Intelligent monitoring system for early fire prevention of power distribution room with multi-modal fusion

Abstract

The application discloses a kind of early fire prevention intelligent monitoring systems of multi-modal fusion power distribution room, specifically related to fireproof intelligent monitoring technical field, the present application obtains heat release particle concentration by calculating the light intensity of particle angular scattering, combined with particle concentration and diameter dynamic adjustment scattering angle, when heat release particle concentration exceeds preset threshold, constructs double-branch symmetric fusion network, and fuses visible light and infrared image branch.By global attention mechanism, the spatial dependence of power equipment is captured, and the segmentation result of the fusion network is output.The system also determines whether to enhance the visible light image according to the particle concentration and the picture light intensity.Finally, the segmentation result, particle concentration trend and temperature change data are fused to trigger fire warning.The system realizes efficient early fire warning for power distribution room, and improves the intelligence and precision of fire monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire prevention monitoring technology, and more specifically, to an intelligent monitoring system for early fire prevention in power distribution rooms with multimodal fusion. Background Technology

[0002] Research in the field of artificial intelligence in the power sector primarily focuses on the detection and identification of common power scenario defect types or specific problems. As a critical node in power transmission and distribution, the fire safety of equipment in power distribution rooms is paramount, directly impacting the stable operation and reliability of the power system. Therefore, effective fire prevention in power distribution rooms is of utmost importance.

[0003] In existing technologies, the scattering angle is usually set to a fixed value when calculating the concentration of microparticles, which leads to inaccurate calculations of microparticle concentration. At the same time, when inputting visible light images, features may not be fully extracted due to fire or dust.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an intelligent monitoring system with multimodal fusion power distribution room early fire prevention to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An intelligent monitoring system for early fire prevention in power distribution rooms with multimodal fusion includes the following steps:

[0008] First, the intensity of scattered light at the particle angle is calculated, and the trend of change of pyroelectric particle concentration is obtained based on the relationship between particle concentration and intensity of scattered light at the particle angle.

[0009] When the concentration of pyroelectric microparticles exceeds the system's preset threshold, a bi-branch symmetric fusion network is constructed. The bi-branch symmetric fusion network is used to fuse the visible light branch and the infrared branch, capture the spatial dependency of the power equipment through a global attention mechanism, and output the segmentation result of the bi-branch symmetric fusion network.

[0010] Then, based on the concentration of pyroelectric microparticles and the intensity of ambient light, it is determined whether image enhancement of the visible light image is necessary; finally, the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and temperature change data are fused to trigger an early warning.

[0011] In a preferred embodiment, based on Mie scattering theory, the number of pyroelectric microparticles is calculated by measuring the incident and scattered intensities of the light beam before and after passing through the microparticles; when the microparticle concentration is N, the angular scattered light intensity is: In the formula, θ is the scattering angle; I(θ) is the intensity of the scattered light at the scattering angle θ; λ is the wavelength of the light source; I0 is the incident light intensity; d is the particle diameter; m is the refractive index of light; and r is the distance from the scattering body.

[0012] Therefore, a constant C is defined: Therefore, the concentration of microparticles and the intensity of angularly scattered light have the following relationship: I(θ)=NC(1+cos 2 θ).

[0013] In a preferred embodiment, θ is dynamically updated; an update coefficient is calculated using a weighted summation based on the particle concentration and particle diameter; if the update coefficient is less than a system preset threshold, θ does not need to be updated; if the update coefficient is greater than the system preset threshold, θ is updated, and θ is adjusted according to the value of the update coefficient; and the particle concentration is recalculated.

[0014] In a preferred embodiment, when the concentration of pyroelectric microparticles exceeds a preset threshold of the system, a bi-branch symmetric fusion network is constructed to locate possible ignition points.

[0015] In a preferred embodiment, the dual-branch symmetric fusion network uses a visual Transformer as a basic block to extract features from the visible light branch and the infrared branch. Then, based on the existing dual-branch feature extraction architecture, an efficient multi-scale feature fusion layer is designed to deeply mine complementary information from another modality in the feature modeling relationship between different modalities, and finally obtain the fused features.

