Hydraulic power plant medium leakage identification method and system

Abstract

The invention provides a hydraulic power plant medium leakage identification method and system, and relates to the technical field of hydropower leakage identification, and the method comprises the steps: deploying a high-definition video monitoring device and an operation state sensor module, carrying out the synchronous registration, configuring infrared channel enhancement and polarized light interference suppression, and carrying out the spatial filtering noise reduction and adaptive adjustment; and configuring a multi-modal enhanced input tensor. High-definition video monitoring equipment and an operation state sensor module are deployed in an operation area of a hydraulic power plant, and technologies of infrared channel enhancement, polarized light interference suppression, spatial filtering noise reduction and the like are combined, so that a multi-mode enhanced input tensor is constructed; according to the method, a clear leakage area identification effect can still be kept in a night low-illumination environment, water vapor and fog interference and metal reflection conditions, so that the problem that the identification precision is reduced due to the fact that the prior art depends on single-mode input is effectively solved, and stable distinguishing of oil medium leakage and water medium leakage is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water and electricity leakage identification. Specifically, it relates to a water and electricity plant medium leakage identification method and system. BACKGROUND

[0002] During long-term high-load operation of a water and electricity plant, oil and water media may leak at equipment pipelines, valves and key operating parts. Traditional leakage monitoring methods rely on manual inspection, single sensor monitoring or conventional camera image acquisition. These methods have obvious limitations in complex working conditions, such as low-illumination environments at night, water vapor and fog interference, and strong reflection effects on metal surfaces of equipment, which may result in the inability to accurately identify the leakage area. In addition, the real-time determination of the leakage type and degree is insufficient, and it is often difficult to support the maintenance personnel to take timely joint measures, which poses a great risk to the operation safety.

[0003] However, the existing leakage monitoring technology has two shortcomings. First, the identification of the leakage area relies on single modal input, which may result in a decrease in identification accuracy in a strong interference environment and the inability to stably distinguish between oil and water media. Second, the determination of the leakage degree is based on experience or a single threshold, lacks fine quantitative indicators and grading mechanisms, and is difficult to achieve scientific early warning triggering and differentiated response. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a water and electricity plant medium leakage identification method and system. By introducing a multi-modal enhanced input tensor and a deep neural network model, the present application achieves high-precision identification of the leakage area in complex working conditions and can distinguish between oil and water media. In addition, by configuring quantitative indicators such as area growth rate and shape factor and combining a grading threshold system, the present application can grade and determine the leakage degree in real time, significantly improving the accuracy of monitoring and the timeliness of response, thereby effectively improving the safety assurance capability and the level of automation management during the operation of the water and electricity plant.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions: A water and electricity plant medium leakage identification method and system, comprising the following steps: Step S100, deploy high-definition video monitoring equipment and operating state sensor modules, and perform synchronous registration. Configure infrared channel enhancement and polarization light interference suppression, and perform spatial filtering noise reduction and adaptive adjustment. Configure a multi-modal enhanced input tensor.

[0006] Step S200, configure a deep neural network model, and introduce a time consistency constraint mechanism, configure a segmentation mask output and a geometric boundary parameter output structure, and calculate a liquid leakage area parameter and a shape factor parameter, and configure a leakage area state parameter set.

[0007] Step S300, configure a leakage type discrimination deep neural network model, and introduce a liquid leakage medium category label, configure an optical texture feature vector, and extract an environmental interaction feature vector of the liquid leakage area, fuse the optical texture feature vector and the environmental interaction feature vector, and perform liquid leakage medium category discrimination, output the liquid leakage medium category label, and configure a medium category discrimination confidence parameter, and configure a liquid leakage medium label result set.

[0008] Step S400, statistics liquid leakage area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold, output leakage degree grade label, and trigger grading early warning, set cache leakage degree grading result, and configure leakage degree quantization and grading early warning result set.

[0009] Step S500, configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, and perform pre-training parameter transfer, introduce an incremental learning mechanism, and configure an online learning strategy and execute model parameter update.

[0010] Step S600, configure a remote alarm and information push channel, configure a linkage control logic unit, and trigger automatic execution, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

[0011] As a preferred scheme of the present application, the step S100, specifically: Step S100.1, deploy high-definition video monitoring equipment and running state sensor modules, and perform synchronous registration.

[0012] Step S100.2, configure infrared channel enhancement and polarized light interference suppression, and perform spatial filtering noise reduction and adaptive adjustment, and configure a multi-modal enhanced input tensor.

[0013] When the brightness of the video image collected by the high-definition video monitoring equipment is lower than the set brightness threshold 200 Lux, the water and power plant medium leakage identification system automatically enables the infrared imaging module, the infrared imaging module uses an infrared image sensor with a 5-12 micron waveband to obtain infrared image data, performs thermal radiation enhancement collection on the monitoring area, and extracts the thermal anomaly distribution characteristics of the potential liquid leakage area.

[0014] In the daytime strong light irradiation environment, if the video image exists the mirror surface reflection interference generated by the metal equipment surface and the ground water, the water and electricity plant medium leakage identification system implements the polarization light interference suppression on the video image, obtains the linear polarization image by using the polarization filter, and reduces the reflection interference by the polarization removal algorithm.

[0015] When the video image is disturbed by water mist, airflow or dust and appears noise, the system implements spatial filtering denoising on the video image, the denoising method adopts Gaussian filtering, the filtering radius is set to 3 pixels, the standard deviation is set to 1.2, the image background is smoothed, and the integrity of the liquid leakage edge texture structure is maintained.

[0016] In the case that the overall gray scale distribution of the video image is uneven, the system implements adaptive adjustment of brightness and contrast on the video image, adopts a local histogram equalization algorithm to improve the gray scale difference between the liquid leakage area and the background area, and the adjustment target is to maintain the global contrast of the image in the gray scale variance ratio interval of 1.5 to 2.5.

[0017] After completing the infrared channel enhancement, polarization light interference suppression, spatial filtering denoising and adaptive adjustment of brightness and contrast of the video image, the enhanced video image frame, the oil pressure change parameter collected by the oil pressure sensor, the environmental temperature change parameter collected by the temperature sensor and the air humidity change parameter collected by the humidity sensor are packaged into a multi-modal enhanced input tensor.

[0018] The multi-modal enhanced input tensor includes: image pixel matrix dimension, oil pressure change parameter dimension, environmental temperature change parameter dimension, air humidity change parameter dimension and timestamp dimension, and is encoded into a unified format data input into the leakage area intelligent positioning deep neural network model.

[0019] As a preferred scheme of the present application, the step S200 specifically comprises: Step S200.1, configuring a deep neural network model and introducing a time sequence consistency constraint mechanism.

[0020] Step S200.2, configuring a segmentation mask output and a geometric boundary parameter output structure, and calculating a liquid leakage area parameter and a shape factor parameter, and configuring a leakage area state parameter set.

[0021] The leakage area intelligent positioning deep neural network model outputs, when processing each frame of multi-modal enhanced input tensor, includes: a pixel-level segmentation mask image and a geometric boundary parameter of the liquid leakage area.

[0022] The pixel-level segmentation mask image is a binary image, the part with a pixel value of 1 represents the liquid leakage area, and the part with a pixel value of 0 represents a non-leakage area, and the geometric boundary parameter includes the center coordinates of the minimum circumscribed rectangle, the rectangle width, the rectangle height and the rectangle rotation angle.

[0023] After obtaining the pixel-level segmentation mask image, the liquid leakage area parameter and the liquid leakage area shape factor parameter are calculated based on the image binary analysis method, the liquid leakage area parameter is calculated by pixel counting method, and is converted into actual area with unit of cm² by combining video image resolution, and the liquid leakage area shape factor parameter is used to describe the compactness of the area.

[0024] Based on the pixel-level segmentation mask image, the geometric boundary parameter, the liquid leakage area parameter and the liquid leakage area shape factor parameter, a leakage area state parameter set is constructed, and the leakage area state parameter set includes: liquid leakage area mask image data, liquid leakage area bounding box center coordinate data, liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area bounding box rotation angle parameter, liquid leakage area area parameter and liquid leakage area shape factor parameter.

[0025] After the inference of the water and power plant medium leakage identification system is completed, the leakage area state parameter set is cached to the real-time monitoring task data buffer unit, and is sorted in the order of video image timestamp, to generate a structured output data stream.

[0026] As a preferred scheme of the present application, the step S300 specifically comprises: Step S300.1, configuring a leakage type discrimination deep neural network model, and deriving a liquid leakage medium category label, configuring an optical texture feature vector, and extracting an environmental interaction feature vector of the liquid leakage area.

[0027] Step S300.2, fusing the optical texture feature vector and the environmental interaction feature vector, and performing liquid leakage medium category discrimination, outputting the liquid leakage medium category label, configuring a medium category discrimination confidence parameter, and configuring a liquid leakage medium label result set.

