect multi-modal sensing data of a target area, and standardized data are obtained through time synchronization and preprocessing. The improved YOLOv7-tiny and the ResNet-34 network fused with deformable convolution are adopted to extract visible light and infrared features respectively, and fusion is carried out by combining a cross-modal attention mechanism and smoke features. Meanwhile, an environment and sensor states are fused to construct a reinforcement learning state space, a combined strategy network is trained through a near-end strategy optimization algorithm, and the adaptive weight coefficient of each sensor is dynamically output. And finally, calculating the comprehensive fire probability after weight fusion of the sensors, and realizing graded early warning. According to the method, adaptive fusion and dynamic decision of multi-sensor data are realized, and the accuracy, robustness and response speed of fire detection in a complex environment are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of multimodal reinforcement learning technology, and more specifically, to a method and system for detecting fireworks based on deep reinforcement learning and feature fusion. Background Technology

[0002] Forest fires, recognized as one of the eight major natural disasters worldwide, pose a serious threat to the global ecological environment and human life and property. Traditional forest fire monitoring methods mainly rely on manual patrols, observation from lookout towers, and isolated monitoring equipment, which suffer from problems such as delayed response, blind spots, and high false alarm rates. In recent years, with the development of drone technology and sensor technology, drone-based forest fire monitoring methods have gradually been applied.

[0003] Existing drone-based forest fire monitoring methods mostly employ single-type sensors or simple decision rule fusion. For example, some systems use visible light cameras for flame and smoke identification, but their effectiveness drops significantly under low-light conditions (such as nighttime); other systems use infrared thermal imagers to detect temperature anomalies, but are easily affected by sunlight reflection, animal heat sources, and other interferences. Some advanced systems attempt to combine multiple sensors, using a combination of visible light cameras, infrared sensors, and smoke sensors to improve identification accuracy. However, these methods still have significant shortcomings in multi-sensor data fusion:

[0004] The limitations of single-mode detection include: limitations of visible light detection, inadequacies of infrared thermal imaging, and low spatial resolution of gas sensors.

[0005] The difficulties of multi-sensor fusion include: spatiotemporal asynchrony, data heterogeneity, and feature conflicts and weight adjustment.

[0006] Poor adaptability to harsh environments: limitations of traditional image processing methods and inadequacies of fixed threshold strategies. Summary of the Invention

[0007] The main objective of this application is to provide a smoke detection method and system based on deep reinforcement learning and feature fusion, so as to at least solve the problems in the background art mentioned above and improve the detection accuracy, precision and response speed of search and rescue targets.

[0008] To achieve the above objectives, a method and system for fire detection based on deep reinforcement learning and feature fusion are provided.

[0009] Firstly, this application provides a fireworks detection method based on deep reinforcement learning and feature fusion, the method comprising:

[0010] Raw multimodal data of the forest environment are collected using multi-source sensors carried by drones, and the raw multimodal data is processed in time synchronization to obtain standardized multimodal data;

[0011] We use various feature extraction methods to extract different data features from standardized multimodal data, and use a cross-modal attention mechanism to fuse and concatenate the different data features to obtain multimodal fusion features;

[0012] Environmental state information of the forest area and sensor state information of multiple sources are collected. The environmental state information and sensor state information are fused into complete state space information. A combined policy network and reward function based on reinforcement learning network are constructed. The combined policy network is trained using the proximal policy optimization algorithm and the experience replacement mechanism.

[0013] The raw multimodal data is preprocessed for environmental adaptation to obtain the processing results. Based on the processing results, the complete state space information and the reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors.

[0014] Multimodal fusion features are used to identify different detection results using a target recognition algorithm. The comprehensive fire probability is calculated using the different detection results and their corresponding weight coefficients. A fire warning threshold is preset, and the comprehensive fire probability is compared with the fire warning threshold to obtain the comparison result. The corresponding fire warning level is obtained based on the comparison result.

[0015] Preferably, the raw multimodal data of the forest environment is collected using multi-source sensors mounted on a drone, and the raw multimodal data is time-synchronized to obtain standardized multimodal data, including:

[0016] The drone was equipped with a full-color camera to collect raw visible light image data of the forest environment, a dual-spectrum thermal imager to collect raw infrared data of the forest environment, and a laser gas sensor to collect raw smoke concentration data of the forest environment.

[0017] The GPS_Sync function is used to perform time synchronization processing on the raw visible light image data, raw infrared data, and raw smoke concentration data to obtain standardized visible light image data, standardized infrared data, and standardized smoke concentration data, respectively.

[0018] Preferably, different data features of standardized multimodal data are extracted using multiple feature extraction methods, and cross-modal attention mechanisms are used to fuse and concatenate the different data features to obtain multimodal fused features, including:

[0019] Visible light features of standardized multimodal data are extracted using a static feature extraction network;

[0020] Infrared features of standardized multimodal data are extracted using an infrared feature extraction network;

[0021] The first fused feature is obtained by fusing visible light features and infrared features using a cross-modal attention mechanism.

[0022] The standardized smoke concentration data is combined with the first fusion feature to obtain the second fusion feature.

[0023] Preferably, visible light features of standardized multimodal data are extracted using a static feature extraction network, including:

[0024] Visible light features are extracted using an improved YOLOv7-tiny network, which combines a backbone network with a channel and spatial attention mechanism and a feature pyramid network.

[0025] The backbone network is the CSPNet network, which is used to extract multi-scale visible light features in smoke and flame regions.

[0026] Channel and spatial attention mechanisms are used to enhance the CSPNet network's ability to extract flame and smoke features from visible light features;

[0027] Feature pyramid networks are used to fuse deep semantic features and shallow detail features from visible light features.

[0028] Preferably, infrared features of standardized multimodal data are extracted using an infrared feature extraction network, including:

[0029] Infrared features were extracted using a ResNet-34 network incorporating deformable convolutions;

[0030] Among them, the ResNet-34 network is used to extract temperature distribution features from infrared features;

[0031] The last layer in the ResNet-34 network is replaced with a deformable convolution, which is used to extract infrared features from ignition regions of different shapes, angles and sizes in normalized multimodal data.

[0032] Preferably, the environmental status information includes at least: the light intensity, severe weather level, and visibility level of the target area;

[0033] Sensor status information includes at least: visible light image quality, infrared data quality, and smoke sensor confidence level.

[0034] Preferably, the reward function is constructed, including:

[0035] Construct rewards based on detection accuracy, efficiency, and robustness.

[0036] The detection accuracy reward includes: a reward for correct detection, a penalty for false alarms, and a penalty for missed detection. The reward for correct detection is +1 when a fire is correctly detected; the penalty for false alarms is -0.5 when a fire is incorrectly detected; and the penalty for missed detection is -1 when a fire is not detected.

[0037] The efficiency bonus is -0.1 for each multi-source sensor used;

[0038] The robustness bonus is +0.5 when the detection accuracy is greater than or equal to 90%.

[0039] Preferably, a combination policy network based on reinforcement learning is constructed, and the combination policy network is trained using a proximal policy optimization algorithm and an experience replacement mechanism, including:

[0040] Step 1: Initialize the reinforcement learning network;

[0041] Step 2: Construct an experience replacement buffer, which is used to store state transition tuples generated during training;

[0042] Step 3: Obtain training data through sampling and interaction;

[0043] Step 4: Randomly sample a portion of the data from the empirical replacement buffer and optimize the objective function through the near-end policy to update the parameters of the Actor network in the combined policy network;

[0044] Step 5: Update the parameters of the Critic network in the combined policy network using the temporal difference error and mean square error loss functions;

[0045] Step 6: Repeat steps 3 to 5 until the combined policy network converges to obtain a trained combined policy network that can output weight coefficients.

