Transformer fire stage identification method based on image identification

Abstract

The invention relates to the technical field of image recognition, and particularly discloses a transformer fire stage recognition method based on image recognition. Comprising the following steps of 1, collecting real-time image data of a transformer through a monitoring camera, and monitoring the operation state of the transformer; 2, preprocessing the image data, including noise removal, brightness normalization and target area extraction, so as to improve the accuracy of subsequent recognition; step 3, extracting depth features of the image based on a convolutional neural network, extracting feature parameters such as color distribution, smoke form, flame texture and the like, and performing feature fusion to form a multi-modal feature vector; according to the technical scheme, fine recognition of the fire stages is achieved, the transformer fire development process is divided into six stages by introducing a deep learning model and an image recognition algorithm, fine-grained recognition of the whole fire process is achieved, and the bottleneck that the fire stages are difficult to distinguish through a traditional monitoring means is broken through.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for identifying the stage of a transformer fire based on image recognition. Background Technology

[0002] As core equipment in the power system, transformers play a crucial role in voltage transformation and power transmission, and their operational safety directly affects the continuity and reliability of power supply. A fire caused by a transformer malfunction can not only lead to equipment damage and power grid paralysis, but also cause enormous economic losses and social impact. Therefore, achieving early detection, accurate identification, and timely response to transformer fire hazards is a critical issue in current power safety operation and maintenance.

[0003] Currently, transformer fire monitoring mainly relies on traditional sensors such as those for temperature, smoke, and gas. While these methods can play a role, they suffer from problems such as slow response times, high false alarm and false negative rates, and difficulty in reflecting the dynamic evolution of a fire, especially in the early or transitional stages. Furthermore, existing monitoring methods lack the ability to precisely differentiate between different stages of fire development, thus failing to support effective tiered response and emergency decision-making.

[0004] With the development of artificial intelligence and computer vision technology, fire detection based on image recognition has become a new direction for development. By acquiring image data through devices such as high-definition cameras and infrared thermal imaging, and combining it with deep learning to analyze features such as flames, smoke, and temperature changes, non-contact, multi-angle, and all-weather fire monitoring can be achieved. It is superior to traditional methods in terms of response speed and perception accuracy, and has stronger environmental adaptability.

[0005] However, current research largely focuses on scenarios such as forests and buildings, lacking dedicated models for the specific environment of transformers. Especially in fine-grained identification of fire stages, a mature system has not yet been established, making it difficult to accurately classify and judge the entire fire process from its initial stage to deflagration. Limited by the complex operating environment of transformers and significant background interference, existing image recognition algorithms face problems such as low accuracy, poor generalization, and insufficient real-time performance in practical applications.

[0006] Therefore, there is an urgent need to develop an intelligent image recognition method that is geared towards transformer scenarios, integrates multimodal information, and has the ability to identify fire stages, so as to improve the monitoring and emergency response capabilities of transformer fires and provide support for the intelligent power operation and maintenance system.

[0007] Therefore, the aforementioned technical problems need to be solved. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention proposes a transformer fire stage identification method based on image recognition.

[0009] To solve the above-mentioned technical problems, the basic technical solution proposed by this invention is as follows: A transformer fire stage identification method based on image recognition includes the following steps: Step 1: Collect real-time image data of the transformer through surveillance cameras to monitor the transformer's operating status; Step 2: Preprocess the image data, including noise removal, brightness normalization, and target region extraction, to improve the accuracy of subsequent recognition; Step 3: Extract depth features of the image based on convolutional neural network, and extract feature parameters such as color distribution, smoke pattern, and flame texture, and perform feature fusion to form a multimodal feature vector; Step 4: Use a multi-classification model to classify the multimodal feature vectors and output the corresponding fire stage categories, including: normal, overheating, initial smoke, dense smoke, open flame, deflagration or intense combustion; Step 5: Based on the fire stage category, trigger alarm commands of different levels, and perform cross-verification by combining transformer and sensor data to reduce the false alarm rate.

[0010] Preferably, the monitoring camera mentioned in step one includes a high-definition camera and an infrared thermal imager, which can collect image data of the transformer from different angles and in different bands, and realize data synchronization through timestamps.

[0011] Preferably, the noise removal in step two employs a bilateral filtering algorithm, which preserves edge information while removing high-frequency noise; the brightness normalization uses histogram equalization to enhance image details in low-light environments; and the target region extraction uses a pre-trained model to extract the transformer body region, reducing background interference.

