Smoke detection method based on deep learning

Abstract

The application provides a smoke detection method based on deep learning, which comprises the following steps: acquiring RGB image data, environmental humidity data and smoke particle scattering light intensity data of a security monitoring area to obtain multi-source monitoring data; inputting the multi-source monitoring data into a preset lightweight smoke detection model to output a smoke position and a corresponding confidence; the lightweight smoke detection model takes SqueezeNet as a basic framework, and a feature extraction layer comprises a dynamic channel pruning module, a spatial attention mechanism composed of a 3*3 convolution before an output layer and a Sigmoid function; if the confidence reaches a threshold value, a fire risk level is determined in combination with the environmental humidity data; an early warning signal containing the smoke position, the confidence and the fire risk level is generated and pushed to a security terminal. In the application, the defects of current easy missed detection and misjudgment are overcome.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a smoke detection method based on deep learning. Background Technology

[0002] With the widespread adoption of security monitoring technology, smoke detection, as a core component of fire early warning, is increasingly in demand in scenarios such as homes, commercial venues, and industrial parks. Currently, mainstream smoke detection technologies fall into two main categories: one is based on traditional sensors, relying on hardware devices such as smoke particle sensors to detect physical quantities like the concentration of smoke particles in the environment to trigger an alarm. However, this type of technology is susceptible to environmental interference (such as dust), resulting in high false alarm rates, limited detection range, and the inability to pinpoint the location of smoke. The other category is based on computer vision, which uses cameras to collect image data and combines it with deep learning models to achieve smoke recognition and location. Compared to traditional sensor technologies, this approach offers advantages such as wider coverage and visualized positioning, making it the mainstream direction of current technological development.

[0003] However, existing deep learning-based smoke detection technologies still have many technical challenges in practical security applications: First, there is insufficient fusion of multi-source data. Most solutions rely solely on single RGB image data for detection, resulting in low differentiation between smoke and background (such as fog and shadows) in complex environments, easily leading to missed detections or false positives. Summary of the Invention

[0004] The main objective of this invention is to provide a smoke detection method based on deep learning, which aims to overcome the shortcomings of current methods that are prone to missed detections and false positives.

[0005] To achieve the above objectives, this invention provides a smoke detection method based on deep learning, comprising the following steps:

[0006] Acquire RGB image data, ambient humidity data, and smoke particle scattered light intensity data of the security monitoring area to obtain multi-source monitoring data;

[0007] The multi-source monitoring data is input into a preset lightweight smoke detection model, which outputs the smoke location and corresponding confidence level. The lightweight smoke detection model is based on the SqueezeNet framework, with a dynamic channel pruning module in the feature extraction layer and a spatial attention mechanism consisting of a 3×3 convolution and a sigmoid function in the output layer.

[0008] If the confidence level reaches the threshold, the fire risk level is determined by combining the environmental humidity data; an early warning signal containing the smoke location, confidence level, and fire risk level is generated and pushed to the security terminal.

[0009] Furthermore, before obtaining multi-source monitoring data, the following steps are taken:

[0010] An adaptive illumination compensation algorithm is applied to the RGB image to eliminate interference from backlight and low light. The ambient humidity data and the light intensity data scattered by smoke particles are converted into a standardized feature matrix according to the collection timestamp.

[0011] Furthermore, the multi-source monitoring data is input into a preset lightweight smoke detection model, which outputs the smoke location and corresponding confidence level, including:

[0012] The multi-source monitoring data is input into the lightweight smoke detection model. Based on the dynamic channel pruning module of the lightweight smoke detection model, the mean and variance of the BatchNorm layer of each feature channel are calculated. Valid feature channels with variances greater than a preset variance threshold are selected, and only valid channels are retained for feature transfer. Based on the spatial attention mechanism, global spatial information is extracted from the feature map through 3×3 convolution, and then a spatial weight map is generated by the Sigmoid function.

[0013] The bounding box regression algorithm is used to determine the rectangular coordinates of the smoke region in the spatial weight map to represent the smoke location. The classification probability value of the smoke region is calculated by combining the cross-entropy loss function as the confidence score. If there are multiple overlapping smoke location prediction boxes, the non-maximum suppression algorithm is used to remove redundant prediction boxes and output a unique smoke location and its corresponding confidence score.

[0014] Furthermore, the spatial attention mechanism assigns a weight coefficient of 1.2-1.5 to feature points corresponding to smoke regions and a weight coefficient of 0.5-0.8 to feature points in non-smoke regions.

[0015] Furthermore, the lightweight smoke detection model also includes a feature correction submodule;

[0016] The feature correction submodule collects the output values ​​of each effective channel during model inference in real time and calculates the fluctuation range of the output value of the same channel in adjacent frames of data.

[0017] If the fluctuation amplitude exceeds the amplitude threshold, the channel weight is dynamically adjusted, the mean of the BatchNorm layer of that channel is fine-tuned, and the weight coefficient of the corresponding channel feature point in the spatial attention mechanism is increased by a preset range on the original basis.

[0018] Furthermore, it also includes:

[0019] Based on the average image brightness in the RGB image data, the illumination intensity of the security monitoring area is calculated. If the illumination intensity is less than the illumination threshold, the preset variance threshold is lowered to a preset value.

[0020] Furthermore, after outputting the smoke location and corresponding confidence level, it also includes:

[0021] If the confidence level is within the preset range, the original 3×3 convolution kernel is replaced with a 5×5 dilated convolution kernel, and the global spatial features of the suspected smoke region are extracted again to generate a secondary weight map.

[0022] For the area overlap rate of the smoke region in the spatial weight map and the secondary weight map, if the overlap rate is greater than 0.8, the original confidence level is increased by 0.1-0.2; if the overlap rate is ≤0.8, it is judged as environmental interference, and the original confidence level is decreased by 0.1-0.15.

[0023] The present invention also provides a smoke detection device based on deep learning, comprising:

[0024] The acquisition unit is used to acquire RGB image data, ambient humidity data, and smoke particle scattered light intensity data of the security monitoring area to obtain multi-source monitoring data.

[0025] The output unit is used to input the multi-source monitoring data into a preset lightweight smoke detection model and output the smoke location and corresponding confidence level. The lightweight smoke detection model is based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a spatial attention mechanism composed of a 3×3 convolution and a sigmoid function.

