Electric arc fault fire early warning method, system, medium and equipment in cable tunnel

Abstract

The invention discloses an early warning method, system, medium and equipment for cable arc fault fire in a cable tunnel, and the method comprises the steps: collecting fire images, wind speed, smoke color and concentration at different fire development stages, and enabling the fire images to comprise flame and smoke; the fire image, the wind speed, the smoke color and the concentration form a data set, the data set is divided into a training set and a test set according to a proportion, a YOLOv8 neural network is trained based on the training set to obtain an optimized YOLOv8 neural network early warning model, and the optimized YOLOv8 neural network early warning model is tested by using the test set. When the accuracy rate, the recall rate and the precision value meet preset requirements, determining the optimized YOLOv8 neural network early warning model; and inputting a to-be-tested fire image, wind speed, smoke color and concentration into the optimized YOLOv8 neural network early warning model to obtain a fire development stage.

Description

Technical Field

[0001] The present invention relates to the technical field of cable arc fault fire monitoring, and in particular to an early warning method, system, medium and equipment for cable arc fault fire in a cable tunnel. Background Art

[0002] Cables are critical energy transmission equipment in power systems, characterized by high output and widespread installation. Their safe and reliable operation is directly linked to the stability of the power grid. However, due to the complex and variable operating environments of cables and the flammability of their insulation materials, fires are frequent, resulting in significant economic losses and severely impacting social production and residents' lives. Fire alarm systems are essential to prevent these serious incidents. However, the causes of power cable ignition vary greatly, and the development and combustion process of cable ignitions vary greatly under different circumstances. Therefore, the question of whether fire warning devices can provide warnings in different situations is urgently needed.

[0003] The main causes of cable ignition include unreliable connections, carbonization paths, air arcs, external heat sources, overheating of the cable core, short circuits, etc. The process that cables must go through to eventually ignite is the generation of arcs on the cable surface or inside. Arcs can be divided into internal arcs and external arcs based on their position relative to the cable. External arcs can be considered as discharge arcs caused by cable sections around the target cable or other discharge arcs caused by insulation problems in equipment such as poles and towers, which burn the target cable and cause it to ignite. Internal arcs can be divided into parallel arcs and series arcs. The main cause of series arcs is unreliable connections in weak links such as joints, or cable cores are broken due to external forces. Once a series arc occurs, it will not only ignite a single cable, but also act as an external arc to burn surrounding cables. Parallel arcs are caused by discharge from the core to the ground or between cores after the cable insulation fails. Distribution cables and joints have a high failure rate and are very prone to single-phase grounding faults. According to statistics, over 80% of single-phase grounding faults are arc-type. When a single-phase grounding fault occurs in a system with an ungrounded neutral point and continues to operate for one to two hours, the current flowing through the grounding point during the fault is three times the per-phase capacitance current to ground during normal operation, generating a parallel arc at the grounding point. Once a cable is ignited, the flame pattern and smoke generated vary with the duration of the cable combustion. The duration and intensity of the combustion determine the extent of erosion and damage to the power cable material.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The present invention provides an early warning method, system, medium and equipment for cable arc fault fire in a cable tunnel, which can evaluate the severity of cable fire and provide early warning.

[0006] An early warning method for cable arc fault fire in a cable tunnel includes:

[0007] Collecting fire images, wind speed, smoke color and concentration at different fire development stages, wherein the fire images include flames and smoke;

[0008] Fire images, wind speed, smoke color and concentration constitute the data set and are divided into training set and test set according to the proportion.

[0009] The YOLOv8 neural network is trained based on the training set to obtain an optimized YOLOv8 neural network warning model. The optimized YOLOv8 neural network warning model is tested using the test set. When the accuracy, recall rate, and precision values meet the predetermined requirements, the optimized YOLOv8 neural network warning model is confirmed.

[0010] The fire image, wind speed, smoke color and concentration to be tested are input into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

[0011] In the aforementioned early warning method for cable arc fault fires in cable tunnels, the accuracy, recall, and precision of image detection are used as indicators for evaluating image recognition performance. The accuracy P refers to the ratio of detected object samples to actual object samples:

[0012]

[0013] TP is the number of samples that are actually correct among the results that are detected as correct; FP is the number of samples that are actually wrong among the results that are detected as correct.

[0014] Recall rate refers to the ratio of the number of correctly detected object samples to the number of all actual object samples.

[0015]

[0016] Where FN is the number of correct samples that were not detected;

[0017] The precision value refers to the ratio of the number of correctly detected samples to the total number of samples.

[0018]

[0019] Where ALL is the total number of samples.

[0020] In the aforementioned early warning method for cable arc fault fire in a cable tunnel, a high-speed camera captures fire images at different fire development stages.

