Dike patrol management method and system based on video identification AI algorithm

Abstract

The invention discloses an embankment patrol management method and system based on a video identification AI algorithm, particularly relates to industrial-grade system application of deep integration of the embankment video AI algorithm, a hydraulic engineering embankment video monitoring system and patrol management business, and relates to the technical field of information. Comprising the following steps: taking historical and real-time video monitoring data obtained by an existing dike video monitoring system as basic data, carrying out learning training through a convolutional network algorithm, constructing a video identification AI model, and obtaining a training sample library; and performing patrol management on the dike by dynamically collecting video data of the dike and utilizing a video identification AI model according to a training sample library based on a patrol business logic process of the dike. Informatization and intelligentization of dike patrol management are realized, and the dike patrol maintenance capability is improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and more particularly, to a levee inspection management method and system based on a video recognition AI algorithm. Background Art

[0002] In the levee business management work of the water conservancy department, levee safety monitoring and levee inspection are key contents of the levee management of water conservancy projects.

[0003] The levee safety monitoring system involves various monitoring means such as displacement monitoring, seepage pressure monitoring, seepage flow monitoring, and video monitoring. Due to historical reasons and the influence of current technical factors, only the video monitoring system is relatively complete. Due to historical factors and construction planning factors, the inspection management in the safety management of levee projects still stays at backward means such as paper file records and telephone notifications, with low inspection efficiency and low informatization level. Therefore, the water conservancy levee project urgently needs a perfect monitoring means and an algorithm for analyzing and using monitoring data, and combines the algorithm and the monitoring system to improve the levee management business application. In the prior art, the timeliness of data acquisition is low and the cost is high. At the same time, the algorithms applied in the prior art require long-term training and do not have the practicality and professionalism required by the levee management business. Summary of the Invention

[0004] In order to solve the above problems, the object of the present invention is to provide a levee inspection management technology based on a video recognition AI algorithm, aiming to realize the informatization and intelligentization of levee inspection work at low cost, meet the requirements of various services in the levee inspection management work, and form a levee inspection management system that meets the informatization and intelligentization of levee inspection.

[0005] In order to achieve the above technical object, the present application provides a levee inspection management method based on a video recognition AI algorithm, characterized by including the following steps:

[0006] Taking the historical and real-time video monitoring data obtained from the existing video monitoring system of the levee as basic data, learning and training through a convolutional network algorithm, constructing a video recognition AI model, and obtaining a training sample library;

[0007] Based on the inspection business logic process of the levee, dynamically collecting the video data of the levee, using the video recognition AI model, and based on the training sample library, conducting inspection management on the levee.

[0008] Preferably, when conducting inspection and management of dikes, according to the inspection business logic and combined with the video recognition AI model, an inspection business logic process is constructed, including: the video surveillance AI robot inspection business process and the manual inspection business process. Among them, the video surveillance AI robot inspection business process represents the process of performing unmanned automatic inspection by carrying a video acquisition device, and the manual inspection business process represents the process of assisting manual dike inspection through the video recognition AI model.

[0009] Preferably, when performing model training and learning, by configuring a distributed training network, the convolutional network algorithm is trained based on the basic data to construct a video recognition AI model.

[0010] Preferably, when constructing a training sample library, after annotating and enhancing the basic data, a training sample library is generated through the video recognition AI model. Among them, annotation is performed through pixel-level segmentation and multi-modal alignment, and data enhancement processing is performed through infrared enhancement, environment simulation, and multi-spectral fusion.

[0011] Preferably, when constructing a training sample library, a training sample library is constructed by obtaining a dam seepage video sample library, a dam damage sample library, a dam shelter forest damage sample library, and a personnel behavior recognition library sample library.

[0012] Preferably, when constructing a video recognition AI model, after optimizing the model through hyperparameter search space definition, the gradient health check and weight histogram are used to check the model training process.

[0013] Preferably, when performing model training, after saving the function for checking points and the function for cleaning old check points, the check point file is loaded and the model parameters, optimizer state, and learning rate scheduler state are restored; by constructing a TensorRT engine and setting a version control policy and an automated update pipeline, the model is trained and the results of the model training are stored in the database.