[0028] The optical texture feature vector and the environmental interaction feature vector are spliced into a fused feature vector, and are input into a fusion decision subnetwork for liquid leakage medium category discrimination, and the fusion decision subnetwork outputs an oil medium prediction probability and a water medium prediction probability.

[0029] The leakage type discrimination deep neural network model outputs the medium category discrimination confidence parameter while outputting the liquid leakage medium category label, the medium category discrimination confidence parameter is composed of the maximum prediction probability value, and the value range is 0-1, representing the credibility of the liquid leakage area medium category discrimination result, when the medium category discrimination confidence parameter is lower than 0.6, the water and power plant medium leakage identification system automatically triggers a candidate area review strategy, marks the liquid leakage area as a to-be-confirmed category and outputs a warning prompt, for review and confirmation by the operation and maintenance personnel or the artificial intelligence labeling module.

[0030] Based on the liquid leakage medium category label, the medium category discrimination confidence parameter and the liquid leakage area shape factor parameter, a liquid leakage medium label result set is constructed, which includes liquid leakage area mask map data, liquid leakage area bounding box center coordinate data, liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area area parameter, liquid leakage area shape factor parameter, liquid leakage medium category label and medium category discrimination confidence parameter. The water power plant medium leakage identification system outputs the liquid leakage medium label result set to the leakage degree grading module and the multi-stage linkage response module, forming a logically closed loop leakage identification and classification process.

[0031] As a preferred scheme of the present application, the step S400 specifically includes: Step S400.1, statistics liquid leakage area area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold.

[0032] Step S400.2, output leakage degree grade label, trigger grading early warning, set cache leakage degree grading result, and configure leakage degree quantization and grading early warning result set.

[0033] As a preferred scheme of the present application, the step S500 specifically includes: Step S500.1, configure historical sample library and artificial correction result set, configure transfer learning mechanism, and perform pre-training parameter transfer.

[0034] Step S500.2, introduce incremental learning mechanism, configure online learning strategy and perform model parameter update.

[0035] In the running process of the water power plant medium leakage identification system, when the liquid leakage medium label result set generated in the monitoring task data buffer unit and the artificial correction result set are different, and the liquid leakage medium category label misjudgment rate exceeds the 3% threshold, the system automatically triggers the incremental learning mechanism. The incremental learning mechanism first supplements the newly collected liquid leakage area mask map data, liquid leakage area bounding box parameter and liquid leakage area area parameter to the training sample set, and performs small batch online training on the supplemented training sample set. The number of samples in each batch is set to 64.

[0036] After the incremental learning mechanism completes parameter optimization, the water power plant medium leakage identification system executes the online learning strategy. The online learning strategy directly updates the optimized model parameters to the running instances of the leakage area intelligent positioning deep neural network model and the leakage type discrimination deep neural network model, realizing instant parameter replacement.

[0037] As a preferred scheme of the present application, the step S600 specifically includes: Step S600.1, configure the remote alarm and information push channel, configure the linkage control logic unit, and trigger the automatic execution.

[0038] Step S600.2, configure the alarm record and event backtracking mechanism, configure the feedback channel and permission control mechanism, and configure the system linkage and remote interaction result set.

[0039] The water power plant medium leakage identification system caches the alarm information and linkage execution result to the real-time monitoring task data buffer unit and writes it into the historical sample library when triggering the linkage control logic unit to execute the action. The historical sample library stores the following contents: alarm trigger timestamp, liquid leakage area state parameter set, liquid leakage medium label result set, liquid leakage degree level label, and linkage execution record.

[0040] After the alarm information and linkage execution result are written into the historical sample library, they are simultaneously fed back to the adaptive optimization closed loop. The adaptive optimization closed loop is a dynamic optimization module in the water power plant medium leakage identification system. When false alarms are triggered three times in a row under the same working condition, the adaptive optimization closed loop automatically increases the liquid leakage area growth rate threshold by 0.5 cm² / s. When there is a continuous missed alarm, the adaptive optimization closed loop automatically reduces the liquid leakage area threshold by 20 cm².

[0041] The water power plant medium leakage identification system interacts with the water power plant operation and maintenance control system through a remote interaction interface, which is a communication port deployed by the water power plant medium leakage identification system. An encryption communication method based on the Transport Layer Security protocol is used. The permission control mechanism sets three types of operation permissions: read-only permission, alarm response permission, and control permission.

[0042] The water power plant medium leakage identification system constructs a system linkage and remote interaction result set, which includes: liquid leakage area state parameter set, liquid leakage medium label result set, liquid leakage degree quantification and grading early warning result set, alarm push record, linkage execution record, and feedback adjustment record. The system linkage and remote interaction result set is output through a multi-level linkage response module.

[0043] A water power plant medium leakage identification system includes the following modules: A video sensing and acquisition module is used to deploy high-definition video monitoring equipment and running state sensor modules, and to perform synchronous registration. It configures infrared channel enhancement and polarized light interference suppression, and performs spatial filtering noise reduction and adaptive adjustment. It configures a multi-modal enhanced input tensor.

[0044] The leakage area identification module is configured to configure a deep neural network model, introduce a time consistency constraint mechanism, configure a segmentation mask output and a geometric boundary parameter output structure, calculate a liquid leakage area parameter and a shape factor parameter, and configure a leakage area state parameter set.

[0045] The leakage category discrimination module is configured to configure a leakage type discrimination deep neural network model, introduce a liquid leakage medium category label, configure an optical texture feature vector, extract an environmental interaction feature vector of the liquid leakage area, fuse the optical texture feature vector and the environmental interaction feature vector, perform liquid leakage medium category discrimination, output the liquid leakage medium category label, configure a medium category discrimination confidence parameter, and configure a liquid leakage medium label result set.

[0046] The leakage quantification grading module is configured to statistically calculate a liquid leakage area parameter and an area growth rate, configure a leakage degree quantification index set, configure a leakage degree grading threshold, output a leakage degree grade label, trigger a grading early warning, set a buffer leakage degree grading result, and configure a leakage degree quantification and grading early warning result set.

[0047] The optimization learning training module is configured to configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, perform pre-training parameter transfer, introduce an incremental learning mechanism, configure an online learning strategy, and perform model parameter update.

[0048] The remote linkage control module is configured to configure a remote alarm and information push channel, configure a linkage control logic unit, trigger automatic execution, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

[0049] Compared with the prior art, the present application has the following advantages: 1. By deploying high-definition video monitoring equipment and operating state sensor modules in the operating area of the hydropower plant, and combining infrared channel enhancement, polarized light interference suppression, and spatial filtering noise reduction technologies, a multi-modal enhanced input tensor is constructed, which can still maintain clear leakage area identification effect in night low-illumination environment, water vapor mist interference, and metal reflection conditions, thereby effectively overcoming the problem of reduced identification accuracy caused by relying on single modal input in the prior art, and realizing stable differentiation of oil and water medium leakage.

[0050] 2. Through the collaborative processing of the intelligent positioning of the deep neural network model and the deep neural network model for distinguishing the leakage type, combined with the optical texture feature vector, the environmental interaction feature vector and the shape factor parameter, the liquid leakage area is accurately distinguished, and the area growth rate and the leakage degree quantization index are introduced to establish a hierarchical determination threshold system, which can realize the quantitative discrimination of different levels of slight, moderate and severe leakage, and effectively solve the problem that the prior art only relies on experience or single threshold determination and lacks a fine grading mechanism.

[0051] 3. By introducing a historical sample library, an artificial correction result set, a transfer learning mechanism, an incremental learning mechanism and an online learning strategy, a self-adaptive optimization closed loop is formed, which can dynamically improve the accuracy of leakage area recognition and leakage type discrimination under complex working conditions, and automatically trigger retraining when the recognition accuracy is lower than the set threshold, thereby ensuring the reliability and stability of long-term operation of the system. At the same time, the system realizes automatic execution and multi-level linkage response through a remote alarm channel and a linkage control logic unit, greatly improving the real-time early warning capability and emergency disposal efficiency of the hydropower plant leakage accident. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A flowchart of a hydropower plant medium leakage identification method provided by an embodiment of the present application.

[0053] Figure 2 A framework diagram of a hydropower plant medium leakage identification system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] Please refer to Figure 1 , Figure 1 A flowchart of a hydropower plant medium leakage identification method provided by an embodiment of the present application.

[0056] In the present embodiment, a hydropower plant medium leakage identification method and system can include steps S100, S200, S300, S400, S500 and S600. Step S100, deploy high-definition video monitoring equipment and running state sensor modules, and perform synchronous registration, configure infrared channel enhancement and polarization light interference suppression, and perform spatial filtering noise reduction and adaptive adjustment, and configure multi-modal enhanced input tensor.

[0057] Step S200, configure a deep neural network model and introduce a timing consistency constraint mechanism, configure a segmentation mask output and a geometric boundary parameter output structure, and calculate a liquid leakage area parameter and a shape factor parameter, and configure a leakage area state parameter set.

