Nuclear power plant underground pipe gallery abnormal state detection method

Abstract

The invention relates to the technical field of nuclear power plant underground pipe gallery detection, and provides a nuclear power plant underground pipe gallery abnormal state detection method comprising the following steps: establishing an underground pipe gallery normal working state video image data set; target detection and classification are realized according to an image preprocessing method and a YOLO algorithm; integrating the optimized target detection model and the classification model to construct a deep learning model conforming to an underground pipe gallery monitoring scene; training a deep learning model by using the video image data set; constructing an underground pipe gallery anomaly detection system by deploying the trained deep learning model; connecting a camera to obtain a real-time video stream; and performing abnormal state detection on the video stream by using the deployed model. According to the detection method, intelligent monitoring and early warning of the underground pipe gallery area of the nuclear power plant are achieved, the overall safety protection capacity of the nuclear power plant is improved, the safety protection level of the nuclear power plant is improved, potential safety hazards are reduced, and safe and efficient operation of the nuclear power plant is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of underground pipe gallery detection in nuclear power plants, and in particular to a method for detecting abnormal states of underground pipe gallery in nuclear power plants. Background Art

[0002] With the rapid development of my country's nuclear power industry, the number of nuclear power plants is increasing. The safe and stable operation of nuclear power plants is directly related to the safety of people's lives and property and the sustainable development of social economy. Among them, effective monitoring and security measures in various areas within the nuclear power plant are particularly important. The pipeline corridor of a nuclear power plant connects different areas and plays a pivotal role in ensuring the normal flow of personnel, equipment and materials between different areas. Therefore, timely detection of abnormal conditions and situations in the pipeline corridor and real-time monitoring of dynamic changes in the pipeline corridor are the key to ensuring the safe operation of nuclear power plants.

[0003] At present, the pipe gallery area of ​​nuclear power plants is mainly monitored by security personnel. However, as the scale of nuclear power plants continues to expand, manual inspections alone have certain blind spots and cannot achieve all-weather, all-round monitoring, posing a great safety hazard. In order to solve the problem of unattended operation, existing methods mainly rely on dedicated detection sensors, which not only leads to high equipment prices, but also requires frequent maintenance. Moreover, such methods often use non-intelligent single indicators, resulting in the singleness of the detection results and the limitation of indicators.

[0004] With the rapid progress of computer vision, video image processing and other technologies, they have shown great application prospects in the field of intelligent video surveillance. However, traditional visual methods such as density estimation and one-class classification usually require manual design of features, which have limited performance in complex underground pipe gallery environments. In addition, in dealing with complex scenes, lighting changes or occlusions, these methods often face the challenges of high computational complexity and poor generalization ability, which leads to a decrease in the accuracy of abnormal state detection. Summary of the invention

[0005] The purpose of the present invention is to provide a method for detecting abnormal conditions of an underground pipe gallery in a nuclear power plant. The detection method realizes intelligent monitoring and early warning of the underground pipe gallery area of ​​the nuclear power plant, thereby improving the overall safety protection capability of the nuclear power plant and thus improving the security level of the nuclear power plant, reducing safety hazards, and ensuring the safe and efficient operation of the nuclear power plant.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting abnormal state of an underground pipe gallery in a nuclear power plant, comprising:

[0008] Step 1: Establish a video image dataset of the normal working status of the underground pipeline gallery;

[0009] Step 2: Implement target detection and classification based on image preprocessing methods (such as denoising, contrast enhancement) and the YOLO algorithm;

[0010] Step 3: Integrate the optimized target detection model and classification model to build a deep learning model that meets the underground pipeline corridor monitoring scenario;

[0011] Step 4: Train the deep learning model using the video image dataset;

[0012] Step 5: Build an underground pipeline corridor anomaly detection system by deploying the trained deep learning model;

[0013] Step 6: Connect the camera to get real-time video stream;

[0014] Step 7: Use the deployed model to detect abnormal conditions in the video stream.

[0015] In some embodiments, based on the 3D structural design of the underground tunnel, a computer vision algorithm is used to calculate key positions, gun-type cameras and spherical cameras are installed, high-definition video of the underground tunnel area is continuously recorded for a period of time, and a video processing algorithm is used to extract key frames at a frequency of 0.5 seconds per frame to form an image sequence data set.

