A method and system for indoor fire detection and automatic fire extinguishing based on deep learning

Abstract

The present invention discloses a method and system for indoor fire detection and automatic fire extinguishing based on deep learning. The system includes an upper computer and a lower computer. The upper computer includes an image acquisition module, which uses a binocular camera to collect image data of an indoor monitoring area in real time; an image detection module, which performs flame detection on the collected image data based on a trained deep learning network model. If a flame is detected, the flame area is marked in the image using a frame, and the pixel position information (x, y) of the flame center is determined according to the center point of the frame; a flame positioning module, which uses a binocular camera positioning mapping method to map the pixel position information of the flame center to obtain the spatial position information (x, y, z) corresponding to the flame center; and the lower computer includes an automatic fire extinguishing system, which is used to receive the spatial position information of the flame center sent by the upper computer, and start the fire extinguisher to extinguish the fire after tracking. The system can quickly detect flames and extinguish them.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and artificial intelligence technology, and in particular to a method and system for indoor fire detection and automatic fire extinguishing based on deep learning. Background Art

[0002] Fire is one of the most common disasters in daily life and nature. If a fire cannot be detected and controlled early, it will cause serious loss of life and property, and the consequences are very serious. In the early stage of a fire, the burning area of ​​​​combustible materials is small, the flames are not high, the radiant heat is not strong, and the fire develops slowly, which is the best time to extinguish the fire. Therefore, early detection and control of fire is particularly important.

[0003] Traditional flame detection is based on physical sensors such as temperature and smoke. However, these methods are highly dependent on the environment and have a limited range of recognition. Sometimes it is difficult to effectively identify the flame, or when the fire is identified, it is difficult to control and the alarm cannot be issued in time.

[0004] The flame detection method based on machine vision uses the flame's own characteristics to detect the presence of flames in the image. In the early days of machine vision detection, flame features were mainly designed and extracted manually, mainly including flame color, flame shape, flame edge and other features. Due to the diversity of flames and the complexity of flame combustion, it is difficult to accurately identify various types of flames with manually extracted flame features, especially for small target flames in the early stage of a fire. The recognition rate is low.

[0005] With the development of artificial intelligence technology, flame recognition based on deep learning technology has achieved remarkable results. However, in flame detection based on deep learning technology, the number of samples and the selection and integrity of positive and negative samples have a great impact on the detection results.

[0006] In addition, after detecting a fire, the current system mainly warns through an alarm, but there will be a delay from the time relevant personnel receive the alarm to the time they take action, which will cause the fire to spread rapidly and fail to control the fire in a timely manner.

[0007] In summary, it is crucial to overcome the shortcomings of existing flame detection and propose an indoor fire detection and automatic fire extinguishing system and method based on deep learning. Summary of the invention

[0008] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method and system for indoor fire detection and automatic fire extinguishing based on deep learning.

[0009] To achieve the above object, the technical solution adopted by the present invention is: a method for indoor fire detection and automatic fire extinguishing based on deep learning, comprising the following steps:

[0010] Step 1: calibrate the binocular camera. Use the left and right cameras to shoot the black and white grid calibration plate images respectively. Use MATLAB's stereoCameraCalibrator or write your own code to calibrate. Delete the images with the largest errors in turn and calibrate until the average error of the remaining images is ≤0.3. Finally, export and save the parameters of the left and right cameras. The error refers to the difference between the corner point coordinates calculated by the black and white grid calibration plate image and the actual corner point coordinates of the black and white grid calibration plate.

[0011] Step 2, collect image data of the indoor monitoring area through a binocular camera, and correct the left and right images respectively according to the camera parameters of step 1 to eliminate distortion, convert the image into a grayscale image, increase the weight of the R channel in the RGB channel, and the weight size is R>G>B. Compare the grayscale images, blur the background part of the original image, distinguish the flame area from the background, and improve the recognition of the flame;

[0012] Step 3, input the image corrected in step 2 into the improved Yolo V8 model for target detection. In the BackBone part, the model outputs the image at different scales through convolution and sends them to the Neck part respectively. In the Neck part, features are further extracted and deep information is extracted by upsampling. The images extracted at different depths are input to the Head part for prediction. If flames are detected in the image, the circularity is calculated, and the flame image data with a circularity within 0.3-0.7 is extracted. The flame category and prediction box position information are saved, and the operation of step 4 is entered.