[0021] In the described early warning method for cable arc fault fire in a cable tunnel, a high-speed camera is used to capture the flame combustion morphology change process and smoke diffusion image as a fire image while ensuring that the distance between the high-speed camera and the cable ignition point is fixed at 2m.

[0022] In the early warning method for cable arc fault fire in a cable tunnel, the YOLOv8 neural network includes an input, a backbone network, a neck network, and a head network.

[0023] In the aforementioned early warning method for cable arc fault fires in cable tunnels, the input includes a convolution layer, a batch normalization layer, and a SiLU activation function. The convolution layer uses a convolution kernel to extract local features. The formula is: Y = X*W+b, where X is the input image feature, W is the convolution kernel weight, * represents the convolution operation, and b is the bias term. The convolution uses a 3×3 or 1×1 convolution kernel and adopts a step size of 1 or 2 for feature extraction and dimensionality reduction. The batch normalization layer standardizes the data distribution and accelerates the convergence speed. The formula is: Among them, μ and σ 2 are the mean and variance of the data, ∈ is a small constant; the SiLU activation function enhances the ability to distinguish between flames, smoke and other forms. The formula is:

[0024] In the described early warning method for cable arc fault fires in cable tunnels, a backbone network extracts deep features of fire images. The first layer of the backbone network is a first-layer CBS structure of convolution, normalization, and SiLU activation, and the second layer is a second-layer CBS structure of convolution, normalization, and SiLU activation. The resolution is reduced by convolution with a step size of 2, the number of channels is increased, and the C2f module performs Bottleneck calculation to improve feature extraction capabilities. The method includes dividing the input feature map into two parts according to the channel, one part directly jumps to the input feature segmentation unit of the Bottleneck calculation and multiple Bottleneck structures for extracting deep features, and splicing the features of the jump connection and the features extracted by the Bottleneck to form a feature fusion structure for the final output.

[0025] An early warning system for implementing the method includes:

[0026] an acquisition unit for acquiring fire images, wind speed, smoke color, and concentration at different fire development stages, wherein the fire images include flames and smoke;

[0027] A processing unit is used to construct a data set based on fire images, wind speed, smoke color and concentration and divide it into a training set and a test set according to the proportion.

[0028] A training unit is used to train the YOLOv8 neural network based on the training set to obtain an optimized YOLOv8 neural network early warning model, and use the test set to test the optimized YOLOv8 neural network early warning model. When the accuracy, recall rate and precision values meet the predetermined requirements, the optimized YOLOv8 neural network early warning model is confirmed;

[0029] The test unit is used to input the fire image, wind speed, smoke color and concentration to be tested into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

[0030] A computer storage medium includes computer instructions, which, when executed on a computer, cause the computer to execute the method described above.

[0031] An electronic device, comprising:

[0032] A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein:

[0033] When the processor executes the program, the method described is implemented.

[0034] Compared with existing technologies, this invention has the following advantages: It uses a high-speed camera, an infrared thermal imager, and an electronic scale to record the flame morphology changes and smoke diffusion after cable ignition, as well as the temperature changes around the cable and the progress of cable material loss. Finally, it uses a feature-based YOLO image recognition method for early warning of power cable fires to assess the severity of cable fires and provide early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are intended only to illustrate preferred embodiments and are not to be construed as limiting the present invention. It should be understood that the drawings described below are merely examples of the present invention, and that those skilled in the art will be able to derive other drawings from these drawings without inventive effort. Throughout the drawings, identical reference numerals are used to denote identical components.

[0036] In the attached figure:

[0037] Figure 1 It is a schematic diagram of the process of the present invention.

[0038] The present invention will be further explained below with reference to the accompanying drawings and embodiments. DETAILED DESCRIPTION

[0039] Specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0040] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the invention. The scope of protection of the present invention shall be as defined in the attached claims.

[0041] To facilitate understanding of the embodiments of the present invention, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0042] like Figure 1 As shown, the early warning method for cable arc fault fire in a cable tunnel includes the following steps:

[0043] Collecting fire images, wind speed, smoke color and concentration at different fire development stages, wherein the fire images include flames and smoke;

[0044] Fire images, wind speed, smoke color and concentration constitute the data set and are divided into training set and test set according to the proportion.

[0045] The YOLOv8 neural network is trained based on the training set to obtain an optimized YOLOv8 neural network warning model. The optimized YOLOv8 neural network warning model is tested using the test set. When the accuracy, recall rate, and precision values meet the predetermined requirements, the optimized YOLOv8 neural network warning model is confirmed.