[0016] In some embodiments, a labeling tool and a conventional convolutional neural network target detection model are used to interactively identify and label target objects in each key frame.

[0017] In some embodiments, step 2 comprises:

[0018] Step 2.1: Remove noise from underground pipeline gallery images based on pixel correlation;

[0019] Step 2.2: Enhance the image quality based on the contrast optimization algorithm of histogram analysis;

[0020] Step 2.3: Construct a light normalization neural network to eliminate the impact of non-uniform light in the underground tunnel area.

[0021] In some embodiments, local pixel correlation of an image is analyzed and denoising is performed using an adaptive non-local denoising model.

[0022] In some embodiments, a correlation analysis algorithm is used to evaluate the contribution of each channel in the target detection model to the loss function, and non-critical channels are dynamically removed to obtain a compressed lightweight detection model.

[0023] In some embodiments, attention modules with learnable weights are added to the high and low layers of the feature pyramid structure to enhance the detection of small objects.

[0024] In some embodiments, a loss function with adaptive weight adjustment is designed to dynamically learn the weights of different tasks, while considering the regression loss of the detection head prediction box parameters and the classification loss of the classification head prediction category to achieve joint optimization of detection and classification.

[0025] In some embodiments, step 4 includes:

[0026] Step 4.1: Based on single-precision calculation, half-precision is used for model forward propagation to reduce video memory consumption;

[0027] Step 4.2: Build a dynamic loss scaling module to automatically adjust the scaling ratio of the loss function according to the gradient change to optimize the numerical stability during mixed precision training. At the same time, in order to reduce GPU memory consumption, set the model to perform forward calculations at half precision (FP16), making the training process more efficient and ensuring the stability and accuracy of the training, which is suitable for a variety of complex scenarios including underground pipe gallery monitoring.

[0028] Step 4.3: Design cyclic learning rate to achieve finer-grained hyperparameter control.

[0029] In some embodiments, a model deployment tool is used to optimize the detection and classification models and save them in a standard format, configure a high-performance GPU server as a system inference server, deploy the model service, and write a calling interface.

[0030] Compared with the prior art, the abnormal state detection method of the underground pipe gallery of a nuclear power plant provided by the present invention has the following beneficial effects:

[0031] This method collects and establishes a video image dataset for monitoring underground pipe corridors in nuclear power plants, studies computer vision and machine learning algorithms applicable to underground pipe corridor scenarios, and designs an algorithm model that can accurately detect abnormal conditions such as water accumulation, rats, and falling down in real time. This method improves the overall safety protection capability of nuclear power plants by realizing intelligent monitoring and early warning of underground pipe corridor areas in nuclear power plants, thereby improving the security level of nuclear power plants, reducing safety hazards, and ensuring the safe and efficient operation of nuclear power plants.

[0032] The present invention uses video image processing technology and deep learning networks to identify and detect monitoring images, which can automatically learn abstract representations in data, avoid relying on manually designed features, and better adapt to different data patterns. Therefore, the development of automatic detection technology for abnormal conditions of underground pipe galleries in nuclear power plants based on deep neural network models not only reduces costs and improves intelligence, but also has stronger anti-interference performance, greatly improving the safety monitoring level of nuclear power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1The present invention provides a flow chart of a method for detecting abnormal conditions of underground pipe galleries in a nuclear power plant. DETAILED DESCRIPTION

[0034] The following is further explained in detail through specific implementation methods.

[0035] like Figure 1 As shown, the present invention provides a method for detecting abnormal state of an underground pipe gallery in a nuclear power plant, comprising:

[0036] Step 1: Establish a video image dataset of the normal working state of the underground pipe gallery. Set up a high-definition camera in the underground pipe gallery area, record long-term video images of the underground pipe gallery in normal working state, remove invalid frames, and construct an image dataset.