[0012] Preferably, the convolutional neural network in step three includes YOLOv8 and EfficientNet-B3. YOLOv8 is used to detect flame or smoke feature regions and crop out the region of interest, while EfficientNet-B3 is used to extract multi-level features of the region of interest.

[0013] Preferably, the multi-classification model in step four includes either multi-classification ResNet or lightweight EfficientNet, and the criteria for determining the fire stage category include: Normal state: No obvious texture or color abnormalities; Overheating state: Localized high temperature in the infrared enhanced image, but no smoke; Initial smoke: Changes in image grayscale distribution and sparse texture; Dense smoke: The overall image is dark with dense texture; Open flame combustion: a large area of ​​high saturation and rapid change; Deflagration or violent combustion: The flame area expands dramatically and exhibits obvious dynamic characteristics.

[0014] Preferably, the different levels of alarm commands mentioned in step five include: Normal state: Records the normal state, with no alarms; Overheating status: Push overheating warning to the operation and maintenance terminal to remind the user to pay attention to the load; Initial smoke: Activate the primary audible and visual alarm; on-site inspection is recommended. Dense smoke: Activate the intermediate-level audible and visual alarm and coordinate with the fire emergency plan. Open flame combustion: Activates advanced audible and visual alarm, triggers automatic fire extinguishing system, and remotely pushes alarm information; In case of deflagration or violent combustion: cut off the power supply to the transformer, activate all fire extinguishing equipment, and trigger the emergency response mechanism.

[0015] Preferably, the sensor data in step five includes temperature sensor data and gas sensor data. The temperature sensor has a measurement range of -50℃ to 300℃ and an accuracy of ±1℃; the gas sensor detects CO concentrations of 0-1000ppm and has an accuracy of ±5ppm.

[0016] Preferably, the model training and optimization steps are also included: collecting images of transformer fire simulation experiments and on-site operation to construct a dataset, using random rotation, random scaling, adding Gaussian noise, color perturbation and simulated smoke / flame overlay to perform data augmentation, training the model through the Adam optimizer, and selecting the optimal model based on the F1 score of the validation set.

[0017] Preferably, the method can be deployed on an edge computing platform to achieve millisecond-level fast inference response, and can be interfaced with a substation automation monitoring system to achieve data sharing and coordinated control.

[0018] Preferably, the area of ​​the region of interest extracted from the target region in step two is not less than 60% of the original image, and the smoke density extracted in step three is based on the gray-level co-occurrence matrix feature of texture change, and the flame outline dynamics are based on the outline change rate of Canny edge detection.

[0019] The beneficial effects of this invention are: The technical solution of this invention enables refined identification of fire stages: by introducing deep learning models and image recognition algorithms, the development process of transformer fires is divided into six stages, achieving fine-grained identification of the entire fire process and breaking through the bottleneck of traditional monitoring methods that are difficult to distinguish fire stages.

[0020] Employing advanced neural network models such as YOLOv8 and EfficientNet, combined with image preprocessing and feature fusion technologies, it effectively improves the accuracy of fire feature extraction and classification; it can be deployed on edge computing platforms to achieve millisecond-level fast inference response, meeting the needs of real-time on-site monitoring. Attached Figure Description

[0021] Figure 1 This is a flowchart of the transformer fire stage identification method based on image recognition according to the present invention; Figure 2 This is a system block diagram of the transformer fire stage identification method based on image recognition according to the present invention. Detailed Implementation

[0022] The following will be combined with the appendix Figure 1 To be continued Figure 2 The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] A transformer fire stage identification method based on image recognition includes the following modules and steps: Image acquisition module Surveillance cameras (including high-definition cameras, infrared thermal imagers, etc.) are installed around the transformer area to collect on-site image data at regular intervals and synchronize them according to timestamps.

[0024] Image preprocessing module The acquired images are processed, including noise removal (using a bilateral filtering algorithm to remove high-frequency noise while preserving edge information), brightness normalization (applying histogram equalization to enhance image details in low-light environments), and target region extraction (extracting the transformer body region through a pre-trained model to reduce background interference), in order to improve the accuracy of subsequent recognition.

[0025] Feature extraction module The image depth features are extracted using a convolutional neural network (CNN); simultaneously, feature parameters such as color distribution, smoke morphology, and flame texture are extracted; and feature fusion is performed to form a multimodal feature vector.