[0026] The generation unit is used to determine the fire risk level by combining the environmental humidity data if the confidence level reaches a threshold; generate an early warning signal containing the smoke location, confidence level and fire risk level, and push it to the security terminal.

[0027] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0028] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0029] This invention provides a deep learning-based smoke detection method, comprising: acquiring RGB image data, ambient humidity data, and smoke particle scattering intensity data of a security monitoring area to obtain multi-source monitoring data; inputting the multi-source monitoring data into a preset lightweight smoke detection model, and outputting the smoke location and corresponding confidence level; the lightweight smoke detection model is based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a spatial attention mechanism composed of a 3×3 convolution and a sigmoid function; if the confidence level reaches a threshold, the fire risk level is determined in conjunction with the ambient humidity data; a warning signal containing the smoke location, confidence level, and fire risk level is generated and pushed to the security terminal. In this invention, RGB image data, ambient humidity data, and smoke particle scattering intensity data are fused, and then based on the lightweight smoke detection model based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a 3×3 convolution and a sigmoid function, the multi-source data is fused and analyzed, overcoming the current defects of easy missed detection and false judgment. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the steps of a smoke detection method based on deep learning in one embodiment of the present invention;

[0031] Figure 2 This is a structural block diagram of a smoke detection device based on deep learning in one embodiment of the present invention;

[0032] Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0033] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0034] Reference Figure 1 An embodiment of the present invention provides a smoke detection method based on deep learning, comprising the following steps:

[0035] Step S1: Obtain RGB image data, ambient humidity data, and smoke particle scattered light intensity data of the security monitoring area to obtain multi-source monitoring data;

[0036] Step S2: Input the multi-source monitoring data into a preset lightweight smoke detection model and output the smoke location and corresponding confidence level; the lightweight smoke detection model is based on SqueezeNet, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a spatial attention mechanism composed of a 3×3 convolution and a sigmoid function.

[0037] Step S3: If the confidence level reaches the threshold, determine the fire risk level by combining the environmental humidity data; generate an early warning signal containing the smoke location, confidence level and fire risk level, and push it to the security terminal.

[0038] It is particularly important to note that all technical steps, algorithm applications, and parameter settings in the technical solution of this application have clear technical objectives and application value. They do not utilize complex steps and algorithmic formulas to achieve simple functions. To provide detailed explanations of each step and avoid ambiguity, some conventional algorithms are used for illustration. However, this does not mean that the algorithms and technical features listed herein are the only way to implement the technical solution of this application, nor is it intended to limit the scope of protection of this application. This application is not a combination or stacking of the listed algorithms and technical features; its essence is to exemplify the implementation methods of this application to fully explain it. It does not pursue formal complexity by adding meaningless technical steps, nor does it involve the accumulation of technologies divorced from practical needs; it conforms to the conventional logic of technical improvement and design.

[0039] In this embodiment, as described in step S1 above, during the specific implementation process, it is first necessary to deploy suitable data acquisition equipment in the security monitoring area. RGB image data is acquired through a monitoring camera with high-definition imaging capabilities. This camera must support real-time frame acquisition (frame rate not less than 25fps) to ensure that the dynamic characteristics of smoke morphology changes can be captured. Ambient humidity data is acquired through a high-precision humidity sensor integrated into the monitoring equipment or deployed independently. The acquisition accuracy must be controlled within ±2%RH, and the sampling period must be synchronized with the image acquisition period (e.g., once per second) to avoid affecting the subsequent fusion effect due to data timing deviation. Smoke particle scattered light intensity data is acquired through a laser scattering smoke sensor. This sensor detects the scattered light intensity of the smoke particles by emitting a laser of a specific wavelength (e.g., 650nm). The output light intensity signal must be converted into a digital value (unit: dB), and the sampling frequency must be consistent with the humidity data.

[0040] After completing the real-time acquisition of the three types of data, the raw data needs to be preliminarily processed to construct a multi-source monitoring dataset. This process requires establishing a data association mechanism, using the acquisition timestamp as a unique identifier to associate and bind RGB image frames, ambient humidity values, and smoke particle scattered light intensity values ​​acquired at the same time, forming a complete multi-source monitoring data record. Introducing RGB image data can intuitively reflect the spatial distribution and morphological characteristics of smoke; introducing ambient humidity data can help distinguish smoke from water vapor (water vapor is easily misidentified as smoke in high humidity environments); introducing smoke particle scattered light intensity data can quantify the concentration of smoke particles from a physical property perspective. The synergistic effect of these three factors can effectively compensate for the limitations of single-data-dimensional detection and improve the accuracy of subsequent detection.

[0041] As described in step S2 above, a preset lightweight smoke detection model is used to extract and analyze features from multi-source monitoring data, ultimately outputting the smoke location and corresponding confidence level. During the model input stage, the constructed multi-source monitoring data must first be converted into a data format suitable for the model input format: RGB image data needs to be scaled to the model's preset input size (e.g., 320×320 pixels) and normalized (mapping pixel values ​​from 0-255 to the 0-1 range); environmental humidity data and smoke particle scattering light intensity data need to be standardized separately (converting values ​​to a standard normal distribution based on the mean and standard deviation of historical datasets), then integrated into a 1×2 dimensional numerical vector. This vector is then fused with the processed RGB image data through feature concatenation to form an input feature map with dimensions of 320×320×3+1×2, ensuring that multi-source data can be simultaneously received and processed by the model.

[0042] The model inference process relies on a lightweight detection model based on the SqueezeNet framework. This framework, through the Squeeze-Excitation structure, reduces the number of model parameters (only 1 / 50th of the traditional AlexNet) while maintaining detection accuracy, making it suitable for the limited computing power of terminal devices in security scenarios. The dynamic channel pruning module of the feature extraction layer plays a crucial role: during model inference, this module calculates the mean and variance of the BatchNorm layer for each feature channel in real time. It filters effective feature channels using a preset variance threshold (e.g., 0.1). Channels with variance greater than the threshold are identified as containing valid smoke features and retained for subsequent feature transfer; channels with variance less than the threshold are identified as containing redundant information (e.g., background interference features) and are masked to reduce computational resource consumption. Through dynamic channel pruning, the computational load during model inference can be further reduced without losing key features, improving real-time performance.