[0058] Step S300, configure a leakage type discrimination deep neural network model, introduce a liquid leakage medium category label, configure an optical texture feature vector, extract an environmental interaction feature vector of the liquid leakage area, fuse the optical texture feature vector and the environmental interaction feature vector, and perform liquid leakage medium category discrimination, output the liquid leakage medium category label, and configure a medium category discrimination confidence parameter, and configure a liquid leakage medium label result set.

[0059] Step S400, statistics liquid leakage area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold, output leakage degree grade label, and trigger grading early warning, set cache leakage degree grading result, and configure leakage degree quantization and grading early warning result set.

[0060] Step S500, configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, and perform pre-training parameter transfer, introduce an incremental learning mechanism, and configure an online learning strategy and execute model parameter update.

[0061] Step S600, configure a remote alarm and information push channel, configure a linkage control logic unit, and trigger automatic execution, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

[0062] In some embodiments, the step S100, in particular: Step S100.1, deploy high-definition video monitoring equipment and running state sensor modules, and perform synchronous registration.

[0063] Set up a hydropower plant medium leakage identification system, and deploy high-definition video monitoring equipment and running state sensor modules in the hydropower plant operation area.

[0064] The high-definition video monitoring equipment is a camera unit fixedly installed in the main plant building, oil pipeline gallery and water turbine room, with the following collection parameters: minimum resolution of 1920x1080 pixels, frame rate not less than 30 frames per second, and wide dynamic range imaging function to adapt to the strong light and dark contrast working environment.

[0065] The operating state sensor module comprises an oil pressure sensor arranged on the main transformer oil pipe water power plant medium leakage identification system, a temperature sensor arranged inside the water turbine room, and an air humidity sensor arranged in a possible leakage diffusion area, which respectively collect oil pressure change parameters, environmental temperature change parameters, and air humidity change parameters.

[0066] The joint deployment of the high-definition video monitoring device and the operating state sensor module ensures that data is still obtained under the working conditions of high load operation of the device, insufficient night lighting, or water vapor mist.

[0067] A unified time unit is introduced in the high-definition video monitoring device and the operating state sensor module. The unified time unit is calibrated based on a global positioning water power plant medium leakage identification system, and the time accuracy error is controlled within ±0.05 seconds. Through the unified time unit, an independent time stamp identifier is configured for each frame of video image, and a synchronous time stamp identifier is configured for the oil pressure change parameters collected by the oil pressure sensor, the environmental temperature change parameters collected by the temperature sensor, and the air humidity change parameters collected by the humidity sensor.

[0068] The video image and the operating state sensor parameter are obtained through the high-definition video monitoring device and the operating state sensor module, and the video image and the operating state sensor parameter are time-registered based on the time stamp identifier to construct a time-space fusion feature set after synchronous registration.

[0069] Step S100.2, configure infrared channel enhancement and polarized light interference suppression, and perform spatial filtering noise reduction and adaptive adjustment, configure a multi-modal enhancement input tensor.

[0070] When the brightness of the video image collected by the high-definition video monitoring device is lower than the set brightness threshold 200 Lux, the water power plant medium leakage identification system automatically enables the infrared imaging module. The infrared imaging module uses an infrared image sensor in the 5-12 micron wave band to obtain infrared image data, implements thermal radiation enhancement collection on the monitoring area, and extracts the thermal anomaly distribution features of the potential liquid leakage area.

[0071] In the daytime strong light irradiation environment, if the video image has specular reflection interference caused by the metal equipment surface and the ground water, the water power plant medium leakage identification system implements polarized light interference suppression on the video image, uses a polarization filter to obtain a linearly polarized image, and reduces the reflection interference through a polarization removal algorithm.

[0072] When the video image is disturbed by water mist, airflow, or dust and appears noise, the system performs spatial filtering noise reduction on the video image. The noise reduction method uses Gaussian filtering, the filtering radius is set to 3 pixels, and the standard deviation is set to 1.2, the image background is smoothed, and the integrity of the liquid leakage edge texture structure is maintained.

[0073] In the case of uneven overall gray scale distribution of the video image, the system performs adaptive adjustment of brightness and contrast on the video image, uses a local histogram equalization algorithm to improve the gray scale difference between the liquid leakage area and the background area, and the adjustment target is to keep the global contrast of the image in the gray scale variance ratio interval of 1.5 to 2.5.

[0074] After completing the infrared channel enhancement, polarized light interference suppression, spatial filter noise reduction, and brightness and contrast adaptive adjustment of the video image, the enhanced video image frame, the oil pressure change parameter collected by the oil pressure sensor, the environmental temperature change parameter collected by the temperature sensor, and the air humidity change parameter collected by the humidity sensor are packaged as a multi-modal enhanced input tensor.

[0075] The multi-modal enhanced input tensor includes an image pixel matrix dimension, an oil pressure change parameter dimension, an environmental temperature change parameter dimension, an air humidity change parameter dimension, and a timestamp dimension, and is encoded into a unified format data input into the leakage area intelligent positioning deep neural network model.

[0076] Through the multi-modal enhanced input tensor, the leakage area intelligent positioning deep neural network model can maintain the recognition accuracy of the liquid leakage area in complex background and low contrast scenes, providing reliable input for leakage type discrimination and leakage degree quantification.

[0077] In some specific embodiments, the step S200, in particular: Step S200.1, configure a deep neural network model and introduce a time consistency constraint mechanism.

[0078] In the medium leakage identification system of the hydropower plant, a leakage area intelligent positioning deep neural network model for liquid leakage target identification is configured, and the leakage area intelligent positioning deep neural network model is built based on an improved instance segmentation framework and consists of a feature extraction subnetwork, a region proposal submodule, a pixel-level segmentation submodule, and a bounding box regression submodule.

[0079] The feature extraction subnetwork is used for high-level semantic feature extraction on the image pixel matrix dimension in the multi-modal enhanced input tensor, the region proposal submodule generates candidate regions based on the extracted semantic features and determines whether the candidate regions contain liquid leakage areas, the pixel-level segmentation submodule performs fine segmentation on the candidate regions and outputs a liquid leakage area mask image, and the bounding box regression submodule performs geometric constraint regression on the liquid leakage area and outputs a set of liquid leakage area bounding box parameters.

[0080] The leakage area intelligent positioning deep neural network model is configured with a time consistency constraint mechanism, the optical flow of consecutive video image frames is estimated, the displacement vector field of corresponding pixel points between frames is calculated, and the optical flow field is expressed as: wherein: , is the spatial coordinate of a pixel point of a video image, is the motion component of a pixel point in the horizontal direction, is the motion component of a pixel point in the vertical direction, is the displacement vector of a pixel point in consecutive video image frames.

[0081] After the optical flow estimation is completed, the leakage region intelligent positioning deep neural network model performs projection matching on the liquid leakage region mask between consecutive frames, and calculates the intersection over union. When the intersection over union of the same target between consecutive frames is less than the set threshold 0.65, the leakage region intelligent positioning deep neural network model triggers the mask correction operation.

[0082] Step S200.2, configure the segmentation mask output and geometric boundary parameter output structure, and calculate the liquid leakage region area parameter and shape factor parameter, configure the leakage region state parameter set.

[0083] The leakage region intelligent positioning deep neural network model outputs, when processing each frame of multi-modal enhanced input tensor, include: a pixel-level segmentation mask image and geometric boundary parameters of the liquid leakage region.

[0084] The pixel-level segmentation mask image is a binary image, the part with a pixel value of 1 represents the liquid leakage region, and the part with a pixel value of 0 represents the non-leakage region. The geometric boundary parameters include the center coordinates of the minimum circumscribed rectangle, the rectangle width, the rectangle height, and the rectangle rotation angle.

[0085] After obtaining the pixel-level segmentation mask image, the liquid leakage region area parameter and the liquid leakage region shape factor parameter are calculated based on the image binary analysis method. The liquid leakage region area parameter is calculated by the pixel counting method, and is converted into the actual area in combination with the video image resolution, with the unit of cm². The liquid leakage region shape factor parameter is used to describe the compactness of the region, and the calculation is specifically as follows: wherein: A is the pixel area of the liquid leakage region, P is the contour pixel perimeter of the liquid leakage region, is the structural compactness index of the liquid leakage region, and the value range is , the value closer to 1 indicates that the shape tends to be circular, and the smaller the value, the more elongated or irregular the shape.

[0086] ​​​Based on the pixel-level segmentation mask map, the geometric boundary parameter, the liquid leakage area area parameter and the liquid leakage area shape factor parameter, a leakage area state parameter set is constructed, and the leakage area state parameter set includes: liquid leakage area mask map data, liquid leakage area bounding box center coordinate data, liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area bounding box rotation angle parameter, liquid leakage area area parameter and liquid leakage area shape factor parameter.

[0087] After the inference of the water and electricity plant medium leakage identification system is completed, the leakage area state parameter set is cached to the real-time monitoring task data buffer unit, and is sorted in the order of video image timestamp to generate a structured output data stream.

