Power transmission line forest fire identification system and method fusing multi-scale sub-band features and MDS dimension reduction

Abstract

The invention relates to the technical field of power transmission line forest fire monitoring, in particular to a power transmission line forest fire recognition system fusing multi-scale sub-band features and MDS dimensionality reduction, which comprises an image acquisition module used for acquiring image data of an area along a power transmission line, and an image transmission module in communication connection with the image acquisition module and used for transmitting the image data in real time, and the image processing and analyzing module is in communication connection with the image transmission module and is used for converting the image data into a grayscale image. Two-dimensional discrete wavelet transform (2DDWT) is introduced into the field of transmission line forest fire recognition, an image is decomposed into an approximate sub-band (LL) and three detail sub-bands (HL, LH and HH) through a Daubechies-4 filter bank, accurate feature separation of flame and smoke under different frequency scales is achieved, the approximate sub-band retains core structure information, and the accuracy of recognition of forest fire is improved. The detail sub-bands capture multi-directional edge and texture features, and the defect that in the prior art, only original image input is relied on, and frequency domain information is not fully mined is overcome.

Description

Technical Field

[0001] This invention relates to the field of power transmission line wildfire monitoring technology, specifically to a power transmission line wildfire identification system and method that integrates multi-scale sub-band features and MDS dimensionality reduction. Background Technology

[0002] With the rapid development of the power system and the continuous advancement of smart grid construction, the safe and stable operation of transmission lines has become a core link in ensuring the continuity and reliability of power supply. Among the various factors affecting the safety of transmission lines, wildfires, as a natural disaster with strong suddenness, wide destructive range, and rapid spread, pose a particularly prominent threat to transmission channels. In areas with dense forest cover, complex terrain, and dry climate, frequent wildfires along transmission lines can easily lead to serious accidents such as line tripping and equipment damage, directly affecting the normal operation of the power system. Therefore, it is urgent to build an efficient, accurate, and real-time wildfire monitoring and early warning technology system. At present, a variety of technical solutions have been developed in the field of wildfire monitoring of transmission lines, which can be mainly divided into two categories: traditional monitoring methods and intelligent detection methods based on computer vision. However, both of them have significant limitations.

[0003] However, traditional methods of wildfire monitoring mainly rely on ground-based lookout towers, satellite remote sensing, infrared sensors, and video surveillance, which have significant shortcomings: Ground-based lookout tower monitoring has limited line of sight and is affected by terrain obstruction, resulting in high deployment and maintenance costs and an inability to achieve large-scale comprehensive coverage; Satellite remote sensing monitoring, although with a wide coverage area, has a long image update cycle and low resolution, making it difficult to meet real-time response requirements and easily missing early response opportunities; Infrared sensors and video surveillance have reduced recognition capabilities at night and in smoke-covered scenarios, and video surveillance is easily affected by changes in lighting and complex backgrounds, resulting in a high false alarm rate.

[0004] Early manual feature extraction methods relied on manually designed features such as color space analysis and texture feature extraction, combined with traditional classifiers for recognition. These methods had weak feature representation and generalization capabilities, high false detection rates in complex backgrounds, and low efficiency. Traditional deep learning methods, such as convolutional neural networks (CNNs), could automatically extract high-level semantic features, but most of them directly used the original image as input, failing to fully explore the potential information at different frequency scales. Their detection performance was insufficient in scenes with mixed flames and smoke or complex backgrounds. Feature redundancy issues: the feature vectors extracted by deep CNNs had high dimensionality and a lot of redundant information, resulting in wasted computational resources and reduced classification efficiency. Limitations of wavelet transform applications: although wavelet transforms can capture multi-scale local features, there has been no research on its fusion with deep CNN features and combined with dimensionality reduction techniques in the field of power transmission channel wildfire monitoring, failing to leverage the synergistic advantages of multi-scale features and deep features.

[0005] In summary, existing technologies for detecting wildfires on power transmission lines in complex environments generally suffer from problems such as low recognition accuracy, weak anti-interference ability, and slow response speed. Therefore, there is a lack of new wildfire recognition technologies that can integrate multi-scale feature extraction, deep feature fusion, and efficient dimensionality reduction to improve detection performance in complex scenarios, achieve early identification, rapid response, and accurate early warning of wildfire hazards, and ensure the safe operation of power transmission lines. Summary of the Invention

[0006] This invention aims to provide a transmission line wildfire identification system and method that integrates multi-scale sub-band features and MDS dimensionality reduction. It is mainly used to solve the problems of low identification accuracy, weak anti-interference ability, and slow response speed in the detection of wildfires on transmission lines in complex environments. Therefore, there is a lack of new wildfire identification technology that can integrate multi-scale feature extraction, deep feature fusion and efficient dimensionality reduction to improve the detection performance in complex scenarios, realize early identification, rapid response and accurate warning of wildfire hazards, and ensure the safe operation of transmission lines.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A power transmission line wildfire identification system integrating multi-scale sub-band features and MDS dimensionality reduction includes:

[0009] The image acquisition module is used to acquire image data of the area along the power transmission line;

[0010] The image transmission module is connected in communication with the image acquisition module and is used to transmit image data in real time;

[0011] The image processing and analysis module, which communicates with the image transmission module, is used to convert image data into grayscale images. It performs a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale images to obtain four sub-bands. It extracts the depth feature vectors of each sub-band through a convolutional neural network (CNN), concatenates them to form a comprehensive feature vector, and uses multidimensional scaling (MDS) technology to reduce the dimensionality of the comprehensive feature vector. It then uses a classifier to determine whether there are wildfire areas.