[0046] Preferably, the raw multimodal data undergoes environmentally adaptive preprocessing to obtain the processing result, including:

[0047] Calculate the darkness index of the original visible light image data; perform dynamic temperature compensation on the original infrared data based on ambient temperature and humidity to obtain compensated infrared data; use Kalman filtering to remove random noise from the original smoke concentration data to obtain noise-reduced smoke concentration data.

[0048] Complete state-space information and reward function are used for:

[0049] In step 3, the complete state space information is used as the input to the Actor network and combined with the processing results to calculate the sensor weight allocation action in the current state;

[0050] After sampling and interaction, the reward function is called to calculate the immediate reward value and store the immediate reward value in the state transition tuple of the experience putback buffer;

[0051] In steps 4 and 5, some sample data contain complete state space information, and the temporal difference error and mean square error loss functions are calculated through the reward function.

[0052] Secondly, this application provides a fireworks detection system based on deep reinforcement learning and feature fusion, which is applied to the method described in the first aspect above. The system includes:

[0053] The multimodal data acquisition and preprocessing module is used to collect raw multimodal data of the forest environment using multi-source sensors carried by the UAV, and to perform time synchronization processing on the raw multimodal data to obtain standardized multimodal data.

[0054] The multimodal feature fusion module is connected to the multimodal data acquisition and preprocessing module. The multimodal feature fusion module is used to extract different data features of standardized multimodal data using multiple feature extraction methods, and to fuse and splice different data features using a cross-modal attention mechanism to obtain multimodal fused features.

[0055] The combined policy network construction module is connected to the multimodal feature fusion module. The combined policy network construction module is used to collect environmental state information of the forest environment and sensor state information of multiple sources, fuse the environmental state information and sensor state information into complete state space information, construct a combined policy network and reward function based on reinforcement learning network, and train the combined policy network using the proximal policy optimization algorithm and experience replacement mechanism.

[0056] The adaptive sensor weight decision module is connected to the combined policy network construction module. The adaptive sensor weight decision module is used to perform environmental adaptive preprocessing on the original multimodal data to obtain the processing results. Based on the processing results, the complete state space information and reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors.

[0057] The fire probability comprehensive assessment and early warning module is connected to the adaptive sensor weight decision module. The fire probability comprehensive assessment and early warning module is used to identify multimodal fusion features using a target recognition algorithm to obtain different detection results. It calculates the comprehensive fire probability using different detection results and corresponding weight coefficients, presets a fire early warning threshold, compares the comprehensive fire probability with the fire early warning threshold to obtain a comparison result, and obtains the corresponding fire early warning level based on the comparison result.

[0058] This application provides a smoke and fire detection method and system based on deep reinforcement learning and feature fusion. The method includes: simultaneously collecting raw multimodal data of a forest environment using a UAV equipped with multiple source sensors, and preprocessing this data to obtain standardized multimodal data; extracting features from visible light and infrared data using an improved YOLOv7-tiny network and a ResNet-34 network fused with deformable convolutions, respectively, and fusing these features with smoke concentration data through a cross-modal attention mechanism to generate multimodal fusion features; simultaneously, constructing a complete state space by fusing environmental state and sensor state information, training a combined policy network based on a near-end policy optimization algorithm to make decisions on the environmentally adaptive preprocessed data, and dynamically outputting adaptive weight coefficients for each sensor; calculating the comprehensive fire probability based on the target recognition results and corresponding weight coefficients of the multimodal fusion features, and classifying fire warning levels through preset thresholds; finally, dispatching UAVs for fire confirmation and intelligent fire suppression according to the warning level. This method achieves adaptive optimization fusion of multi-sensor data in complex environments and efficient fire identification, significantly improving the accuracy, robustness, and response speed of the detection system. Attached Figure Description

[0059] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0060] Figure 1 A flowchart of a fireworks detection method based on deep reinforcement learning and feature fusion provided for this application;

[0061] Figure 2 The overall flowchart of the fireworks detection method based on deep reinforcement learning and feature fusion provided in this application;

[0062] Figure 3 The flowchart of the target detection and recognition process based on deep reinforcement learning and feature fusion for the fireworks detection method provided in this application is shown.

[0063] Figure 4 A schematic diagram of the structure of a fireworks detection system based on deep reinforcement learning and feature fusion provided in this application;

[0064] Figure 5 The image shows the detection and recognition effect of a fireworks detection method based on deep reinforcement learning and feature fusion provided in this application. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein.

[0067] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0068] This application provides a smoke and fire detection method and system based on deep reinforcement learning and feature fusion. The method includes: firstly, simultaneously acquiring multimodal information such as visible light images, infrared data, and smoke concentration in a forest area using a full-color camera, dual-spectrum thermal imager, and laser gas sensor mounted on a UAV, and performing standardized preprocessing. Then, an improved YOLOv7-tiny network with an attention mechanism is used to extract visible light features, and a ResNet-34 network with deformable convolutions is used to extract infrared temperature features. These features are then fused with smoke features through a cross-modal attention mechanism. Simultaneously, the system collects real-time environmental conditions such as illumination and weather, as well as the sensor's own state, constructing a complete state space. A reinforcement learning network is trained based on a near-end policy optimization algorithm to form an adaptive decision model that dynamically adjusts the weight coefficients of the visible light, infrared, and smoke sensors according to the environment. Finally, the system comprehensively weights the recognition results of each modality to calculate the overall fire probability and automatically generates graded early warnings based on preset thresholds, thereby guiding the UAV in accurate fire confirmation and fire suppression scheduling. This method effectively realizes intelligent fusion and adaptive decision-making of multi-source information in complex forest environments, comprehensively improving the accuracy of fire detection, environmental robustness, and emergency response efficiency.

[0069] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0070] Figure 1 The flowchart of a smoke detection method based on deep reinforcement learning and feature fusion provided in this application includes: collecting raw multimodal data of a forest environment using multi-source sensors mounted on a UAV; performing time synchronization processing on the raw multimodal data to obtain standardized multimodal data; extracting different data features from the standardized multimodal data using various feature extraction methods; and fusing and stitching the different data features using a cross-modal attention mechanism to obtain multimodal fusion features; collecting environmental state information of the forest environment and sensor state information from the multi-source sensors; fusing the environmental state information and sensor state information into complete state space information; and constructing a reinforcement learning-based... The combined policy network and reward function of the network are trained using a proximal policy optimization algorithm and an experience replacement mechanism. The original multimodal data undergoes environmentally adaptive preprocessing to obtain the processing results. Based on the processing results, the complete state space information and reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors. The multimodal fusion features are then identified using a target recognition algorithm to obtain different detection results. The comprehensive fire probability is calculated using the different detection results and the corresponding weight coefficients. A fire warning threshold is preset, and the comprehensive fire probability is compared with the fire warning threshold to obtain the comparison results. Based on the comparison results, the corresponding fire warning level is obtained.