[0014] Preferably, when constructing the video surveillance AI robot inspection business process, the video surveillance AI robot inspection business process includes the following steps:

[0015] Step 1.1: Formulate video robot inspection tasks: Combine daily management work to formulate regular video robot inspection tasks and generate daily inspection paths, and combine daily monitoring and early warning information to formulate special video robot inspection tasks and generate special inspection paths; Archive the generated tasks separately to form corresponding inspection tasks;

[0016] Step 1.2: Start inspection tasks: According to the requirements of regular inspection tasks, complete the inspection according to the regular inspection task inspection path and archive the inspection record information; According to the requirements of special inspection tasks, complete the special inspection according to the special inspection task inspection path and archive the special inspection record information;

[0017] Step 1.3: Automatically call the video AI recognition algorithm to analyze and judge the video image information collected during the video robot's patrol mission;

[0018] Step 1.4: Complete the inspection and handling work according to the analysis and judgment results: If there is no problem, the inspection record will be directly archived and the inspection task is completed; if there is a problem, it will be reported to the administrator based on the problem identified by the video, and the administrator will assign personnel to handle it;

[0019] Among them, general problems are assigned to the inspection team for inspection and handling, and professional problems are assigned to the professional handling team for professional problem handling.

[0020] Preferably, when constructing a manual inspection business flow, the manual inspection business flow includes the following steps:

[0021] Step 2.1: Formulate manual inspection tasks: formulate manual inspection tasks in combination with daily management work and generate daily manual inspection plans;

[0022] Step 2.2: The inspectors conduct inspections according to the manual inspection plan. The inspectors complete the inspection judgment based on the actual inspection situation. If the inspection is determined not to be carried out, the inspection process ends. If the inspection is determined to be carried out, the normal inspection process will be entered into the next step.

[0023] Step 2.3: Patrol personnel conduct normal patrols: Patrol personnel conduct normal patrols as required. During the patrol process, the administrator delegates the use of the video surveillance system and video AI intelligent analysis algorithm to patrol personnel to assist in manual patrol work.

[0024] Step 2.4: The inspectors make judgments on the problems found during the inspection: if a problem is found and the inspectors can handle it themselves, the inspectors will handle it according to the daily management regulations, and record and archive the handling situation, and then the manual inspection work ends; if a problem is found and the inspectors cannot handle it themselves, the inspectors will report it to the administrator according to the management regulations, and the administrator will assign a professional team to handle it. The handling situation will be summarized as required and collected by the reviewers. The reviewers will manually archive it according to the handling results of the professional team or automatically archive it according to the time regulations, and then the entire inspection process ends.

[0025] The present invention also discloses a dike inspection and management system based on a video recognition AI algorithm. The system is used in the above-mentioned dike inspection and management method based on a video recognition AI algorithm. The system includes:

[0026] An intelligent model training module, which uses the historical and real-time video monitoring data obtained from the existing video monitoring system of the dike as basic data, conducts learning and training through a convolutional network algorithm, constructs a video recognition AI model, and obtains a training sample library;

[0027] An inspection management module, which is used to manage the inspection of the dike based on the inspection business logic process of the dike, dynamically collect the video data of the dike, utilize the video recognition AI model, and based on the training sample library.

[0028] The present invention discloses the following technical effects:

[0029] Based on the historical and real-time video monitoring data obtained from the existing video monitoring system of the dike management unit, on the basis of making full use of the existing resources, the present invention constructs a video recognition AI algorithm based on the dike business. On top of the video recognition AI algorithm, it integrates the dike inspection management business, develops an information-based system for dike inspection management that combines the dike video recognition algorithm and the dike inspection business, realizes the informatization and intelligence of dike inspection management, and improves the ability of dike inspection, maintenance and repair. Brief Description of the Drawings

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0031] Figure 1 It is the video monitoring AI inspection business flow described in the present invention;

[0032] Figure 2 It is the manual inspection business flow described in the present invention;

[0033] Figure 3 It is the video AI algorithm described in the present invention;

[0034] Figure 4 It is the dike inspection management system described in the present invention. Detailed Embodiments

[0035] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Usually, the components of the embodiments of this application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents the selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative efforts belong to the scope of protection of this application.

[0036] As Figures 1-4 shown, the present invention provides a levee inspection management technology based on video recognition AI algorithms, which uses the existing video monitoring system as the data basis to solve the informatization and intelligentization problems of levee inspection at low cost. The levee inspection management system includes the design of the inspection business logic process, video recognition AI algorithms, and levee inspection business functions; specifically, it includes the design of the inspection business logic process, the construction of video recognition AI algorithms, and the construction of levee inspection business functions. The content is as follows:

[0037] According to the inspection business logic and combined with the achievements of the construction of video recognition AI algorithms, an inspection business logic process is constructed, including the inspection business flow of video surveillance AI robots and the manual inspection business flow.