[0016] In a preferred embodiment, the concentration of pyroelectric microparticles and the light intensity of the image are determined, and the enhancement coefficient is calculated using a weighted summation formula. If the weight coefficient is greater than the system threshold, then the visible light branch of the symmetric fusion network that requires input dual branches needs to be enhanced; otherwise, it is not necessary.

[0017] In a preferred embodiment, after obtaining multi-scale fusion features, scale restoration and detail restoration are performed on the fusion features at different scales to determine the segmentation result of the bi-branch symmetric fusion network.

[0018] In a preferred embodiment, the module includes the following modules: an optical detection and microparticle analysis module, a bi-branch symmetric fusion network module, a multi-scale decoding and output module, and a decision-making and early warning module.

[0019] The optical detection and microparticle analysis module is used to measure the intensity of angular scattered light from microparticles in real time based on Mie scattering theory using a laser emitter and photoelectric sensor, and to calculate the concentration of pyroelectric microparticles based on the intensity of angular scattered light from microparticles; when the concentration of pyroelectric microparticles exceeds the preset threshold of the system, a two-branch symmetric fusion network is constructed to locate the possible ignition point;

[0020] The dual-branch symmetric fusion network module is used to fuse the visible light branch and the infrared branch, and captures the spatial dependencies of power equipment through a global attention mechanism.

[0021] The multi-scale decoding and output module is used to recover feature scales step by step from high to low levels, and retains global context through cross-layer decoding; a channel attention mechanism is introduced to calculate global discriminative weights and enhance the feature representation of key regions;

[0022] The decision-making and early warning module is used to integrate the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and the temperature change data; if the segmented region abnormally overlaps with the microparticle concentration / temperature, an early warning is triggered.

[0023] In a preferred embodiment, the optical detection and microparticle analysis module includes a dynamic scattering angle adjustment module. The dynamic scattering angle adjustment module is used to calculate an update coefficient based on the microparticle concentration and microparticle diameter using a weighted summation. If the update coefficient is less than a system preset threshold, the scattering angle does not need to be updated. If the update coefficient is greater than the system preset threshold, the scattering angle is updated and adjusted according to the value of the update coefficient. The microparticle concentration is then recalculated.

[0024] In a preferred embodiment, the dual-branch symmetric fusion network module includes an image enhancement module, which is used to determine whether to enhance the visible light image based on the concentration of pyroelectric microparticles and the intensity of light in the image.

[0025] The technical effects and advantages of this invention are as follows:

[0026] This invention first calculates the angular scattered light intensity of the microparticles, and then dynamically adjusts the scattering angle based on the microparticle concentration and the microparticle diameter, so that the pyroelectric microparticle concentration obtained based on the relationship between the microparticle concentration and the angular scattered light intensity of the microparticles is more accurate.

[0027] When the concentration of pyroelectric microparticles exceeds the system's preset threshold, a bi-branch symmetric fusion network is constructed. This bi-branch symmetric fusion network is used to fuse the visible light branch and the infrared branch, and captures the spatial dependencies of power equipment through a global attention mechanism, outputting the segmentation result fused from the bi-branch symmetric fusion network. When inputting the bi-branch symmetric fusion network, it is first necessary to determine whether to enhance the visible light image based on the pyroelectric microparticle concentration and the image illumination intensity, making the input more accurate and improving the fault tolerance of the bi-branch symmetric fusion network's segmentation result.

[0028] Finally, the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and the temperature change data are integrated to trigger an early warning. Attached Figure Description

[0029] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0030] Figure 1 This is a flowchart illustrating an intelligent monitoring system for early fire prevention in a power distribution room with multimodal fusion, as described in this invention.