[0013] Step 4: If a flame is detected, the alarm and reminder functions are triggered, and the distance measurement function is called at the same time to perform epipolar correction on the corrected left and right images to obtain a mapping table, that is, the corresponding pixel points in the left image in the right image;

[0014] Step 5: During stereo matching, the disparity map is obtained by mapping and filtering, sparse disparity is converted into dense disparity, and a mathematical model depth=(f*b) / (d+(c r -c l )), where depth is the depth, f is the focal length, b is the baseline length, d is the parallax, and c is the column coordinates of the left and right camera principal points;

[0015] Step 6, using the predicted box position information of step 3, average pooling is performed on the boxed area to obtain the regional average value, and the pixel point closest to the regional average value is found as the coordinate of the flame center, and the pixel coordinates of the flame center are converted into three-dimensional coordinates through the reprojection matrix;

[0016] Step 7: Draw the predicted box of the flame, and draw the category probability and distance together;

[0017] Step 8: After receiving the flame spatial position information, the lower computer drives the fire extinguishing device to track to the designated position to realize automatic fire extinguishing.

[0018] As a specific implementation, the improved Yolo V8 model includes:

[0019] The feature extraction layer uses a feature pyramid network, which is built by a CSPDarkNet53 convolutional neural network, and adds an ECA attention mechanism based on CSPDarkNet53;

[0020] In the network layer, PANet is used to take the features from the 4th, 6th, and 10th layers of the feature extraction layer as input, and after upsampling and channel fusion, they are output to the Head for classification prediction;

[0021] At the output network layer, the decision part of the Decoupled Head target detection model of YOLOX is used as the Head of the Yolo V8 model to separate the regression branch and the prediction branch.

[0022] Loss function, the CIOU loss function is modified to WIOU, and the dynamic non-monotonic focusing mechanism anchor box is used for quality assessment;

[0023] Fast spatial pyramid pooling layer, modify the SPP module to the SPPELAN module.

[0024] As a specific implementation method, the improved Yolo V8 model training process is as follows:

[0025] Step 1: collect flame data, flame-like data, and non-flame data, and calibrate the data to form model training and test data sets, and use 90% of them as training sets and 10% as test sets;

[0026] Step 2: Build a deep learning network, add ECA attention mechanism, WIOU, SPPFELAN modules and image preprocessing code to the official source code, and change CIOU to WIOU and SPPF module to SPPELAN module;

[0027] Step 3: Create a new model.yaml file, write the modified model network, set the model width and depth parameters, create a new training file, import the modified library file, write the training code and set the hyperparameters to start training;

[0028] Step 4: In each cycle, the images of the training set are mosaic processed with the set batch size and sent to the model for forward propagation, prediction and loss function. The predicted results are then back-propagated to optimize the network weights. This process includes the regression of categories and prediction boxes.

[0029] Step 5: Disable mosaic preprocessing for the last 10 epochs and perform training. When the loss function approaches 0 infinitely and the prediction accuracy is increased by no more than 10%, the training is terminated and the obtained model file is saved.

[0030] Another object of the present invention is to provide an indoor fire detection and automatic fire extinguishing system based on deep learning, comprising a host computer and a slave computer, wherein the host computer comprises:

[0031] Image acquisition module, using binocular camera to collect image data of indoor monitoring area in real time;

[0032] The image detection module performs flame detection on the collected image data based on the trained deep learning network model. If a flame is detected, the flame area is marked in the image with a border, and the pixel position information (x, y) of the flame center is determined according to the center point of the border. The deep learning network model adopts an improved Yolo V8 model.

[0033] The flame positioning module uses the binocular camera positioning mapping method to map the pixel position information of the flame center and obtain the spatial position information (x, y, z) corresponding to the flame center;

[0034] The lower machine includes:

[0035] The automatic fire extinguishing system is used to receive the spatial position information of the flame center from the host computer, and activate the fire extinguisher to extinguish the fire after tracking.