[0088] In some embodiments, the step S300, in particular: Step S300.1, configure a leakage type discrimination deep neural network model, and derive a liquid leakage medium category label, configure an optical texture feature vector, and extract an environmental interaction feature vector of the liquid leakage area.

[0089] In the water and electricity plant medium leakage identification system, a leakage type discrimination deep neural network model is configured, and the liquid leakage medium category is discriminated according to the liquid leakage area shape factor parameter. The leakage type discrimination deep neural network model outputs a liquid leakage medium category label, and the liquid leakage medium category label includes: an oil medium category label and a water medium category label.

[0090] The leakage type discrimination deep neural network model includes a dual-channel feature extraction structure and a fusion decision subnetwork. The dual-channel feature extraction structure extracts an optical texture feature vector and an environmental interaction feature vector respectively, and the fusion decision subnetwork performs liquid leakage medium category probability prediction based on a fusion feature vector.

[0091] The optical texture feature extraction channel in the dual-channel feature extraction structure performs convolutional coding operation on the image pixel information contained in the liquid leakage area mask map data, extracts the optical texture feature vector of the liquid leakage area, and the optical texture feature vector includes: color gradient distribution characteristics, edge texture complexity index and region internal gray homogeneity index.

[0092] The optical texture of the oil medium is a rainbow halo, low-contrast flow mark and edge diffusion blur, and the optical texture of the water medium is transparency, mirror reflection and edge sharp transition. The optical texture feature vector is uniformly standardized and coded into a one-dimensional vector after extraction, and is input into the fusion decision subnetwork.

[0093] The environmental interaction feature extraction channel in the dual-channel feature extraction structure analyzes the adhesion and interaction performance of the liquid leakage area in the spatial environment. By jointly analyzing the liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area area parameter, and liquid leakage area shape factor parameter, an environmental interaction feature vector is constructed.

[0094] The oil medium generally exhibits a sagging adhesion form on an inclined surface or a vertical surface, with a large difference between the bounding box width parameter and the bounding box height parameter, forming an elongated strip-shaped structure. The water medium exhibits a radial expansion distribution or a high-reflectivity droplet distribution on the surface of the equipment or the ground, with the liquid leakage area shape factor parameter close to 1, indicating a compact structure.

[0095] Step S300.2, fuse the optical texture feature vector and the environmental interaction feature vector, and perform liquid leakage medium category discrimination, output the liquid leakage medium category label, and configure the medium category discrimination confidence parameter and the liquid leakage medium label result set.

[0096] The optical texture feature vector and the environmental interaction feature vector are spliced into a fused feature vector, which is input into a fusion decision sub-network for liquid leakage medium category discrimination. The fusion decision sub-network outputs an oil medium prediction probability and a water medium prediction probability. The liquid leakage medium category label discrimination is specifically: In the formula: is the prediction probability of the liquid leakage area being an oil medium, is the prediction probability of the liquid leakage area being a water medium, is the medium category label output result of the liquid leakage area, and the final discrimination result is determined by selecting the category label corresponding to the maximum probability.

[0097] The leakage type discrimination deep neural network model outputs the medium category discrimination confidence parameter while outputting the liquid leakage medium category label. The medium category discrimination confidence parameter is composed of the maximum prediction probability value and has a value range of 0-1, representing the credibility of the liquid leakage area medium category discrimination result. When the medium category discrimination confidence parameter is less than 0.6, the hydropower plant medium leakage identification system automatically triggers the candidate area review strategy, marks the liquid leakage area as a to-be-confirmed category, and outputs a warning prompt for review and confirmation by the operation and maintenance personnel or the artificial intelligence labeling module.

[0098] Based on the liquid leakage medium category label, the medium category discrimination confidence parameter and the liquid leakage area shape factor parameter, a liquid leakage medium label result set is constructed, which includes: liquid leakage area mask map data, liquid leakage area bounding box center coordinate data, liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area area parameter, liquid leakage area shape factor parameter, liquid leakage medium category label and medium category discrimination confidence parameter. The water power plant medium leakage identification system outputs the liquid leakage medium label result set to the leakage degree grading module and the multi-stage linkage response module, forming a logically closed loop leakage identification and classification process.

[0099] In some embodiments, the step S400, in particular: Step S400.1, statistics liquid leakage area area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold.

[0100] In the leakage area intelligent positioning deep neural network model output leakage area state parameter set, the liquid leakage area area parameter is quantified in cm² unit, and the water power plant medium leakage identification system takes the video image timestamp as the basis to implement difference calculation on the continuous frame liquid leakage area area parameter, and obtains the area growth rate per unit time, which is expressed in cm² / s and represents the diffusion speed of liquid leakage. The calculation of the area growth rate is specifically: In the formula: is the area growth rate of the liquid leakage area, and the unit is , is the liquid leakage area parameter at the current time, and the unit is , is the liquid leakage area parameter at the previous time, and the unit is , is the time difference between video image frames, and the unit is second.

[0101] The water power plant medium leakage identification system constructs a leakage degree quantization index set based on the liquid leakage area area parameter and the liquid leakage area area growth rate, and the leakage degree quantization index set includes: liquid leakage area area parameter, liquid leakage area area growth rate parameter and liquid leakage medium category label. The leakage degree quantization index set comprehensively represents the liquid leakage scale and development trend.

[0102] According to the water power plant operation safety specification and equipment operation limit, the leakage degree grading threshold is set, which includes: liquid leakage area area parameter threshold and liquid leakage area area growth rate threshold.

[0103] When the liquid leakage area parameter is less than 50 cm2 and the liquid leakage area growth rate is less than 2 cm2 / s, it is a slight leakage level; when the liquid leakage area parameter is in the range of 50-200 cm2 or the liquid leakage area growth rate is in the range of 2-5 cm2 / s, it is a moderate leakage level; when the liquid leakage area parameter is greater than 200 cm2 or the liquid leakage area growth rate is greater than 5 cm2 / s, it is a serious leakage level.

[0104] Step S400.2, output the leakage level label, trigger the grading early warning, set the leakage level grading result, and configure the leakage level quantification and grading early warning result set.

[0105] The leakage level quantification index set is input into the leakage level grading discrimination module, and the leakage level grading discrimination module outputs the liquid leakage level label. The liquid leakage level label includes a slight leakage level label, a moderate leakage level label, and a serious leakage level label. After the water and power plant medium leakage identification system outputs the liquid leakage level label, when the liquid leakage level label is a moderate leakage level or a serious leakage level, an alarm module is triggered immediately. The alarm module pushes the leakage position, the liquid leakage medium category label, and the liquid leakage level label. The output channels of the alarm module include a control room alarm terminal, a mobile terminal message push, and a linkage control interface.

[0106] The leakage level grading discrimination module outputs the liquid leakage level label at the same time, and caches the liquid leakage area mask map data, the liquid leakage area bounding box center coordinate data, the liquid leakage area bounding box width parameter, the liquid leakage area bounding box height parameter, the liquid leakage area parameter, the liquid leakage area growth rate parameter, and the liquid leakage medium category label to the real-time monitoring task data buffer unit together with the liquid leakage level label, and writes them into the historical sample library.

[0107] The historical sample library serves as a training data source for subsequent optimization and iteration of the leakage area intelligent positioning deep neural network model and the leakage type discrimination deep neural network model.

[0108] Based on the leakage area state parameter set, the liquid leakage medium label result set, and the leakage level label, a leakage level quantification and grading early warning result set is constructed. The leakage level quantification and grading early warning result set includes the liquid leakage area mask map data, the liquid leakage area bounding box parameter, the liquid leakage area parameter, the liquid leakage area growth rate parameter, the liquid leakage medium category label, the liquid leakage level label, and the alarm push record.

[0109] The water power plant medium leakage identification system quantifies the leakage degree and outputs a hierarchical early warning result set to a multi-stage linkage response module, realizing a complete closed-loop processing link from liquid leakage area identification, liquid leakage medium type discrimination to leakage degree hierarchical early warning.

[0110] In some embodiments, the step S500, in particular: Step S500.1, configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, and perform pre-training parameter transfer.

[0111] In the long-term operation of the water power plant medium leakage identification system, the liquid leakage area state parameter set, the liquid leakage medium label result set and the leakage degree grade label output by the monitoring task data buffer unit are extracted regularly and stored in the historical sample library. The historical sample library includes: video image frame data, liquid leakage area mask image data, liquid leakage area bounding box parameters, liquid leakage area area parameters, liquid leakage area shape factor parameters, liquid leakage medium type label and liquid leakage degree grade label. The operation and maintenance personnel implement review and artificial labeling based on the leakage identification result to form an artificial correction result set, and establish an index mapping relationship between the artificial correction result set and the historical sample library.