[0038] To achieve the informatization and intelligentization of inspection management, the inspection management business flow is sorted out, and the specific business implementation is as follows:

[0039] 1. Inspection business flow of video surveillance AI robots

[0040] Formulate the inspection tasks of video robots, start the inspection tasks, automatically call the video AI recognition algorithm, and complete the inspection and disposal work according to the analysis and judgment results. If there are no problems, directly file the inspection records to complete the inspection tasks; if there are problems, report them to the administrator according to the problems identified by the video, and the administrator assigns personnel to handle them.

[0041] Manual inspection business flow

[0042] Formulate manual inspection tasks. The inspection personnel conduct inspections according to the requirements of the manual inspection plan. The inspection personnel conduct normal inspections and determine the problems found in the inspection tasks. If a problem is found and the inspection personnel can handle it themselves, the inspection personnel shall handle it in accordance with the daily management regulations, record the handling situation and file it. At this time, this manual inspection work ends; if a problem is found and the inspection personnel cannot handle it themselves, the inspection personnel shall report it to the administrator in accordance with the management regulations. The administrator assigns a professional team to handle it. The handling situation is summarized and collected to the review personnel as required. The review personnel conduct manual filing based on the handling results of the professional team or conduct automatic filing according to the time regulations. At this time, the entire inspection process ends.

[0043] 2. Construction of video recognition AI algorithms. Further, the steps of the algorithm construction plan include the following:

[0044] Construct a training environment based on convolutional network algorithms to form a basic environment for model algorithm training;

[0045] Construct a training sample library as the basic database for model training. The further steps of the plan are as follows:

[0046] First, construct a vortex recognition library for the water body in front of the dam, including data collection work, data strategy enhancement work, data storage and warehousing, etc.

[0047] Second, construct a seepage recognition library for the dam body, including data collection work, data strategy enhancement work, data storage and warehousing, etc.

[0048] Third, damage to the dam body (signs and marks), including data collection work, data strategy enhancement work, data storage and warehousing, etc.

[0049] Fourth, damage to the shelter forest on the dam body, including data collection work, data strategy enhancement work, data storage and warehousing, etc.;

[0050] Fifth, recognition of personnel behavior, including data collection work, data strategy enhancement work, data storage and warehousing, etc.

[0051] Adjust the weights of the training algorithm, adjust the parameters of the constructed algorithm, and improve the algorithm accuracy. The further steps of the plan are as follows:

[0052] First, select the model architecture. Select a relatively common and mature algorithm. In this patent, ResNet-50 is selected to construct the model framework and optimize it separately to improve efficiency.

[0053] Second, define the hyperparameter search space. In machine learning and deep learning, defining the hyperparameter search space is an important step in optimizing algorithms, which is used to find the best model parameter configuration, including determining the hyperparameters to be optimized, automatic parameter tuning, and training process monitoring.

[0054] Model training and warehousing: After training the constructed model, the algorithm results of the model are warehoused for easy calling during the business application stage. The process includes training process automation, deployment optimization and monitoring, and model version and update management.

[0055] 3. Combine with the video recognition AI algorithm, inspect the business logic, and build the embankment management inspection business function.

[0056] Combined with the inspection business logic process, building the embankment management inspection business function includes the video surveillance AI robot inspection function and the manual inspection function. Further, the sub-functions at all levels of the embankment management inspection system are as follows:

[0057] The video surveillance AI robot inspection function includes sub-items such as inspection tasks, inspection analysis, and alarm handling.

[0058] The manual inspection function includes sub-items such as inspection systems, manual inspections, and inspection statistics.

[0059] Embodiment: The present invention provides a method for embankment inspection management based on a video recognition AI algorithm, which specifically includes the following contents:

[0060] The first step: Sort out and design the inspection business logic process

[0061] For the inspection management work, improve the business process design. There are two directions. One is the video surveillance AI robot inspection. Manually set the video inspection path and start the AI inspection. Combine the video algorithm to identify the safety conditions of various water conservancy facilities on the inspection path. The other is the manual inspection. The video surveillance system combines the inspection tasks of the manual inspectors to assist the manual inspection work. When the inspector is within the video surveillance range, the inspector can directly call the video control authority within the range and use the video surveillance equipment to view the behaviors within the range, and quickly analyze various embankment danger situations and project conditions within the inspection range.