[0043] After feature extraction, a spatial attention mechanism embedded before the model's output layer optimizes the extracted feature map. This mechanism first extracts global spatial information from the feature map using a 3×3 convolutional kernel. The choice of a 3×3 kernel balances the capture capabilities of local and global features, avoiding overfitting of local information due to an excessively small kernel size. Then, the convolutionally processed feature map is input into a sigmoid function to generate a spatial weight map with values ​​ranging from 0 to 1. Regions with higher values ​​in the weight map correspond to regions with more significant smoke features in the feature map. Through this spatial attention mechanism, higher weights are assigned to feature points corresponding to smoke regions, while lower weights are assigned to feature points in non-smoke regions (such as walls and furniture), thus strengthening the response intensity of smoke features and reducing background interference.

[0044] Finally, the model completes the result calculation through two branches: the regression branch uses a bounding box regression algorithm, which fits the smallest rectangle that can encompass the smoke area based on the coordinate information of the high-weight region in the spatial weight map. The coordinates of the top left corner (x1, y1) and the bottom right corner (x2, y2) of this rectangle jointly represent the smoke location; the classification branch combines the cross-entropy loss function to calculate the probability value of the smoke area belonging to the smoke category. This probability value is the confidence score of the corresponding smoke location (ranging from 0 to 1). The higher the confidence score, the stronger the reliability of the model in determining that the area is smoke. If the model outputs multiple smoke location prediction boxes that overlap (the overlap area ratio exceeds the preset IOU threshold, such as 0.5), a non-maximum suppression algorithm is used to remove redundant prediction boxes, retaining only the prediction box with the highest confidence score as the final output result, ensuring the uniqueness and accuracy of the smoke location output.

[0045] As in step S3 above, in the risk level determination stage, a confidence threshold must first be set (this threshold can be dynamically adjusted based on the accuracy requirements of the historical detection dataset, such as setting it to 0.7): if the confidence level output in step S2 is lower than this threshold, it is determined that there is no smoke in the current monitoring area or the smoke concentration does not reach the warning standard, and the risk level determination is not triggered; if the confidence level reaches or exceeds this threshold, it is determined that there is smoke in the current area, and the fire risk level needs to be further determined by combining the environmental humidity data collected in step S1.

[0046] Risk level determination requires establishing a correlation rule between humidity and fire risk: Since ambient humidity directly affects the combustion and diffusion rate of smoke (dry air in low humidity environments allows smoke to quickly ignite surrounding combustibles, resulting in a higher fire risk; humid air in high humidity environments slows down the combustion and diffusion rate of smoke, resulting in a relatively lower fire risk), humidity data needs to be divided into different ranges and matched with corresponding risk levels. For example, humidity <40%RH is defined as a low humidity range, corresponding to a high fire risk level (even a small amount of smoke can quickly ignite a fire, requiring an emergency warning); humidity between 40%-60%RH is defined as a medium humidity range, corresponding to a medium fire risk level (smoke combustion and diffusion rate is moderate, requiring timely warnings and monitoring of fire development); humidity >60%RH is defined as a high humidity range, corresponding to a low fire risk level (smoke combustion and diffusion rate is slow, requiring a warning but not emergency response). By combining confidence levels with ambient humidity data, the problems of over-warning or delayed warnings caused by judging risk solely based on the presence or absence of smoke can be avoided, achieving accurate quantification of risk levels.

[0047] During the early warning signal generation phase, the early warning signal content must be constructed, containing key information: This signal must at least cover three core types of information: smoke location (the coordinates of the rectangle output in step S2), confidence level (the probability value output in step S2), and fire risk level (a level determined based on humidity). Supplementary information such as data collection timestamps and monitoring area numbers can also be added to ensure the integrity and traceability of the early warning signal. The format of the early warning signal must be adapted to the receiving and display requirements of the security terminal. For example, using JSON format to encapsulate the data facilitates parsing and display by the terminal device. For terminals that support image display, the rectangle representing the smoke location must be overlaid on the corresponding RGB image frame to generate a visual image with smoke markers, intuitively presenting the specific location of the smoke in the monitored area.

[0048] Finally, the generated warning signals are pushed to designated security terminals via preset communication protocols (such as MQTT and HTTP). These terminals include the large display screen in the security monitoring center, mobile terminals for management personnel (such as mobile apps), and on-site audible and visual alarms. Signals pushed to the monitoring center screen must be displayed in real time to allow management personnel to quickly grasp the situation on-site; signals pushed to mobile terminals must be presented as message notifications to ensure management personnel can receive warning information promptly even when not at the monitoring center; and signals pushed to on-site audible and visual alarms must trigger alarms (such as emitting a high-decibel alarm sound and flashing red warning lights) to alert on-site personnel to evacuate or take appropriate action. Through multi-terminal collaborative pushing, comprehensive coverage of warning information can be achieved, maximizing security response efficiency.

[0049] In one embodiment, before obtaining multi-source monitoring data, the process includes:

[0050] An adaptive illumination compensation algorithm is applied to the RGB image to eliminate interference from backlight and low light. The ambient humidity data and the light intensity data scattered by smoke particles are converted into a standardized feature matrix according to the collection timestamp.

[0051] In this embodiment, the aim is to eliminate interference from the original data and unify the data format, providing high-quality input for subsequent multi-source data fusion and model inference, and avoiding the impact of original data quality issues on detection accuracy. Specifically, this includes two key operations:

[0052] To address lighting interference such as backlighting (where excessively dark foreground causes smoke to blend into a dark background) and low light (where low overall brightness obscures smoke details), the lighting type is first determined through grayscale histogram analysis, followed by differentiated processing. Backlighting scenes utilize zoned dynamic exposure (16×16 grid division, gamma correction in dark areas, and gain reduction in bright areas), while low light scenes employ multi-scale Retinex enhancement (decomposing illumination and reflection components), ultimately achieving uniform image brightness and clear smoke details.

[0053] To address the significant differences in the dimensions and numerical ranges between humidity (0-100% RH) and the intensity of light scattered by smoke particles (-50 to 0dB), Z-score standardization was first used to eliminate these differences (converting historical data by mean and standard deviation to a standard normal distribution with a mean of 0 and a standard deviation of 1). Then, a two-dimensional feature matrix was constructed based on the collection timestamp (rows correspond to timestamps, and columns correspond to the two types of standardized data), achieving data dimension unification and temporal correlation, which can be accurately matched with RGB images at the same timestamp.

[0054] The combination of these two operations enables multi-source data to possess high quality, structure, and strong correlation characteristics, laying a reliable data foundation for subsequent model inference.