[0012] The early warning and feedback module communicates with the image processing and analysis module and is used to generate and push alarm information when a wildfire area is determined to exist;

[0013] The data storage and management module is used to store image data, classification results, and alarm records.

[0014] A method for identifying wildfires on transmission lines that integrates multi-scale sub-band features and MDS dimensionality reduction includes the following steps:

[0015] S1. Collect image data of the area along the power transmission line;

[0016] S2. Preprocess the image data and convert it into a grayscale image to reduce interference from changes in illumination;

[0017] S3. Perform a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale image to obtain four sub-bands. Input each sub-band into a convolutional neural network (CNN) to extract the corresponding depth feature vectors.

[0018] S4. Concatenate the four deep feature vectors to obtain the comprehensive feature vector;

[0019] S5. Use multidimensional scaling (MDS) technology to reduce the dimensionality of the comprehensive feature vector, obtain a low-dimensional feature vector, and input it into the classifier to determine whether there is a wildfire area.

[0020] S6. If a wildfire area is identified, an alarm message is generated and pushed to relevant terminal devices.

[0021] Working principle and beneficial effects of the present invention:

[0022] 1. Working principle: Through the collaborative design of five major modules, multi-scale feature extraction, deep feature fusion and efficient dimensionality reduction technology are organically integrated to solve the technical pain points of low identification accuracy and weak anti-interference ability of existing wildfire monitoring systems in complex backgrounds. Through real-time data transmission and processing of each module, early identification and rapid warning of wildfire hazards on power transmission lines can be achieved, ensuring the stability of power supply.

[0023] The method addresses the problems of insufficient exploitation of image frequency domain information and low efficiency caused by feature redundancy in existing wildfire identification methods through a coherent process of "acquisition-preprocessing-feature extraction-fusion-dimensionality reduction-classification-early warning". By fusing multi-scale decomposition of wavelet transform with deep feature extraction of CNN, the method improves the feature discrimination ability in complex backgrounds. Combined with MDS dimensionality reduction to optimize computational efficiency, the method ultimately achieves accurate and rapid identification and early warning of wildfires.

[0024] 2. Beneficial effects:

[0025] (1) By introducing two-dimensional discrete wavelet transform (2DDWT) into the field of wildfire identification of transmission lines, and decomposing the image into approximate sub-band (LL) and three detail sub-bands (HL, LH, HH) through the Daubechies-4 filter bank, the features of flames and smoke at different frequency scales are accurately separated. The approximate sub-band retains the core structural information, while the detail sub-band captures multi-directional edge and texture features. This solves the problem that the existing technology only relies on the original image input and does not fully explore the frequency domain information. Combined with the InceptionV3 or MobileNet model to extract and splice the deep features of each sub-band, the comprehensive feature vector covers both spatial structure and frequency domain detail information, making the wildfire feature representation more comprehensive. The recognition accuracy in complex backgrounds (cloudy, foggy, shadowy) and mixed flame and smoke scenes significantly surpasses the traditional manual feature extraction method and single CNN model, effectively reducing the false detection rate and false negative rate.

[0026] (2) By adopting the standard grayscale conversion formula (y=0.2989R+0.5870G+0.1140B) and standardization processing in the image preprocessing stage, the interference of illumination changes on feature extraction is reduced; the multi-subband feature fusion strategy combines the powerful spatial modeling capability of CNN with the time-frequency analysis advantage of wavelet transform, enabling the model to adapt to the diverse interference in application scenarios such as complex terrain and dense forest coverage along the transmission line. It can still effectively identify wildfire features under extreme conditions such as smoke obscuration and nighttime monitoring. Compared with existing infrared sensors and video surveillance systems, the anti-interference capability is enhanced, and the generalization performance covers the monitoring needs of transmission channels with different climates and terrains.

[0027] (3) By employing multidimensional scaling (MDS) technology for feature dimensionality reduction, and through distance matrix construction, inner product matrix calculation, and eigenvalue decomposition, invalid and redundant information is eliminated while strictly preserving sample similarity relationships. Compared with traditional PCA and LDA methods, this method has a better compression effect on nonlinear features, reducing the comprehensive feature vector dimension from 4× ( ≥4096) is significantly reduced to k (k≪4) The dimensionality-reduced feature vectors are input into the SVM classifier to reduce computational resource consumption. Combined with the dual deployment mode of embedded edge devices and cloud servers, it not only meets the remote data processing needs of large-scale inspections, but also reduces transmission latency through localized response. This solves the pain points of low computational efficiency and difficulty in adapting to real-time monitoring in existing deep learning methods, and ensures rapid response for wildfire identification and early warning.

[0028] (4) The image acquisition module uses a drone equipped with a high-definition camera or thermal imaging equipment to overcome the limitations of limited line of sight and high deployment cost of ground observation towers, and realize large-scale and flexible image acquisition along the power transmission line. The system supports two modes: remote monitoring and local rapid response. It can be deployed on cloud servers or embedded edge devices according to actual needs, and adapt to the monitoring needs of different application scenarios such as dense forest areas and complex terrain. At the same time, the data storage and management module completely retains historical images, detection results and alarm records, providing data support for subsequent analysis of wildfire occurrence patterns and optimization of monitoring strategies, forming a closed-loop management of "acquisition-analysis-early warning-tracing".