[0031] Figure 2 This is a schematic diagram of a symmetrical fusion network structure with two branches;

[0032] Figure 3 This is a schematic diagram of the fusion layer structure;

[0033] Figure 4 This is a schematic diagram of the decoding process;

[0034] Figure 5 This is a schematic diagram of the structure of an intelligent monitoring system for early fire prevention in a power distribution room with multimodal fusion according to the present invention; Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention first calculates the angular scattered light intensity of the particles, and then dynamically adjusts the scattering angle based on the particle concentration and particle diameter, making the pyroelectric particle concentration obtained based on the relationship between particle concentration and angular scattered light intensity more accurate. When the pyroelectric particle concentration exceeds a preset threshold, a bi-branch symmetric fusion network is constructed. This bi-branch symmetric fusion network is used to fuse the visible light branch and the infrared branch, and captures the spatial dependencies of power equipment through a global attention mechanism, outputting the segmentation result fused from the bi-branch symmetric fusion network. When inputting the bi-branch symmetric fusion network, it is first necessary to determine whether image enhancement of the visible light image is required based on the pyroelectric particle concentration and the image illumination intensity, making the input more accurate and improving the fault tolerance of the bi-branch symmetric fusion network segmentation result.

[0037] Finally, the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and the temperature change data are integrated to trigger an early warning.

[0038] Example 1

[0039] This invention discloses an intelligent monitoring system for early fire prevention in power distribution rooms with multimodal fusion, such as... Figure 1 The specific steps shown are as follows:

[0040] First, the intensity of scattered light at the particle angle is calculated, and the trend of change of pyroelectric particle concentration is obtained based on the relationship between particle concentration and intensity of scattered light at the particle angle.

[0041] When the concentration of pyroelectric microparticles exceeds the system's preset threshold, a bi-branch symmetric fusion network is constructed. The bi-branch symmetric fusion network is used to fuse the visible light branch and the infrared branch, capture the spatial dependency of the power equipment through a global attention mechanism, and output the segmentation result of the bi-branch symmetric fusion network.

[0042] Then, based on the concentration of pyroelectric microparticles and the intensity of ambient light, it is determined whether image enhancement of the visible light image is necessary; finally, the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and temperature change data are fused to trigger an early warning.

[0043] Specific;

[0044] When there are no thermal hazards, the concentration of pyroelectric particles in a closed clean space will not change much and can be considered a constant value. However, when the substance reaches the critical point where a thermal hazard may occur, the concentration of pyroelectric particles will increase significantly. Based on this, the occurrence of a fire can be predicted in advance based on the concentration of pyroelectric particles.

[0045] Based on the Mie scattering theory, when a beam of light shines on a particle, it scatters in all directions. By measuring the incident intensity and scattering intensity of the beam before and after it passes through the particle, the number of pyroelectric particles can be calculated. Finally, the fire status can be accurately determined based on the particle concentration and its changing trend.

[0046] When the particle concentration is N, the particle angular scattered light intensity is: In the formula, θ is the scattering angle; I(θ) is the intensity of the scattered light at the scattering angle θ; λ is the wavelength of the light source; I0 is the incident light intensity; d is the particle diameter; m is the refractive index of light; and r is the distance from the scatterer.

[0047] Furthermore, during the detection process, λ, I0, d, m, and r can all be considered constant values.

[0048] Furthermore, we define a constant C: Therefore, the concentration of microparticles and the intensity of angularly scattered light have the following relationship: I(θ)=NC(1+cos 2 θ).

[0049] To simplify the calculation, a three-dimensional angle is usually used to receive the light intensity, i.e., θ is approximately 0°. After calculating the particle concentration, the update coefficient is calculated by weighted summation based on the particle concentration and particle diameter, as shown in the following formula: G = a*N + b*d; where D represents the update coefficient; N represents the particle concentration; d represents the particle diameter; and a and b are the weighting coefficients for particle concentration and particle diameter, respectively.

[0050] If the update coefficient is less than the system preset threshold, θ does not need to be updated. If the update coefficient is greater than the system preset threshold, θ is updated and adjusted according to the value of the update coefficient; and the particle concentration is recalculated.

[0051] The particle diameter is obtained by laser diffraction. When the particle diameter is close to or exceeds the incident light wavelength (e.g., particle size > 1 μm), the scattered light distribution will show a significant forward scattering enhancement characteristic (Mie scattering dominates). When the particle concentration is extremely high (e.g., > 10^6 particles / CC), the photon may undergo multiple scatterings, which will cause the single angle measurement of θ = 0° to deviate from the linear relationship. Therefore, the value of θ needs to be updated. In general, θ can be approximated as 0° for calculation.