[0046] The fire image, wind speed, smoke color and concentration to be tested are input into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

[0047] In a preferred embodiment of the method for early warning of cable arc fault fire in a cable tunnel, the accuracy, recall rate, and precision of image detection are used as indicators for evaluating image recognition performance. The accuracy rate P refers to the ratio of detected object samples to actual object samples:

[0048]

[0049] TP is the number of samples that are actually correct among the results that are detected as correct; FP is the number of samples that are actually wrong among the results that are detected as correct.

[0050] Recall rate refers to the ratio of the number of correctly detected object samples to the number of all actual object samples.

[0051]

[0052] Where FN is the number of correct samples that were not detected;

[0053] The precision value refers to the ratio of the number of correctly detected samples to the total number of samples.

[0054]

[0055] Where ALL is the total number of samples.

[0056] In a preferred embodiment of the method for early warning of cable arc fault fire in a cable tunnel, a high-speed camera captures fire images at different fire development stages.

[0057] In a preferred embodiment of the early warning method for cable arc fault fire in a cable tunnel, a high-speed camera is used to capture the flame combustion morphology change process and smoke diffusion image as a fire image while ensuring that the distance between the high-speed camera and the cable ignition point is fixed at 2m.

[0058] In a preferred embodiment of the early warning method for cable arc fault fire in a cable tunnel, the YOLOv8 neural network includes an input, a backbone network, a neck network and a head network.

[0059] In a preferred embodiment of the early warning method for cable arc fault fire in a cable tunnel, the input includes a convolution layer, a batch normalization layer, and a SiLU activation function. The convolution layer uses a convolution kernel to extract local features. The formula is: Y = X*W+b, where X is the input image feature, W is the convolution kernel weight, * represents the convolution operation, and b is the bias term. The convolution uses a 3×3 or 1×1 convolution kernel and uses a step size of 1 or 2 for feature extraction and dimensionality reduction. The batch normalization layer standardizes the data distribution and accelerates the convergence speed. The formula is: Among them, μ and σ 2 are the mean and variance of the data, ∈ is a small constant; the SiLU activation function enhances the ability to distinguish between flames, smoke and other forms. The formula is:

[0060] In a preferred embodiment of the early warning method for cable arc fault fire in a cable tunnel, a backbone network extracts deep features of the fire image. The first layer of the backbone network is a first-layer CBS structure of convolution, normalization and SiLU activation, and the second layer is a second-layer CBS structure of convolution, normalization and SiLU activation. The resolution is reduced by convolution with a step size of 2, the number of channels is increased, and the C2f module performs Bottleneck calculation to improve the feature extraction capability. The method includes dividing the input feature map into two parts according to the channel, one part is directly jump-connected, and the other part enters the input feature segmentation unit of the Bottleneck calculation and multiple Bottleneck structures for extracting deep features, and splicing the jump-connected features and the features extracted by Bottleneck to form a feature fusion structure for the final output.

[0061] An early warning system for implementing the method includes:

[0062] an acquisition unit for acquiring fire images, wind speed, smoke color, and concentration at different fire development stages, wherein the fire images include flames and smoke;

[0063] A processing unit is used to construct a data set based on fire images, wind speed, smoke color and concentration and divide it into a training set and a test set according to the proportion.

[0064] A training unit is used to train the YOLOv8 neural network based on the training set to obtain an optimized YOLOv8 neural network early warning model, and use the test set to test the optimized YOLOv8 neural network early warning model. When the accuracy, recall rate and precision values meet the predetermined requirements, the optimized YOLOv8 neural network early warning model is confirmed;

[0065] The test unit is used to input the fire image, wind speed, smoke color and concentration to be tested into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

[0066] Fire severity assessment and early warning. To deepen understanding and recognition of the fire severity assessment and early warning method of the present invention, we first introduce the image recognition method based on YOLO. To quickly and accurately identify targets, YOLO divides the image into grids and simultaneously predicts bounding boxes and class probabilities. Bounding box coordinates and class probabilities are generated by convolutional layers after feature extraction by a deep convolutional neural network (CNN). YOLO improves the detection of targets of different sizes by using multi-scale anchor boxes. The final detection results are optimized using non-maximum suppression (NMS), which filters out redundant and low-confidence predictions, making YOLO an efficient and reliable target detection method. After conducting series and parallel arc tests on a resonant platform, the YOLO neural network can divide the acquired flame combustion morphology and smoke diffusion photos into grids, identify a large number of grids, demarcate the grid area where the flame or smoke is located, predict the bounding box of the complete flame or smoke, and generate the bounding box coordinates. During the internal arc ignition process, the cable sheath is initially well sealed, but only after a period of ablation does a noticeable gap appear, allowing air to enter. As the burning material fully contacts the air, the flame grows significantly, and the smoke diffusion rate increases. As combustion progresses, the remaining material decreases, and the arc distance increases, causing the flame to decrease again, and the smoke diffusion rate to decrease. This indicates that the flame morphology and smoke diffusion range at each stage of the cable ablation process have distinct characteristics. Because the camera distance from the test cable remains constant during the experiment, after the fault is identified, the grid boundary size can indicate the relative size of the actual flame and the diffusion range of the smoke. This corresponds to the flame morphology and smoke range at different times after cable ignition, thereby identifying the time of the cable fire disaster and serving as an assessment criterion for the cable ablation condition and fire severity.