[0029] (5) Through the coherent technical solution of “multi-scale feature extraction - deep fusion - efficient dimensionality reduction - accurate classification”, the early identification, rapid response and accurate early warning of wildfire hazards of transmission lines can be realized. This effectively makes up for the shortcomings of long update cycle of satellite remote sensing images and the susceptibility of infrared sensors to obstruction. Its accurate identification effect and real-time early warning mechanism can help maintenance personnel to deal with wildfires in a timely manner in the early stage of wildfire spread and reduce the occurrence rate of power line tripping and power outage accidents caused by wildfires.

[0030] Preferably, the image acquisition module is a high-definition camera or thermal imaging device mounted on a drone, and the image data is an RGB image or an infrared image. The system can be deployed on a cloud server or an embedded edge device, supporting remote monitoring and localized rapid response modes. The acquisition device mounted on the drone can adapt to application scenarios with dense forests and complex terrain along power transmission lines, achieving large-scale and flexible image acquisition. The dual deployment mode of cloud and embedded edge devices not only meets the remote data processing needs of large-scale inspections, but also reduces data transmission latency through localized response, adapting to real-time monitoring requirements in different scenarios and improving the practicality and adaptability of the system.

[0031] Preferably, the two-dimensional discrete wavelet transform employs the Daubechies-4 filter bank, achieving multi-scale decomposition through a combination of horizontal filtering, vertical filtering, and downsampling. The four sub-bands are the approximate sub-band (LL), horizontal detail sub-band (HL), vertical detail sub-band (LH), and diagonal detail sub-band (HH). The Daubechies-4 filter bank has excellent time-frequency localization characteristics, making it suitable for multi-scale feature separation of wildfire images. Through combined horizontal and vertical filtering and downsampling, it can accurately extract low-frequency structural information and multi-directional high-frequency detail information of the image. The division of labor and cooperation among the four sub-bands ensures comprehensive capture of flame and smoke features at different frequency levels, laying the foundation for subsequent deep feature extraction.

[0032] Preferably, the convolutional neural network is an InceptionV3 or MobileNet model, and the deep feature vector is extracted from the flattened layer of the model; the multidimensional scaling (MDS) achieves dimensionality reduction by constructing a distance matrix, calculating the inner product matrix, and eigenvalue decomposition; the classifier is a support vector machine (SVM); the InceptionV3 model has powerful multi-scale feature extraction capabilities, while the MobileNet model balances feature extraction performance with lightweight deployment requirements; both can extract deep semantic features from sub-band images; compared with traditional PCA and LDA methods, MDS technology can more efficiently preserve the similarity relationship between feature samples and remove redundant information; the SVM classifier performs stably in small sample and high-dimensional data classification; the combination of these three ensures a balance between feature extraction accuracy, computational efficiency, and classification accuracy.

[0033] Preferably, the grayscale image conversion formula is y=0.2989R+0.5870G+0.1140B, where R, G, and B are the pixel values ​​of the red, green, and blue channels, and y is the grayscale value. Preprocessing also includes standardizing the grayscale image to improve the adaptability of subsequent algorithms. This grayscale conversion formula is a standard conversion method commonly used in the industry, capable of accurately fusing RGB three-channel information and effectively reducing the interference of light intensity changes on image features. Standardization processing, by unifying the image pixel value scale, avoids the impact of data differences under different lighting and shooting conditions on CNN feature extraction, ensuring the stability of subsequent multi-scale decomposition and feature extraction, and providing a guarantee for overall recognition accuracy.

[0034] Preferably, the approximate sub-band (LL) retains the main structural information of the image, the horizontal detail sub-band (HL) reflects the horizontal edge information, the vertical detail sub-band (LH) reflects the vertical edge information, and the diagonal detail sub-band (HH) reflects the diagonal texture information. The four sub-bands correspond to different feature dimensions of the image. The approximate sub-band captures the overall contour structure of the wildfire area, while the three detail sub-bands extract the edge morphology and texture features of the flames and smoke, respectively. This separation and extraction of multi-dimensional features avoids the omission of feature information caused by a single image input, and enables wildfire-related features to be fully represented at different scales and in different directions, thereby improving the distinguishability of features from background interference.

[0035] Preferably, the dimension of the single subband depth feature vector output by the convolutional neural network is... ≥4096, the feature splicing is performed by concatenating sub-bands in sequence, forming a dimension of The comprehensive feature vector; A dimension setting of ≥4096 ensures the richness of the depth features of a single sub-band, fully covering the semantic information related to wildfires. The concatenation method according to the sub-band order can completely preserve the feature details of each sub-band, avoiding information loss during feature fusion. The resulting comprehensive feature vector contains both the low-frequency structure of the image and high-frequency detail information from multiple directions, significantly improving the discriminative ability of the features and providing support for subsequent accurate classification. .

[0036] Preferably, the specific steps of the multidimensional scaling (MDS) are: constructing a distance matrix of the comprehensive feature vectors. , For the first The and the first Distance between samples; construct matrix ; Calculate the inner product matrix ,in for The Row mean for The Column mean for The global mean; perform eigenvalue decomposition on matrix B, and select the top... The projection matrix is ​​composed of the largest eigenvalues ​​and their corresponding eigenvectors. and diagonal matrix ; Calculate the dimensionality-reduced eigenvectors ,and Much smaller than the original dimension of the comprehensive feature vector; the step-by-step implementation process of MDS strictly follows the core principle of "preserving sample similarity", accurately representing the correlation between samples through distance matrix construction and inner product matrix transformation; in the eigenvalue decomposition process, the first... The largest eigenvalue can be used to filter out the core features with the strongest discriminative power and eliminate redundant information. The design, which is much smaller than the original dimension, significantly reduces the computational burden of subsequent classifiers, and while ensuring recognition accuracy, it significantly improves the real-time response speed of the method.