[0052] When the concentration of pyroelectric microparticles exceeds the system's preset threshold, a bi-branch symmetric fusion network is constructed to locate potential ignition points.

[0053] like Figure 2As shown, a dual-branch symmetric fusion network is first constructed, using a visual Transformer as the basic block to extract features from both branches, better capturing the global dependencies of substation and distribution network equipment to better represent the unique features of power equipment. Then, based on the existing dual-branch feature extraction architecture, an efficient multi-scale feature fusion layer is designed. This layer delves into the complementary information from another modality within the feature modeling relationships of different modalities to compensate for the shortcomings of the current modality itself, such as the lack of strict texture and contour information in the infrared modality and the weak feature representation ability caused by the decrease in visible light visibility. Compared to the simple splicing and mutual attention calculations of previous multimodal fusion models, the fusion layer proposed in this paper improves cross-modal fusion learning from multiple levels. Furthermore, an efficient feature decoding module is designed, which can effectively recover the detailed information of infrared and visible light from the fused features for outputting the final segmentation result.

[0054] Furthermore, to achieve accurate and efficient fusion at different scales, this paper proposes a multi-scale feature fusion layer, where the visible light and infrared fusion layers at each scale are as follows: Figure 3 As shown. Considering that accurate segmentation requires clear contours and pixel accuracy, this invention aims to fully utilize the correlation between visible light mode and infrared mode features to complete feature fusion.

[0055] To determine whether image enhancement of the visible light image is necessary, the concentration of pyroelectric particles and the image illumination intensity are determined. Specifically, the enhancement coefficient is calculated using a weighted summation formula: F = h * N + i * gz; where F represents the enhancement coefficient, gz represents the image illumination intensity, and h and i represent the weighting coefficients for the pyroelectric particle concentration and the image illumination intensity, respectively.

[0056] If the weight coefficients are greater than the system threshold, then image enhancement is required for the visible light image that requires input dual-branch symmetric fusion network; otherwise, it is not required.

[0057] High concentrations of microparticles (e.g., >10^6 particles / CC) significantly scatter visible light, causing haze or halo effects in images. Key textures (such as insulator cracks and equipment corrosion) become blurred due to scattered light interference, requiring the visible light branch to extract features from low signal-to-noise ratio data.

[0058] In the fusion layer, each visible light feature and infrared feature is represented as V. n and I nWhere n represents the nth stage of feature extraction, n = {1, 2, 3, 4}. This paper designs an adaptive cross-modal correlation network to calculate the modal correlation matrix. The adaptive cross-modal correlation is optimized based on the different features of visible light and infrared images to achieve accurate and efficient fusion at multiple scales. This network can automatically adjust weights based on the specific characteristics of the input data, effectively utilizing the high-resolution texture information of visible light and the temperature information of infrared images, thereby enhancing the model's scene resolution capabilities.

[0059] Specifically, four convolutional layers are first used to dynamically learn representative features of the visible light and infrared modes respectively: V n ′=f(V n ), I n =g(I n Then the feature maps are concatenated to obtain the representative features after concatenation: Here, || denotes concatenating matrices along their columns. In this case, F n Each channel of each pixel in the image corresponds to the representative infrared and visible light features of each location in the original image. Next, multiple MLP layers are used to filter out the important complementary features F. n For power equipment, the complementary features highlight the clearer boundary features in the visible light image and also enhance the heating area of ​​the power equipment in the infrared image.