[0062] 1. Video surveillance AI robot inspection process

[0063] The first step: Formulate video robot inspection tasks. Combine the daily management work to formulate regular video robot inspection tasks and generate daily inspection paths, and combine the daily monitoring and early warning information to formulate special video robot inspection tasks and generate special inspection paths; archive the generated tasks respectively to form corresponding inspection tasks.

[0064] The second step: Start the inspection task. According to the requirements of the regular inspection task, complete the inspection according to the regular inspection task inspection path and archive the inspection record information; according to the requirements of the special inspection task, complete the special inspection according to the special inspection task inspection path and archive the special inspection record information.

[0065] Step 3: The system automatically calls the video AI recognition algorithm to analyze and judge the video image information collected during the video robot's patrol mission.

[0066] Step 4: Complete the inspection and handling work according to the analysis and judgment results. If there are no problems, the inspection record will be directly archived and the inspection task is completed. If there are problems, they will be reported to the administrator based on the problems identified in the video, and the administrator will assign personnel to handle them.

[0067] a. General problems will be assigned to the inspection team for inspection and handling. If the inspection personnel can handle it, it will be automatically archived after handling and the entire video AI robot inspection process will end. If the problem cannot be handled by the inspection personnel, it will be fed back to the problem reporting process of the next level, and the administrator will assign the corresponding personnel to handle it.

[0068] b. Professional issues are assigned to professional handling teams for handling. After handling, the professional team will archive the handling information and submit it to the reviewer. The reviewer can archive and end the entire process after manual review. The reviewer can also set automatic archiving to end the entire process according to the effective time period.

[0069] 2. Manual inspection process

[0070] Step 1: Develop manual inspection tasks. Develop manual inspection tasks in conjunction with daily management work and generate a daily manual inspection plan.

[0071] Step 2: The inspector conducts the inspection according to the manual inspection plan. The inspector completes the inspection judgment work according to the actual inspection situation. If the inspection is determined not to be carried out, the inspection process ends; if the inspection is determined to be carried out, the normal inspection process will be entered into the next step.

[0072] Step 3: Inspectors conduct normal inspections. Inspectors conduct normal inspections as required. During the inspection process, the administrator delegates access to the video surveillance system and video AI intelligent analysis algorithms to inspectors to assist with manual inspections.

[0073] Step 4: The inspector determines any issues discovered during the inspection. If an issue is discovered and the inspector can handle it themselves, they will handle it according to daily management regulations, record the handling and file it, and the manual inspection process ends at this point. If an issue is discovered and the inspector cannot handle it themselves, the inspector will report it to the administrator according to management regulations. The administrator will assign a professional team to handle it. The handling results will be summarized and submitted to the reviewer as required. The reviewer will determine whether to manually file the information based on the professional team's handling results or automatically file it according to the time limit. This concludes the entire inspection process.

[0074] Step 2: Build a video recognition AI algorithm

[0075] For the embankment management work, improve the development of video algorithms. There are two directions. One is the embankment safety video monitoring algorithm, which covers the identification of water body vortices in front of the dam, the identification of seepage in the dam body, the identification of dam body damage (signs), the identification of damage to the protective forest on the dam body, etc.; the other is the algorithm for identifying the behavior of embankment inspection personnel.

[0076] The embankment safety video monitoring algorithm is to build algorithms for common embankment business scenarios. The main monitoring data related to the dam body concerned in common embankment management work mainly include seepage in the dam body, dam body damage (signs), damage to the protective forest on the dam body, water body vortices in front of the dam, and personnel behavior identification, etc. Algorithms dedicated to embankment safety video monitoring are built for the above scenarios; the algorithm construction adopts the existing common AI algorithm construction process, mainly including the construction of the convolutional network algorithm training environment, the construction of the sample training library, the tuning of the training algorithm weights, and model training.

[0077] Step 1: Build a training environment based on the convolutional network algorithm

[0078] Step ①: Hardware configuration and verification

[0079] The effective construction of any AI recognition algorithm should be based on a hardware environment that can fully utilize its algorithm performance. Since it is for the use of embankment management departments. Therefore, this hardware scenario makes full use of the existing computer room equipment of embankment management departments to build a computing power environment, and supplements equipment such as GPU servers that meet the computing efficiency. The hardware operation environment system is required to support the CUDA module.