[0112] The water power plant medium leakage identification system introduces a transfer learning mechanism. The transfer learning mechanism takes the feature extraction sub-network parameters pre-trained on a general image recognition data set as initial parameters, and transfers them to the leakage area intelligent positioning deep neural network model and the leakage type discrimination deep neural network model. Through the transfer learning mechanism, the deep neural network model can quickly converge under limited sample conditions, shorten the training period and improve the generalization ability in oil and water medium discrimination tasks. The parameter configuration of the transfer learning mechanism is: the initial learning rate is 0.001, and the learning rate is exponentially decayed by a decay factor of 0.9 every 10 training periods.

[0113] Step S500.2, introduce an incremental learning mechanism, configure an online learning strategy and perform model parameter update.

[0114] During the operation of the water power plant medium leakage identification system, when the liquid leakage medium label result set generated in the monitoring task data buffer unit and the artificial correction result set differ, and the liquid leakage medium type label misjudgment rate exceeds the 3% threshold, the system automatically triggers the incremental learning mechanism. The incremental learning mechanism first supplements the newly collected liquid leakage area mask image data, liquid leakage area bounding box parameters and liquid leakage area area parameters to the training sample set, and performs small batch online training on the supplemented training sample set. The number of samples in each batch is set to 64. During the training process, the loss function uses a cross-entropy loss function: wherein: is a cross-entropy loss function, is a true liquid leakage medium class label, is a liquid leakage medium class probability predicted by a deep neural network model, is a number of mini-batch training samples.

[0115] After the incremental learning mechanism completes parameter optimization, the hydropower plant medium leakage identification system executes an online learning strategy, which directly updates the optimized model parameters to the running instances of the leakage area intelligent positioning deep neural network model and the leakage type discrimination deep neural network model, realizing immediate parameter replacement.

[0116] The updated model is put into operation in real time in the inference stage, which can significantly improve the accuracy of the liquid leakage area mask map data, the accuracy of the liquid leakage area parameter calculation, and the accuracy of the liquid leakage medium class label discrimination. In order to suppress the risk of overfitting, the online learning strategy introduces an L2 regularization term in the training process, and the regularization coefficient is set to 0.001.

[0117] Under the joint action of the historical sample library, the artificial correction result set, the transfer learning mechanism, the incremental learning mechanism, and the online learning strategy, the hydropower plant medium leakage identification system forms a model self-adaptive optimization closed loop. The model self-adaptive optimization closed loop continuously monitors the recognition accuracy of the liquid leakage area state parameter set and the liquid leakage medium label result set. When the recognition accuracy is lower than the 95% threshold, the retraining process is automatically triggered to iteratively update the deep neural network model parameters.

[0118] Through the self-adaptive optimization closed loop, the leakage area intelligent positioning deep neural network model and the leakage type discrimination deep neural network model can maintain high recognition rates under complex working conditions such as night low illumination, heavy fog high humidity, and water vapor mist diffusion, thereby ensuring the long-term stable operation and reliability of the hydropower plant medium leakage identification system under all working conditions.

[0119] In some specific embodiments, the step S600 specifically comprises: Step S600.1, configure a remote alarm and information push channel, configure a linkage control logic unit, and trigger automatic execution.

[0120] When the liquid leakage degree level label is a serious leakage level, the hydropower plant medium leakage identification system sends an alarm information to the hydropower plant operation and maintenance control system through an encrypted communication interface, the encrypted communication interface is a data transmission channel configured between the hydropower plant medium leakage identification system and the hydropower plant operation and maintenance control system, and the alarm information includes: liquid leakage area mask map data, liquid leakage area boundary box parameters, liquid leakage area area parameters, liquid leakage area area growth rate parameters, liquid leakage medium category label and liquid leakage degree level label.

[0121] The alarm information is pushed to the on-duty terminal, the dispatching center terminal and the remote maintenance platform terminal at the same time, so as to ensure that the operation and maintenance personnel at all levels can immediately obtain the leakage state.

[0122] The hydropower plant medium leakage identification system and the hydropower plant operation and maintenance control system establish a linkage control logic unit, the linkage control logic unit is an independent module carrying out control strategies, and the control strategies are executed based on the liquid leakage medium label result set and the liquid leakage degree level label: when the liquid leakage medium category label is an oil medium category label and the liquid leakage degree level label is a serious leakage level, the linkage control logic unit issues a shutdown instruction, closes the related valves of the oil pipeline corridor, and starts the oil collection system; when the liquid leakage medium category label is a water medium category label and the liquid leakage degree level label is a serious leakage level, the linkage control logic unit starts the plant drainage system and triggers the emergency shutdown device to protect the safety of the water turbine operation; when the liquid leakage degree level label is a moderate leakage level, the linkage control logic unit pushes a linkage reminder signal to prompt the on-duty personnel to manually confirm and intervene.

[0123] Step S600.2, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

[0124] When triggering the linkage control logic unit to execute actions, the hydropower plant medium leakage identification system caches the alarm information and the linkage execution result to a real-time monitoring task data buffer unit, and writes into a historical sample library, and the historical sample library stores contents including: alarm trigger timestamp, liquid leakage area state parameter set, liquid leakage medium label result set, liquid leakage degree level label and linkage execution record.

[0125] The monitoring task data buffer unit is an operation storage module of the hydropower plant medium leakage identification system, used for temporarily saving real-time task output, and the alarm record and event backtracking mechanism are used for event traceability and disposal effect evaluation.

[0126] The alarm information and the linkage execution result are fed back to the adaptive optimization closed loop after being written into the historical sample library. The adaptive optimization closed loop is a dynamic optimization module in the medium leakage identification system of the hydropower plant. When false alarms are triggered for three times in succession under the same working condition, the adaptive optimization closed loop automatically increases the liquid leakage area growth rate threshold by 0.5 cm² / s. When there is a continuous false alarm, the adaptive optimization closed loop automatically reduces the liquid leakage area threshold by 20 cm².

[0127] The medium leakage identification system of the hydropower plant interacts with the operation and maintenance control system of the hydropower plant through a remote interaction interface. The remote interaction interface is a communication port deployed by the medium leakage identification system of the hydropower plant. A three-type operation permission is set by using an encryption communication method based on a transmission layer security protocol: read-only permission, alarm response permission and control permission.

[0128] The medium leakage identification system of the hydropower plant constructs a system linkage and remote interaction result set, which includes a liquid leakage area state parameter set, a liquid leakage medium label result set, a liquid leakage degree quantification and grading early warning result set, an alarm push record, a linkage execution record and a feedback adjustment record. The system linkage and remote interaction result set is output through a multi-level linkage response module.

[0129] Please refer to Figure 2 , Figure 2 A flowchart of a medium leakage identification system of a hydropower plant is provided for the embodiments of the present application.

[0130] A medium leakage identification system of a hydropower plant includes the following modules: A video sensing and collecting module is configured to deploy high-definition video monitoring equipment and running state sensor modules, and to perform synchronous registration, configure infrared channel enhancement and polarization light interference suppression, and perform spatial filtering and noise reduction and adaptive adjustment, and configure a multi-modal enhanced input tensor.

[0131] A leakage area identification module is configured to configure a deep neural network model, introduce a time sequence consistency constraint mechanism, configure a segmentation mask output and a geometric boundary parameter output structure, and calculate a liquid leakage area parameter and a shape factor parameter, and configure a leakage area state parameter set.

[0132] A leakage category discrimination module is configured to configure a leakage type discrimination deep neural network model, introduce a liquid leakage medium category label, configure an optical texture feature vector, extract an environmental interaction feature vector of a liquid leakage area, fuse the optical texture feature vector and the environmental interaction feature vector, and perform liquid leakage medium category discrimination, output the liquid leakage medium category label, and configure a medium category discrimination confidence parameter, and configure a liquid leakage medium label result set.

[0133] The leakage quantification grading module is configured to count the liquid leakage area parameter and the area growth rate, configure a leakage degree quantification index set, configure a leakage degree grading judgment threshold, output a leakage degree grade label, trigger a grading early warning, set a buffer leakage degree grading result, and configure a leakage degree quantification and grading early warning result set.

[0134] The optimization learning training module is configured to configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, perform pre-training parameter transfer, introduce an incremental learning mechanism, configure an online learning strategy, and perform model parameter update.

[0135] The remote linkage control module is configured to configure a remote alarm and information push channel, configure a linkage control logic unit, trigger automatic execution, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

[0136] In the above content, in actual application, first, in the water and power plant operation area, deploy high-definition video monitoring equipment and operation state sensor modules, and perform high-precision time calibration by the unified time unit, the high-definition video monitoring equipment is fixedly installed in the main plant house, the oil pipeline corridor and the water turbine room, the minimum resolution is set to 1920x1080 pixels, the frame rate is not less than 30 frames / second, and the wide dynamic range imaging function is provided, the operation state sensor modules include the oil pressure sensor arranged in the oil pipeline corridor, the temperature sensor arranged in the water turbine room and the air humidity sensor arranged in the diffusion area, the oil pressure change parameters, the environmental temperature change parameters and the air humidity change parameters collected by various sensors are time stamped by the unified time unit, and the error is controlled within ±0.05 seconds, based on the configuration, the system can still maintain the integrity of data acquisition under the condition of night low illumination and high humidity.