[0071] This application provides a smoke and fire detection method based on deep reinforcement learning and feature fusion. The method uses a drone equipped with multi-source sensors to collect raw multimodal data of a forest environment, which is then standardized through time synchronization and targeted preprocessing. Visible light and infrared features are extracted using an improved YOLOv7-tiny network and a ResNet-34 network with deformable convolutions, respectively, and combined with smoke concentration features through an adaptive fusion mechanism via cross-modal attention. Simultaneously, a complete state space is constructed based on ambient light, weather conditions, and real-time sensor status. A near-end policy optimization algorithm is used to train a combined policy network to dynamically optimize the allocation of weights for visible light, infrared, and smoke sensors. The results of each modality are then weighted and fused to calculate the overall fire probability, generating a tiered early warning based on preset thresholds. Finally, after the early warning is confirmed, a firefighting plan is intelligently generated based on real-time multi-source data, and firefighting drones are dispatched for precise firefighting.

[0072] This method enables adaptive collaborative perception and decision-making based on multi-source information in complex forest environments. By dynamically adjusting sensor weights through reinforcement learning, it effectively overcomes dependence on a single environment and improves the system's detection accuracy and robustness under different lighting and weather conditions. Cross-modal feature fusion technology fully mines and correlates multi-dimensional information such as visible light texture, infrared temperature, and smoke concentration, significantly enhancing the early fire identification capability and anti-interference ability. Furthermore, the integrated design of the entire process, from multi-modal data acquisition, intelligent feature fusion, dynamic weight decision-making to tiered early warning and drone-based joint firefighting, constructs a complete "perception-decision-response" closed loop, significantly shortening fire emergency response time and comprehensively improving the intelligence level and practical application effectiveness of forest fire monitoring and early warning.

[0073] Figure 2 The overall flowchart of the fireworks detection method based on deep reinforcement learning and feature fusion provided in this application includes: collecting multi-source heterogeneous datasets using multi-source sensors, performing spatiotemporal alignment of the multi-source data and then extracting features, fusing the extracted features and inputting them into a policy network, performing target detection and recognition under the influence of environmental conditions, updating the MAPPO policy simultaneously, and finally obtaining the recognition result.

[0074] Specifically, the raw multimodal data of the forest environment is collected using multi-source sensors mounted on the UAV. The raw multimodal data is then processed in time synchronization to obtain standardized multimodal data. This includes: collecting raw visible light image data of the forest environment using a full-color camera mounted on the UAV; collecting raw infrared data of the forest environment using a dual-spectrum thermal imager mounted on the UAV; and collecting raw smoke concentration data of the forest environment using a laser gas sensor mounted on the UAV. The GPS_Sync function is then used to process the raw visible light image data, raw infrared data, and raw smoke concentration data in time synchronization to obtain standardized visible light image data, standardized infrared data, and standardized smoke concentration data, respectively.

[0075] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. The method uses a drone equipped with a full-color camera, a dual-spectrum thermal imager and a laser gas sensor to collect visible light images, infrared data and smoke concentration data of forest areas, respectively. The original multimodal data is transformed into spatiotemporally aligned standardized multimodal data through a GPS time synchronization preprocessing step.

[0076] Visible light image acquisition: A high-sensitivity full-color camera was used under low-light conditions (illuminance as low as...). The camera acquires 1920×1080 resolution video with an image transmission latency of no more than 40 milliseconds. The camera lens is mounted at an angle downwards towards the forest to ensure coverage of key monitoring areas.

[0077] Infrared data acquisition: A FLIR Vue TZ20 dual-spectrum thermal imager was used, with a detection wavelength range of 7.5-13.5μm and a temperature detection range of -20°C to 550°C. It can detect temperature differences of 0.05°C within a 5km range. The infrared sensor acquires temperature distribution data at a frequency of 10Hz.

[0078] Smoke concentration data acquisition: Sensirion SCD40 laser gas sensor, based on tunable diode laser absorption spectroscopy (TDLAS) technology, was used to monitor CO, Concentration, response time less than 10 seconds, detection limit as low as 0.1 ppm.

[0079] Hardware timestamp injection:

[0080] When acquiring data, each sensor records a precise GPS timestamp (format: UTC time + nanosecond offset) for each data frame by connecting to the PPS (pulse per second) signal and NMEA time message output by the U-BloxZED-F9P GPS module. Specifically:

[0081] Visible light camera: Records a timestamp at the end of each frame exposure. ;

[0082] Infrared thermal imager: Records a timestamp when outputting each frame of the temperature matrix. ;

[0083] Smoke sensor: Records a timestamp upon completion of each concentration sampling. .

[0084] Call the GPS_Sync function to align the time window:

[0085] The raw data streams from each sensor are input into the GPS_Sync function, which performs the following operations:

[0086] Reference time selection: The visible light image time series is used as the reference time axis. (Usually due to the highest frame rate);

[0087] Missing frame interpolation: For infrared and smoke data, if a certain reference time... If no corresponding sampling point is found, linear interpolation is used to generate the synchronization value:

[0088]

[0089] in, Indicates at the reference time Synchronized data values, Indicates time Sensor sampling values ​​at the location, Indicates time Sensor sampling values ​​at the location, Represents the first on the reference time axis At that moment, Distance Most recent before and after sampling time, For sensor data (infrared temperature matrix or smoke concentration value);

[0090] Data resampling: unifying all sensor data to the same time series. Generate a time-aligned data stream.

[0091] Synchronization accuracy verification and anomaly handling:

[0092] Calculate the maximum deviation between each sensor timestamp and the reference time axis. :

[0093]

[0094] in, This indicates the maximum deviation between each sensor's timestamp and the reference time axis. Indicates the timestamp recorded by the sensor. Represents the first on the reference time axis At that moment, if (The threshold is usually set to 1ms), then a synchronization anomaly flag is triggered, and the data in that time window is discarded or a prediction compensation mechanism is enabled. This indicates the time synchronization deviation threshold.

[0095] Output standardized multimodal data:

[0096] The synchronized data is packaged into a unified structure to form a time-aligned multimodal data matrix:

[0097]

[0098] in, Represents a multimodal data matrix. For the first Visible light image with timestamps For synchronous infrared temperature matrix, To synchronize the smoke concentration vector (including CO, concentration), This represents the total number of sampling time points.

[0099] This method effectively overcomes the limitations of a single sensor under low light, temperature interference, and noise by using multi-sensor collaborative acquisition and targeted preprocessing, thereby improving data quality and consistency. GPS time synchronization ensures accurate alignment of multi-source data in the spatiotemporal dimension, providing a reliable data foundation for subsequent multimodal feature fusion and fire identification, and significantly improving the monitoring stability and accuracy of the system in complex forest environments.

[0100] Specifically, various feature extraction methods are used to extract different data features from standardized multimodal data, and a cross-modal attention mechanism is used to fuse and concatenate the different data features to obtain multimodal fusion features. These include: using a static feature extraction network to extract visible light features from standardized multimodal data; using an infrared feature extraction network to extract infrared features from standardized multimodal data; using a cross-modal attention mechanism to fuse visible light features and infrared features to obtain a first fusion feature; and concatenating standardized smoke concentration data with the first fusion feature to obtain a second fusion feature.

[0101] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. The method first extracts spatial and semantic features of visible light images by using an improved YOLOv7-tiny network combined with the CBAM attention mechanism. At the same time, it uses a ResNet-34 network with fused deformable convolutions to extract temperature distribution and shape adaptive features of infrared images. Then, a cross-modal attention module is designed to achieve deep interaction and complementary fusion of visible light and infrared features through learnable query-key-value mapping to generate a first fused feature. Finally, the time-series smoke concentration data is used as a one-dimensional feature vector and concatenated with the first fused feature to form a second fused feature containing visual, thermal radiation and gas information, i.e., a multimodal fused feature.