[0037] Preferably, the classifier is a Support Vector Machine (SVM), and the classification judgment function is: C(x)=1 if f(x)>T; C(x)=0 otherwise; where C(x)=1 indicates the presence of a wildfire area, C(x)=0 indicates the absence of a wildfire area, and T is a preset classification threshold. The SVM classifier has good generalization ability and robustness, and can effectively process high-dimensional feature data after dimensionality reduction, accurately distinguishing wildfire areas from background interference. The preset classification threshold T can be dynamically adjusted according to actual application scenarios (such as different climate and terrain conditions), flexibly balancing the false detection rate and the false negative rate, ensuring the adaptability of the method in complex environments, and realizing accurate determination and reliable early warning of wildfire areas. Attached Figure Description

[0038] Figure 1 The flowchart shows the transmission line wildfire identification method of the present invention, which integrates multi-scale sub-band features and MDS dimensionality reduction.

[0039] Figure 2 This is a two-dimensional discrete wavelet transform step diagram of the power transmission line wildfire identification system that integrates multi-scale sub-band features and MDS dimensionality reduction according to the present invention.

[0040] Figure 3 This diagram illustrates the feature extraction steps of the power transmission line wildfire identification system that integrates multi-scale sub-band features and MDS dimensionality reduction, as presented in this invention. Detailed Implementation

[0041] 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.

[0042] Example 1, such as Figures 2-3 As shown, the power transmission line wildfire identification system, which integrates multi-scale sub-band features and MDS dimensionality reduction, includes:

[0043] The image acquisition module is used to collect image data along the power transmission line. The image acquisition module is a high-definition camera or thermal imaging device mounted on a drone, and the image data is RGB or infrared images. The system can be deployed on a cloud server or embedded edge device, supporting remote monitoring and localized rapid response modes.

[0044] Hardware configuration: Using a drone as a mobile platform, two sets of data collection devices are configured to complement each other's functions.

[0045] Visible light acquisition unit: High-definition camera, used to capture RGB images of the area surrounding the power transmission line, adapted for image acquisition under normal daylight conditions;

[0046] Thermal imaging acquisition unit: Thermal imaging equipment used to capture wildfire heat source signals in low light or complex environments such as nighttime or when smoke obscures the view, making up for the limitations of visible light acquisition.

[0047] Data collection parameter settings: Data collection is carried out along the transmission line according to the preset inspection route. The route covers key monitoring areas on both sides of the transmission line. The collection frequency is set according to the inspection requirements to ensure no monitoring blind spots. During the collection process, basic information such as the transmission line section corresponding to the image and the collection time are recorded simultaneously to provide a basis for subsequent identification and traceability.

[0048] The image transmission module communicates with the image acquisition module and is used for real-time transmission of image data. The transmission method utilizes a wireless communication network to achieve real-time data transmission, supporting two transmission paths:

[0049] Remote transmission: The collected image data is transmitted back to the ground control center in real time to meet the centralized data processing needs of large-scale inspections;

[0050] Localized transmission: In edge computing device deployment scenarios, data is directly transmitted to the nearest edge computing node, reducing data transmission latency;

[0051] Data protection: It has a data caching function. When the wireless communication network is interrupted, it automatically stores the collected image data and automatically retransmits it after the network is restored to avoid data loss. The image data is transmitted in its original format to ensure the accuracy of subsequent processing.

[0052] The image processing and analysis module, which communicates with the image transmission module, is used to convert image data into grayscale images. It performs a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale images to obtain four sub-bands. A convolutional neural network (CNN) is used to extract the depth feature vectors of each sub-band, which are then concatenated to form a comprehensive feature vector. Multidimensional scaling (MDS) technology is used to reduce the dimensionality of the comprehensive feature vector. A classifier is used to determine whether wildfire areas exist. The two-dimensional discrete wavelet transform uses a Daubechies-4 filter bank, and multi-scale decomposition is achieved through a combination of horizontal filtering, vertical filtering, and downsampling. The four sub-bands are the approximate sub-band (LL), horizontal detail sub-band (HL), vertical detail sub-band (LH), and diagonal detail sub-band (HH).

[0053] The convolutional neural network is either InceptionV3 or MobileNet, and the deep feature vector is extracted from the flattened layer of the model. Multidimensional scaling (MDS) reduces dimensionality by constructing a distance matrix, calculating the inner product matrix, and decomposing eigenvalues. The classifier is a support vector machine (SVM).

[0054] Deployment Modes: Supports two deployment models to adapt to different application scenarios:

[0055] Cloud deployment: Deployed on a cloud server, it is suitable for centralized data processing of large-scale power transmission line inspections and can process image data transmitted by multiple drones simultaneously.

[0056] Edge deployment: Deployed on embedded edge devices, suitable for localized monitoring scenarios with high response speed requirements, enabling local processing of image data and rapid decision-making.