[0055] In one embodiment, the multi-source monitoring data is input into a preset lightweight smoke detection model, and the smoke location and corresponding confidence level are output, including:

[0056] The multi-source monitoring data is input into the lightweight smoke detection model. Based on the dynamic channel pruning module of the lightweight smoke detection model, the mean and variance of the BatchNorm layer of each feature channel are calculated. Valid feature channels with variances greater than a preset variance threshold are selected, and only valid channels are retained for feature transfer. Based on the spatial attention mechanism, global spatial information is extracted from the feature map through 3×3 convolution, and then a spatial weight map is generated by the Sigmoid function.

[0057] The bounding box regression algorithm is used to determine the rectangular coordinates of the smoke region in the spatial weight map to represent the smoke location. The classification probability value of the smoke region is calculated by combining the cross-entropy loss function as the confidence score. If there are multiple overlapping smoke location prediction boxes, the non-maximum suppression algorithm is used to remove redundant prediction boxes and output a unique smoke location and its corresponding confidence score.

[0058] In this embodiment, preprocessed multi-source monitoring data is input into a lightweight smoke detection model, which is first processed by a dynamic channel pruning module. The dynamic channel pruning module calculates the mean and variance of the BatchNorm layer for each feature channel. The mean and variance reflect the dispersion of channel features; the larger the variance, the richer the effective smoke features contained in the channel. A preset variance threshold (e.g., 0.1) is set, and channels with variances greater than this threshold are selected as effective feature channels, while redundant channels (containing invalid information such as background interference) with variances less than the threshold are removed. Only the features transmitted by effective channels are retained, which reduces the computational load of the model, improves real-time performance, and avoids interference from redundant information, laying the foundation for subsequent accurate feature extraction.

[0059] Based on the selected effective feature channels, the model optimizes the feature map through a spatial attention mechanism. First, a 3×3 convolution is used to extract global spatial information from the feature map. The 3×3 convolution kernel balances local and global feature capture, avoiding local overfitting and fully capturing the spatial distribution information of smoke. Then, the convolved feature map is input into a Sigmoid function to generate a spatial weight map with values ​​ranging from 0 to 1. Higher weight values ​​correspond to more significant smoke features. This mechanism strengthens feature points in smoke areas and weakens them in non-smoke areas, highlighting key smoke features and further reducing the impact of background interference.

[0060] The bounding box regression algorithm is used to analyze the spatial weight map and fit the smallest rectangular box that encloses the high-weight smoke region. The coordinates of the top-left corner (x1, y1) and bottom-right corner (x2, y2) of the rectangle represent the smoke location. Simultaneously, the cross-entropy loss function is used to calculate the classification probability value of this region belonging to the smoke category, which serves as the confidence score. The higher the probability, the stronger the model's reliability. This step transforms feature analysis into specific detection results, providing core data support for subsequent early warnings and ensuring the accuracy of location and confidence score outputs.

[0061] If the model outputs multiple predicted smoke locations with overlapping areas (e.g., the overlapping area exceeds a preset IOU threshold of 0.5), a non-maximum suppression algorithm is used. This algorithm retains the highest-confidence box among the overlapping boxes and removes the remaining redundant boxes, preventing the same smoke area from being detected multiple times. Finally, it outputs a unique smoke location and its corresponding confidence level, ensuring the uniqueness and accuracy of the detection results and meeting the precise smoke localization requirements of security scenarios.

[0062] In one embodiment, the spatial attention mechanism assigns a weight coefficient of 1.2-1.5 to feature points corresponding to smoke regions and a weight coefficient of 0.5-0.8 to feature points in non-smoke regions.

[0063] In this embodiment, during the optimization of feature maps using the spatial attention mechanism, a differentiated weighting coefficient allocation strategy is adopted for feature points in smoke and non-smoke regions: feature points identified as smoke regions are assigned a weighting coefficient of 1.2-1.5; feature points identified as non-smoke regions (such as walls, furniture, etc. in the background) are assigned a weighting coefficient of 0.5-0.8. The core purpose of this coefficient setting is to strengthen smoke features and suppress background interference: a high weighting coefficient of 1.2-1.5 can amplify the role of smoke region feature points in subsequent model inference, making the model more focused on key features such as the shape and edges of the smoke; a low weighting coefficient of 0.5-0.8 can weaken the influence of non-smoke region feature points, avoiding background information interfering with the model's judgment of smoke.

[0064] By clearly allocating coefficient intervals, the contribution of features in different regions can be precisely controlled, further improving the spatial weight map's accuracy in identifying smoke regions. This provides a more reliable feature basis for subsequent smoke location and confidence calculation, helping the model reduce the false positive rate.

[0065] In one embodiment, the lightweight smoke detection model further includes a feature correction submodule;

[0066] The feature correction submodule collects the output values ​​of each effective channel during model inference in real time and calculates the fluctuation range of the output value of the same channel in adjacent frames of data.

[0067] If the fluctuation amplitude exceeds the amplitude threshold, the channel weight is dynamically adjusted, the mean of the BatchNorm layer of that channel is fine-tuned, and the weight coefficient of the corresponding channel feature point in the spatial attention mechanism is increased by a preset range on the original basis.

[0068] In this embodiment, the feature correction submodule first collects the output values ​​of each effective feature channel in real time during model inference. Effective channels are those filtered by the dynamic channel pruning module (with variance greater than a preset threshold), and their output values ​​directly reflect the smoke feature information extracted by the current channel. Data acquisition must be synchronized with model inference to ensure data timeliness. Subsequently, the module calculates the fluctuation range of the output value of the same effective channel in adjacent frames (e.g., 5 consecutive frames). By calculating the average of the absolute values ​​of the differences between the channel's output values ​​in adjacent frames, the stability of the channel's feature output is quantified. This step provides data for subsequent judgments on whether channels need adjustment, avoiding feature extraction deviations caused by channel output fluctuations.

[0069] When the calculated fluctuation amplitude exceeds a preset threshold (e.g., 15%), it indicates poor stability of the feature output of the effective channel, which may be affected by environmental interference (e.g., sudden changes in illumination), requiring dynamic adjustment of the channel weights. The adjustment consists of two parts: first, fine-tuning the mean of the BatchNorm layer for that channel (e.g., ±8%) to optimize the channel feature standardization effect and improve feature consistency; second, in the spatial attention mechanism, increasing the weight coefficients of the corresponding channel feature points by a preset range (e.g., 0.1-0.3) to further enhance the response of the effective smoke features of that channel. If the fluctuation amplitude does not exceed the threshold, the channel parameters remain unchanged, ensuring efficient inference under stable conditions and balancing accuracy and computational efficiency.