[0102] An improved YOLOv7-tiny network is used as the static feature extraction network, with its backbone employing a CSPNet (Cross-Stage Partial Network) structure to extract multi-scale features from visible light images. A CBAM (Convolutional Block Attention Module) is embedded in the network, consisting of two stages:

[0103] Channel attention: Each channel of the feature map is weighted by importance, calculated using the following formula:

[0104]

[0105] in, This represents the channel attention weight matrix. Indicates the input feature map, This represents the Sigmoid activation function. This represents a multilayer perceptron. Indicates global average pooling. This indicates global max pooling.

[0106] Spatial attention: Weights are applied to the spatial locations of the feature maps; the calculation formula is as follows:

[0107]

[0108] in, Represents the spatial attention weight matrix. This represents a 7×7 convolution operation. This indicates splicing along the channel dimension.

[0109] CBAM output is ,in, This represents the feature map after double attention weighting. This indicates element-wise multiplication.

[0110] A visible light feature map is generated by fusing deep semantic features and shallow detail features using a Feature Pyramid Network (FPN). , Represents the visible light characteristic map. Representing a three-dimensional tensor, , and These represent the height, width, and number of channels of the feature map, respectively.

[0111] ResNet-34 was used as the infrared feature extraction network, with a standardized infrared image as input and high-altitude temperature distribution features as output.

[0112] The last layer of ResNet-34 is replaced with a Deformable Convolution, whose output features are:

[0113]

[0114] in, Indicates the output feature map at position The value, Indicates the position of the convolution kernel The weight, Indicates the input feature map, Indicates the output position. Indicates the sampling position of the convolution kernel. For learnable offsets, This represents the set of real numbers. The structure can adapt to fire point regions of different shapes and angles, outputting infrared feature maps. , Represents infrared feature map, Representing a three-dimensional tensor, , and These represent the height, width, and number of channels of the feature map, respectively.

[0115] The process of generating the first fusion feature is as follows:

[0116] Using 1×1 convolution and Mapped to the same number of channels :

[0117] ;

[0118] ;

[0119] in, Indicates the characteristics of visible light. Indicates infrared characteristics, This represents the weight matrix of the 1×1 convolution kernel used to map visible light features. Let represent the weight matrix of the 1×1 convolutional kernel used to map infrared features. A cross-modal attention module is designed, where:

[0120] Visible light features are used as the query, and infrared features are used as the key and value:

[0121]

[0122] in, This represents the feature matrix mapped from visible light features, used as the query vector. This represents the softmax function.

[0123] The specific implementation is as follows:

[0124]

[0125] in, Indicates the characteristics after fusion. This represents the visible light characteristics after mapping. Represents the mapped infrared features. For learnable parameters, For feature dimensions.

[0126] The attention-weighted features are residually concatenated with the original visible light features to obtain the first fused feature. .

[0127] The process of generating the second fusion feature is as follows:

[0128] Time-series smoke concentration data Mapped to feature vectors through fully connected layers , Represents the set of real numbers. Indicates the dimension of the feature vector.

[0129] Will By copying in the spatial dimension, we obtain... , with the first fusion feature Stitching along the channel dimension:

[0130]

[0131] in, Indicates multimodal fusion features, Indicates the first fusion feature, express The result of replication in the spatial dimension, It represents the set of real numbers, which includes information on three modes: visible light, infrared, and smoke.

[0132] This method achieves semantic-level deep fusion of visible light and infrared features through a cross-modal attention mechanism, effectively utilizing the complementary information of the two modalities to enhance the ability to identify early smoke and low-temperature fire points. At the same time, smoke concentration features are introduced into the fusion process, providing key chemical evidence for fire detection and further improving the system's perception comprehensiveness and judgment reliability in complex environments. The resulting multimodal fusion features combine spatial details, thermal distribution trends, and gas concentration time-series information, providing rich and consistent feature representations for subsequent adaptive weighted decision-making.

[0133] Specifically, visible light features of standardized multimodal data are extracted using a static feature extraction network, including: extracting visible light features using an improved YOLOv7-tiny network, wherein the improved YOLOv7-tiny network combines a backbone network with channel and spatial attention mechanisms and a feature pyramid network; wherein the backbone network is a CSPNet network, which is used to extract multi-scale visible light features in smoke and flame regions; the channel and spatial attention mechanisms are used to enhance the ability of the CSPNet network to extract flame and smoke features from visible light features; and the feature pyramid network is used to fuse deep semantic features and shallow detail features from visible light features.

[0134] This application provides a smoke and fire detection method based on deep reinforcement learning and feature fusion. The method uses an improved YOLOv7-tiny network to extract visible light features. Its backbone is based on the CSPNet structure, which can adaptively extract multi-scale spatial features from smoke and flame regions. On this basis, a dual attention mechanism of CBAM channels and space is introduced to dynamically weight the channel importance and key spatial regions of the feature map, thereby enhancing the ability to focus on flame and smoke features. Furthermore, a feature pyramid network (FPN) is used to fuse shallow detail features and deep semantic features to form a visible light feature representation that takes into account both local texture and global semantics.

[0135] The advantages of this method are as follows: the CSPNet backbone effectively improves the efficiency and receptive field coverage of multi-scale feature extraction, and enhances the model's adaptability to flame and smoke regions of different sizes; the CBAM attention mechanism significantly improves the network's sensitivity and selection ability to fire-related features through channel and spatial dual-path weighting, and suppresses background interference; the feature pyramid network structure realizes the natural fusion of shallow details and deep semantics, ensuring that the feature semantic expression ability is enhanced while maintaining edge clarity, and providing a robust and rich visible light feature foundation for subsequent cross-modal fusion.

[0136] Specifically, infrared features of standardized multimodal data are extracted using an infrared feature extraction network, including: extracting infrared features using a ResNet-34 network fused with deformable convolutions; wherein, the ResNet-34 network is used to extract temperature distribution features in the infrared features; and the last layer in the ResNet-34 network is replaced with a deformable convolution, which is used to extract infrared features on ignition regions of different shapes, angles and sizes in the standardized multimodal data.

[0137] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. The method uses a ResNet-34 network with deformable convolution to extract infrared features. The ResNet-34 network serves as the basic feature extraction structure to effectively extract temperature distribution features related to the ignition point from infrared images. By replacing the last layer of the ResNet-34 network with a deformable convolution layer, the network can adaptively adjust the receptive field shape and spatial sampling position of the convolution kernel, thereby enabling more accurate feature modeling of ignition areas with different shapes, angles, and sizes.

[0138] The advantages of this method are as follows: ResNet-34 network, with its deep residual structure, can stably extract temperature gradients and distribution patterns in infrared images, effectively characterizing thermal radiation features; the introduction of deformable convolution enhances the network's geometric adaptability to irregular and multi-morphological fire point regions, overcomes the limitations of traditional convolution in extracting deformable target features, improves the characterization ability and generalization performance of infrared features in complex fire scene scenarios, and provides more discriminative thermal imaging features for subsequent cross-modal fusion.

[0139] Specifically, the environmental status information includes at least: the light intensity of the target area, the severity of severe weather, and the visibility level; the sensor status information includes at least: the visible light image quality, the infrared data quality, and the confidence level of the smoke sensor.

[0140] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. This method collects environmental state information (including light intensity, severe weather level and visibility level) and sensor state information (including visible light image quality, infrared data quality and smoke sensor confidence) of the target area to construct a multi-dimensional state space to describe the current working conditions of the system, providing comprehensive context-aware input for subsequent reinforcement learning decisions.