[0036] As a specific implementation, the binocular camera positioning mapping method comprises the following steps:

[0037] For any flame point P in space, the projection on the image plane taken by the first camera C1 is point P1, and the projection on the image plane taken by the second camera C2 is point P2. The epipolar constraint is used to find the corresponding point P2' of P1 in C2. At this time, the stereo matching imaging algorithm is used to obtain the parallax, and the spatial information of point P can be obtained by calculating it with the focal length obtained during camera calibration.

[0038] As a specific implementation, the host computer also includes an alarm module. When a flame is detected, the alarm module alarms through an alarm and sends a text message to a preset responsible person via GPRS, and at the same time sends the spatial location information of the flame to the automatic fire extinguishing system in the slave computer through the network.

[0039] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art:

[0040] 1) Use the improved Yolo V8 deep learning model to implement flame detection, combine the ECA attention mechanism, and modify the CIOU loss function in the model to WIOU and other improvement measures to improve the accuracy of deep learning model in detecting flames;

[0041] 2) Combined with the binocular camera positioning algorithm, the spatial position information corresponding to the center position of the flame can be obtained through mapping, thereby driving the lower computer to track to the specified position, and controlling the lower computer to carry the fire extinguisher to the specified height for automatic fire extinguishing. The present invention realizes a complete indoor fire detection and automatic fire extinguishing system integrating flame detection, alarm and automatic fire extinguishing. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the overall structural diagram of the indoor fire detection and automatic fire extinguishing system based on deep learning described in the present invention;

[0043] Figure 2 It is the structural diagram of the lower computer in the present invention;

[0044] Figure 3 It is a structural schematic diagram of the deep learning network model described in the present invention;

[0045] Figure 4 Schematic diagram of the binocular camera positioning and mapping method in the present invention;

[0046] Among them: 1. Trolley; 11. Fire extinguisher; 12. Threaded rod; 13. Control unit; 14. Various sensors; DETAILED DESCRIPTION

[0047] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, the present invention discloses an indoor fire detection and automatic fire extinguishing system based on deep learning, which mainly includes two parts: an upper computer and a lower computer. The upper computer includes an image acquisition module, an image detection module, an alarm module, and a flame positioning module; the lower computer mainly includes an automatic fire extinguishing module.

[0049] Specifically, the image acquisition module uses a binocular camera to collect image data of the indoor monitoring area in real time; the image detection module implements flame detection on the collected image data based on the pre-trained Yolo V8 network model. If a flame is detected, the flame area is marked in the image with a frame, and the pixel position information (x, y) of the flame center is determined according to the center point of the frame; the flame positioning module obtains the relationship τ between the image coordinates and the spatial coordinates through the binocular camera positioning mapping method, and converts the pixel position information of the flame center determined in the image detection module into the corresponding spatial coordinates through τ, that is, the spatial position information (x, y, z) of the flame is determined; the alarm module, when a flame is detected, reminds the responsible person of the fire through two methods: an alarm and a text message sent through GPRS, and sends the spatial position information of the flame to the lower computer automatic fire extinguishing system through the network; the automatic fire extinguishing system is used to receive the spatial position information of the flame center sent by the upper computer, and reach the flame position after tracking through the trolley 1 with the fire extinguisher 11, adjust the threaded rod 12 on the trolley 1, raise the fire extinguisher 11 to a given height, and start the fire extinguisher 11 to extinguish the fire. The lower machine structure is as follows Figure 2 As shown, it includes a trolley 1 with a fire extinguisher 11, and a control system 13 for starting the fire extinguisher to extinguish a fire is arranged on the trolley. The control system receives signals sent from a host computer through various sensors 14 and controls the lifting and lowering of a threaded rod 12 to adjust the height of the fire extinguisher 11.

[0050] Here, after the image acquisition module acquires the image data, it will pre-process the image data. By analyzing the color and shape of the flame, we can get:

[0051] (1) In the RGB channels of the flame, the pixel value of the R channel is the most prominent, much higher than that of other channels. When converting the solid image to a grayscale image, the weight of the R channel can be increased, and the weight size is R>G>B, so that the flame area and the background can be distinguished; (2) Compared with other interference factors, such as lights and the sun, the illumination of the interference factors tends to be more circular, so it can be judged through mathematical modeling, C = (4*π*S) / L 2 , where C is the circularity, S is the area of ​​the region, and L is the perimeter of the region contour. Experiments have shown that the flame has a circularity between 0.3 and 0.7.