[0070] In one embodiment, it further includes:

[0071] Based on the average image brightness in the RGB image data, the illumination intensity of the security monitoring area is calculated. If the illumination intensity is less than the illumination threshold, the preset variance threshold is lowered to a preset value.

[0072] In this embodiment, the feature correction submodule also combines RGB image data to achieve lighting scene adaptation. Specifically, it first calculates the average brightness of the image based on the RGB image data: by extracting the R, G, and B channel values ​​of all pixels in each frame of the RGB image, the brightness value of each pixel is calculated according to the formula (R+G+B) / 3, and then the average brightness value of all pixels in the entire frame is taken to obtain the average brightness of the frame, thereby indirectly quantifying the light intensity of the security monitoring area (the lower the average brightness value, the weaker the light intensity).

[0073] The calculated light intensity is then compared with a preset light threshold (e.g., a brightness average of 50). If the light intensity is less than this threshold (indicating a low-light scene), the preset variance threshold of the dynamic channel pruning module is lowered to a preset value (e.g., from 0.1 to 0.08). This relaxes the effective channel selection criteria in low-light scenes, retaining more channels that may contain faint smoke features, avoiding misclassification as redundant channels due to weakened feature signals caused by insufficient light, ensuring that the model can still effectively extract smoke features in low-light environments, and improving the detection adaptability to different lighting scenes.

[0074] In one embodiment, after outputting the smoke location and corresponding confidence level, the method further includes:

[0075] If the confidence level is within the preset range, the original 3×3 convolution kernel is replaced with a 5×5 dilated convolution kernel, and the global spatial features of the suspected smoke region are extracted again to generate a secondary weight map.

[0076] For the area overlap rate of the smoke region in the spatial weight map and the secondary weight map, if the overlap rate is greater than 0.8, the original confidence level is increased by 0.1-0.2; if the overlap rate is ≤0.8, it is judged as environmental interference, and the original confidence level is decreased by 0.1-0.15.

[0077] In this embodiment, after outputting the smoke location and corresponding confidence level, it is first determined whether the confidence level falls within a preset range (e.g., 0.5-0.7, which represents a low-confidence suspected smoke range, requiring further verification of its reliability). If it falls within this range, it indicates that the current smoke determination is not reliable enough, and secondary feature extraction needs to be initiated. Specifically, the original 3×3 convolutional kernel in the spatial attention mechanism is replaced with a 5×5 dilated convolutional kernel (dilation rate set to 2). The 5×5 dilated convolutional kernel can expand the receptive field of feature extraction without increasing computational cost, more fully capturing the global spatial features of the suspected smoke region, and avoiding feature omissions caused by the limited receptive field of the original 3×3 convolutional kernel. Based on the new convolutional kernel, the feature map of the suspected smoke region is reprocessed to generate a secondary spatial weight map, providing a more comprehensive feature basis for subsequent confidence level calibration.

[0078] After generating the secondary weighted map, the overlap rate of the smoke region area between the initial output spatial weighted map and the secondary weighted map is calculated. The overlap rate is obtained by calculating the ratio of the intersection area to the union area of ​​high-weight smoke regions (e.g., regions with a weight value ≥ 0.6) in the two weighted maps. If the overlap rate is greater than 0.8, it indicates a high degree of consistency between the two feature extraction results, and the original smoke judgment is highly reliable. The original confidence level is increased by 0.1-0.2 (e.g., from 0.6 to 0.7-0.8) to enhance the reliability of the smoke judgment. If the overlap rate is ≤ 0.8, it indicates a significant difference between the two feature extractions, and the original judgment may be affected by environmental interference (e.g., dust, light spots). The original confidence level is decreased by 0.1-0.15 (e.g., from 0.6 to 0.45-0.5) to reduce the risk of misjudgment. Through dynamic adjustment, the confidence level is made to better reflect the actual smoke situation, improving detection accuracy.

[0079] In one embodiment, after generating an early warning signal containing the smoke location, confidence level, and fire risk level and pushing it to the security terminal, the system further includes:

[0080] Obtain the hardware identifier of each security terminal, and generate a character matrix from the characters in the hardware identifier; obtain the number of each type of security terminal, and generate a curve based on the number;

[0081] Based on the character matrix, the mean value of the numbers in each matrix element is calculated; the mean value is used as the x-axis of the curve to obtain the y-axis value;

[0082] Obtain a preset encoding algorithm matrix; wherein the encoding algorithm matrix includes multiple encoding algorithm numeric identifiers, and each encoding algorithm numeric identifier corresponds to one encoding algorithm;

[0083] In the coding algorithm matrix, obtain the coding algorithm digital identifier that is closest to the vertical coordinate value, and use it as the target digital identifier. Then, use the coding algorithm corresponding to the target digital identifier as the target coding algorithm.

[0084] Based on the curve, the encoding parameters of the target encoding algorithm are optimized to obtain an optimized encoding algorithm, which is used to encode and store RGB image data, environmental humidity data, and smoke particle scattering light intensity data of the security monitoring area.

[0085] In this embodiment, firstly, through the communication links established with each security terminal, the unique hardware identifiers (such as device serial numbers, network adapter MAC addresses, and other unique identifiers) of all online security terminals are acquired in batches. For each hardware identifier, the characters contained therein (including letters, numbers, and symbols) are extracted, and a character matrix is ​​generated according to the character length of the hardware identifier, following the principle of uniform distribution and regular row and column arrangement.

[0086] Simultaneously, all online security terminals are categorized and statistically analyzed according to their functional types (e.g., monitoring terminals responsible for image acquisition, alarm terminals for risk alerts, relay terminals for data forwarding, etc.), recording the real-time online quantity of each type of terminal. Using terminal type as the horizontal axis (e.g., acquisition terminals, alarm terminals, relay terminals arranged sequentially) and the corresponding online quantity of terminals as the vertical axis, a data fitting algorithm (e.g., linear fitting or cubic polynomial fitting) is employed to smooth the statistical data, simulating and generating terminal quantity curves that reflect the distribution and changing trends of each type of terminal quantity. This provides an intuitive numerical mapping for subsequent parameter calculations.