[0141] State-space design:

[0142] Environmental conditions :

[0143] Light intensity Light intensity affects the effectiveness of visible light sensors; in low light conditions, infrared sensors may be necessary.

[0144] Rain, fog, and haze levels Severe weather conditions (such as fog, haze, rain, and snow) can affect the quality of visible light images, so the adaptability of the sensor needs to be considered.

[0145] visibility When visibility is low, additional infrared or smoke sensors may be needed to compensate for the shortcomings of visual sensors.

[0146] Sensor status :

[0147] Visible light image quality : Represents the sharpness of visible light images, affecting the effectiveness of visible light sensors.

[0148] Infrared data quality : Represents the clarity and accuracy of infrared data. Infrared sensors typically perform better in low light or inclement weather.

[0149] Smoke sensor confidence This represents the confidence level of the smoke sensor in detecting smoke and reflects the reliability of smoke detection.

[0150] Complete state space The state space is a combination of the environmental state and the sensor state, and can be represented as:

[0151]

[0152] in, , .

[0153] Action space design: The action space defines the possible actions of the agent at each time step. Here, actions are the weighted assignments of sensors, determining the importance of each sensor in the final decision.

[0154] Sensor weight allocation The weighting represents the relative importance of the three sensors (visible light, infrared, and smoke sensors).

[0155]

[0156] in, and This indicates that the sum of all weights is 1, ensuring that the weight allocation at each time step is effective and normalized. These represent the weights of the visible light, infrared, and smoke sensors, respectively.

[0157] The advantages of this method are as follows: by fusing real-time state information from both the environment and sensors, the system can accurately perceive changes in the external environment and its own working state, providing sufficient and structured decision-making basis for adaptive sensor weight allocation; this multi-dimensional state representation not only enhances the system's adaptability to complex weather and lighting conditions, but also realizes dynamic evaluation of data credibility through sensor quality feedback, thereby significantly improving the intelligence, robustness and decision reliability of the multi-sensor fusion strategy in real complex scenarios.

[0158] Specifically, a reward function is constructed, including: a detection accuracy reward, an efficiency reward, and a robustness reward. The detection accuracy reward includes: a correct detection reward, a false alarm penalty, and a false negative penalty. The correct detection reward is +1 when a fire is correctly detected; the false alarm penalty is -0.5 when a fire is incorrectly detected; the false negative penalty is -1 when a fire is not detected; the efficiency reward is -0.1 for each multi-source sensor used; and the robustness reward is +0.5 when the detection accuracy is greater than or equal to 90%.

[0159] This application provides a smoke and fire detection method based on deep reinforcement learning and feature fusion. The reward function designed by this method consists of three parts: detection accuracy reward, which encourages correct fire detection with a positive reward (+1) and suppresses incorrect judgments with a false alarm penalty (-0.5) and a false negative penalty (-1); efficiency reward, which applies a slight negative reward (-0.1) for each use of the sensor to guide the system to complete the task with the least sensor resources; and robustness reward, which provides additional positive incentives (+0.5) when the detection accuracy reaches more than 90% to enhance the stable performance of the system in complex environments.

[0160] This method employs multi-objective joint optimization to encourage high detection accuracy while balancing resource utilization efficiency and system robustness. The accuracy reward and punishment mechanism effectively distinguishes between the severity of false alarms and false negatives, guiding the strategy to focus more on avoiding false negatives. Efficiency incentives enable the system to learn to simplify sensor usage while ensuring performance, reducing power consumption and computational burden. Robustness rewards further enhance the system's stable output capability under harsh conditions, making the trained strategy more valuable for practical deployment.

[0161] Specifically, a combined policy network based on reinforcement learning is constructed, and trained using a proximal policy optimization algorithm and an experience replacement mechanism, including: Step 1: Initialize the reinforcement learning network; Step 2: Construct an experience replacement buffer to store state transition tuples generated during training; Step 3: Obtain training data through sampling and interaction; Step 4: Randomly sample a portion of the data from the experience replacement buffer and update the parameters of the Actor network in the combined policy network using the proximal policy optimization objective function; Step 5: Update the parameters of the Critic network in the combined policy network using temporal difference error and mean squared error loss functions; Step 6: Repeat steps 3 to 5 until the combined policy network converges to obtain a trained combined policy network capable of outputting weight coefficients.

[0162] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. The method constructs a reinforcement learning combined policy network containing an Actor network and a Critic network, and trains it using the Proximal Policy Optimization (PPO) algorithm and an experience replay mechanism. First, the network parameters and experience buffer are initialized. Then, state transition data is collected by interacting with the environment and stored in the buffer. Batch data is then sampled from the buffer. The Actor network policy is updated using the PPO pruning objective function. At the same time, the value estimation of the Critic network is updated using temporal difference error and mean square error. The interaction-sampling-update process is repeated until the network converges. Finally, a policy network that can adaptively output sensor weight coefficients according to the environment and sensor state is obtained.

[0163] Training is performed using the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm. MAPPO can handle policy optimization for multiple agents (sensors) and effectively avoids drastic policy updates, thereby improving training stability.

[0164] Actor Network:

[0165] Input layer: Receives the complete state (6-dimensional).

[0166] Hidden layers: Two fully connected network layers, each with 128 neurons, using the ReLU activation function.

[0167] Output layer: 3 neurons, using the Softmax activation function, the output is the sensor weight allocation, ensuring that the sum of the output weights is 1.

[0168] Critic Network:

[0169] Evaluation state value function The value of the current state is used to guide the updates of the Actor network.

[0170] The structure is similar to that of the Actor network, but the output is a single value representing the state value.

[0171] Training process:

[0172] During training, through multiple rounds of sampling and updating, the agent gradually learns how to reasonably allocate sensor weights under different environmental conditions.

[0173] Experience replay mechanism: store and transfer tuples Training is performed by sampling batches of data from the replay buffer.

[0174] Advantage function estimation: The effectiveness of the current strategy is determined by calculating the advantage function (i.e., the advantage of a certain action relative to the state value function).

[0175] Proximal Policy Optimization (PPO) pruning objective function:

[0176]

[0177] in, Describe the objective function. Represents the expectation operator, This indicates the current state of the policy network (Actor network). Select action The probability, This indicates that the old policy (the policy network before the last update) is in state. Select action The probability, Indicates the proportion of strategy change. Indicates the state Select action Compared to the quality of average movements, This represents the clipping function, which takes the input value as the clipping value. Limited to the range Inside, This is the pruning parameter, typically set to 0.2. This formula ensures the stability of the training process by limiting the update magnitude of the policy.

[0178] Training the Critic network: The parameters of the Critic network are updated by minimizing the temporal difference error using the mean squared error (MSE) loss function.

[0179] The advantages of this method are as follows: the PPO algorithm effectively limits the magnitude of policy updates through the policy proportional pruning mechanism, ensuring the stability and convergence efficiency of the training process; the experience replay mechanism breaks the temporal correlation between samples, improves data utilization, and reduces training variance; the Actor-Critic framework realizes the synergistic optimization of policy and value functions, enabling the network to learn efficient sensor weight allocation strategies and accurately evaluate state values ​​to guide policy improvement; the finally trained policy network has strong generalization ability and real-time decision-making ability, and can realize intelligent dynamic adjustment of multi-sensor fusion weights in changing environments.