[0057] Software environment: Equipped with wavelet transform processing library, deep learning framework and machine learning tool library, core supports the operation of core algorithms such as two-dimensional discrete wavelet transform (2DDWT), convolutional neural network (CNN) feature extraction, multidimensional scaling (MDS) dimensionality reduction and support vector machine (SVM) classification;

[0058] Core algorithm configuration: Wavelet transform: Using the Daubechies-4 filter bank, a first-level two-dimensional discrete wavelet transform is performed to achieve multi-scale decomposition of the image;

[0059] CNN model: InceptionV3 or MobileNet model is selected to extract the depth feature vectors of each sub-band image;

[0060] Dimensionality reduction and classification: The dimensionality of the fused feature vectors is reduced using MDS technology, and finally an SVM classifier is used to complete the wildfire identification and judgment.

[0061] The early warning and feedback module communicates with the image processing and analysis module to generate and push alarm information when a wildfire area is determined to exist. Early warning information generation: When the classification result determines that a wildfire area exists, standardized alarm information is automatically generated. The information includes core content such as the transmission line section where the suspected wildfire is located, image evidence, and identification confidence level, to ensure that maintenance personnel can quickly grasp the key situation.

[0062] Information push and linkage: Alarm information is pushed to relevant operation and maintenance personnel through terminal devices. It also supports linkage with the fire resource dispatch system to push wildfire-related information in real time when needed, providing data support for emergency response.

[0063] The data storage and management module is used to store image data, classification results, and alarm records. Storage content includes: categorized storage of various types of data, including:

[0064] Raw data: Visible light images and thermal images collected by the drone;

[0065] Process data: preprocessed grayscale image, sub-band image after wavelet decomposition, and extracted feature vectors;

[0066] Results data: wildfire identification results, alarm records, operation and maintenance records, etc.;

[0067] Management functions: Supports data retrieval and querying by transmission line section, collection time, identification result and other dimensions, which facilitates historical data tracing, analysis and subsequent model optimization training; adopts standardized data storage format to ensure long-term effective data retention.

[0068] Example 2, as follows Figure 1 As shown, the method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction includes the following steps:

[0069] S1. Collect image data along the transmission line and obtain image data of the area surrounding the transmission line, including RGB visible light images and thermal imaging images. The images cover key monitoring areas on both sides of the transmission line to ensure that potential wildfire hazards are included.

[0070] S2. Preprocess the image data by converting it to a grayscale image to reduce interference from changes in lighting. The grayscale conversion formula is y = 0.2989R + 0.5870G + 0.1140B, where R, G, and B are the pixel values ​​of the red, green, and blue channels, and y is the grayscale value. Preprocessing also includes standardizing the grayscale image to improve the adaptability of subsequent algorithms. Standardizing the converted grayscale image unifies the pixel value scale, reduces interference from different lighting conditions and shooting angles, and improves the adaptability and stability of subsequent algorithms.

[0071] S3. Perform a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale image to obtain four sub-bands. Input each sub-band into a convolutional neural network (CNN) to extract the corresponding depth feature vectors. The approximate sub-band (LL) retains the main structural information of the image, the horizontal detail sub-band (HL) reflects the horizontal edge information, the vertical detail sub-band (LH) reflects the vertical edge information, and the diagonal detail sub-band (HH) reflects the diagonal texture information.

[0072] The Daubechies-4 filter bank was selected to perform horizontal and vertical filtering operations on the grayscale image.

[0073] By combining downsampling processing, four frequency sub-band images are finally obtained, each sub-band corresponding to image features in different dimensions:

[0074] Approximate subband (LL): Contains low-frequency components of the image, preserving the main structural information of the wildfire area and its surrounding environment;

[0075] Horizontal detail subband (HL): Focuses on the horizontal edge features of the image, capturing the morphological changes of flames and smoke in the horizontal direction;

[0076] Vertical detail subband (LH): Reflects the vertical edge information of the image, representing the contour features of flames and smoke in the vertical direction;

[0077] Diagonal detail subband (HH): Extracts texture information along the diagonal direction of the image, capturing detailed features such as flame texture and smoke drift trajectory;

[0078] The dimension of the single subband depth feature vector output by the convolutional neural network ≥4096, feature splicing is done by concatenating subbands in sequence, forming a dimension of The comprehensive feature vector;

[0079] The sub-band images are processed sequentially through the core layers of the CNN, such as convolutional layers and pooling layers, to gradually extract the deep semantic features of the images;

[0080] The flatten layer preceding the fully connected layer outputs the depth feature vector corresponding to each sub-band. The feature vector dimension is 4096 or higher to ensure sufficient coverage of wildfire-related feature information in the sub-band images. The feature vectors corresponding to the four sub-bands are denoted as follows: ,in The feature dimension of a single subband;

[0081] S4. Concatenate the four deep feature vectors to obtain a comprehensive feature vector. The deep feature vectors of the four sub-bands are then fused using a feature concatenation method to form the comprehensive feature vector. The fusion formula is as follows: ,in The vertical concatenation of the representation vectors results in a fused comprehensive feature vector that simultaneously encompasses the structural information of the approximate subband and the edge and texture information of the three detailed subbands, achieving a comprehensive representation of wildfire features and enhancing the discriminative ability of the features.