[0137] Then, in the monitoring process, when the brightness of the video image collected by the high-definition video monitoring equipment is lower than 200 Lux, the system automatically enables the infrared imaging module, the infrared sensor works in the 5-12 micrometer wave band to enhance the thermal radiation feature, when the video image has strong reflection interference, the polarized light filter is enabled, and the reflection effect is eliminated by combining the polarization removal algorithm, when the monitoring image has water mist, air flow or dust interference, the system adopts Gaussian filtering for spatial noise reduction, the filtering radius is set to 3 pixels, and the standard deviation is 1.2, when the gray scale distribution is uneven, the local histogram equalization is used to adjust the contrast, the global gray scale variance ratio is maintained between 1.5-2.5, the processed image frame and the sensor parameter are jointly packaged as a multi-modal enhanced input tensor, which is used as the input of the deep neural network model.

[0138] Secondly, in the liquid leakage recognition link, the leakage area intelligent positioning deep neural network model extracts features and targets based on the multi-modal enhanced input tensor. The model performs temporal consistency constraint on consecutive video frames through optical flow estimation, calculates the pixel displacement vector field, and triggers mask correction when the intersection over union is less than 0.65. The system outputs the liquid leakage area segmentation mask graph and geometric boundary parameters, including the center coordinates, width, height, and rotation angle of the rectangular frame. The liquid leakage area parameter is converted to cm² by pixel counting combined with resolution, and the shape factor parameter is used to measure the compactness of the area. The value closer to 1 indicates that the structure is closer to a circle. Finally, the leakage area state parameter set is formed and cached to the real-time monitoring task data buffer unit.

[0139] After that, in the leakage type discrimination link, the leakage type discrimination deep neural network model uses a double-channel structure to extract optical texture feature vectors and environmental interaction feature vectors. The optical texture feature vectors include color gradient distribution, edge texture complexity, and gray homogeneity, which can distinguish the oil medium's rainbow halo, low-contrast flow marks, and fuzzy diffusion, and the water medium's transparency, strong reflection, and sharp boundary. The environmental interaction feature vectors are constructed by analyzing the boundary box width parameter, boundary box height parameter, area parameter, and shape factor parameter, which are used to represent the medium's adhesion and diffusion method in the environment. The fusion decision sub-network outputs the oil medium probability and water medium probability. When the confidence is less than 0.6, the candidate area review strategy is triggered, and finally the liquid leakage medium label result set is generated.

[0140] Then, in the leakage degree grading link, the system calculates the liquid leakage area parameter and area growth rate, which is expressed in cm² / s and used to measure the diffusion speed. When the liquid leakage area is less than 50 cm² and the area growth rate is less than 2 cm² / s, it is determined as a minor leakage level. When the liquid leakage area is between 50-200 cm² or the area growth rate is between 2-5 cm² / s, it is determined as a moderate leakage level. When the liquid leakage area is greater than 200 cm² or the area growth rate is greater than 5 cm² / s, it is determined as a serious leakage level. After the leakage degree level label is output, the system immediately pushes the alarm information to the control room terminal, mobile terminal, and linkage control interface, and generates the leakage degree quantification and grading early warning result set.

[0141] Finally, in the optimization learning and remote linkage link, the historical sample library continuously stores video image frames, mask maps, boundary box parameters and artificial correction result sets, and realizes model optimization combined with transfer learning and incremental learning mechanism. When the liquid leakage medium category label misjudgment rate exceeds 3%, a small batch of online training is automatically triggered, the batch sample is set to 64, and the cross entropy loss function is used to optimize the model parameters. The optimized model is updated in real time through the online learning strategy to ensure that the recognition accuracy is not less than 95%. At the same time, the system pushes the alarm information through the remote encryption channel under the serious leakage level, and cooperates with the linkage control logic unit of the water and power plant operation and maintenance control system to execute the control strategies such as shutdown, valve closing or drainage system starting. All alarm and control results are fed back to the historical sample library to form a complete closed loop.

[0142] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for identifying a medium leakage in a hydroelectric plant, characterized by, Comprise the following steps: S100, deploy high-definition video monitoring equipment and running state sensor module, and carry out synchronous registration, configure infrared channel enhancement and polarization light interference suppression, and carry out spatial filtering noise reduction and adaptive adjustment, configure multi-modal enhancement input tensor; S200, configure deep neural network model, and introduce time sequence consistency constraint mechanism, configure segmentation mask output and geometric boundary parameter output structure, and calculate liquid leakage area parameter and shape factor parameter, configure leakage area state parameter set; S300, configure leakage type discrimination deep neural network model, and introduce liquid leakage medium category label, configure optical texture feature vector, and extract environmental interaction feature vector of liquid leakage area, fuse optical texture feature vector and environmental interaction feature vector, and carry out liquid leakage medium category discrimination, output liquid leakage medium category label, and configure medium category discrimination confidence parameter, configure liquid leakage medium label result set; S400, statistics liquid leakage area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold, output leakage degree grade label, and trigger grading early warning, set cache leakage degree grading result, and configure leakage degree quantization and grading early warning result set; S500, configure historical sample library and artificial correction result set, configure transfer learning mechanism, and carry out pre-training parameter transfer, introduce incremental learning mechanism, and configure online learning strategy and execute model parameter update; S600, configure remote alarm and information push channel, configure linkage control logic unit, and trigger automatic execution, configure alarm record and event backtracking mechanism, configure feedback channel and permission control mechanism, configure system linkage and remote interaction result set.

2. A method of identifying a medium leak in a hydroelectric plant as claimed in claim 1, characterized in that, The S100, specifically: S100.1, deploy high-definition video monitoring equipment and running state sensor module, and carry out synchronous registration; Set up water and power plant medium leakage identification system, and arrange high-definition video monitoring equipment and running state sensor module in water and power plant operation area; The high-definition video monitoring equipment is: fixedly installed camera unit in main plant house, oil pipeline gallery and water turbine room, and the collection parameters are set as: minimum resolution 1920*1080 pixels, frame rate not less than 30 frames per second; The running state sensor module comprises: oil pressure sensor arranged on main transformer oil pipe water and power plant medium leakage identification system, temperature sensor arranged in water turbine room, and air humidity sensor arranged in possible leakage diffusion area, which respectively collect oil pressure change parameter, environmental temperature change parameter and air humidity change parameter; A unified time unit is introduced in high-definition video monitoring equipment and running state sensor module, the unified time unit adopts time source calibration based on global positioning water and power plant medium leakage identification system, the time error is controlled within ±0.05 seconds, through the unified time unit, independent time stamp identifier is configured for each frame of video image, and synchronous time stamp identifier is configured for oil pressure change parameter collected by oil pressure sensor, environmental temperature change parameter collected by temperature sensor and air humidity change parameter collected by humidity sensor; The video image and the running state sensor parameter are acquired through the high-definition video monitoring device and the running state sensor module, and the time-space fusion feature set after synchronization and registration is constructed based on the time stamp identification of the video image and the running state sensor parameter; S100.2, configure infrared channel enhancement and polarization light interference suppression, and perform spatial filtering noise reduction and adaptive adjustment, configure multi-modal enhancement input tensor; When the brightness of the video image collected by the high-definition video monitoring device is lower than the set brightness threshold 200 Lux, the water and power plant medium leakage identification system automatically enables the infrared imaging module, the infrared imaging module uses an infrared image sensor with a 5-12 micron wave band to acquire infrared image data, performs thermal radiation enhancement collection on the monitoring area, and extracts the thermal anomaly distribution characteristics of the potential liquid leakage area; In the daytime strong light irradiation environment, if the video image has specular reflection interference caused by the surface of metal equipment and the ground water, the water and power plant medium leakage identification system performs polarization light interference suppression on the video image, uses a polarization filter to acquire a linear polarization image, and reduces the reflection interference through a polarization removal algorithm; When the video image is disturbed by water mist, airflow or dust and appears noise, the system performs spatial filtering noise reduction on the video image, the noise reduction method adopts Gaussian filtering, the filtering radius is set to 3 pixels, the standard deviation is set to 1.2, the image background is smoothed, and the integrity of the liquid leakage edge texture structure is maintained; In the case of uneven overall gray scale distribution of the video image, the system performs brightness contrast adaptive adjustment on the video image, uses a local histogram equalization algorithm to improve the gray scale difference between the liquid leakage area and the background area, and adjusts the target to maintain the global contrast of the image in the gray scale variance ratio interval of 1.5 to 2.5; After completing the infrared channel enhancement, polarization light interference suppression, spatial filtering noise reduction and brightness contrast adaptive adjustment of the video image, the enhanced video image frame, the oil pressure change parameter collected by the oil pressure sensor, the environmental temperature change parameter collected by the temperature sensor and the air humidity change parameter collected by the humidity sensor are packaged as a multi-modal enhancement input tensor; The multi-modal enhancement input tensor includes: image pixel matrix dimension, oil pressure change parameter dimension, environmental temperature change parameter dimension, air humidity change parameter dimension and time stamp dimension, and is encoded into a unified format data input into the leakage area intelligent positioning deep neural network model.