[0180] Specifically, environmentally adaptive preprocessing is performed on the original multimodal data to obtain the processing results, including: calculating the darkness index of the original visible light image data; performing dynamic temperature compensation on the original infrared data according to the ambient temperature and humidity to obtain compensated infrared data; and using Kalman filtering to remove random noise from the original smoke concentration data to obtain noise-removed smoke concentration data.

[0181] The complete state space information and reward function are used in the following ways: In step 3, the complete state space information is used as the input and processing result of the Actor network to calculate the sensor weight allocation action in the current state; after sampling and interaction, the reward function is called to calculate the instant reward value and store the instant reward value in the state transition tuple of the experience putback buffer; in steps 4 and 5, some sample data contain complete state space information, and the temporal difference error and mean square error loss functions are calculated through the reward function.

[0182] This application provides a smoke detection method based on deep reinforcement learning and feature fusion. This method performs refined environmentally adaptive preprocessing on the raw multimodal data: it assesses visible light imaging conditions by calculating the darkness index, performs dynamic temperature compensation on infrared data based on ambient temperature and humidity to eliminate environmental interference, and uses Kalman filtering to denoise smoke concentration data to improve the signal-to-noise ratio. These processing results, together with real-time acquired environmental conditions such as illumination and weather, as well as the sensor's own state information, constitute a complete system state space. Subsequently, a combined policy network trained based on a proximal policy optimization algorithm is activated. It takes this state space as input and aims to maximize the reward function defined by detection accuracy, efficiency, and robustness, calculating and outputting the optimal multi-source sensor weight allocation scheme in real time. Finally, the system uses these dynamic weights to perform weighted fusion of the preliminary identification results of each modality obtained based on cross-modal fusion features, deriving a comprehensive fire probability and triggering a corresponding level of warning.

[0183] For a visible light image, calculate the grayscale value of all pixels. The average difference between the value and 128 yields the Darkness Index. The calculation formula is:

[0184]

[0185] in, Indicates the darkness index. This represents the grayscale values ​​of all pixels in a visible light image. This indicates the number of pixels.

[0186] For infrared data, dynamic compensation is performed based on ambient temperature and humidity to eliminate the influence of atmospheric attenuation on temperature measurement. Based on the infrared sensor's operating wavelength (7.5µm–13.5µm) and real-time environmental parameters, the average spectral transmittance under current atmospheric conditions is calculated using an atmospheric radiative transfer model (a simplified MODTRAN model). :

[0187]

[0188] in, Indicates the average transmittance of the spectrum. The transmittance function can be obtained through a pre-defined lookup table or empirical formula. Indicates the ambient temperature of the target area. Indicates relative humidity. Indicates atmospheric pressure. This indicates the observation distance from the infrared sensor to the target.

[0189] According to ambient temperature and atmospheric transmittance Calculate the radiation intensity of the atmospheric path itself in the infrared band. :

[0190]

[0191] in, Indicates radiation intensity. This represents the integral intensity of the Planck blackbody radiation function within the sensor band.

[0192] Let the measured radiation intensity of the infrared sensor be... The true radiation intensity of the target surface The following radiative transfer equation can be used to solve it:

[0193]

[0194] in, Indicates the actual radiation intensity. This indicates the measured radiation intensity.

[0195] Based on the sensor calibration curve, Convert to the target's true temperature :

[0196]

[0197] in, Represents the actual temperature. This represents the inversion function based on Planck's law, which can be obtained through sensor factory calibration or experimental fitting.

[0198] The above steps are repeated at preset time intervals or frame rates to achieve frame-by-frame dynamic compensation of infrared data, ensuring the accuracy and consistency of temperature measurement results under different environmental conditions.

[0199] For smoke concentration data, Kalman filtering is used to eliminate random noise and improve the signal-to-noise ratio.

[0200] Establishing a state-space model: Defining the system's state vectors With observation vector :

[0201]

[0202] in, express The actual smoke concentration at any given time (e.g., CO or...) Concentration, unit: ppm This indicates the rate of change in concentration (unit: ppm / s). express The raw concentration observation values ​​directly read by the time sensor. Let k represent the state vector at time k. The state equation is:

[0203]

[0204] in, Here is the state transition matrix. Let k be the state vector at time k-1. The sampling time interval, The noise is a process noise that follows a Gaussian distribution. covariance matrix Calibration is based on the stability of concentration changes.

[0205] The observation equation is:

[0206]

[0207] in, Let k represent the observation function at time k. For the observation matrix, To observe the noise, it follows a Gaussian distribution. ,variance Determined by the measurement accuracy of the sensor.

[0208] Filter initialization:

[0209] Assume initial state estimation and its error covariance matrix for:

[0210]

[0211] in, The process noise covariance matrix The variance term corresponding to the rate of change, This represents the observation function at time 0.

[0212] Iterative filtering process:

[0213] For every moment The following two phases are executed:

[0214] a. Prediction phase:

[0215] ;

[0216] ;

[0217] in, This represents the prior state estimate at time k. This represents the posterior state estimate at time k-1. Represents the state transition matrix. This represents the prior error covariance at time k. This represents the posterior error covariance at time k-1. express transpose, This represents the process noise covariance.

[0218] b. Update phase:

[0219] Calculate Kalman gain :

[0220]

[0221] in, Indicates Kalman gain, Represents the observation matrix. express transpose, This represents the variance of the observation noise.

[0222] State estimation update:

[0223]

[0224] in, This represents the posterior state estimate at time k. This represents the prior state estimate at time k. This represents the observed value.

[0225] Error covariance update:

[0226]

[0227] in, Let represent the posterior error covariance matrix at time k. Let the prior error covariance matrix at time k be denoted as . Represents the identity matrix. Indicates Kalman gain, Represents the observation matrix. Let represent the prior error covariance matrix.

[0228] Output filtering result:

[0229] The concentration component in the state vector is taken as the final smoke concentration value after denoising.

[0230]

[0231] in, This represents the filtered smoke concentration value. This represents the posterior concentration estimate. Represents the selection matrix. This represents the posterior state estimation vector.

[0232] Parameter calibration and adaptive adjustment.

[0233] Process noise covariance With observation noise variance The sensor can be calibrated through static testing and dynamic response experiments.

[0234] In some embodiments, the confidence index of the sensor can be used. Dynamic adjustment To achieve adaptive filtering:

[0235]

[0236] in, This represents the variance of the observed noise at the current moment. This represents the variance of the basic observation noise. This indicates the confidence level of the smoke sensor.

[0237] The GPS_Sync function is used to perform time synchronization processing on the raw visible light image data, standardized infrared data, and standardized smoke concentration data to obtain standardized multimodal data.

[0238] This method achieves end-to-end environmental adaptation and intelligence, from data preprocessing to final decision-making. First, targeted preprocessing (dynamic temperature compensation, Kalman filtering) directly corrects the interference of environmental variables on the raw data, improving data quality from the source. Second, it innovatively introduces reinforcement learning into sensor fusion decision-making, enabling the system to autonomously and dynamically adjust the confidence weights of each sensor based on real-time environmental conditions (such as decreased visible light quality due to severe weather) and task objectives (balancing accuracy and efficiency), rather than employing a fixed fusion strategy. This ensures optimal detection performance in complex and ever-changing environments. This closed-loop mechanism of "perceiving the environment - assessing the state - optimizing the decision" significantly enhances the system's robustness, adaptability, and overall reliability in real-world scenarios.