[0080] Step ②: Determine the development language

[0081] Since most of the mainstream machine learning and image learning model libraries in the industry are in Python language, the programming language of this project adopts Python language, and all AI recognition technology algorithms related to the algorithm and the program algorithms of the convolutional part are programmed and developed using this language.

[0082] Step ③: Deploy the software environment

[0083] The operating system of the running environment adopts the mainstream Linux server operating system, and the kernel of the running system in the current running environment is optimized in advance and the CUDA environment is installed.

[0084] A. Kernel optimization (improve video processing efficiency)

[0085] Set the maximum number of file handles that can be allocated by the system kernel (i.e., the upper limit of the number of files / connections opened simultaneously)

[0086] Default value: Varies depending on the system (usually about 80,000).

[0087] Why is it set to 10,000,000?

[0088] It is applicable to scenarios that need to handle high-concurrency I / O throughput to avoid errors such as Too many open files caused by the algorithm program exceeding the default limit.

[0089] It is necessary to adjust the ulimit limit of the user / process simultaneously (through / etc / security / limits.conf).

[0090] Control the eagerness of the kernel to swap memory data to disk (swap).

[0091] The lower the value, the more inclined to keep data in memory; the higher the value, the more actively use disk swapping.

[0092] Value range: 0 (try to avoid swapping) to 100 (actively swap).

[0093] Default value: Usually 60.

[0094] Why is it set to 10?

[0095] Give priority to using physical memory rather than swapping to disk, suitable for the following scenarios:

[0096] Systems with sufficient memory.

[0097] Applications sensitive to disk I / O latency (such as databases, cache services).

[0098] When using SSDs (to reduce disk wear caused by frequent swapping).

[0099] fs.file-max = 10000000

[0100] vm.swappiness = 10

[0101] CUDA Environment Installation

[0102] In the Linux environment, for different Linux system versions, select different corresponding Linux driver packages from the NVIDIA official website for download and run the installed driver in the Linux environment.

[0103] Step ④, Configure the Distributed Training Network

[0104] Configure the distributed training network for the project to improve communication efficiency for multi-server training.

[0105] NCCL Configuration

[0106] Enable NCCL Debug Log

[0107] It is recommended to turn it off after debugging to avoid excessive logs

[0108] Analyze the communication topology by viewing the collnet section in the log

[0109] Specify the communication network card

[0110] First, confirm the physical network card name through ifconfig or ipa (modern systems may use ens160 / enp0s3, etc.)

[0111] It must be specified in the scenario of multiple network cards to prevent the automatic selection of inefficient paths

[0112] SSH passwordless intercommunication

[0113] Generate a key pair (execute on all nodes)

[0114] Press Enter 3 times continuously to generate a passwordless key (if high security is required, a passphrase can be set)

[0115] The default save paths for the keys are: ~ / .ssh / id_rsa (private key) and ~ / .ssh / id_rsa.pub (public key)

[0116] Copy the public key to the target node (execute on the control node)

[0117] The password of the target node needs to be entered for the first connection

[0118] The public key will be automatically appended to ~ / .ssh / authorized_keys of the target node

[0119] Verify passwordless login

[0120] Step 2: Build the training sample library

[0121] The video AI model based on the convolutional network algorithm should have a training sample library optimized for this algorithm. When building the training library, it will include the collection and construction of standard samples in multiple scenarios such as vortices in front of the dam, seepage of the dam body, damage to the dam body (signs), damage to the dam body's shelter forest, and identification of human behavior. At the same time, different training sample data from multiple angles and dimensions will be actively collected to form a basic training Train library that meets the model training requirements, ensuring the high quality and diversity of the standard data set in the sample library, covering extreme scenarios and long-tail distributions. During the general training process, the data in the small-scale Validation verification sample library will be repeatedly used for training correction. A part of the original Train library data can be cut to create a separate Validation library. The data ratio of the Train library to the Validation library in this case is 32:1

[0122] Step: Build the sample library of vortices in front of the dam

[0123] Data collection

[0124] a. Source

[0125] The data source includes real data and synthetic data. The real data is obtained by continuously recording videos with a dam monitoring camera, and the real sample data is formed by taking frames from the real videos and converting and retaining them; the synthetic data is the simulation videos and sample results of different flow velocities and vortex forms constructed and generated based on the computational fluid dynamics model (similar to OpenFOAM).

[0126] b. Annotation requirements

[0127] Temporal annotation: Annotate the start frame and end frame when the vortex appears (time accuracy ±0.1 second).