[0067] The YOLOv8 network is a highly efficient object detection architecture within the YOLO image recognition method. It primarily consists of the input layer, backbone network, neck network, and head network. The input layer preprocesses and enhances the raw input image data. When the input dataset is small, limited, and lacks a corresponding set of trial-and-error data, data augmentation techniques such as random cropping and horizontal flipping are used to expand the number of different samples in the dataset. Other techniques include mosaic enhancement, hybrid enhancement, and spatial perturbation. The backbone network is the feature extraction component, extracting features from the image enhanced by the input layer. The neck network refines and combines features extracted from different layers of the backbone network to create a rich and informative feature representation, enhancing the model's object recognition capabilities. Finally, the head network performs the actual prediction and outputs the bounding box, class probability, and objectness score of the recognized object.

[0068] The specific steps for using the YOLOv8 neural network to identify flame combustion morphology and smoke diffusion photos include:

[0069] The flame combustion morphology and smoke diffusion photos under different arc conditions are input into the YOLO algorithm. The input image is divided into grids of equal size using the SXS grid. The flame and smoke borders are predicted by calculation, and the output coordinates are:

[0070] Z x =κ(r x )+p x

[0071] Z y =κ(r y )+p y

[0072] Z w =b w χ w

[0073] Z h =b h χ h

[0074] Where Z x 、Z y 、Z w 、Z h are the predicted output coordinates after grid division, representing the size and position of the input image; κ(r x ),κ(r y ) is the center point of the grid; p x 、p y are the length and width of the grid; w , χ h is the degree of size scaling; b w 、b h Indicates the predicted length and width of the grid.

[0075] After the grid predicts the borders of flames and smoke, the confidence score of the border is calculated. This reflects whether the grid contains the predicted object, such as flames, and can also accurately predict the location of the flame grid. Obviously, when there is no flame in the grid, the confidence score is zero and the corresponding grid is not considered. Otherwise, the coverage score of the predicted border and the actual area surrounded by the flame is corresponding to each other. The larger the numerical value, the larger the overlapping area of the two. The coverage score reflects the similarity between the predicted grid area in the image (that is, the area surrounded by the predicted border of the recognized image) and the actual grid area (that is, the area surrounded by the actual border of the recognized image):

[0076]

[0077]

[0078] Where I(δ) and U(δ) are the predicted bounding boxes and the actual bounding boxes; e is the image dataset; ω represents a picture obtained from the dataset; is the output of the image; for the actual allocation of the image.

[0079] When the coverage score is the largest, it is considered that the flame shape identified at this time corresponds best to the flame combustion shape at the actual combustion time.

[0080] When using the image recognition method based on the YOLOv8 network for fire warning, the accuracy, recall rate, and precision of image detection are used as the evaluation indicators for image recognition performance:

[0081] 1) Accuracy

[0082] The accuracy is referred to as P, which refers to the ratio of samples detected as objects to samples that are actually objects:

[0083]

[0084] TP is the number of samples that are actually correct among the results detected as correct; FP is the number of samples that are actually wrong among the results detected as correct.

[0085] 2) Recall rate

[0086] Recall is the ratio of the number of correctly detected object samples to the number of all actual object samples.

[0087]

[0088] Where FN is the number of correct samples that were not detected.

[0089] 3) Accuracy value

[0090] The precision value refers to the ratio of the number of correctly detected samples to the total number of samples.

[0091]

[0092] Where ALL is the total number of samples.

[0093] After identification is complete, images of the flame morphology and smoke diffusion range at each stage of cable ignition during the arc development process are obtained for identification and reference. This allows the flame morphology or smoke diffusion range images to be used to determine the cable ignition stage and assess the extent of cable damage and fire severity. Cable ignition caused by internal arcs can also be identified in the early stages of a fire, when flames have not yet appeared on the surface and only a small amount of smoke or other phenomena remain, allowing for early warning or fault elimination. Fire fault information is transmitted to a host computer, providing early warning of power cable fires.