[0060] Specifically, for splicing features, the network compares the feature responses of the infrared and visible light modes channel by channel, calculates the cross-modal weights of each region in that feature channel, and weights the same region in that feature channel to form a new cross-modal feature. n ′=MLP(F n The specific implementation of the above cross-modal feature acquisition is first achieved through: transformation function φ: splicing feature F n Mapping to a new feature space R: R = φ(F); this transformation can be achieved by a parameterized linear transformation (MLP is chosen in this paper), and then the response of each feature vector is mapped to cross-modal weights using a normalization function: Next, the cross-modal weights are divided into two parts, which are used to weight the infrared modal features and the visible light modal features respectively. We obtain the modal correlation matrix of visible light and infrared images by selecting the most discriminative feature for each location pair using complementary features: M = max 1<d<D F' n (i,j,d); where D represents F n The number of channels, therefore, by obtaining F nThe maximum feature value at height i and width j on channel d represents the correlation between the infrared image and the visible light image. In this way, the obtained modal correlation matrix can dynamically acquire the statistical properties of different modal features. For each channel d at height i and width j, its parameters are adaptively selected to maximize the expressive power of the fused features.

[0061] This mechanism enables the network to adaptively adjust its processing strategy based on the specific content of each input sample, rather than operating according to static preset rules. For example, in regions with rich features, the network may increase the weight of visible light image features; while in regions with significant thermal information, it may strengthen the feature representation of infrared images. The modality correlation matrix not only optimizes the information utilization rate of the fusion process but also significantly improves the robustness and accuracy of the model in handling diverse scenarios.

[0062] Ultimately, we obtained the characteristics after fusion: in By learning the correlation between visible light and infrared images, the discriminative feature representation in the infrared image is fused into the visible light feature. Therefore, it has the outline and texture information of visible light, and can be supplemented by the temperature distribution characteristics of the infrared image when affected by lighting and other factors.

[0063] In obtaining multi-scale fusion features With n = {1, 2, 3, 4}, this paper designs a multi-scale attention decoder to perform scale recovery and detail restoration on the fused features at different scales. The specific structure of the decoder is as follows: Figure 4 As shown. Since higher-level features often possess more global information, this paper fuses features at each level. We employ a step-by-step decoding approach, recovering features from higher to lower levels and concatenating them with more detailed features from the previous level. In the multi-scale decoding framework, we use a cross-layer decoding layer D to recover fused features from adjacent scales and a channel attention layer A to enhance more discriminative feature representations. To demonstrate the effectiveness of our multi-scale decoding design, we avoid directly referencing specific network layers, instead abstracting the entire decoding process through multiple decoding layers and channel attention. This highlights the mathematical essence of our method, as well as the flexibility and scalability of the decoding layers. Based on this, for fused features at consecutive levels... and Cross-layer decoding features are represented as: in, This represents the transformation process from high-level features to low-level features, including but not limited to linear mappings, nonlinear mappings, or other composite functions. The most common approach is to use multilayer perceptrons (MLPs) and two-dimensional convolutions. A represents the attention mechanism, where channel attention can be represented as follows.

[0064] First, global average pooling is calculated to obtain a global description of the fused features, i.e., which features are more discriminative in segmentation. For each channel c, the global discriminative weights G are calculated. c : Where FC represents a fully connected layer, F i,h,w The feature map representing the attention to be calculated is typically... and Given n = {1, 2, 3}, the characteristic for calculating attention is: F a =F i,h,w ·G; G represents the global discriminative weight of all channels, ultimately, After post-processing, the final segmentation result is obtained. In the multi-scale decoder designed in this paper, More mapping methods can be used in the future, such as the Transformer block and the Mamba module. A can be designed with a more scale-sensitive attention mechanism for other tasks in the power scenario, such as small object detection.

[0065] In this embodiment, a combined loss function is chosen to optimize our multi-scale feature fusion network to address image segmentation tasks in complex power environments. The key to designing the loss function is its ability to effectively handle class imbalance and promote accurate detail prediction across various scales. Specifically, our loss function combines cross-entropy loss and Dice loss, each with its own advantages, and their combination has proven highly effective in various image segmentation tasks. The cross-entropy loss is calculated as follows: Where C represents the total number of categories, y o,c It is an indicator function, indicating whether category c is the correct classification for pixel set o, p o,c This represents the probability that o is predicted to be category c.