[0082] S5. Multidimensional scaling (MDS) is used to reduce the dimensionality of the comprehensive feature vector, resulting in a low-dimensional feature vector. This low-dimensional feature vector is then input into the classifier to determine if wildfire areas exist. The specific steps of MDS are: constructing the distance matrix of the comprehensive feature vector. , For the first The and the first Distance between samples; construct matrix ; Calculate the inner product matrix ,in for The Row mean for The Column mean for The global mean; perform eigenvalue decomposition on matrix B, and select the top... The projection matrix is ​​composed of the largest eigenvalues ​​and their corresponding eigenvectors. and diagonal matrix ; Calculate the dimensionality-reduced eigenvectors ,and Much smaller than the original dimension of the composite feature vector This process retains key discrimination information while removing redundancy, thus reducing the complexity of subsequent classification calculations.

[0083] The classifier is a support vector machine (SVM), and the classification judgment function is: C(x)=1 if f(X)>T; C(x)=0 otherwise; where C(x)=1 indicates the existence of wildfire areas, C(x)=0 indicates the absence of wildfire areas, and T is the preset classification threshold;

[0084] The classifier learns the differences between wildfire and non-wildfire characteristics during the training process to establish a classification model; the classification decision function is: ,in =1 indicates that there is a wildfire area. =0 indicates an area free of wildfires;

[0085] S6. If a wildfire area is determined to exist, an alarm message is generated and pushed to the relevant terminal devices. If the classification result is C(x)=1, the system automatically triggers the early warning mechanism, generates an alarm message containing information such as suspected wildfire areas, power line sections, and image evidence, and pushes it to maintenance personnel through the terminal devices. If the classification result is C(x)=0, only the image data and processing results are stored in the data storage and management module for subsequent historical data tracing and model iteration optimization.

[0086] As can be seen from the above, the specific embodiments of the present invention are as follows:

[0087] I. System Deployment and Hardware Configuration:

[0088] The image acquisition module uses a DJI Matrice 350RTK drone as a mobile acquisition platform, equipped with two sets of acquisition devices to complement each other. The visible light acquisition device is a Zenmuse P1 full-frame camera with a resolution of 5472×3648 pixels, a frame rate of 30fps, a lens focal length of 24mm, and supports automatic exposure adjustment, adapting to different lighting scenarios such as strong daylight and weak evening light.

[0089] Thermal imaging acquisition equipment: FLIRVueProR infrared thermal imager, with a temperature measurement range of -20℃ to 150℃, thermal sensitivity ≤50mK, and resolution of 640×512 pixels. It can capture wildfire heat source signals at night and in smoke-covered environments. Drone flight planning is achieved using WaypointPlanner software, setting inspection routes along the power transmission line with a 50-meter interval between routes and a flight altitude of 100 meters, ensuring coverage of the monitoring area within 500 meters on both sides of the power transmission line. One visible light image and one infrared image are captured every 30 seconds.

[0090] The image transmission module supports dual deployment on cloud servers and edge computing devices to meet the needs of different scenarios:

[0091] Cloud deployment (large-scale inspection scenario):

[0092] Hardware configuration: Industrial-grade server (Intel Xeon Gold 6330 processor × 2, 64GB DDR4 memory, 4TB NVMe SSD, NVIDIA A100 GPU × 2).

[0093] Software environment: Ubuntu 20.04 LTS operating system, TensorFlow 2.8 deep learning framework, PyWavelets 1.4.1 wavelet transform library, scikit-learn 1.0.2 machine learning library;

[0094] Edge deployment (localized, rapid response scenario):

[0095] Hardware configuration: NVIDIA Jetson AGXXavier edge computing box (8-core Carmel ARM CPU, 512-core Volta GPU, 32GB memory).

[0096] Software environment: Ubuntu 18.04 Jetson version, TensorRT 8.4 inference acceleration engine, other library versions are consistent with the cloud version, and inference speed is improved through model quantization (INT8);

[0097] Early warning and feedback module:

[0098] Alarm information generation includes: wildfire location (drone GPS positioning, accuracy ±1m), confidence level (classifier output probability), image screenshot (visible light + infrared comparison image), and warning level (divided into Level 1 ≥90%, Level 2 70%-89%, and Level 3 50%-69% based on confidence level).

[0099] Push methods: Push to the maintenance personnel's dedicated APP (supports Android / iOS), WeChat official account template messages, SMS (China Mobile / China Unicom / China Telecom corporate SMS channel), and email (corporate email) via RESTful API interface;

[0100] Firefighting coordination: It connects with the local fire command platform and pushes information such as the location of wildfires, the name of power transmission lines, and the surrounding terrain through the GB / T28181 protocol to support the dispatch of firefighting resources;

[0101] Data storage and management module:

[0102] Structured data storage: Utilizes a MySQL 8.0 database to store alarm records (location, time, level, processing result) and device status (drone, camera, transmission module operating parameters). A single table supports tens of millions of data storage entries, with time-based indexing.

[0103] Unstructured data storage: Uses MongoDB 5.0 database to store original images, sub-band images, and feature vectors, and supports retrieval by shooting time, geographical location, and image type (visible light / infrared);

[0104] Data backup: A dual backup strategy of "local + off-site" is adopted. Local storage retains 1 year of data, and off-site cloud storage (Alibaba Cloud OSS) retains 3 years of data, supporting historical data traceability and model iteration training.

[0105] II. Specific Implementation Steps of the Method:

[0106] Image acquisition: Using the aforementioned UAV platform, the UAV inspects the power transmission line along a preset route, simultaneously acquiring visible light RGB images and infrared images. During the acquisition process, the UAV automatically records metadata such as GPS coordinates, shooting time, light intensity (acquired by the built-in sensor of the visible light camera), and ambient temperature (acquired by the thermal imager). The images are uniformly saved in JPEG format, and the metadata is stored in JSON format in association with the image file.