[0094] Cable tunnel fires have several unique characteristics: The ventilation environment is complex. Transformers are typically enclosed, while cable tunnels often have duct ventilation systems, which means that flame and smoke spread is affected by wind speed. Flames and smoke in cable tunnels may spread along the wind flow, which differs from the flame spread characteristics in static environments. Complex flame spread paths: In cable tunnel fires, flames spread along cable bundles and bridges, rather than burning in a fixed area. Due to the long and narrow tunnel structure, flames and smoke spread in an irregular pattern, rather than the relatively uniform combustion pattern seen in transformer fires. Diverse fire sources: While the primary combustible in transformer fires is transformer oil, the primary combustible in cable tunnels is polymer materials such as polyvinyl chloride (PVC) and cross-linked polyethylene (XLPE) cable sheathing, which produce large amounts of toxic smoke. These different burning speeds and smoke characteristics require different fire identification models. Different heat conduction characteristics: In cable tunnel fires, flames conduct heat along the cables themselves, causing the fire to spread along the tunnel. The specific variables of cable tunnel fire are combined with the YOLO recognition algorithm to make it suitable for cable tunnel fire monitoring. The wind speed variable is introduced to optimize the YOLO target detection so that it considers the flame shape and smoke diffusion pattern under the influence of wind speed. The wind speed variable is added to the training data set, that is, under different wind speed conditions V 风速 Images of flame and smoke diffusion are collected at speeds of 0, 1, and 3 m / s, creating a dataset more consistent with cable tunnel environments. Wind speed information is added to the YOLO prediction model, enabling the neural network to adjust flame and smoke detection thresholds based on wind speed, thus avoiding false alarms caused by wind speed. Specifically, wind speed is first determined for the input neural network data before accuracy and other data are used in Detect.

[0095] No wind (0m / s): Flames rise and smoke spreads vertically.

[0096] Low wind speed (1-2 m / s): The flame is slightly tilted, and the direction of smoke diffusion is offset.

[0097] High wind speed (>3m / s): The flame is blown obliquely and the smoke flows along the tunnel direction.

[0098] The flame spread direction variable is introduced. The flame propagation pattern of cable tunnel fires is different from that of ordinary fires: the flames of cable tunnel fires will propagate vertically along the cable path. During training, YOLO learns the longitudinal propagation characteristics of flames. If the flame direction is detected to be mainly along the cable (rather than spreading uniformly outward), it can be marked as a cable fire mode. When calculating the flame bounding box, the directional offset weight is increased to improve the adaptability to cable fires. The smoke color and concentration variables are introduced as judgment conditions to distinguish the stage of the fire. The smoke of cable tunnel fires has unique color characteristics:

[0099] PVC cable burning produces thick black smoke, which is different from the gray-white smoke commonly seen in ordinary fires.

[0100] The burning of XLPE cables produces white or light yellow smoke, which is different from the smoke produced by burning ordinary wood or plastic.

[0101] YOLO is used to identify the color of fire smoke, using black smoke (PVC) and yellow smoke (XLPE) as key indicators of cable tunnel fires to improve detection accuracy. This is performed within the head network's object classification. This not only identifies and classifies cable tunnel fires at different stages to determine their severity, but also now requires identifying and classifying them at different stages under stagnant airflow, determining the direction of fire spread, and identifying the color of the smoke produced.

[0102] In one embodiment, Figure 1 The figure below illustrates the complete process of the YOLOv8 object detection algorithm, which includes the input, backbone, neck, and head networks. This structure is explained in detail below, along with the specific characteristics of cable tunnel fire detection. w is the width scaling factor, which controls the number of channels in each convolution layer (i.e., feature map depth), affecting the model's parameter count and computational complexity. r is the resolution scaling factor, which controls the resolution of the input image and affects detection accuracy and computational complexity. It is typically set to 1.

[0103] 1. Input

[0104] effect:

[0105] Input a 640×640×3 (RGB three-channel) fire image (which can be flames or smoke in a cable tunnel).

[0106] ●Normalized (pixel value 0255 is mapped to 01).

[0107] ●Perform data enhancement (random flipping, brightness adjustment, etc.).

[0108] Operations in the picture:

[0109] The initial size is 640×640×3, and after CBS (convolution + normalization + activation) processing, it becomes 320×320×64.

[0110] The CBS module (Convolution+BatchNorm+SiLU) is a basic feature extraction module commonly used in YOLOv8 and other deep learning models. It mainly consists of three parts:

[0111] a) Convolution

[0112] b) Batch Normalization (BN)

[0113] c) SiLU activation function (Sigmoid Linear Unit, SiLU)

[0114] Each CBS module consists of the following three main steps:

[0115] 1) Convolutional layer (Conv):

[0116] ○Function: Use convolution kernel to extract local features, such as the edge of flames, the shape of smoke, etc.

[0117] ○Formula: Y=X*W+b

[0118] ○Where X is the input image feature, W is the convolution kernel weight, * represents the convolution operation, and b is the bias term.