[0026] Stage Identification Module A multi-classification model (such as multi-class ResNet or lightweight EfficientNet) is used to classify the feature vectors; the output fire stage category is then defined, including: Phase 1: Normal state (no obvious texture or color abnormalities); Phase 2: Overheating state (localized high temperature in infrared enhanced image, but no smoke); Stage 3: Initial smoke (changes in image grayscale distribution, sparse texture); Phase 4: Dense smoke (overall image grayscale is dark, with dense texture); Stage 5: Open flame combustion (large area of ​​high saturation and rapid change); Stage 6: Deflagration or violent combustion (the flame area expands dramatically and exhibits obvious dynamic characteristics).

[0027] Decision and Alarm Module Based on the identified fire stage, different levels of alarm commands are triggered; cross-validation is performed by combining data from other transformer sensors (such as temperature, humidity, and gas detection) to further reduce the false alarm rate.

[0028] The specific implementation method is as follows: System Deployment and Image Acquisition High-definition industrial cameras (with automatic exposure and night vision functions) and infrared thermal imagers are deployed around the transformer area to collect on-site image data at regular intervals (25fps), and data synchronization between multiple devices is achieved through timestamps.

[0029] Image preprocessing The acquired raw images are processed according to the following steps: Noise removal: A bilateral filtering algorithm is used, with the spatial Gaussian standard deviation set to 10 and the grayscale value Gaussian standard deviation set to 0.1, to remove high-frequency noise while preserving edge information; Brightness normalization: Apply adaptive histogram equalization to low-light images, limit contrast gain to 2.0, and enhance image details; Target region extraction: The transformer body is detected by a pre-trained YOLOv8 model, and the region of interest (ROI) is cropped to reduce background interference. The area of ​​the ROI is no less than 60% of the original image.

[0030] Feature extraction and stage recognition Preliminary detection: The YOLOv8 model was used to detect flame or smoke feature regions in the image, with a confidence threshold of 0.7. The detected regions were then cropped into their respective regions of interest (ROI). Deep feature extraction: The ROI is input into the EfficientNet-B3 network to extract multi-level features, including: smoke density (gray-level co-occurrence matrix features based on texture changes), flame contour dynamics (contour change rate based on Canny edge detection), color distribution (HSV channel statistics of high saturation areas), and infrared temperature features (temperature mean and range of local high-temperature areas). Stage classification: Multi-dimensional features are fused through a fully connected layer, and the softmax function is used to output classification probabilities. The category with the highest probability is selected as the fire stage. Phase 1 (Normal): No flame / smoke features, no abnormalities in texture and color, and a classification probability ≥ 0.9; Stage 2 (Overheating): Local high-temperature areas (temperature > 120℃) exist in the infrared image, with no smoke features, and a classification probability ≥ 0.85; Stage 3 (initial smoke): grayscale distribution standard deviation > 30, sparse texture (entropy value < 5), no flame features, classification probability ≥ 0.8; Stage 4 (Dense Smoke): The overall grayscale mean of the image is <80, the texture is dense (entropy value ≥7), there are no flame features, and the classification probability is ≥0.8; Phase 5 (Open Flame): High saturation areas (saturation > 0.7) account for 10%-50% of the ROI, the contour change rate is > 0.3 / s, and the classification probability is ≥ 0.85; Stage 6 (Deflagration): The area of ​​the flame region expands by more than 50% within 1 second, the outline change rate is >0.8 / s, and the classification probability is ≥0.9.

[0031] Assisted judgment and alarm linkage Auxiliary judgment: Combining data from temperature sensor (measurement range -50℃ to 300℃, accuracy ±1℃) and gas sensor (detection of CO concentration 0-1000ppm, accuracy ±5ppm), the reliability of the result is improved when the image recognition result is consistent with the trend of sensor data (e.g., stage 2 corresponds to temperature >120℃, stage 5 corresponds to CO concentration >50ppm); if the data are contradictory, a verification mechanism is triggered (re-acquire the image and perform secondary recognition).

[0032] Alarm linkage: Different responses are triggered depending on the stage of the fire. Phase 1: Record normal status, no alarms; Phase 2: Push overheat warnings to the operation and maintenance terminal to remind users to pay attention to the load; Phase 3: Activate the primary audible and visual alarm (volume 80dB, flashing frequency 2Hz), on-site inspection is recommended; Phase 4: Activate the intermediate-level audible and visual alarm (volume 100dB, flashing frequency 5Hz) and prepare the fire emergency plan accordingly; Phase 5: Activate advanced audible and visual alarms, trigger automatic fire extinguishing systems (such as dry powder fire extinguishers), and remotely push alarm information to the operation and maintenance center; Phase 6: Disconnect the transformer power supply, activate all fire extinguishing equipment, and trigger the emergency response mechanism.