[0037] According to the 3D structural design of the underground pipe gallery, image preprocessing methods (such as denoising and contrast enhancement) are used, and key positions are calculated through the YOLO algorithm. 50 gun-type cameras and 5 spherical cameras are installed to ensure full coverage of the key areas of the underground pipe gallery. High-definition videos of the underground pipe gallery area are recorded continuously for one week, including scenes of different working time periods. Key frames are extracted at a frequency of 0.5 seconds per frame using a video encoding and decoding algorithm (video frame extraction) to form an image sequence dataset. Annotation tools and convolutional neural network target detection models are used to interactively identify and annotate target objects in each key frame. Data enhancement processing such as contour drawing, rotation, and noise addition is performed on the identified target object images. A large underground pipe gallery working environment image dataset containing enhanced samples of target objects is generated.

[0038] Step 2: According to the designed image processing algorithm (as described below), realize the functions of target detection, image segmentation, classification, etc. Use convolutional neural network model for image processing to realize target detection and classification.

[0039] The adaptive non-local denoising model of the underground pipe gallery anomaly detection system specifically includes:

[0040] Step 2.1: When calculating the similarity between pixels, the weights are dynamically adjusted according to the local features of the image. For example, in areas with more noise, pixels with lower similarity are given smaller weights, while in areas with rich textures, similar pixels are determined by increasing the weights. In addition, not only the similarity within the neighborhood of the pixel is considered, but also the non-local similarity within the entire image range is used to better capture the global structural information of the image, thereby maintaining the overall consistency of the image while removing noise, thereby effectively removing noise from underground pipeline gallery images.

[0041] Step 2.2: Use the contrast optimization algorithm of histogram analysis to enhance the image quality. Specifically, calculate the image histogram parameters, design an optimization operator that matches the target distribution, enhance the contrast and reduce the overexposed area.

[0042] Step 2.3: Construct a lighting normalization neural network to eliminate the impact of non-uniform lighting in the underground corridor area. The network consists of three core modules: encoding, dimming, and decoding. The encoding module extracts high-level features of the input image to capture structural and detail information; the dimming module, as the core, uses a deep learning algorithm to intelligently analyze and balance the light intensity of each area, especially for the common strong and weak light areas in the corridor. The decoding module fuses the adjusted lighting features with the original image features to reconstruct the normalized image. This process reduces the impact of non-uniform lighting on image quality, improves the accuracy of subsequent processing and analysis, and provides more reliable image data support for the monitoring, maintenance, and management of underground corridors.

[0043] Through this adaptive non-local denoising model, the image quality in the underground pipeline corridor environment can be effectively improved, noise interference can be eliminated, contrast can be adjusted, and non-uniform illumination can be normalized, providing higher quality input images for subsequent detection and classification.

[0044] Step 3: Build a deep learning model that integrates target detection and image classification. Integrate the optimized target detection model and classification model to build a deep learning model that meets the underground pipe gallery monitoring scenario. The optimized target detection module can accurately locate the pipeline interface and equipment status even in the face of complex lighting and occlusion, ensuring high recognition accuracy; the image classification module deepens the understanding of the attributes of objects in the scene, accurately distinguishes various types of equipment and abnormal conditions, and lays a solid foundation for monitoring analysis and emergency response. Based on this integration, the processing efficiency can be improved, the computing overhead can be reduced, and the synergy between models can be enhanced through shared feature representation and joint optimization, making the analysis of underground pipe gallery monitoring scenarios more comprehensive and in-depth.

[0045] In step 3, the correlation analysis algorithm is used to evaluate the contribution of each channel in the target detection model to the loss function, and non-critical channels are dynamically removed to obtain a compressed lightweight detection model. The importance of each channel in the target detection model is calculated, and non-critical channels are dynamically removed to achieve model compression.

[0046] In step 3, attention modules with learnable weights are added to the high and low layers of the feature pyramid structure to strengthen the detection of small targets. An attention module is added between the detection model and the classification model to output the fusion weights of the detection and classification features to strengthen the fusion of the two features. A knowledge distillation framework is constructed to first train the large model to obtain knowledge and transfer it to the small model to improve the effect of the small model. The detection capability of small targets is enhanced by adding attention modules and learnable weights. By introducing a lightweight detection model with an attention mechanism, knowledge transfer between the large model and the small model is achieved.