[0119] ○ In YOLOv8, convolution typically uses 3×3 or 1×1 convolution kernels and adopts a stride of 1 or 2 for feature extraction and dimensionality reduction.

[0120] 2) Batch Normalization Layer (BatchNorm):

[0121] ○ Function: Standardize data distribution, speed up convergence, reduce gradient vanishing or gradient exploding problems, and improve model stability.

[0122] ○Formula:

[0123] Among them, μ and σ 2 are the mean and variance of the data, and ∈ is a small constant to prevent division by zero errors.

[0124] 3) SiLU (Swish) activation function:

[0125] ○ Function: Introducing nonlinearity enables the neural network to learn complex features and enhance the ability to distinguish forms such as flames and smoke.

[0126] ○Formula:

[0127] Role in Cable Tunnel Fire Detection

[0128] In cable tunnel fires, the flames and smoke can be very complex (unlike the stable combustion of transformer fires). CBS can extract local features of flames and smoke (such as edges, color, and shape), ensuring that the model can identify flame / smoke targets at both high and low resolutions. It also enhances feature expression capabilities and prevents misjudgment due to complex backgrounds (such as cables and walls). 2. Backbone Network

[0129] effect:

[0130] Extract the deep features of flames and smoke, reduce the spatial resolution, and increase the number of channels to achieve stronger expression capabilities.

[0131] C2f (CSP structure) and SPPF (spatial pyramid pooling) are used to further enhance feature extraction.

[0132] Process analysis in the figure:

[0133] 1) The first layer CBS (320×320×64), basic convolution + normalization + SiLU activation.

[0134] 2) The second layer CBS (160×160×128) uses convolution with a stride of 2 to reduce the resolution and increase the number of channels.

[0135] 3) C2f (160×160×128→80×80×256), performs Bottleneck calculation to improve feature extraction capabilities.

[0136] C2f (Cross Stage Partial Fusion) is a core module used by YOLOv8 to enhance feature extraction capabilities. It contains multiple Bottleneck structures to reduce the amount of computation while maintaining strong feature expression capabilities.

[0137] C2f consists of the following parts:

[0138] a) Partial Split: The input feature map is divided into two parts by channel. One part is directly skipped (Shortcut), and the other part enters the Bottleneck calculation.

[0139] b) Multiple Bottleneck calculations (core part): used to extract deep features.

[0140] c) Feature Fusion: Concatenate the features from the skip connection and the features extracted from the bottleneck to form the final output.

[0141] In cable tunnel fire detection, the Bottleneck structure helps extract flame and smoke features of different scales and improve detection capabilities; it can also reduce the computational complexity of C2f and increase detection speed.

[0142] 4) Multi-layer CBS+C2f continues dimensionality reduction:

[0143] a)80×80×256→40×40×512

[0144] b)40×40×512→20×20×512

[0145] 5) SPPF (Spatial Pyramid Pooling)

[0146] a) 20×20×512 The maximum pooling layer extracts global information, enhances detection capabilities, and finally outputs 20×20×512.

[0147] In the Backbone (backbone network) part of YOLOv8, SPPF is connected to the end of Backbone and is used for:

[0148] Extract global features

[0149] Increase receptive field

[0150] ●Reduce the amount of calculation and improve the reasoning speed

[0151] Spatial pyramid pooling can optimize cable tunnel fire detection capabilities:

[0152] (1) Increase the receptive field

[0153] Due to the use of multi-scale pooling, the network can focus on a wider range of information, such as:

[0154] ○ In cable tunnel fire detection, it can help the model better identify smoke spread trends.

[0155] ○ Let the model capture the global characteristics of flames and smoke, even if the flames are relatively small in the local image.

[0156] (2) Higher computational efficiency

[0157] Traditional SPP (Spatial Pyramid Pooling) requires multiple pooling kernels of different sizes (such as 1×1, 5×5, and 9×9), while SPPF only uses 5×5 and reuses the results, reducing the amount of computation and improving inference speed.

[0158] (3) Adapting to targets of different scales

[0159] In cable tunnel fire detection scenarios:

[0160] The shape and size of flames and smoke are constantly changing. SPPF can extract information at different scales through multi-scale pooling, improving detection capabilities.

[0161] ○ Enhanced small target detection capabilities to ensure that even small flames or smoke can be identified.

[0162] 3. Neck

[0163] effect:

[0164] Fusion of features at different scales ensures that YOLOv8 can detect both large flames and small smoke.

[0165] ●Adopt FPN (feature pyramid) + PAN (path enhancement) structure.

[0166] Process analysis in the figure:

[0167] 1) After SPPF, the 20×20×512 image is upsampled to 40×40×512.

[0168] 2) Concatenate with the 40×40×512 image extracted by Backbone and perform C2f calculation.