[0066] Dice loss, based on the Dice coefficient, is a common similarity metric used in medical image segmentation. This paper chooses this loss to eliminate imbalances in electrical samples, and its calculation method is as follows: Among them, y i It is a binary tag, p i This is the predicted probability, a small constant used to avoid a zero denominator. Our model employs a linear combination of cross-entropy loss and Dice loss, which helps to simultaneously optimize both class discriminative power and the geometric consistency of the predicted regions. The combined loss function is expressed as: L = α × L ce +(1-α)×L diceThe weight factor was set to 0.7, a value chosen based on experimental results, with the aim of combining the two losses for model training.

[0067] The segmentation results, particle concentration trends, and temperature change data of the bi-branch symmetric fusion network are integrated; if the segmented region abnormally overlaps with the particle concentration / temperature, it is judged as an "early fire" and an early warning is triggered.

[0068] Example 2

[0069] This invention discloses an intelligent monitoring system for early fire prevention in power distribution rooms with multimodal fusion, such as... Figure 5 As shown, it includes the following modules: optical detection and microparticle analysis module, bi-branch symmetric fusion network module, multi-scale decoding and output module, and decision and early warning module;

[0070] The optical detection and microparticle analysis module is used to measure the intensity of angular scattered light from microparticles in real time based on Mie scattering theory using a laser emitter and photoelectric sensor, and to calculate the concentration of pyroelectric microparticles based on the intensity of angular scattered light from microparticles; when the concentration of pyroelectric microparticles exceeds the preset threshold of the system, a two-branch symmetric fusion network is constructed to locate the possible ignition point;

[0071] The optical detection and microparticle analysis module includes a dynamic scattering angle adjustment module. This module calculates an update coefficient based on the microparticle concentration and diameter using a weighted summation. If the update coefficient is less than a preset system threshold, the scattering angle does not need to be updated. If the update coefficient is greater than the preset system threshold, the scattering angle is updated and adjusted according to the value of the update coefficient. The microparticle concentration is then recalculated.

[0072] The dual-branch symmetric fusion network module is used to fuse the visible light branch and the infrared branch, and captures the spatial dependencies of power equipment through a global attention mechanism.

[0073] The dual-branch symmetric fusion network module includes an image enhancement module, which is used to determine whether to enhance the visible light image based on the concentration of pyroelectric microparticles and the intensity of light in the image.

[0074] The multi-scale decoding and output module is used to recover feature scales step by step from high to low levels, and retains global context through cross-layer decoding; a channel attention mechanism is introduced to calculate global discriminative weights and enhance the feature representation of key regions;

[0075] The decision-making and early warning module is used to integrate the segmentation results of the bi-branch symmetric fusion network, the microparticle concentration trend, and the temperature change data; if the segmented region abnormally overlaps with the microparticle concentration / temperature, an early warning is triggered.

[0076] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0078] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0080] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent monitoring method for early fire prevention of a multi-modal fusion power distribution room, characterized in that, The method comprises the following steps: First, based on Mie scattering theory, the number of pyroelectric microparticles is calculated by measuring the incident and scattered intensities of the light beam before and after passing through the microparticles; when the microparticle concentration is N, the angular scattered light intensity is: In the formula, θ is the scattering angle; I(θ) is the intensity of the scattered light at the scattering angle θ; and λ is the wavelength of the light source. d is the incident light intensity; d is the particle diameter; m is the refractive index of light; r is the distance from the scattering body; thus, a constant is defined. : Therefore, the concentration of microparticles and the intensity of angularly scattered light have the following relationship: ; When the concentration of thermoparticles exceeds a preset threshold of the system, a double-branch symmetric fusion network is constructed; the double-branch symmetric fusion network is used for fusing a visible light branch and an infrared branch, capturing spatial dependence of the power equipment through a global attention mechanism, and outputting a segmentation result of the fused double-branch symmetric fusion network; Then, whether picture enhancement needs to be performed on the visible light image is judged according to the concentration of thermoparticles and the picture light intensity; finally, the segmentation result of the fused double-branch symmetric fusion network, the concentration trend of thermoparticles and the temperature change data are fused; and the prewarning is triggered.