[0239] Fire detection and early warning output primarily achieves accurate and real-time response through weighted decision fusion of sensor data and early warning level classification, thereby providing fire detection and early warning output. The following is a detailed analysis of each part:

[0240] Weighted decision fusion:

[0241] Weighted decision fusion is the core of this step. By combining the identification results of each sensor (visible light, infrared, and smoke sensors) with their corresponding weights, a comprehensive fire probability is obtained. This process effectively integrates information from different sensors, and especially in the case of multimodal sensor combinations, it can improve the robustness and accuracy of the system.

[0242]

[0243] in, Indicates the probability of a fire. This represents the probability of a fire based on a visible light image. This represents the probability of a fire based on infrared images. This represents the fire probability based on smoke sensors. The obtained sensor weights represent the relative importance of each sensor to the final fire determination.

[0244] This weighted decision-making process can dynamically adjust the influence of each sensor based on the current environment and sensor status, ensuring that fire detection results have good accuracy and robustness under different conditions.

[0245] Warning level classification:

[0246] Based on the calculated fire probability The fire identification results are categorized into different warning levels. This step ensures that the system can respond promptly after a fire occurs and escalates the intensity of handling and response at each level.

[0247] No fire: If If no fire is detected, the system determines that there is no fire and does not issue a warning. This indicates that the probability of a fire is low, and the system can continue monitoring.

[0248] Level 1 Warning (Suspected Fire): If If the fire is detected, a Level 1 warning is triggered, indicating a suspected fire. At this point, the system will instruct the drone to approach the fire area for close observation, collecting more data to confirm the authenticity and extent of the fire. This is to ensure timely collection of more information and avoid misjudgments.

[0249] Level 2 warning (fire confirmed): If If the fire is detected, it is considered a confirmed fire, and the system will immediately send an alarm message to the command center, including the specific location of the fire, the extent of the fire, and a preliminary estimate of the fire intensity. This is the highest emergency response level, ensuring rapid dispatch of firefighting resources.

[0250] Linkage control:

[0251] Once a fire is confirmed, the system will initiate coordinated control and automatically take a series of measures to ensure the effective extinguishing of the fire and the conduct of rescue operations.

[0252] Automatic sensor activation: After a fire is confirmed, the system automatically activates visible light cameras and relevant sensors to collect data, obtaining on-site images and wind information. This information is crucial for determining the fire's spread and formulating firefighting strategies.

[0253] Generating Firefighting Plans: Based on the location and spread trend of the fire, the system can intelligently generate the optimal firefighting plan. This plan includes not only firefighting strategies but also the allocation of firefighting resources and the planning of action routes.

[0254] Dispatching firefighting drones: The system automatically dispatches firefighting drones to the fire scene based on the optimal firefighting plan. Drones can conduct efficient firefighting operations on-site, and are particularly suitable for hard-to-reach or dangerous areas.

[0255] Dynamic strategy adjustment: During the firefighting process, the system continuously monitors changes in the fire scene (such as the spread of the fire) and dynamically adjusts the firefighting strategy based on real-time feedback. This adaptive capability ensures that the firefighting strategy can cope with various emergencies in the course of the fire.

[0256] Figure 3 The flowchart of the target detection and recognition process based on deep reinforcement learning and feature fusion for fireworks detection provided in this application includes:

[0257] The task type is analyzed, the initial weights of each sensor are combined, the weights of each sensor are updated using multi-source data, and the multi-source data is fused and processed to obtain the target detection and recognition results.

[0258] Figure 4 A schematic diagram of a fireworks detection system based on deep reinforcement learning and feature fusion provided in this application includes:

[0259] The multimodal data acquisition and preprocessing module is used to collect raw multimodal data of the forest environment using multi-source sensors carried by the UAV, and to perform time synchronization processing on the raw multimodal data to obtain standardized multimodal data.

[0260] The multimodal feature fusion module is connected to the multimodal data acquisition and preprocessing module. The multimodal feature fusion module is used to extract different data features of standardized multimodal data using multiple feature extraction methods, and to fuse and splice different data features using a cross-modal attention mechanism to obtain multimodal fused features.

[0261] The combined policy network construction module is connected to the multimodal feature fusion module. The combined policy network construction module is used to collect environmental state information of the forest environment and sensor state information of multiple sources, fuse the environmental state information and sensor state information into complete state space information, construct a combined policy network and reward function based on reinforcement learning network, and train the combined policy network using the proximal policy optimization algorithm and experience replacement mechanism.

[0262] The adaptive sensor weight decision module is connected to the combined policy network construction module. The adaptive sensor weight decision module is used to perform environmental adaptive preprocessing on the original multimodal data to obtain the processing results. Based on the processing results, the complete state space information and reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors.

[0263] The fire probability comprehensive assessment and early warning module is connected to the adaptive sensor weight decision module. The fire probability comprehensive assessment and early warning module is used to identify multimodal fusion features using a target recognition algorithm to obtain different detection results. It calculates the comprehensive fire probability using different detection results and corresponding weight coefficients, presets a fire early warning threshold, compares the comprehensive fire probability with the fire early warning threshold to obtain a comparison result, and obtains the corresponding fire early warning level based on the comparison result.

[0264] This application provides a smoke and fire detection system based on deep reinforcement learning and feature fusion. First, a multimodal data acquisition and preprocessing module simultaneously collects and standardizes raw data acquired by a full-color camera, thermal imager, and gas sensor mounted on a drone. Next, a multimodal feature fusion module extracts visible light and infrared features using an improved YOLOv7-tiny network and a ResNet-34 network with deformable convolutions, respectively, and fuses them with smoke concentration features through a cross-modal attention mechanism. Simultaneously, an adaptive sensor weight decision module fuses environmental states (such as illumination and weather) and sensor states into a complete state space, driving a pre-trained reinforcement learning network to dynamically assign weights to visible light, infrared, and smoke sensors. Finally, a comprehensive fire probability assessment and early warning module calculates the comprehensive fire probability based on the recognition results and weights of each modality and automatically generates tiered early warnings based on preset thresholds.

[0265] This system achieves a deep integration of data-driven and decision-making intelligence. First, through modular multimodal feature fusion and dynamic weighted decision-making, the system can intelligently adjust its reliance on different sensors based on real-time environmental changes (e.g., automatically increasing the weight of infrared and gas sensors at night or in smoky environments). This effectively overcomes the limitations of single sensors in complex scenarios, significantly improving detection accuracy and environmental adaptability. Second, the fully automated closed-loop design, from raw data processing and cross-modal feature extraction to adaptive decision-making and early warning response, not only enhances the system's robustness and interpretability but also significantly shortens the response time from fire awareness to early warning decision-making, providing a complete technical solution for efficient and accurate early monitoring and emergency response to forest fires.

[0266] Figure 5 The image shows the detection and recognition effect of a fireworks detection method based on deep reinforcement learning and feature fusion provided in this application.

[0267] To verify the effectiveness of the method of this invention, experiments were conducted on wildfire datasets across multiple sample sets, covering different weather conditions, light intensities, and fire development stages. The method of this invention was compared with existing typical methods, and the results are as follows:

[0268] Table 1 Comparison of Experimental Methods

[0269]

[0270] Experimental results show that the method of this invention outperforms existing typical methods in all indicators. Especially in harsh environments, the method of this invention, through a sensor combination strategy based on deep reinforcement learning, can dynamically adjust the dominant sensor and maintain high detection accuracy, achieving a 7.5 percentage point improvement over decision-level fusion methods.