[0047] In step 3, a joint optimization loss function for integrated detection and classification tasks is provided, and a loss function for adaptive weight adjustment is designed to dynamically learn the weights of different tasks, while considering the regression loss of the detection head prediction box parameters and the classification loss of the classification head prediction category, to achieve joint optimization of detection and classification. It also includes the regression loss of target detection prediction and the cross entropy loss of classification prediction, to achieve joint optimization of detection and classification. This integrated framework can achieve effective fusion learning of detection and classification, and through the collaboration of heterogeneous models and dynamic adjustment of weights, the detection and classification capabilities are improved at the same time, enhancing the generalization performance of the model.

[0048] Step 4: Use the video image dataset to train the deep learning model. Use the image dataset in the previous step and iteratively train the deep learning model using the trainer. Use the underground pipeline corridor anomaly detection model optimization method based on mixed precision training (described below) to iteratively read the video images and process them, including:

[0049] Step 4.1: Based on single-precision calculation, half-precision is used for model forward propagation to reduce video memory consumption; in the hardware environment of 4 1080Ti, the batch size is set to 256 to maximize the parallel computing capability of the GPU.

[0050] Step 4.2: Set up a dynamic loss scaling module to optimize numerical stability under mixed precision; set half-precision FP16 for model forward calculation to reduce GPU memory consumption and support larger batch size; use the dynamic loss scaling module to automatically adjust the loss function according to gradient changes to stabilize mixed precision training. Automatically adjust the scaling ratio of the loss function according to gradient changes to optimize numerical stability during mixed precision training; at the same time, in order to reduce GPU memory consumption, set the model to perform forward calculations under half-precision (FP16), making the training process more efficient, ensuring the stability and accuracy of the training, and suitable for a variety of complex scenarios including underground pipe gallery monitoring.

[0051] Step 4.3: Design a cyclic learning rate to achieve finer-grained hyperparameter control. Use a cyclic learning rate change strategy to perform a learning rate decay and increase cycle every 50 iters to achieve finer-grained control.

[0052] Step 4 makes full use of mixed precision training technology to accelerate the reasoning of the underground pipeline corridor anomaly detection model and conduct a more detailed search in the hyperparameter space to obtain better model results.

[0053] Step 5: Build an underground utility corridor anomaly detection system by deploying the trained deep learning model. Use the deep network deployment tool to deploy the trained deep learning model to the detection system server.

[0054] Use the model deployment tool to optimize the detection and classification models and save them in a standard format. Configure a high-performance GPU server as the system inference server, deploy the model service, and write the calling interface.

[0055] Step 6: Connect the camera to obtain real-time video stream. Connect the camera in the underground corridor area through the communication protocol to obtain real-time video stream.

[0056] Select high-definition network cameras that support standard video stream output protocols such as RTSP and ONVIF. The detection system connects to the cameras in the underground corridor area through the RTSP protocol and calls the interface to obtain real-time video streams.

[0057] Step 7: Use the deployed model (the above parts) to detect abnormal conditions in the video stream. Perform model inference on the collected video stream frame images to detect abnormal conditions.

[0058] After the detection system obtains the video stream frame, it calls the model service interface, inputs each frame image separately and obtains the detection box and classification result. It determines whether it is abnormal based on the classification result, and triggers the system alarm once the preset abnormal category is detected.

[0059] In summary, the video image processing algorithm provided by the present invention can accurately detect various abnormal conditions including water accumulation, rats, and falling down in the underground pipe gallery area; based on the deep learning model, it can distinguish between normal and abnormal conditions and realize automatic identification of abnormal conditions; the constructed underground pipe gallery abnormality detection system realizes all-weather and all-round monitoring of the underground pipe gallery area, and always grasps the dynamics in the underground pipe gallery; once an abnormal state is detected, an early warning can be quickly issued to improve the event response speed; by adopting the online learning algorithm, the system can be continuously optimized to adapt to the changes in the working environment of the underground pipe gallery; the security capability and safety management level of the underground pipe gallery of the nuclear power plant are significantly improved, as follows:

[0060] Accurate anomaly detection in underground utility corridors integrates advanced target detection and classification algorithms to achieve more accurate detection of abnormal conditions in underground utility corridors. It provides solid data support for model development by collecting and constructing specific image and video datasets of utility corridor scenes.