[0087] For each generated character matrix, all elements (i.e., each character) are traversed one by one, and numeric characters within each element are selected (letters and symbols are filtered out). These numeric characters are then converted into their corresponding numerical values. After obtaining the numerical values ​​corresponding to all numbers in the matrix, the arithmetic mean of these values ​​(i.e., the sum of all values ​​divided by the total number of values) is calculated to obtain the numerical mean of the character matrix. This numerical mean is used as the x-axis of a curve, substituted into the curve's function equation (such as the fitted polynomial equation), and the y-axis value corresponding to this x-axis is calculated. This y-axis value is related to both the characteristics of the terminal hardware identifier and the distribution information of the number of terminals, providing a core reference indicator for subsequent encoding algorithm matching.

[0088] A verified encoding algorithm matrix is ​​pre-stored. This matrix is ​​a two-dimensional structured data. Each row or column of the matrix corresponds to an encoding algorithm digital identifier (composed of integers, such as 1, 2, 3, etc., which are unique). The corresponding position in the matrix records the detailed information of the encoding algorithm associated with the digital identifier (including algorithm name, core principle, applicable scenarios and basic parameters, etc.).

[0089] For example, in the encoding algorithm matrix, number 1 corresponds to the Huffman coding algorithm, suitable for scenarios with high data repetition rates; number 2 corresponds to the LZW coding algorithm, suitable for text or image data; and number 3 corresponds to the arithmetic coding algorithm, suitable for scenarios with high compression ratio requirements. This matrix-based storage method allows for the rapid establishment of a mapping relationship between encoding algorithm numbers and specific encoding algorithms, ensuring that the target algorithm can be quickly located based on numerical values.

[0090] Iterate through all the encoding algorithm identifiers in the preset encoding algorithm matrix, calculate the difference between the value of each identifier and the calculated ordinate value, and obtain the absolute difference between each identifier and the ordinate value. Compare the magnitudes of all absolute differences, and select the encoding algorithm identifier with the smallest absolute difference as the target identifier. If two or more identifiers have the same absolute difference with the ordinate value (i.e., the minimum difference is repeated), select the identifier with the highest success rate of the corresponding encoding algorithm in historical applications as the target identifier.

[0091] Based on the preset correspondence between the target digital identifier and the encoding algorithm, the encoding algorithm associated with the digital identifier is extracted from the encoding algorithm matrix and identified as the target encoding algorithm. This algorithm is the basic encoding algorithm that is most suitable for the current scenario and is selected based on the terminal hardware and quantitative characteristics.

[0092] Furthermore, the characteristics of the generated curves are analyzed, including the slope of the curve (reflecting the increasing or decreasing trend of the number of different types of terminals), the peaks and troughs of the curve (reflecting the extreme value distribution of the number of terminals), and the smoothness of the curve (reflecting the stability of the number of terminals). Based on these characteristics, the basic parameters of the target encoding algorithm are optimized accordingly.

[0093] If the curve slope is positive (indicating an increasing number of terminals of a certain type, suggesting a potential increase in subsequent data transmission volume), increase the number of parallel processing threads in the target encoding algorithm while appropriately reducing the data compression ratio to improve the algorithm's processing speed. If the curve has a significant peak (indicating a much higher number of terminals of a certain type than others, with relatively concentrated data sources), adjust the algorithm's data block size, setting the data block size proportional to the number of terminals of that type to optimize data processing efficiency. If the curve is generally smooth (indicating a stable number of terminals), maintain the algorithm's basic compression ratio while increasing the data verification frequency to improve the accuracy of data encoding.

[0094] After parameter optimization, an optimized encoding algorithm was obtained and applied to multi-source data processing in security monitoring areas. RGB image data, ambient humidity data, and smoke particle scattering light intensity data were encoded in real time. The encoded data streams were then categorized and stored in a designated database according to data type and acquisition time, ensuring both the security and integrity of data storage and the efficiency of subsequent data retrieval.

[0095] In one embodiment, after generating an early warning signal containing the smoke location, confidence level, and fire risk level and pushing it to the security terminal, the system further includes:

[0096] Obtain the hardware identifier of each security terminal, and generate a character matrix from the characters in the hardware identifier; obtain the number of each type of security terminal, and generate a curve based on the number;

[0097] Obtain the matrix features of the character matrix and the curve features of the curve, and adjust the preset encoding table based on the matrix features and curve features to obtain the adjusted encoding table;

[0098] Based on the character matrix and curve features, a first subkey is generated; based on the curve and matrix features, a second subkey is generated.

[0099] Based on the adjustment mapping table, the first subkey and the second subkey are encoded and combined to obtain the target key, which is used to encrypt and store the RGB image data, environmental humidity data and smoke particle scattering light intensity data of the security monitoring area.

[0100] In this embodiment, a data interaction link is first established with all online security terminals through an encrypted communication protocol to obtain the unique hardware identifier of each terminal in batches (such as device MAC address, factory serial number, or chip unique identification code) to ensure the uniqueness and integrity of the identification information. For each hardware identifier, valid characters (including letters and numbers, and filtering out meaningless symbols) are extracted, and a character matrix is ​​generated according to the principle of balanced row and column length based on the total character length.

[0101] Meanwhile, security terminals are categorized into different types based on their functional attributes (such as image acquisition terminals, audible and visual alarm terminals, data relay terminals, and control terminals). The real-time online count of each type of terminal is statistically analyzed through a terminal status reporting mechanism. Using the terminal type name as the horizontal axis and the corresponding online terminal count as the vertical axis, a nonlinear fitting algorithm is employed to smooth the statistical data, eliminating random fluctuations and generating a terminal count curve that accurately reflects the distribution and trends of each type of terminal count. This curve must include specific numerical points for the number of terminals of each type and the slope of the count change between adjacent types, providing a continuous numerical basis for subsequent feature calculations.

[0102] For the generated character matrix, a multi-dimensional feature extraction method is used to obtain matrix features: First, all elements of the matrix are traversed, numeric characters are selected and converted into numerical values, and the arithmetic mean of all numerical values ​​is calculated (denoted as M1), reflecting the degree of concentration of the numerical distribution in the matrix; Second, the frequency of different letter characters in the matrix is ​​counted, and the sum of the ASCII code values ​​corresponding to the three most frequent letters is selected (denoted as M2), reflecting the differences in character features of the matrix; Third, the sum of the ASCII codes of the characters in each row of the matrix is ​​calculated, and the variance of the sum of each row is taken (denoted as M3), representing the feature dispersion between rows of the matrix; M1, M2, and M3 are weighted and calculated to obtain the final matrix feature value, which integrates the numerical, character, and structural features of the matrix and has uniqueness.