Claims

1. A smoke detection method based on deep reinforcement learning and feature fusion, characterized in that, The method includes: Raw multimodal data of the forest environment are collected using multi-source sensors carried by drones, and the raw multimodal data is processed by time synchronization to obtain standardized multimodal data; Different data features of the standardized multimodal data are extracted using multiple feature extraction methods, and the different data features are fused and concatenated using a cross-modal attention mechanism to obtain multimodal fusion features; Collect environmental state information of the forest area and sensor state information of the multi-source sensors, fuse the environmental state information and sensor state information into complete state space information, construct a combined policy network and reward function based on reinforcement learning network, and train the combined policy network using proximal policy optimization algorithm and experience replacement mechanism; The original multimodal data is subjected to environmental adaptive preprocessing to obtain the processing result. Based on the processing result, the complete state space information and the reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors. The multimodal fusion features are identified using a target recognition algorithm to obtain different detection results. The comprehensive fire probability is calculated using the different detection results and the corresponding weight coefficients. A fire warning threshold is preset, and the comprehensive fire probability is compared with the fire warning threshold to obtain a comparison result. The corresponding fire warning level is obtained based on the comparison result.

2. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 1, characterized in that, The process involves using multi-source sensors mounted on a drone to collect raw multimodal data of the forest environment, and then performing time-synchronized processing on the raw multimodal data to obtain standardized multimodal data, including: The drone uses a full-color camera to collect raw visible light image data of the forest environment, a dual-spectrum thermal imager to collect raw infrared data of the forest environment, and a laser gas sensor to collect raw smoke concentration data of the forest environment. The GPS_Sync function is used to perform time synchronization processing on the original visible light image data, the original infrared data, and the original smoke concentration data to obtain standardized visible light image data, standardized infrared data, and standardized smoke concentration data, respectively.

3. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 2, characterized in that, The process involves extracting different data features from the standardized multimodal data using multiple feature extraction methods, and then fusing and concatenating these different data features using a cross-modal attention mechanism to obtain a multimodal fused feature, including: Visible light features of the standardized multimodal data are extracted using a static feature extraction network; Infrared features of the standardized multimodal data are extracted using an infrared feature extraction network; The visible light feature and the infrared feature are fused using a cross-modal attention mechanism to obtain a first fused feature; The standardized smoke concentration data is combined with the first fusion feature to obtain the second fusion feature.

4. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 3, characterized in that, The extraction of visible light features from the standardized multimodal data using a static feature extraction network includes: The visible light features are extracted using an improved YOLOv7-tiny network, wherein the improved YOLOv7-tiny network combines a backbone network with a channel and spatial attention mechanism and a feature pyramid network. The backbone network is a CSPNet network, which is used to extract multi-scale visible light features in the smoke and flame regions. The channel and spatial attention mechanisms are used to enhance the CSPNet network's ability to extract flame and smoke features from the visible light features; The feature pyramid network is used to fuse deep semantic features and shallow detail features from the visible light features.

5. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 3, characterized in that, The process of extracting infrared features from the standardized multimodal data using an infrared feature extraction network includes: The infrared features were extracted using a ResNet-34 network incorporating deformable convolutions; The ResNet-34 network is used to extract the temperature distribution features from the infrared features; The last layer in the ResNet-34 network is replaced with the deformable convolution, which is used to extract the infrared features from ignition regions of different shapes, angles and sizes in the standardized multimodal data.

6. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 1, characterized in that, The environmental status information includes at least: the light intensity, severe weather level, and visibility level of the target area; The sensor status information includes at least: visible light image quality, infrared data quality, and smoke sensor confidence level.

7. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 1, characterized in that, The construction of the reward function includes: Construct rewards based on detection accuracy, efficiency, and robustness. The detection accuracy reward includes: a correct detection reward, a false alarm penalty, and a false alarm penalty. The correct detection reward is +1 when the fire is correctly detected; the false alarm penalty is -0.5 when the fire is incorrectly detected; and the false alarm penalty is -1 when the fire is not detected. The efficiency bonus is -0.1 for each of the multi-source sensors used; The robustness bonus is: when the detection accuracy is greater than or equal to 90%, the bonus is +0.

5.

8. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 4, characterized in that, The construction of the combined policy network based on reinforcement learning, and the training of the combined policy network using the proximal policy optimization algorithm and the experience replacement mechanism, includes: Step 1: Initialize the reinforcement learning network; Step 2: Construct an experience replacement buffer, which is used to store state transition tuples generated during training; Step 3: Obtain training data through sampling and interaction; Step 4: Randomly sample a portion of the sample data from the experience replacement buffer, and optimize the objective function through the proximal policy to update the parameters of the Actor network in the combined policy network; Step 5: Update the parameters of the Critic network in the combined policy network using the temporal difference error and mean square error loss functions; Step 6: Repeat steps 3 to 5 until the combined policy network converges to obtain the trained combined policy network that can output the weight coefficients.

9. The smoke detection method based on deep reinforcement learning and feature fusion according to claim 8, characterized in that, The environmental adaptation preprocessing of the original multimodal data to obtain the processing result includes: Calculate the darkness index of the original visible light image data; perform dynamic temperature compensation on the original infrared data according to the ambient temperature and humidity to obtain compensated infrared data; use Kalman filtering to eliminate random noise in the original smoke concentration data to obtain noise-reduced smoke concentration data. The complete state space information and the reward function are used for: In step 3, the complete state space information is used as the input to the Actor network and combined with the processing result to calculate the sensor weight allocation action in the current state; After the sampling and interaction are completed, the reward function is called to calculate the immediate reward value and the immediate reward value is stored in the state transition tuple of the experience replacement buffer; In steps 4 and 5, the partial sample data contains the complete state space information, and the temporal difference error and the mean square error loss function are calculated through the reward function.

10. A smoke detection system based on deep reinforcement learning and feature fusion, characterized in that, The system is applied to the method according to any one of claims 1 to 9, the system comprising: A multimodal data acquisition and preprocessing module is used to acquire the raw multimodal data of the forest environment using the multi-source sensors carried by the UAV, and to perform time synchronization processing on the raw multimodal data to obtain the standardized multimodal data. A multimodal feature fusion module is connected to the multimodal data acquisition and preprocessing module. The multimodal feature fusion module is used to extract different data features of the standardized multimodal data using multiple feature extraction methods, and to fuse and splice the different data features using the cross-modal attention mechanism to obtain the multimodal fused feature. A combined policy network construction module is connected to the multimodal feature fusion module. The combined policy network construction module is used to collect the environmental state information of the forest environment and the sensor state information of the multi-source sensors, fuse the environmental state information and the sensor state information into the complete state space information, construct the combined policy network and the reward function based on the reinforcement learning network, and train the combined policy network using the proximal policy optimization algorithm and the experience replacement mechanism. An adaptive sensor weight decision module is connected to the combined policy network construction module. The adaptive sensor weight decision module is used to perform the environmental adaptive preprocessing on the original multimodal data to obtain the processing result. Based on the processing result, the complete state space information and the reward function are input into the trained combined policy network to calculate the weight coefficients of the multi-source sensors. A fire probability comprehensive assessment and early warning module is connected to the adaptive sensor weight decision module. The fire probability comprehensive assessment and early warning module is used to identify the multimodal fusion features using the target recognition algorithm to obtain different detection results, calculate the comprehensive fire probability using the different detection results and the corresponding weight coefficients, preset the fire early warning threshold, compare the comprehensive fire probability with the fire early warning threshold to obtain the comparison result, and obtain the corresponding fire early warning level based on the comparison result.

Images