[0033] Model training and optimization Dataset Construction: Images of transformer fire simulation experiments and on-site operation were collected, including samples under different lighting (day / night) and weather (sunny / rainy / foggy) conditions, totaling 100,000 images, which were divided into training set, validation set and test set in an 8:1:1 ratio; Data augmentation: The dataset was augmented using random rotation (±30°), random scaling (0.8x-1.2x), Gaussian noise (σ=0.01), color perturbation (brightness ±20%, saturation ±15%), and simulated smoke / flame overlay. Model training: The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and 50 training rounds. The optimal model was selected based on the F1 score on the validation set (F1 ≥ 0.92).

[0034] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A method for identifying the stage of a transformer fire based on image recognition, characterized in that, Includes the following steps: Step 1: Collect real-time image data of the transformer through surveillance cameras to monitor the transformer's operating status; Step 2: Preprocess the image data, including noise removal, brightness normalization, and target region extraction, to improve the accuracy of subsequent recognition; Step 3: Extract depth features of the image based on convolutional neural network, and extract feature parameters such as color distribution, smoke pattern, and flame texture, and perform feature fusion to form a multimodal feature vector; Step 4: Use a multi-classification model to classify the multimodal feature vectors and output the corresponding fire stage categories, including: normal, overheating, initial smoke, dense smoke, open flame, deflagration or intense combustion; Step 5: Based on the fire stage category, trigger alarm commands of different levels, and perform cross-verification by combining transformer and sensor data to reduce the false alarm rate.

2. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The monitoring camera mentioned in step one includes a high-definition camera and an infrared thermal imager, which can collect image data of the transformer from different angles and in different bands, and realize data synchronization through timestamps.

3. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The noise removal in step two uses a bilateral filtering algorithm to remove high-frequency noise while preserving edge information; the brightness normalization uses histogram equalization to enhance image details in low-light environments; and the target region extraction uses a pre-trained model to extract the transformer body region to reduce background interference.

4. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The convolutional neural network mentioned in step three includes YOLOv8 and EfficientNet-B3. YOLOv8 is used to detect the feature regions of flames or smoke and crop out the regions of interest, while EfficientNet-B3 is used to extract multi-level features of the regions of interest.

5. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The multi-classification model mentioned in step four includes either multi-classification ResNet or lightweight EfficientNet, and the criteria for determining the fire stage category include: Normal state: No obvious texture or color abnormalities; Overheating state: Localized high temperature in the infrared enhanced image, but no smoke; Initial smoke: Changes in image grayscale distribution and sparse texture; Dense smoke: The overall image is dark with dense texture; Open flame combustion: a large area of ​​high saturation and rapid change; Deflagration or violent combustion: The flame area expands dramatically and exhibits obvious dynamic characteristics.

6. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The different levels of alarm commands mentioned in step five include: Normal state: Records the normal state, with no alarms; Overheating status: Push overheating warning to the operation and maintenance terminal to remind the user to pay attention to the load; Initial smoke: Activate the primary audible and visual alarm; on-site inspection is recommended. Dense smoke: Activate the intermediate-level audible and visual alarm and coordinate with the fire emergency plan. Open flame combustion: Activates advanced audible and visual alarm, triggers automatic fire extinguishing system, and remotely pushes alarm information; In case of deflagration or violent combustion: cut off the power supply to the transformer, activate all fire extinguishing equipment, and trigger the emergency response mechanism.

7. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The sensor data in step five includes temperature sensor data and gas sensor data. The temperature sensor has a measurement range of -50℃ to 300℃ and an accuracy of ±1℃. The gas sensor detects CO concentration from 0 to 1000ppm and has an accuracy of ±5ppm.

8. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, It also includes model training and optimization steps: collecting images of transformer fire simulation experiments and on-site operation to build a dataset, using random rotation, random scaling, adding Gaussian noise, color perturbation and simulated smoke / flame overlay to perform data augmentation, training the model through the Adam optimizer, and selecting the optimal model based on the F1 score of the validation set.

9. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, The method can be deployed on an edge computing platform to achieve millisecond-level fast inference response, and can be interfaced with a substation automation monitoring system to achieve data sharing and coordinated control.

10. The transformer fire stage identification method based on image recognition according to claim 1, characterized in that, In step two, the area of ​​the region of interest extracted from the target region is not less than 60% of the original image. In step three, the smoke density extracted is based on the gray-level co-occurrence matrix feature of texture change, and the flame outline dynamics are based on the outline change rate of Canny edge detection.

Images