3. A method of identifying a medium leak in a hydroelectric plant as claimed in claim 1, wherein, The S200, specifically: S200.1, configure a deep neural network model and introduce a time sequence consistency constraint mechanism; In the water and power plant medium leakage identification system, a leakage area intelligent positioning deep neural network model for liquid leakage target identification is configured, the leakage area intelligent positioning deep neural network model is constructed based on an improved instance segmentation framework, and is composed of a feature extraction subnetwork, a region suggestion submodule, a pixel-level segmentation submodule and a bounding box regression submodule; The feature extraction sub-network is configured to perform high-level semantic feature extraction on the image pixel matrix dimension in the multi-modal enhanced input tensor, the region proposal submodule is configured to generate a candidate region based on the extracted semantic features and determine whether the candidate region contains a liquid leakage region, the pixel-level segmentation submodule is configured to finely segment the candidate region and output a liquid leakage region mask, and the bounding box regression submodule is configured to perform geometric constraint regression on the liquid leakage region and output a set of liquid leakage region bounding box parameters; The intelligent leakage region positioning deep neural network model is configured with a time consistency constraint mechanism, which estimates the optical flow of consecutive video image frames to calculate the displacement vector field of corresponding pixel points between frames. After the optical flow estimation is completed, the intelligent leakage region positioning deep neural network model performs projection matching on the liquid leakage region masks between consecutive frames and calculates the intersection-over-union ratio. When the intersection-over-union ratio of the same target between consecutive frames is less than a set threshold of 0.65, the intelligent leakage region positioning deep neural network model triggers a mask correction operation. S200.2, configure the segmentation mask output and geometric boundary parameter output structure, and calculate the liquid leakage region area parameter and shape factor parameter, configure the leakage region state parameter set; The intelligent leakage region positioning deep neural network model outputs, when processing each frame of multi-modal enhanced input tensor, include: a pixel-level segmentation mask and geometric boundary parameters of the liquid leakage region; The pixel-level segmentation mask is a binary image, with pixel values of 1 representing the liquid leakage region and pixel values of 0 representing the non-leakage region. The geometric boundary parameters include the center coordinates of the minimum bounding rectangle, the width of the rectangle, the height of the rectangle, and the rotation angle of the rectangle. After obtaining the pixel-level segmentation mask, the liquid leakage region area parameter and the liquid leakage region shape factor parameter are calculated based on image binary analysis method. The liquid leakage region area parameter is calculated by pixel counting method and converted into actual area with unit of cm² combined with video image resolution. The liquid leakage region shape factor parameter is used to describe the compactness of the region. Based on the pixel-level segmentation mask, the geometric boundary parameters, the liquid leakage region area parameter, and the liquid leakage region shape factor parameter, a leakage region state parameter set is constructed, which includes: liquid leakage region mask data, liquid leakage region bounding box center coordinate data, liquid leakage region bounding box width parameter, liquid leakage region bounding box height parameter, liquid leakage region bounding box rotation angle parameter, liquid leakage region area parameter, and liquid leakage region shape factor parameter. After the inference is completed, the leakage region state parameter set is cached to the real-time monitoring task data buffer unit, sorted in the order of video image timestamp, and a structured output data stream is generated.

4. A method of identifying a medium leak in a hydroelectric plant as recited in claim 1, wherein, The S300, specifically: S300.1, configure a leakage type discrimination deep neural network model and derive a liquid leakage medium category label, configure an optical texture feature vector, and extract an environmental interaction feature vector of the liquid leakage region; In the medium leakage identification system of the hydropower plant, a leakage type discrimination deep neural network model is configured, the liquid leakage medium category is discriminated according to the liquid leakage area shape factor parameters, the leakage type discrimination deep neural network model outputs the liquid leakage medium category label, and the liquid leakage medium category label includes: oil medium category label and water medium category label; The leakage type discrimination deep neural network model includes a double-channel feature extraction structure and a fusion decision subnetwork, the double-channel feature extraction structure extracts optical texture feature vectors and environmental interaction feature vectors respectively, and the fusion decision subnetwork performs liquid leakage medium category probability prediction based on the fusion feature vectors; The optical texture feature extraction channel in the double-channel feature extraction structure performs convolutional coding operation on the image pixel information contained in the liquid leakage area mask image data, extracts the optical texture feature vectors of the liquid leakage area, and the optical texture feature vectors include: color gradient distribution characteristics, edge texture complexity indicators and region internal gray homogeneity indicators; The optical texture of the oil medium is characterized by rainbow halo, low-contrast flow mark and edge diffusion blur, and the optical texture of the water medium is characterized by transparency, mirror reflection and edge sharp transition, the optical texture feature vectors are uniformly standardized to one-dimensional vectors after extraction, and are input into the fusion decision subnetwork; The environmental interaction feature extraction channel in the double-channel feature extraction structure analyzes the adhesion and interaction of the liquid leakage area in the space environment, constructs the environmental interaction feature vector by jointly analyzing the liquid leakage area boundary box width parameter, the liquid leakage area boundary box height parameter, the liquid leakage area area parameter and the liquid leakage area shape factor parameter; The oil medium usually shows a drooping adhesion form on an inclined surface or a vertical surface, the boundary box width parameter and the boundary box height parameter have a large difference, forming an elongated strip structure, and the water medium shows a radial expansion distribution or a high-reflection droplet distribution on the surface of the equipment or the ground, the liquid leakage area shape factor parameter is close to 1, indicating a compact structure; S300.2, fuse the optical texture feature vectors and the environmental interaction feature vectors, and perform liquid leakage medium category discrimination, output the liquid leakage medium category label, and configure the medium category discrimination confidence parameter, and configure the liquid leakage medium label result set; The optical texture feature vectors and the environmental interaction feature vectors are spliced into fusion feature vectors, and are input into the fusion decision subnetwork for liquid leakage medium category discrimination, and the fusion decision subnetwork outputs the oil medium prediction probability and the water medium prediction probability; The leakage type discrimination deep neural network model outputs the medium category discrimination confidence parameter while outputting the liquid leakage medium category label, the medium category discrimination confidence parameter is composed of the maximum prediction probability value, the value range is 0-1, and represents the credibility of the liquid leakage area medium category discrimination result, when the medium category discrimination confidence parameter is lower than 0.6, the hydropower plant medium leakage identification system automatically triggers the candidate area review strategy, marks the liquid leakage area as a to-be-confirmed category and outputs a warning prompt, for review and confirmation by the operation and maintenance personnel or the artificial intelligence labeling module. Based on the liquid leakage medium category label, the medium category discrimination confidence parameter and the liquid leakage area shape factor parameter, a liquid leakage medium label result set is constructed, which includes: liquid leakage area mask map data, liquid leakage area bounding box center coordinate data, liquid leakage area bounding box width parameter, liquid leakage area bounding box height parameter, liquid leakage area area parameter, liquid leakage area shape factor parameter, liquid leakage medium category label and medium category discrimination confidence parameter. The water power plant medium leakage identification system outputs the liquid leakage medium label result set to the leakage degree grading module and the multi-stage linkage response module, forming a logically closed loop leakage identification and classification process.