[0052] Here, the binocular camera positioning mapping method includes the following steps:

[0053] See also Figure 4As shown, for any flame point P in space, the projection on the image plane taken by the first camera C1 is point P1, and the projection on the image plane taken by the second camera C2 is point P2. Point e1 and point e2 are the intersection points of the straight line formed by the camera center 1 and the camera center 2 and the imaging surfaces C1 and C2 respectively. The corresponding point P2' of P1 in C2 is found by using the epipolar constraint. Point P2' falls on the straight line formed by e2 and P2. At this time, the stereo matching imaging algorithm is used to obtain the parallax, and the spatial information of point P can be obtained by calculating with the focal length obtained during camera calibration. Here Figure 4 In the figure, the dotted straight lines between point e1 and point P1 and between point e2 and point P2 represent polar lines.

[0054] Here, the deep learning network model uses the improved Yolo V8 model. For details, see Figure 3 As shown in FIG. 1 , the model consists of a Yolo V8 Backbone network, an ECA attention mechanism module, a Neck network, and a Detect detection network. The present invention mainly adds an ECA attention mechanism module based on the original Yolo V8 network, and modifies the original CIOU loss function to WIOU in the Detect detection network.

[0055] The feature extraction layer uses a feature pyramid network, which is built by the CSPDarkNet53 convolutional neural network, and adds the ECA attention mechanism on the basis of CSPDarkNet53;

[0056] The network layer uses PANet to take the features from the 4th, 6th and 10th layers in the feature extraction layer as input, and outputs them to the Head for classification prediction after upsampling and channel fusion; the output network layer uses the decision part of YOLOX's Decoupled Head target detection model as the Head of the Yolo V8 model, separating the regression branch and the prediction branch; the loss function changes the CIOU loss function to WIOU, and uses the dynamic non-monotonic focusing mechanism anchor frame for quality assessment; the fast spatial pyramid pooling layer changes the SPP module to the SPPELAN module.

[0057] By adding the ECA attention mechanism, the local feature detection capability of the model is improved, making it more sensitive to the characteristics of flames; the CIOU loss function of Yolo V8 is modified to WIOU, and the dynamic non-monotonic focusing mechanism anchor frame is used for quality evaluation, which reduces the competitiveness of high-quality anchor frames while reducing the harmful gradients generated by low-quality examples, so that WIOU can focus on the normal quality anchor frames.

[0058] As an embodiment of the present invention, the steps of training and using the deep learning network model are as follows:

[0059] Training process:

[0060] Step S1, collecting flame data and non-flame data, especially collecting a certain amount of flame-like data, such as light and sunlight, and calibrating the data to form model training and test data sets, and using 90% of them as training sets and 10% as test sets;

[0061] Step S2, build a deep learning network, add ECA attention mechanism, WIOU, SPPFELAN modules and image preprocessing code to the official source code, and change CIOU to WIOU, and SPPF module to SPPELAN module;

[0062] Step S3, create a new model.yaml file, write the modified model network, and set the model width and depth parameters, create a new training file, import the modified library file, write the training code and set the hyperparameters to start training;

[0063] Step S4: In each cycle, the images of the training set are mosaic processed with the set batch size and sent to the model for forward propagation to make predictions and obtain the loss function. The predicted results are then back-propagated to optimize the network weights. This process includes the regression of categories and prediction boxes.

[0064] Step S5, in the last 10 epochs, mosaic preprocessing is turned off for training. When the loss function approaches 0 infinitely and the prediction accuracy has not been significantly improved, the training is terminated and the obtained model file is saved.

[0065] Identification and ranging process:

[0066] Step 1: calibrate the binocular camera. Use the left and right cameras to shoot the black and white grid calibration plate images respectively. Use MATLAB's stereoCameraCalibrator or write your own code to calibrate. Delete the images with the largest errors in turn and calibrate until the average error of the remaining images is ≤0.3. Finally, export and save the parameters of the left and right cameras. The error refers to the difference between the corner point coordinates calculated by the black and white grid calibration plate image and the actual corner point coordinates of the black and white grid calibration plate.