[0128] Spatial annotation: Annotate the bounding box of the vortex area for the key frames (COCO format).

[0129] Auxiliary annotation: Use the RAFT optical flow algorithm to extract motion vectors to assist in annotating the water flow direction.

[0130] Data augmentation strategy

[0131] While training the sample data, it is also necessary to introduce data augmentation strategies, which mainly aim to increase the anti-external interference ability of the algorithm. The data augmentation strategies include dynamic perturbation, environmental interference, and resolution multi-scale. Among them, dynamic perturbation mainly requires superimposing a synthetic optical flow field (intensity 0.1 - 0.5) to simulate the dynamic changes of water flow; environmental interference includes adding rain and fog noise (transparency 20% - 80%) and supplementing the dynamic blur (Gaussian kernel) situation, etc.; in addition, its resolution should support multi-scale resolution, and it is required to randomly scale the resolution of the input frame.

[0132] Storage scheme

[0133] The mainstream storage generally uses a non-relational file library form for storage. This scheme uses MinIO to build a private S3 storage to achieve the storage of data objects. And Elasticsearch is used for comprehensive metadata management, and its indexed data attributes include but are not limited to resolution, acquisition time, weather conditions, etc.

[0134] Step ②. Construction of the dam seepage video sample library

[0135] Data collection

[0136] a. Source

[0137] The data sources of dam seepage mainly include two categories of video shooting data: infrared thermal imaging and visible light. The specific resolution is determined by existing video monitoring equipment. The monitoring content includes the surface temperature distribution image of the dam and the appearance changes in the seepage area (such as wetness, moss growth, etc.).

[0138] b. Annotation requirements

[0139] Pixel-level segmentation: Annotate the precise mask of the seepage area.

[0140] Multimodal alignment: The infrared and visible light images need to be time-synchronized (error ≤ 1 second).

[0141] Data augmentation strategy

[0142] The data augmentation strategy includes infrared enhancement, environmental simulation, and multispectral fusion. Among them, infrared enhancement is mainly CLAHE (Contrast Limited Adaptive Histogram Equalization, ClipLimit = 2.0); environmental simulation is mainly adding synthetic humidity effects (simulating different seepage flow levels); multispectral fusion is mainly splicing the infrared and visible light image channels (multi-channel input).

[0143] Storage scheme

[0144] Use MinIO to build a private S3 storage to achieve the storage of data objects and comprehensively manage metadata through Elasticsearch. The indexed data attributes include but are not limited to resolution, acquisition time, weather conditions, etc.

[0145] Step ③. Construction of the sample library for dam damage (sign and label)

[0146] Data collection

[0147] a. Source

[0148] Image acquisition data of signs and labels near the video monitoring station; other video monitoring data such as UAV aerial video image data.

[0149] b. Annotation requirements

[0150] Bounding box detection: Annotate the location of the damaged sign (YOLO format).

[0151] Damage type classification: Multiclass labels (including but not limited to fracture, defacement, missing, etc.).

[0152] Data augmentation strategy

[0153] Data augmentation strategies include occlusion simulation, lighting variation, and affine transformation. Among them, occlusion simulation mainly randomly adds raindrops and mud occlusions (occlusion area ≤ 30%); lighting variation includes adjusting brightness (±30%) and contrast (±20%); affine transformation includes random rotation (-15° to +15°) and perspective distortion.

[0154] Storage scheme

[0155] Use MinIO to build a private S3 storage to achieve the storage of data objects and conduct comprehensive metadata management through Elasticsearch. Its indexed data attributes include but are not limited to resolution, acquisition time, weather conditions, etc.

[0156] Step ④, Construction of the damaged sample library of the dam body protection forest

[0157] Data collection

[0158] a. Source

[0159] Image acquisition data of the protection forest near the video monitoring station; other video monitoring data such as UAV aerial video image data.

[0160] b. Annotation requirements

[0161] Multi-class segmentation: Define vegetation status categories (including healthy forest belt vegetation, pest damage, damage, etc.).

[0162] Time series analysis: Compare multi-period images of the same area and annotate the damaged change areas.

[0163] Data augmentation strategies

[0164] Data augmentation strategies include band enhancement, season simulation, and random cropping. Among them, band enhancement mainly calculates the NDVI index (Normalized Difference Vegetation Index) to highlight vegetation features; season simulation mainly adjusts the RGB channels to simulate withered or newly grown vegetation; random cropping is used to implement multi-scale sliding windows.