[0169] i. Concat (concatenation) is an operation in a neural network that connects two or more feature maps along a certain dimension. In the YOLOv8 architecture, concat is primarily used for feature fusion, specifically integrating features of different scales in the neck region to improve object detection.

[0170] 3) Continue upsampling to obtain 80×80×256, cascade it with the 80×80×256 extracted by Backbone, and then pass it through C2f.

[0171] 4) Get 3 sets of feature maps:

[0172] a)80×80×256

[0173] b)40×40×512

[0174] c)20×20×512

[0175] 5) Bottom-up processing: C2f is calculated for each layer to enhance features.

[0176] 6) Finally, three sets of feature maps are generated for detection.

[0177] 4. Head Network

[0178] effect:

[0179] ●Through the Detect head network, output:

[0180] ○ Category prediction (flame / smoke)

[0181] ○ Bounding box coordinates (x,y,w,h)

[0182] In YOLOv8, the core of target detection is to generate a bounding box to describe the location of flames and smoke, and call the regression branch in the decoupled detection head to generate the coordinates of the target:

[0183] Z x =k(r x )+p x

[0184] Z y =kr y )+p y

[0185] Z w =b w X w

[0186] Z h =b h χ h

[0187] Where Z x 、Z y 、Z w 、Z h are the predicted output coordinates after grid division, representing the size and position of the input image; κ(r x ),κ(r y ) is the center point of the grid; p x 、p y are the length and width of the grid; w , χ h is the degree of size scaling; b w 、b h Indicates the predicted length and width of the grid.

[0188] Confidence

[0189] In YOLOv8, after generating the bounding box, it is necessary to judge the recognition result to adjust the bounding box and call the classification branch (Classification Head) in the decoupled detection head to predict the category information and confidence of the target within the target prediction box:

[0190] Confidence score calculation

[0191] A confidence score is calculated for the bounding box, which reflects whether the grid contains the predicted object, such as a flame, and accurately predicts the location of the grid containing the flame. Obviously, when there is no flame in the grid, the confidence score is zero and the corresponding grid is not considered.

[0192] Coverage score calculation

[0193] The coverage score of the predicted bounding box corresponds to the actual area surrounded by the flame. The larger the value, the greater the overlap area. The coverage score reflects the similarity between the predicted grid area in the image (i.e., the area surrounded by the predicted bounding box of the recognized image) and the actual grid area (i.e., the area surrounded by the actual bounding box of the recognized image):

[0194]

[0195]

[0196] Where I(δ) and U(δ) are the predicted bounding boxes and the actual bounding boxes; e is the image dataset; ω represents a picture obtained from the dataset; Output two for the image; for the actual allocation of the image.

[0197] When B 覆盖 The larger the value is, the closer the predicted border is to the real border and the more accurate the predicted target is. Here, the threshold is set to 0.5. If B 覆盖 If it is greater than 0.5, we consider the sample to be correctly identified, otherwise it is an error. The decoupled detection head will identify and output the corresponding image boundary coordinates.

[0198] Use non-maximum suppression (NMS) to remove duplicate frames.

[0199] In YOLOv8, non-maximum suppression (NMS) is mainly used to remove target boxes with high overlap to ensure that the final detection results are more accurate. 覆盖 Calculate the similarity of candidate frames and remove low-quality frames. The working principle is as follows

[0200] 1. Calculate the confidence scores of all candidate boxes.

[0201] 2. Sort all boxes by confidence in descending order (boxes with higher confidence first).

[0202] 3 Calculate the B of the box and the highest confidence box 覆盖 , if B 覆盖 If the value exceeds the threshold (iou-thres), the low confidence box is removed.

[0203] 4. Keep the final filtered box as the final detection result.

[0204] Process analysis in the figure:

[0205] 1) The head network consists of three Detect layers of different scales:

[0206] a) 80×80×256 → Suitable for small targets (small flames, small smoke)

[0207] b) 40×40×512 → Suitable for medium targets

[0208] c) 20×20×512 → Suitable for large targets

[0209] 2) After passing the Detect layer, the target box and category are output.

[0210] 3) Finally, NMS (non-maximum suppression) is used to filter out low-confidence predictions.

[0211] In one embodiment, the accuracy, recall rate, and precision of image detection are used as indicators for evaluating image recognition performance:

[0212] a) Accuracy

[0213] The accuracy is referred to as P, which refers to the ratio of samples detected as objects to samples that are actually objects:

[0214]

[0215] TP is the number of samples that are actually correct among the results detected as correct; FP is the number of samples that are actually wrong among the results detected as correct.

[0216] b) Recall Recall refers to the ratio of the number of correctly detected object samples to the number of all actual object samples.

[0217]

[0218] Where FN is the number of correct samples that were not detected.