[0061] Deep network optimization for real-time detection capabilities ensures that the system can meet the requirements of real-time detection through the design and optimization of the deep network structure; model pruning and mixed precision training are performed to further improve the performance and efficiency of the model.

[0062] The image processing and model training scheme for underground pipe corridors in nuclear power plants proposes image processing and model training technical methods for underground pipe corridors in nuclear power plants; and designs preprocessing algorithms for image enhancement and illumination normalization to adapt to the complex illumination and other environmental conditions of the pipe corridors.

[0063] The reliability and stability of the underground pipeline corridor anomaly detection system are guaranteed, which realizes the reliability and stability of the detection system and meets the high standards for safe operation of nuclear power plants. Through a series of technical innovations and optimization measures, the stable operation of the detection system in complex environments is ensured.

[0064] The improvement of generalization ability and environmental adaptability has improved the generalization ability and environmental adaptability of the underground pipeline corridor anomaly detection system, enabling it to be more widely used in underground pipeline corridor detection tasks in different scenarios; by enhancing the generalization ability of the model, the need to retrain the model in different environments is reduced, and the practicality and convenience of the detection system are improved.

[0065] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for detecting abnormal state of underground pipe gallery in a nuclear power plant, characterized in that: include: Step 1: Establish a video image dataset of the normal working status of the underground pipeline gallery; Step 2: Perform image processing based on image processing and YOLO algorithm to achieve target detection and classification; Step 3: Integrate the optimized target detection model and classification model to build a deep learning model that meets the underground pipeline corridor monitoring scenario; Step 4: Train the deep learning model using the video image dataset; Step 5: Build an underground pipeline corridor anomaly detection system by deploying the trained deep learning model; Step 6: Connect the camera to get real-time video stream; Step 7: Use the deployed model to detect abnormal conditions in the video stream.

2. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: In step 1, according to the 3D structural design of the underground tunnel, a computer vision algorithm is used to calculate the key positions, and a gun-type camera and a spherical camera are installed to continuously record high-definition video of the underground tunnel area for a period of time. The video processing algorithm is used to extract key frames at a frequency of 0.5 seconds per frame to form an image sequence dataset.

3. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 2, characterized in that: The target objects in each keyframe are interactively identified and labeled using annotation tools and conventional convolutional neural network target detection models.

4. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: Step 2 includes: Step 2.1: Remove noise from underground pipeline gallery images based on pixel correlation; Step 2.2: Enhance the image quality based on the contrast optimization algorithm of histogram analysis; Step 2.3: Construct a light normalization neural network to eliminate the impact of non-uniform light in the underground tunnel area.

5. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 4, characterized in that: In step 2.1, the local pixel correlation of the image is analyzed and denoising is performed using an adaptive non-local denoising model.

6. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: In step 3, the correlation analysis algorithm is used to evaluate the contribution of each channel in the target detection model to the loss function, and non-critical channels are dynamically removed to obtain a compressed lightweight detection model.

7. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: In step 3, attention modules with learnable weights are added to the high and low layers of the feature pyramid structure to enhance the detection of small objects.

8. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: In step 3, a loss function with adaptive weight adjustment is designed to dynamically learn the weights of different tasks. The regression loss of the detection head prediction box parameters and the classification loss of the classification head prediction category are considered at the same time to achieve joint optimization of detection and classification.

9. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: Step 4 includes: Step 4.1: Based on single-precision calculation, half-precision is used for model forward propagation to reduce video memory consumption; Step 4.2: Set up a dynamic loss scaling module to automatically adjust the scaling of the loss function according to the gradient changes to optimize the numerical stability during mixed precision training; set the model to perform forward calculations at half precision; Step 4.3: Design cyclic learning rate to achieve finer-grained hyperparameter control.

10. The method for detecting abnormal state of underground pipe gallery in a nuclear power plant according to claim 1, characterized in that: In step 5, use the model deployment tool to optimize the detection and classification models and save them in a standard format, configure a high-performance GPU server as the system inference server, deploy the model service, and write a calling interface.

Images