[0165] Storage scheme

[0166] Use MinIO to build a private S3 storage to achieve the storage of data objects and conduct comprehensive metadata management through Elasticsearch. Its indexed data attributes include but are not limited to resolution, acquisition time, weather conditions, etc.

[0167] Step ⑤, Construction of the sample library for human behavior recognition

[0168] Data collection

[0169] a. Source

[0170] Image acquisition data of the behavior of people near the video monitoring station; other video surveillance data such as UAV aerial video image data.

[0171] b. Annotation requirements

[0172] Behavior category label: text (such as "walking", "fighting")

[0173] Time interval annotation: JSON format (start frame - end frame)

[0174] Key point / skeleton annotation: COCO or OpenPose format (17 key points)

[0175] Abnormal behavior annotation: binary classification label (normal / abnormal) + abnormal type description

[0176] Data augmentation strategy

[0177] The data augmentation strategy includes time - dimension augmentation and space - dimension augmentation. Among them, time - dimension augmentation includes time cropping, frame rate jitter, time series inversion, and dynamic frame sampling; space - dimension augmentation includes random occlusion, multi - perspective simulation, motion blur, and light perturbation.

[0178] Storage scheme

[0179] Build a private S3 storage using MinIO to store data objects and conduct comprehensive metadata management through Elasticsearch. Its indexed data attributes include but are not limited to resolution, acquisition time, weather conditions, etc.

[0180] Step 3: Tuning the weights of the training algorithm

[0181] Step ①: Selection of model architecture

[0182] Due to different construction ideas, common backbone models include ResNet - 50, EfficientNet - B4, Swin - Tiny, etc. After carefully comparing the parameter effects, this AI algorithm uses ResNet - 50 to construct the model skeleton and conducts separate optimization to improve efficiency; in addition, to improve the practicality of business applications, a multi - task branch design is carried out for multiple business directions to achieve the goal of recognizing multiple results at one time. The specific decomposition is as follows:

[0183] Module decomposition description

[0184] 1. Seepage segmentation branch (seepage_head)

[0185] Enlarge the resolution of the input feature map by 2 times (such as 16x16 → 32x32), and output the seepage area classification result of each pixel, which is suitable for tasks that require fine positioning.

[0186] 2. Vortex Detection Branch (vortex_head)

[0187] Spatial Compression: Extract local features related to vortices through convolution and pooling

[0188] Spatio-Temporal Transformation: Convert a 2D image into a time-series data stream of (h*w) spatial points

[0189] Dynamic Modeling: Use bidirectional analysis to characterize the variation of features over "time" (actually the sequence of spatial positions)

[0190] Global Judgment: Integrate all position information and output a binary classification result indicating whether a vortex is detected

[0191] Data Flow Demonstration

[0192] Assume the input feature map has dimensions [2, 512, 16, 16] (2 samples, 512 channels, 16x16 resolution):

[0193] 1. Seepage Branch:

[0194] 512 → [Convolution] → 256 channels → [Upsampling] → 32x32 → [Convolution] → 10 classes → Output [2, 10, 32, 32]

[0195] 2. Vortex Branch:

[0196] 512 → [Convolution] → 64 channels → [Pooling] → 8x8 → Serialize into 64 points of 64 dimensions → Analyze the sequence → Bidirectional feature concatenation → Linear layer → Output [2, 1] (detection probabilities for two samples)

[0197] Step ②: Definition of the Hyperparameter Search Space

[0198] In machine learning and deep learning, the definition of the hyperparameter search space is an important step in the optimization algorithm for finding the best model parameter configuration.

[0199] a. Determine the hyperparameters to be optimized

[0200] Identify the hyperparameters. According to requirements such as the task objective, the hyperparameters to be optimized this time are the learning rate (learning_rate) and the optimizer. The code is as follows:

[0201] Learning Rate (lr)

[0202] Function: Control the step size of model parameter updates.

[0203] Logic: Sample using a log-uniform distribution between 10^-5 (0.00001) and 10^-3 (0.001).

[0204] For example, the initial lr used in this training is 0.001, rather than values such as 3e-5 or 1e-4 that might be sampled, or values such as 0.0001 or 0.0002 from a uniform distribution.

[0205] Optimizer

[0206] Function: Adam, a variant of the basic optimizer algorithm, can effectively avoid or mitigate the occurrence of gradient descent.

[0207] Logic: Randomly select one from three candidate optimizers:

[0208] AdamW: An improved Adam that correctly handles weight decay.