[0219] c) Precision value The precision value refers to the ratio of the number of correctly detected samples to the total number of samples.

[0220]

[0221] Where ALL is the total number of samples.

[0222] Although there is no convergence problem, in order to make the YOLOv8 neural network training results meet the requirements, the number of iterations n, the threshold B 覆盖 Or adjust the NMS threshold to make modifications until the neural network outputs appropriate results.

[0223] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. A method for early warning of cable arc fault fire in a cable tunnel, characterized in that: The steps include: Collecting fire images, wind speed, smoke color and concentration at different fire development stages, wherein the fire images include flames and smoke; Fire images, wind speed, smoke color and concentration constitute the data set and are divided into training set and test set according to the proportion. The YOLOv8 neural network is trained based on the training set to obtain an optimized YOLOv8 neural network warning model. The optimized YOLOv8 neural network warning model is tested using the test set. When the accuracy, recall rate, and precision values meet the predetermined requirements, the optimized YOLOv8 neural network warning model is confirmed. The fire image, wind speed, smoke color and concentration to be tested are input into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

2. The method for early warning of cable arc fault fire in a cable tunnel according to claim 1, characterized in that: Preferably, the accuracy, recall, and precision of image detection are used as indicators for evaluating image recognition performance. The accuracy P refers to the ratio of samples detected as objects that are actually objects: TP is the number of samples that are actually correct among the results that are detected as correct; FP is the number of samples that are actually wrong among the results that are detected as correct. Recall rate refers to the ratio of the number of correctly detected object samples to the number of all actual object samples. Where FN is the number of correct samples that were not detected; The precision value refers to the ratio of the number of correctly detected samples to the total number of samples. Where ALL is the total number of samples.

3. The method for early warning of cable arc fault fire in a cable tunnel according to claim 1, characterized in that: High-speed cameras capture fire images at different stages of fire development.

4. The method for early warning of cable arc fault fire in a cable tunnel according to claim 3, characterized in that: While ensuring that the distance between the high-speed camera and the cable ignition point is fixed at 2m, the high-speed camera is used to capture the flame combustion morphology change process and smoke diffusion images as fire images.

5. The method for early warning of cable arc fault fire in a cable tunnel according to claim 1, characterized in that: The YOLOv8 neural network consists of input, backbone network, neck network and head network.

6. The method for early warning of cable arc fault fire in a cable tunnel according to claim 5, characterized in that: The input includes a convolutional layer, a batch normalization layer, and a SiLU activation function. The convolutional layer uses a convolution kernel to extract local features. The formula is: Y = X*W+b, where X is the input image feature, W is the convolution kernel weight, * represents the convolution operation, and b is the bias term. Convolution uses a 3×3 or 1×1 convolution kernel and adopts a step size of 1 or 2 for feature extraction and dimensionality reduction. The batch normalization layer standardizes the data distribution and accelerates the convergence speed. The formula is: Among them, μ and σ 2 are the mean and variance of the data, ∈ is a small constant; the SiLU activation function enhances the ability to distinguish between flames, smoke and other forms. The formula is:

7. The method for early warning of cable arc fault fire in a cable tunnel according to claim 1, characterized in that: The backbone network extracts deep features of fire images. The first layer of the backbone network is a first-layer CBS structure with convolution, normalization and SiLU activation, and the second layer is a second-layer CBS structure with convolution, normalization and SiLU activation. It reduces the resolution through convolution with a step size of 2 and increases the number of channels. The C2f module performs Bottleneck calculation to improve the feature extraction capability. It includes dividing the input feature map into two parts according to the channel, one part directly jumps to the other part, and the input feature segmentation unit for Bottleneck calculation and multiple Bottleneck structures for extracting deep features, and splicing the features of the jump connection and the features extracted by Bottleneck to form the final output feature fusion structure.

8. An early warning system for implementing the method according to any one of claims 1 to 7, characterized in that: It includes: an acquisition unit for acquiring fire images, wind speed, smoke color, and concentration at different fire development stages, wherein the fire images include flames and smoke; A processing unit is used to construct a data set based on fire images, wind speed, smoke color and concentration and divide it into a training set and a test set according to the proportion. A training unit is used to train the YOLOv8 neural network based on the training set to obtain an optimized YOLOv8 neural network early warning model, and use the test set to test the optimized YOLOv8 neural network early warning model. When the accuracy, recall rate and precision values meet the predetermined requirements, the optimized YOLOv8 neural network early warning model is confirmed; The test unit is used to input the fire image, wind speed, smoke color and concentration to be tested into the optimized YOLOv8 neural network early warning model to obtain the fire development stage.

9. A computer storage medium, characterized in that The storage medium includes computer instructions, which, when executed on a computer, enable the computer to perform the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

Images