[0103] For the generated curve, curve features are extracted from three dimensions: First, the difference between the maximum and minimum values ​​of the number of all types of terminals on the curve (denoted as C1) is calculated, reflecting the overall fluctuation range of the number of terminals; second, the slope between two adjacent terminal types in the curve is solved, and the sum of the absolute values ​​of all slopes is taken (denoted as C2), reflecting the drastic degree of quantity change; third, the number of terminal types with a quantity greater than the average value in the curve is counted (denoted as C3), characterizing the distribution of high-quantity types; C1, C2, and C3 are weighted and calculated to obtain the final curve feature value, which integrates the fluctuation amplitude, rate of change, and distribution characteristics of the curve, and can accurately reflect the dynamic attributes of the number of terminals.

[0104] A standardized preset encoding table is stored in advance. The encoding table is a key-value pair structure, where the key is the original data identifier and the value is the corresponding basic encoding sequence (a hexadecimal string composed of 0-9 and AF, such as "IMG" corresponding to "1A3F").

[0105] First, the base encoding length of the encoding table is adjusted based on the matrix eigenvalues: if the matrix eigenvalue > 500, the base encoding sequence length corresponding to all original data identifiers is extended by 2 bits (e.g., 1A3F is extended to 1A3F7C); if ≤ 500, it is shortened by 1 bit (e.g., 1A3F is shortened to 1A3), ensuring the encoding length matches the matrix eigenvalues. Second, the character replacement rules of the encoding table are adjusted based on the curve eigenvalues: if the curve eigenvalue > 80, all 3s in the encoding sequence are replaced with D, and 7s with B; if ≤ 80, 2s are replaced with E, and 6s with A, integrating curve feature differences through character replacement. Finally, the last 3 bits of the matrix eigenvalues ​​are concatenated with the last 2 bits of the curve eigenvalues ​​as a check suffix for the encoding table, added to the end of each adjusted encoding sequence to form the final adjusted encoding table. This adjusted encoding table retains the structured advantages of the original encoding table and achieves dynamic adaptation through the integration of matrix and curve features, avoiding the security defects of a fixed encoding table.

[0106] When generating the first subkey, the character matrix and curve feature values ​​are used as the core inputs: First, the first element of each row of the character matrix is ​​extracted to form a character sequence. Each character is converted into its corresponding ASCII code value, and the last 4 bits are taken as binary numbers and concatenated to obtain the binary string S1. Second, the curve feature values ​​are converted into 16-bit binary numbers S2. If the number of bits is less than 16, 0s are added to the high bits. Finally, S1 and S2 are concatenated, and the concatenated binary string is processed by converting every 4 bits into hexadecimal characters to obtain the first subkey of fixed length. This key is directly related to the structural features of the matrix and the numerical features of the curve, and has a strong binding to the terminal hardware.

[0107] When generating the second subkey, the terminal quantity curve and matrix feature values ​​are used as core inputs: First, the numerical values ​​corresponding to the number of terminals of each type are extracted from the terminal quantity curve. These values ​​are converted into binary numbers, and the last 6 bits are taken and concatenated to obtain the binary string S3. Second, the matrix feature values ​​are converted into 24-bit binary numbers S4. If the number exceeds 24 bits, the last 24 bits are truncated. Finally, S3 and S4 are XORed to obtain a new binary string. This string is then processed according to the rule of converting 8 bits into decimal numbers, with each decimal number corresponding to one character, and concatenated to form the second subkey. The first and second subkeys use different generation logics, focusing on different dimensions of the matrix and curve features respectively, providing double security for subsequent key combination.

[0108] Then, the adjusted encoding table obtained in step three is called to encode the first subkey and the second subkey respectively: the original data identifier corresponding to the first subkey in the encoding table is found, its adjusted encoding sequence is obtained, and the first subkey is embedded in the specified position of the encoding sequence (such as between the 3rd and 4th positions) to obtain the encoded first subkey; similarly, the original data identifier corresponding to the second subkey is found, the adjusted encoding sequence is obtained, and the second subkey is embedded in it to obtain the encoded second subkey.

[0109] Subsequently, the two encoded subkeys are combined to form a complete target key. This key has a fixed length and includes encoding table verification information, effectively preventing tampering. Finally, the target key is applied to the encrypted storage of multi-source data in the security monitoring area: a symmetric encryption algorithm is used, with the target key as the encryption key, to perform real-time encryption processing on RGB image data (encrypted by frame segmentation), environmental humidity data (encrypted by time slice grouping), and smoke particle scattering light intensity data (encrypted by acquisition period segmentation). The encrypted data is classified and stored in the encrypted database according to the naming rules of data type, encryption timestamp, and the last 4 digits of the target key. Each time data is read, the matching between the target key and the adjusted encoding table must be verified first to ensure the security of data storage and the legitimacy of access.

[0110] In the above embodiments, this application incorporates some existing algorithms and technical features for explanation and description to make the specification more detailed, clear, and complete, thus complying with the provisions of the Patent Law. However, this is not achieved by using a series of complex steps and algorithmic formulas, nor by complicating the technical solution, nor by combining or stacking conventional or simple features. The existing algorithms and technical features listed are for the purpose of disclosing the specific implementation methods of each step of this application (not to limit this application) and to avoid situations where this application cannot be implemented.

[0111] Reference Figure 2 In another embodiment of the present invention, a smoke detection device based on deep learning is also provided, comprising:

[0112] The acquisition unit is used to acquire RGB image data, ambient humidity data, and smoke particle scattered light intensity data of the security monitoring area to obtain multi-source monitoring data.

[0113] The output unit is used to input the multi-source monitoring data into a preset lightweight smoke detection model and output the smoke location and corresponding confidence level. The lightweight smoke detection model is based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a spatial attention mechanism composed of a 3×3 convolution and a sigmoid function.

[0114] The generation unit is used to determine the fire risk level by combining the environmental humidity data if the confidence level reaches a threshold; generate an early warning signal containing the smoke location, confidence level and fire risk level, and push it to the security terminal.