[0067] Step 2, collect images through a binocular camera, and correct the left and right images respectively according to the camera parameters of step 1, eliminate distortion, and convert the images into grayscale images, wherein the weight of the R channel is increased, the grayscale images are compared, and the background part of the original image is blurred to improve the recognition of the flame;

[0068] Step 3: Input the corrected image into the model for target detection. In the BackBone part, the model outputs the image into three different scales through convolution and sends them to the Neck part respectively. In the Neck part, features are further extracted and deep information is extracted by upsampling. The images extracted at different depths are input to the Head part for prediction. If flames are detected in the image, the circularity is calculated. If the circularity is within 0.3-0.7, the category and prediction box position information are saved.

[0069] Step 4: If a flame is detected, the alarm function is triggered and a GPRS short message reminder function is sent;

[0070] Step 5: If a flame is detected, the distance measurement function is called at the same time, and the corrected left and right images are subjected to epipolar correction to obtain a mapping table, that is, the corresponding pixel points in the left image in the right image;

[0071] Step 6: During stereo matching, the disparity can be easily obtained through the mapping table. At this time, the disparity map can be obtained and filtered to convert the sparse disparity into dense disparity. Through mathematical modeling, depth=(f*b) / (d+(c r -c l )), where depth is the depth, f is the focal length, b is the baseline length, d is the parallax, and c is the column coordinates of the left and right camera principal points;

[0072] Step 7, using the prediction box information of step 3, average pooling is performed on the framed area to obtain the regional average value, and the pixel point closest to the regional average value is found as the coordinate of the flame center, and the flame center pixel coordinates are converted into three-dimensional coordinates through the reprojection matrix;

[0073] Step 8: Draw the predicted box of the flame, and draw the category probability and distance together;

[0074] Step 9: The lower computer receives the flame spatial position information and drives the car to track to the specified position to realize automatic fire extinguishing.

[0075] Here, considering that the irregular shape of the flame means that the center of the flame area is not necessarily the flame core, the average pixel value of the flame area is calculated using deep learning average pooling, and the most similar point in the flame area is found as the coordinate point, which is the pixel coordinate position of the flame core. This method is more generalized than using the center point of the flame area, and the flame core position information obtained is more accurate.

[0076] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for indoor fire detection and automatic fire extinguishing based on deep learning, characterized in that: The following steps are involved: Step 1: calibrate the binocular camera. Use the left and right cameras to shoot the black and white grid calibration plate images respectively. Use MATLAB's stereoCameraCalibrator or write your own code to calibrate. Delete the images with the largest errors in turn and calibrate until the average error of the remaining images is ≤0.