[0107] Image preprocessing: Image format conversion: For visible light RGB images, the grayscale conversion formula is y=0.2989R+0.5870G+0.1140B, where R, G, and B are the original channel pixel values ​​(0-255), and y is the grayscale value (0-255); for infrared images, they are directly input as grayscale images without conversion.

[0108] Standardization: Normalize the pixel values ​​of the grayscale image to the [0,1] interval. The calculation formula is as follows: This eliminates the impact of pixel value scale differences under different lighting conditions on subsequent processing.

[0109] Image cropping: The image is cropped according to the Region of Interest (ROI) of the transmission line, retaining 80% of the central area (removing edge distortion), and the image size is uniformly adjusted to 224×224 pixels after cropping (to adapt to the input requirements of CNN models).

[0110] Wavelet Transform Multi-Scale Decomposition and Depth Feature Extraction: Two-Dimensional Discrete Wavelet Transform: The dwt2 function of the PyWavelets library is called, and the wavelet basis is set to Daubechies-4 (db4). A first-level decomposition is performed on the preprocessed 224×224 grayscale image to obtain four sub-band images (all 112×112 pixels): Approximate sub-band (LL): preserves low-frequency structural information such as the overall outline of the wildfire area; Horizontal detail sub-band (HL): captures the horizontal edge features of flames and smoke; Vertical detail sub-band (LH): captures the vertical edge features of flames and smoke; Diagonal detail sub-band (HH): captures diagonal detail features such as flame texture and smoke dispersion trajectory.

[0111] CNN Feature Extraction: The four sub-band images are input into the pre-trained InceptionV3 model (as described in claim 4). The model is loaded with ImageNet pre-trained weights, the parameters of the first 100 layers are frozen, and the parameters of the last 10 layers are fine-tuned (the training data is a special dataset for power transmission line wildfires). The model's GlobalAveragePooling2D layer is followed by a Flatten layer to output the depth feature vector of each sub-band, with a dimension of 4096. =4096), the four sub-band feature vectors are respectively denoted as ;

[0112] Multi-subband feature fusion: The depth features of four subbands are fused using a feature concatenation method, and are concatenated in the order of "approximate subband → horizontal detail subband → vertical detail subband → diagonal detail subband" to form a comprehensive feature vector. ,in The vertical concatenation of the vectors results in a fused feature vector that contains both low-frequency structure and high-frequency detail information from multiple directions, comprehensively representing the characteristics of wildfires.

[0113] MDS Feature Dimensionality Reduction and Classification: MDS Dimensionality Reduction: The MDS class from the scikit-learn library is used, with parameters n_components=256 (the feature dimension k after dimensionality reduction is 256, much smaller than the original 16384 dimensions) and dissimilarity='euclidean' (using Euclidean distance to construct the distance matrix). The specific implementation steps are as follows: Construct the distance matrix of the comprehensive feature vector. (Test set sample size n=2000) Euclidean distance; calculate matrix ; Calculate the inner product matrix ,in for The Row mean for The Column mean for The global mean of matrix B is obtained by performing eigenvalue decomposition on matrix B, selecting the top 256 largest eigenvalues ​​and their corresponding eigenvectors to form the projection matrix. and diagonal matrix Calculate the dimensionality-reduced feature vectors , SVM classification: The SVC class from the scikit-learn library is called, with the kernel function set to RBF (radial basis function) and the penalty coefficient set to... , The 256-dimensional feature vector after dimensionality reduction is input into the SVM classifier, and the classification threshold is... (Determined through 5-fold cross-validation on the training set), the classification function is: ,in =1 indicates that there is a wildfire area. =0 indicates an area free of wildfires;

[0114] Early warning information generation and push: If the classification result =1, the system automatically determines the confidence level. Determine the warning level, generate alarm information including wildfire location, warning level, image screenshot, and metadata, and push it to the operation and maintenance personnel's terminal through the warning and feedback module, while simultaneously linking with the fire resource dispatch system; if If =1, then only the image data and processing results are stored in the data storage module, and no warning is triggered;

[0115] Experimental conditions: Dataset: Contains 10,000 images related to wildfires along power transmission lines (8,000 training images and 2,000 test images), including:

[0116] Normal scene images: 5000 images (no wildfires, including different environments such as sunny, cloudy, overcast, and nighttime);

[0117] Wildfire scene images: 5000 images (including small flames, thick smoke, mixed flames and smoke, complex background occlusion, etc.);

[0118] Comparison Scheme: Comparison Scheme 1: Traditional CNN method (only input original grayscale image, InceptionV3 model, no wavelet decomposition and MDS dimensionality reduction).

[0119] Comparison Scheme 2: Wavelet decomposition + CNN (no MDS dimensionality reduction, PCA dimensionality reduction is used);

[0120] Evaluation metrics: recognition accuracy, false detection rate (misclassification of normal scenes as wildfires), false negative rate (misclassification of wildfire scenes as normal), and average processing time per image.

[0121] Experimental results: Compared with scheme 1: recognition accuracy 92.1%, false detection rate 6.8%, false negative rate 1%, average processing time 0.8s;

[0122] Comparison with Solution 2: 95.3% accuracy, 3.2% false positive rate, 1.2% false negative rate, and an average processing time of 0.6 seconds;

[0123] This solution achieves a recognition accuracy of 98.2%, a false positive rate of 1.5%, a false negative rate of 0.3%, and an average processing time of 0.4 seconds.