2. The intelligent monitoring method for early fire prevention of a power distribution room with multi-modal fusion according to claim 1, characterized in that: θ is dynamically updated; the update coefficient is calculated by using weighted summation according to the concentration of thermoparticles and the diameter of thermoparticles; if the update coefficient is less than a preset threshold of the system, θ does not need to be updated; if the update coefficient is greater than the preset threshold of the system, θ is updated; the value of θ is adjusted according to the value of the update coefficient; and the concentration of thermoparticles is recalculated. 3.The intelligent monitoring method for early fire prevention of a power distribution room with multi-modal fusion according to claim 1, characterized in that: When the concentration of thermoparticles exceeds a preset threshold of the system, a double-branch symmetric fusion network is constructed to locate the possible fire point.

4. The intelligent monitoring method for early fire prevention of a power distribution room with multi-modal fusion according to claim 3, characterized in that: The double-branch symmetric fusion network uses a visual Transformer as a basic block to extract the features of the visible light branch and the infrared branch, and then designs an efficient multi-scale feature fusion layer based on an existing feature extraction double-branch architecture, deeply mines the complementary information from another modality in the feature modeling relationship of different modalities, and finally obtains the fused features.

5. The intelligent monitoring method for early fire prevention of a power distribution room with multi-modal fusion according to claim 3, characterized in that: The concentration of thermoparticles and the picture light intensity are determined, and a weighting coefficient is calculated by using a weighted summation formula; if the weighting coefficient is greater than a system threshold, the visible light branch of the input double-branch symmetric fusion network needs to be subjected to picture enhancement; otherwise, the visible light branch does not need to be subjected to picture enhancement.

6. The intelligent monitoring method for early fire prevention of a power distribution room with multi-modal fusion according to claim 4, characterized in that: After the multi-scale fused features are obtained, the fused features of different scales are subjected to scale restoration and detail restoration, and the segmentation result of the double-branch symmetric fusion network is determined.

7. An intelligent monitoring system for early fire prevention in a multi-modal fusion power distribution room, characterized in that, The monitoring system is used to implement the method according to any one of claims 1-6, and comprises the following modules: an optical detection and thermoparticle analysis module, a double-branch symmetric fusion network module, a multi-scale decoding and output module, and a decision and prewarning module; The optical detection and thermoparticle analysis module is used for measuring the angular scattering light intensity of thermoparticles in real time through a laser emitter and a photoelectric sensor based on the Mie scattering theory, calculating the concentration of thermoparticles from the angular scattering light intensity of thermoparticles, and locating the possible fire point through a double-branch symmetric fusion network when the concentration of thermoparticles exceeds a preset threshold of the system. The double-branch symmetric fusion network module is used for fusing a visible light branch and an infrared branch, and capturing the spatial dependence of the power equipment through a global attention mechanism. The multi-scale decoding and output module is used for restoring the feature scale step by step from a high layer to a low layer, retaining the global context through cross-layer decoding, introducing a channel attention mechanism, calculating the global discriminative weight, and enhancing the feature expression of the key region. The decision and prewarning module is used for fusing the segmentation result of the double-branch symmetric fusion network, the concentration trend of thermoparticles and the temperature change data, and triggering the prewarning when the segmentation region and the concentration / temperature of thermoparticles are abnormally overlapped.

8. The intelligent monitoring system for early fire prevention of a power distribution room with multi-modal fusion according to claim 7, characterized in that: The optical detection and particle analysis module comprises a dynamic scattering angle adjusting module, which is used to calculate an updating coefficient by weighted summation according to the particle concentration and the particle diameter; if the updating coefficient is less than a system preset threshold, the scattering angle does not need to be updated; if the updating coefficient is greater than the system preset threshold, the scattering angle is updated according to the value of the updating coefficient; and the particle concentration is recalculated.

9. The intelligent monitoring system for early fire prevention of a power distribution room with multi-modal fusion according to claim 7, characterized in that: The dual-branch symmetric fusion network module comprises a picture quality enhancement module, which is used to determine whether picture enhancement is needed for the visible light image according to the thermoluminescent particle concentration and the picture illumination intensity.

Images