3. Finally, export and save the parameters of the left and right cameras. The error refers to the difference between the corner point coordinates calculated by the black and white grid calibration plate image and the actual corner point coordinates of the black and white grid calibration plate. Step 2, collect image data of the indoor monitoring area through a binocular camera, and correct the left and right images respectively according to the camera parameters of step 1 to eliminate distortion, convert the image into a grayscale image, increase the weight of the R channel in the RGB channel, and the weight size is R>G>B. Compare the grayscale images, blur the background part of the original image, distinguish the flame area from the background, and improve the recognition of the flame; Step 3, input the image corrected in step 2 into the improved Yolo V8 model for target detection. In the BackBone part, the model outputs the image at different scales through convolution and sends them to the Neck part respectively. In the Neck part, features are further extracted and deep information is extracted by upsampling. The images extracted at different depths are input to the Head part for detection. If flames are detected in the image, the circularity is calculated, and the flame image data with a circularity within 0.3-0.7 is extracted. The flame category and prediction box position information are saved, and the operation of step 4 is entered. Step 4: If a flame is detected, the alarm and reminder functions are triggered, and the distance measurement function is called at the same time to perform epipolar correction on the corrected left and right images to obtain a mapping table, that is, the corresponding pixel points in the left image in the right image; Step 5: During stereo matching, the disparity map is obtained by mapping and filtering, sparse disparity is converted into dense disparity, and a mathematical model depth=(f*b) / (d+(c r -c l )), where depth is depth, f is focal length, b is baseline length, d is parallax, c is r is the column coordinate of the principal point of the right camera, c l is the column coordinate of the principal point of the left camera; Step 6, using the predicted box position information of step 3, average pooling is performed on the boxed area to obtain the regional average value, and the pixel point closest to the regional average value is found as the coordinate of the flame center, and the pixel coordinates of the flame center are converted into three-dimensional coordinates through the reprojection matrix; Step 7: Draw the predicted box of the flame, and draw the category probability and distance together; Step 8: After receiving the flame spatial position information, the lower computer drives the fire extinguishing device to track to the designated position to realize automatic fire extinguishing; The improved Yolo V8 model includes: The feature extraction layer uses a feature pyramid network, which is built by a CSPDarkNet53 convolutional neural network, and adds an ECA attention mechanism based on CSPDarkNet53; In the network layer, PANet is used to take the features from the 4th, 6th, and 10th layers in the feature extraction layer as input, and after upsampling and channel fusion, they are output to the Head for classification prediction; At the output network layer, the decision part of the Decoupled Head target detection model of YOLOX is used as the Head of the Yolo V8 model to separate the regression branch and the prediction branch. Loss function, the CIOU loss function is modified to WIOU, and the dynamic non-monotonic focusing mechanism anchor box is used for quality assessment; Fast spatial pyramid pooling layer, modify the SPP module to the SPPELAN module, The improved Yolo V8 model training process is as follows: Step 1: collect flame data, flame-like data, and non-flame data, and calibrate the data to form model training and test data sets, and use 90% of them as training sets and 10% as test sets; Step 2: Build a deep learning network, add the ECA attention mechanism, WIOU, SPPELAN modules and image preprocessing code to the official source code, and change CIOU to WIOU and SPP module to SPPELAN module; Step 3: Create a new model.yaml file, write the modified model network, set the model width and depth parameters, create a new training file, import the modified library file, write the training code and set the hyperparameters to start training; Step 4: In each cycle, the images of the training set are mosaic processed with the set batch size and sent to the model for forward propagation, prediction and loss function. The predicted results are then back-propagated to optimize the network weights. This step includes the regression of categories and prediction boxes. Step 5: Disable mosaic preprocessing for the last 10 epochs and perform training. When the loss function approaches 0 infinitely and the prediction accuracy is increased by no more than 10%, the training is terminated and the obtained model file is saved.

2. A deep learning-based indoor fire detection and automatic fire extinguishing system, used to execute the deep learning-based indoor fire detection and automatic fire extinguishing method as claimed in claim 1, characterized in that: The system comprises a host computer and a slave computer, wherein the host computer comprises: Image acquisition module, using binocular camera to collect image data of indoor monitoring area in real time; An image detection module performs flame detection on the collected image data based on a trained deep learning network model. If a flame is detected, the flame area is marked in the image with a border, and the pixel position information (x, y) of the flame center is determined based on the center point of the border. The deep learning network model uses an improved Yolo V8 model. The flame positioning module uses the binocular camera positioning mapping method to map the flame center pixel position information to obtain the spatial position information (x, y, z) corresponding to the flame center; The lower machine includes: The automatic fire extinguishing system is used to receive the spatial position information of the flame center from the host computer, and activate the fire extinguisher to extinguish the fire after tracking.

3. According to claim 2, a deep learning-based indoor fire detection and automatic fire extinguishing system is characterized in that: The binocular camera positioning mapping method comprises the following steps: For any flame point P in space, the projection on the image plane taken by the first camera C1 is point P1, and the projection on the image plane taken by the second camera C2 is point P2. The epipolar constraint is used to find the corresponding point P2' of P1 in C2. At this time, the stereo matching imaging algorithm is used to obtain the parallax, and the spatial information of point P can be obtained by calculating it with the focal length obtained during camera calibration.

4. According to claim 2, a deep learning-based indoor fire detection and automatic fire extinguishing system is characterized in that: The host computer also includes an alarm module. When a flame is detected, the alarm module alarms through an alarm and sends a text message to a preset responsible person through GPRS. At the same time, the spatial location information of the flame is sent to the automatic fire extinguishing system in the slave computer through the network.

Images