[0124] Results analysis: The recognition accuracy improved by 3.9%-6.1%, effectively capturing wildfire characteristics in complex backgrounds and reducing the risk of false detection and false negative detection;

[0125] The processing time for a single image is reduced by 0.2-0.4 seconds, meeting the real-time requirements of power transmission line inspection;

[0126] It is more adaptable to extreme scenarios such as nighttime, smoke-covered, and cloudy conditions, with a false detection rate of only 1.5% and a false negative rate of 0.3%, which is significantly better than traditional methods.

[0127] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A power transmission line wildfire identification system integrating multi-scale sub-band features and MDS dimensionality reduction, characterized in that, include: The image acquisition module is used to acquire image data of the area along the power transmission line; The image transmission module is connected in communication with the image acquisition module and is used to transmit image data in real time; The image processing and analysis module, which communicates with the image transmission module, is used to convert image data into grayscale images. It performs a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale images to obtain four sub-bands. It extracts the depth feature vectors of each sub-band through a convolutional neural network (CNN), concatenates them to form a comprehensive feature vector, and uses multidimensional scaling (MDS) technology to reduce the dimensionality of the comprehensive feature vector. It then uses a classifier to determine whether there are wildfire areas. The early warning and feedback module communicates with the image processing and analysis module and is used to generate and push alarm information when a wildfire area is determined to exist; The data storage and management module is used to store image data, classification results, and alarm records.

2. The power transmission line wildfire identification system integrating multi-scale sub-band features and MDS dimensionality reduction as described in claim 1, characterized in that: The image acquisition module is a high-definition camera or thermal imaging device mounted on a drone, and the image data is an RGB image or an infrared image; the system can be deployed on a cloud server or an embedded edge device, supporting remote monitoring and localized rapid response modes.

3. The power transmission line wildfire identification system integrating multi-scale sub-band features and MDS dimensionality reduction according to claim 1, characterized in that: The two-dimensional discrete wavelet transform uses the Daubechies-4 filter bank, and achieves multi-scale decomposition through a combination of horizontal filtering, vertical filtering and downsampling. The four sub-bands are the approximate sub-band (LL), the horizontal detail sub-band (HL), the vertical detail sub-band (LH), and the diagonal detail sub-band (HH).

4. The power transmission line wildfire identification system integrating multi-scale sub-band features and MDS dimensionality reduction according to claim 1, characterized in that: The convolutional neural network is either InceptionV3 or MobileNet, and the deep feature vector is extracted from the flattened layer of the model. The multidimensional scaling (MDS) achieves dimensionality reduction by constructing a distance matrix, calculating the inner product matrix, and decomposing eigenvalues. The classifier is a support vector machine (SVM).

5. A method for identifying wildfires on transmission lines by integrating multi-scale sub-band features and MDS dimensionality reduction, characterized by: Includes the following steps: S1. Collect image data of the area along the power transmission line; S2. Preprocess the image data and convert it into a grayscale image to reduce interference from changes in illumination; S3. Perform a first-order two-dimensional discrete wavelet transform (2DDWT) on the grayscale image to obtain four sub-bands. Input each sub-band into a convolutional neural network (CNN) to extract the corresponding depth feature vectors. S4. Concatenate the four deep feature vectors to obtain the comprehensive feature vector; S5. Use multidimensional scaling (MDS) technology to reduce the dimensionality of the comprehensive feature vector, obtain a low-dimensional feature vector, and input it into the classifier to determine whether there is a wildfire area. S6. If a wildfire area is determined to exist, generate an alarm message and push it to the relevant terminal devices.

6. The method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction according to claim 5, characterized in that: The grayscale image conversion formula is y=0.2989R+0.5870G+0.1140B, where R, G, and B are the pixel values ​​of the red, green, and blue channels, and y is the grayscale value. The preprocessing also includes standardizing the grayscale image to improve the adaptability of subsequent algorithms.

7. The method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction according to claim 5, characterized in that: The approximate subband (LL) retains the main structural information of the image, the horizontal detail subband (HL) reflects the horizontal edge information, the vertical detail subband (LH) reflects the vertical edge information, and the diagonal detail subband (HH) reflects the diagonal texture information.

8. The method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction according to claim 5, characterized in that: The dimension of the single subband depth feature vector output by the convolutional neural network ≥4096, the feature splicing is performed by concatenating sub-bands in sequence, forming a dimension of The comprehensive feature vector.

9. The method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction according to claim 5, characterized in that: The specific steps of the Multidimensional Scaling (MDS) are as follows: Constructing a distance matrix of the comprehensive feature vectors. , For the first The and the first Distance between samples; construct matrix ; Calculate the inner product matrix ,in for The Row mean for The Column mean for The global mean; perform eigenvalue decomposition on matrix B, and select the top... The projection matrix is ​​composed of the largest eigenvalues ​​and their corresponding eigenvectors. and diagonal matrix ; Calculate the dimensionality-reduced eigenvectors ,and It is much smaller than the original dimension of the comprehensive feature vector.

10. The method for identifying wildfires on transmission lines by fusing multi-scale sub-band features and MDS dimensionality reduction according to claim 5, characterized in that: The classifier is a support vector machine (SVM), and the classification judgment function is: C(x)=1 if f(x)>T; C(x)=0 otherwise; Where C(x)=1 indicates the presence of wildfire areas, C(x)=0 indicates the absence of wildfire areas, and T is the preset classification threshold.

Images