[0115] In this embodiment, the specific implementation of each unit in the above device embodiment is described in the above method embodiment, and will not be repeated here.

[0116] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3 As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0117] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0118] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0119] In summary, the deep learning-based smoke detection method provided in this embodiment of the invention includes: acquiring RGB image data, ambient humidity data, and smoke particle scattering intensity data of a security monitoring area to obtain multi-source monitoring data; inputting the multi-source monitoring data into a preset lightweight smoke detection model, and outputting the smoke location and corresponding confidence level; the lightweight smoke detection model is based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a 3×3 convolution and a sigmoid function forming a spatial attention mechanism; if the confidence level reaches a threshold, the fire risk level is determined in conjunction with the ambient humidity data; a warning signal containing the smoke location, confidence level, and fire risk level is generated and pushed to the security terminal. In this invention, by fusing RGB image data, ambient humidity data, and smoke particle scattering intensity data, and then based on the lightweight smoke detection model based on the SqueezeNet framework, with a feature extraction layer containing a dynamic channel pruning module and an output layer consisting of a 3×3 convolution and a sigmoid function forming a spatial attention mechanism, multi-source data is fused and analyzed, overcoming the current shortcomings of easy missed detections and false judgments.

[0120] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0121] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0122] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based smoke detection method, characterized by, The method comprises the following steps: Obtain RGB image data, environmental humidity data, and smoke particle scattering light intensity data of a security monitoring area to obtain multi-source monitoring data; Input the multi-source monitoring data into a preset lightweight smoke detection model to output a smoke position and a corresponding confidence; When the multi-source monitoring data is input, it needs to be converted into a data form suitable for the input format of the model, including: scaling the RGB image data to the input size preset by the model and performing normalization processing; the environmental humidity data and the smoke particle scattering light intensity data need to be standardized respectively, and then integrated into a numerical vector, fused with the normalized RGB image data through feature splicing operation, and an input feature map is formed; the lightweight smoke detection model is based on a SqueezeNet framework, the feature extraction layer contains a dynamic channel pruning module, an output layer front-embedded 3*3 convolution, and a spatial attention mechanism composed of a Sigmoid function; If the confidence reaches a threshold, determine the fire risk level in combination with the environmental humidity data; generate an early warning signal containing the smoke position, confidence, and fire risk level, and push it to a security terminal. 2.The deep learning-based smoke detection method of claim 1, wherein, Before obtaining the multi-source monitoring data, the following steps are included: Perform adaptive light compensation algorithm on the RGB image to eliminate glare and weak light interference, and convert the environmental humidity data and the smoke particle scattering light intensity data into a standardized feature matrix according to the collection time stamp. 3.The deep learning-based smoke detection method of claim 1, wherein, Input the multi-source monitoring data into a preset lightweight smoke detection model to output a smoke position and a corresponding confidence, including: Input the multi-source monitoring data into the lightweight smoke detection model, calculate the BatchNorm layer mean variance of each feature channel based on the dynamic channel pruning module of the lightweight smoke detection model, select the effective feature channels with variance greater than a preset variance threshold, and only keep the effective channels for feature transmission; based on the spatial attention mechanism, perform global spatial information extraction on the feature map through 3*3 convolution, and then generate a spatial weight map through the Sigmoid function; Determine the rectangular coordinates of the smoke area in the spatial weight map through a bounding box regression algorithm to represent the smoke position, and calculate the classification probability value of the smoke area as the confidence in combination with the cross-entropy loss function; if there are multiple overlapping smoke position prediction boxes, use a non-maximum suppression algorithm to remove redundant prediction boxes, and output a unique smoke position and corresponding confidence. 4.The deep learning-based smoke detection method of claim 1, wherein, The spatial attention mechanism assigns a weight coefficient of 1.2-1.5 to the feature points corresponding to the smoke area, and a weight coefficient of 0.5-0.8 to the feature points of non-smoke areas. 5.The deep learning-based smoke detection method of claim 3, wherein, The lightweight smoke detection model also includes a feature correction sub-module; The feature correction sub-module collects the output values of each effective channel in real time during model inference, and calculates the fluctuation amplitude of the output values of the same channel in adjacent multiple frames of data; If the fluctuation amplitude is greater than an amplitude threshold, trigger dynamic adjustment of the channel weight, fine-tune the BatchNorm layer mean value of the channel, and increase the weight coefficient of the feature points of the corresponding channel in the spatial attention mechanism within a preset range based on the original weight coefficient. 6.The deep learning-based smoke detection method of claim 3, wherein, Also includes: Based on the image brightness mean value in the RGB image data, the light intensity of the security monitoring area is calculated, and if the light intensity is less than the light threshold, the preset variance threshold is lowered to a preset value. 7.The deep learning-based smoke detection method of claim 3, wherein, After outputting the smoke position and the corresponding confidence, the method further comprises: If the confidence is in a preset interval, a 5*5 dilated convolution kernel is used to replace the original 3*3 convolution kernel, and the global spatial features of the smoke suspected area are extracted again to generate a secondary weight map; If the overlap rate of the smoke area in the spatial weight map and the secondary weight map is greater than 0.8, the original confidence is increased by 0.1-0.2; if the overlap rate is less than or equal to 0.8, it is determined as environmental interference, and the original confidence is decreased by 0.1-0.

15. 8.A deep learning-based smoke detection device, characterized by, The method comprises: An acquisition unit is configured to acquire RGB image data, environmental humidity data and smoke particle scattering light intensity data of a security monitoring area to obtain multi-source monitoring data; An output unit is configured to input the multi-source monitoring data into a preset lightweight smoke detection model and output a smoke position and a corresponding confidence; When the multi-source monitoring data is input, it needs to be converted into a data form suitable for the input format of the model, including scaling the RGB image data to the input size preset by the model and performing normalization processing; the environmental humidity data and the smoke particle scattering light intensity data need to be standardized respectively, and then integrated into a numerical vector, fused with the normalized RGB image data through a feature splicing operation, and an input feature map is formed; the lightweight smoke detection model is based on a SqueezeNet framework, and the feature extraction layer contains a dynamic channel pruning module, an output layer, a 3*3 convolution embedded in front of the output layer and a spatial attention mechanism composed of a Sigmoid function; A generation unit is configured to determine a fire risk level in combination with the environmental humidity data if the confidence reaches a threshold value, and generate an early warning signal containing the smoke position, the confidence and the fire risk level and push it to a security terminal.

Images