5. A method for identifying a medium leakage in a hydroelectric plant as claimed in claim 1, characterized in that, The S400 specifically includes: S400.1, statistics liquid leakage area area parameter and area growth rate, configure leakage degree quantization index set, configure leakage degree grading threshold; In the leakage area intelligent positioning deep neural network model output leakage area state parameter set, the liquid leakage area area parameter is quantified in cm² unit, and the water power plant medium leakage identification system takes the video image timestamp as the basis to implement difference calculation on the continuous frame liquid leakage area area parameter, and obtains the area growth rate per unit time, which is expressed in cm² / s and represents the diffusion speed of liquid leakage; The water power plant medium leakage identification system constructs a leakage degree quantization index set based on the liquid leakage area area parameter and the liquid leakage area area growth rate, which includes: liquid leakage area area parameter, liquid leakage area area growth rate parameter and liquid leakage medium category label. The leakage degree quantization index set comprehensively represents the liquid leakage scale and development trend; According to the water power plant operation safety specification and equipment operation limit, set the leakage degree grading threshold, which includes: liquid leakage area area parameter threshold and liquid leakage area area growth rate threshold; When the liquid leakage area area parameter is less than 50 cm² and the liquid leakage area area growth rate is less than 2 cm² / s, it is: slight leakage level, when the liquid leakage area area parameter is in the range of 50-200 cm², or the liquid leakage area area growth rate is in the range of 2-5 cm² / s, it is: moderate leakage level, when the liquid leakage area area parameter is greater than 200 cm², or the liquid leakage area area growth rate is greater than 5 cm² / s, it is: serious leakage level; S400.2, output leakage degree level label, trigger grading early warning, set cache leakage degree grading result, and configure leakage degree quantization and grading early warning result set; The leakage degree quantification index set is input into a leakage degree grading discrimination module, and the leakage degree grading discrimination module outputs a liquid leakage degree grade label, which includes a slight leakage grade label, a moderate leakage grade label, and a severe leakage grade label. The leakage degree grading discrimination module outputs the liquid leakage degree grade label, and simultaneously caches the liquid leakage area mask map data, the liquid leakage area bounding box center coordinate data, the liquid leakage area bounding box width parameter, the liquid leakage area bounding box height parameter, the liquid leakage area area parameter, the liquid leakage area area growth rate parameter, and the liquid leakage medium category label to the real-time monitoring task data buffer unit together with the liquid leakage degree grade label, and writes them into a historical sample library. Based on the leakage area state parameter set, the liquid leakage medium label result set, and the leakage degree grade label, a leakage degree quantification and grading early warning result set is constructed, which includes the liquid leakage area mask map data, the liquid leakage area bounding box parameter, the liquid leakage area area parameter, the liquid leakage area area growth rate parameter, the liquid leakage medium category label, the liquid leakage degree grade label, and an alarm push record.

6. A method for identifying a medium leak in a hydroelectric plant as claimed in claim 1, characterized in that, The S500 specifically includes: S500.1, configuring a historical sample library and an artificial correction result set, configuring a transfer learning mechanism, and performing pre-training parameter transfer; During the long-term operation of the hydropower plant medium leakage identification system, the liquid leakage area state parameter set, the liquid leakage medium label result set, and the leakage degree grade label output from the monitoring task data buffer unit are extracted periodically and stored in the historical sample library, which includes video image frame data, liquid leakage area mask map data, liquid leakage area bounding box parameters, liquid leakage area area parameters, liquid leakage area shape factor parameters, liquid leakage medium category labels, and liquid leakage degree grade labels. The operation and maintenance personnel implement review and artificial labeling based on the leakage identification result to form an artificial correction result set, and establish an index mapping relationship between the artificial correction result set and the historical sample library. The water power plant medium leakage identification system introduces a transfer learning mechanism. The transfer learning mechanism takes the feature extraction subnetwork parameters obtained by pre-training on a general image recognition dataset as initial parameters, and transfers them to the intelligent positioning deep neural network model and the leakage type discrimination deep neural network model. Through the transfer learning mechanism, the deep neural network model can quickly converge under limited sample conditions, shorten the training period, and improve the generalization ability in oil and water medium discrimination tasks. The parameter configuration of the transfer learning mechanism is as follows: the initial learning rate is 0.001, and the learning rate is exponentially decayed by a decay factor of 0.9 every 10 training periods; S500.2, introduce an incremental learning mechanism, configure an online learning strategy and perform model parameter update; During the operation of the water power plant medium leakage identification system, when the liquid leakage medium label result set generated in the monitoring task data buffer unit and the artificial correction result set differ, and the liquid leakage medium category label misjudgment rate exceeds the 3% threshold, the system automatically triggers the incremental learning mechanism. The incremental learning mechanism first supplements the newly collected liquid leakage area mask image data, liquid leakage area bounding box parameters and liquid leakage area area parameters to the training sample set, and performs small batch online training on the supplemented training sample set. The number of samples in each batch is set to 64; After the incremental learning mechanism completes parameter optimization, the water power plant medium leakage identification system executes the online learning strategy. The online learning strategy directly updates the optimized model parameters to the running instances of the intelligent positioning deep neural network model and the leakage type discrimination deep neural network model, realizing instant parameter replacement.

7. A method for identifying a medium leak in a hydroelectric plant as claimed in claim 1, characterized in that, The S600, specifically: S600.1, configure a remote alarm and information push channel, configure a linkage control logic unit, and trigger automatic execution; After the water power plant medium leakage identification system outputs the liquid leakage level label in the leakage level discrimination module, when the liquid leakage level label is a serious leakage level, the water power plant medium leakage identification system sends an alarm information to the water power plant operation and maintenance control system through an encrypted communication interface. The encrypted communication interface is a data transmission channel configured between the water power plant medium leakage identification system and the water power plant operation and maintenance control system. The alarm information includes: liquid leakage area mask image data, liquid leakage area bounding box parameters, liquid leakage area area parameters, liquid leakage area area growth rate parameters, liquid leakage medium category label and liquid leakage level label; The water power plant medium leakage identification system and the water power plant operation and maintenance control system establish a linkage control logic unit, the linkage control logic unit is an independent module bearing a control strategy, and the control strategy is executed based on a liquid leakage medium label result set and a liquid leakage degree grade label; when the liquid leakage medium category label is an oil medium category label and the liquid leakage degree grade label is a serious leakage grade, the linkage control logic unit issues a shutdown instruction, closes related valves of an oil pipeline gallery, and starts an oil collection system; when the liquid leakage medium category label is a water medium category label and the liquid leakage degree grade label is a serious leakage grade, the linkage control logic unit starts a plant drainage system and triggers an emergency shutdown device to protect the safety of a water turbine; when the liquid leakage degree grade label is a moderate leakage grade, the linkage control logic unit pushes a linkage reminder signal to prompt an on-duty personnel to manually confirm and intervene; S600.2, configure an alarm record and an event backtracking mechanism, configure a feedback channel and a permission control mechanism, and configure a system linkage and a remote interaction result set; The water power plant medium leakage identification system caches alarm information and linkage execution results to a real-time monitoring task data buffer unit and writes the alarm information and the linkage execution results into a historical sample library when triggering the linkage control logic unit to execute actions; the historical sample library stores content including an alarm triggering timestamp, a liquid leakage area state parameter set, a liquid leakage medium label result set, a liquid leakage degree grade label, and linkage execution records; After the alarm information and the linkage execution results are written into the historical sample library, the alarm information and the linkage execution results are simultaneously fed back to an adaptive optimization closed loop, which is a dynamic optimization module in the water power plant medium leakage identification system; when false alarms are triggered for three consecutive times under the same working condition, the adaptive optimization closed loop automatically increases a liquid leakage area growth rate threshold value by 0.5 cm² / s; when there are continuous missed alarms, the adaptive optimization closed loop automatically reduces a liquid leakage area threshold value by 20 cm²; The water power plant medium leakage identification system interacts with the water power plant operation and maintenance control system through a remote interaction interface, which is a communication port deployed by the water power plant medium leakage identification system; an encryption communication mode based on a transmission layer security protocol is adopted; and three types of operation permissions are set by a permission control mechanism: a read-only permission, an alarm response permission, and a control permission; The water power plant medium leakage identification system constructs a system linkage and a remote interaction result set, which includes a liquid leakage area state parameter set, a liquid leakage medium label result set, a liquid leakage degree quantification and grading early warning result set, an alarm pushing record, a linkage execution record, and a feedback adjustment record; the system linkage and the remote interaction result set are output through a multi-level linkage response module.

8. A hydroelectric plant medium leak identification system characterized by, The system includes the following modules: A video sensing acquisition module is configured to deploy high-definition video monitoring equipment and a running state sensor module, perform synchronous registration, configure infrared channel enhancement and polarization light interference suppression, perform spatial filtering and noise reduction, and adaptively adjust and configure a multi-modal enhancement input tensor. The leakage area identification module is configured to configure a deep neural network model, introduce a time consistency constraint mechanism, configure a segmentation mask output and a geometric boundary parameter output structure, calculate a liquid leakage area parameter and a shape factor parameter, and configure a leakage area state parameter set. The leakage category discrimination module is configured to configure a liquid leakage medium category label, configure an optical texture feature vector, extract an environmental interaction feature vector of the liquid leakage area, fuse the optical texture feature vector and the environmental interaction feature vector, perform liquid leakage medium category discrimination, output the liquid leakage medium category label, configure a medium category discrimination confidence parameter, and configure a liquid leakage medium label result set. The leakage quantification grading module is configured to count a liquid leakage area parameter and an area growth rate, configure a leakage degree quantification index set, configure a leakage degree grading judgment threshold, output a leakage degree grade label, trigger a grading early warning, set a buffer leakage degree grading result, and configure a leakage degree quantification and grading early warning result set. The optimization learning training module is configured to configure a historical sample library and an artificial correction result set, configure a transfer learning mechanism, perform pre-training parameter transfer, introduce an incremental learning mechanism, configure an online learning strategy, and perform model parameter update. The remote linkage control module is configured to configure a remote alarm and information push channel, configure a linkage control logic unit, trigger automatic execution, configure an alarm record and event backtracking mechanism, configure a feedback channel and permission control mechanism, and configure a system linkage and remote interaction result set.

Images