[0209] RAdam: Adaptively adjusts the learning rate variance to avoid instability in the initial stage of training.

[0210] NAdam: Adam combined with Nesterov momentum.

[0211] Weight decay

[0212] Function: Controls the regularization strength of the model to prevent overfitting.

[0213] Logic: Sample using a uniform distribution between 10^-6 (0.000001) and 10^-3 (0.001).

[0214] For example, values such as 5e-5 or 2e-4 might be sampled.

[0215] Batch size

[0216] Function: When computing power is sufficient, the larger the batch size of the number of samples, the better. However, if the computing power is insufficient, defining an overly large batch size will seriously affect the algorithm speed.

[0217] Logic: Select one from three fixed values (16, 32, 64), with a default of 32.

[0218] b. Automatic hyperparameter tuning

[0219] Adjust parameters through automatic hyperparameter tuning to ensure efficiency. The code is as follows:

[0220] Import dependent libraries

[0221] Function: Import relevant modules of Ray Tune (hyperparameter tuning framework) and PyTorch (deep learning framework).

[0222] Function: Provide tool support for subsequent training and tuning.

[0223] Define the training function train_func

[0224] Function:

[0225] Build a model according to the configuration (config).

[0226] Dynamically select an optimizer (such as Adam or SGD).

[0227] Execute epochs (e.g., 100) of training, evaluate the model after each epoch, and save the checkpoint in the running memory.

[0228] Function: Define the complete training process under a single combination of hyperparameters and provide real-time feedback on performance metrics.

[0229] Define the search space search_space

[0230] Function: Specify the range of hyperparameters to be tuned.

[0231] Function: Define the parameter combination space that Ray Tune needs to explore and guide the automatic search process.

[0232] Configure and run the Tuner

[0233] Function:

[0234] Configure the ASHA scheduler to early-terminate poorly performing trials.

[0235] Specify the resource requirements (GPU) for each trial, the location to save the results, and the checkpoint strategy.

[0236] Function: Automatically manage the hyperparameter search process and efficiently allocate computing resources.

[0237] Result analysis

[0238] Function: Find the result with the highest validation set accuracy from all trials.

[0239] Function: Finally output the model performance corresponding to the optimal hyperparameter combination.

[0240] c. Training process monitoring

[0241] Training process monitoring includes two items: gradient health check and weight histogram recording.

[0242] The logic of the gradient health check is as follows:

[0243] Initialize the total gradient norm

[0244] Initialize total_norm to 0.0 for accumulating the gradient norms of all parameters.

[0245] Traverse the model parameters

[0246] Traverse each parameter p of the model:

[0247] If the parameter p has a gradient (p.grad is not None):

[0248] Calculate the gradient norm of a single parameter

[0249] Calculate the L2 norm (Euclidean norm) of the gradient of this parameter using p.grad.detach().norm(2).

[0250] Accumulate the squared norm

[0251] Accumulate the squared value of each gradient norm into total_norm.

[0252] Calculate the total gradient norm

[0253] After traversing all parameters, take the square root of total_norm to obtain the combined L2 norm of all gradients (i.e., the magnitude of the overall gradient).

[0254] Detect gradient explosion

[0255] If the total gradient norm exceeds the threshold 1e3 (i.e., 1000), trigger a warning Gradient explosion detected!, indicating that gradient explosion may have occurred to the user.

[0256] Weight histogram recording (using TensorBoard):

[0257] Import the TensorBoard tool

[0258] Create a log writer through the SummaryWriter class to generate the log files required by TensorBoard.

[0259] Initialize the log writer

[0260] Specify the log saving path (such as runs / experiment_name), and all recorded data will be stored in this directory.

[0261] Record the weights in the training loop

[0262] Loop through each epoch: At the end of each training cycle, record the model weights.

[0263] Traverse the model parameters: model.named_parameters() obtains the name of each layer of the model (such as layer1.weight) and the corresponding weight tensor.

[0264] Record Histogram: add_histogram writes the distribution of weights to the log in the form of a histogram. The tags are classified by weights / layer name (e.g., weights / layer1.weight), which is convenient for viewing by layer in TensorBoard.

[0265] Optional: Record Gradients

[0266] After uncommenting, the gradient distribution can be additionally recorded to help analyze the problem of gradient vanishing / explosion.

[0267] Close the Writer

[0268] Close the log writer after training to release resources.

[0269] Step 4: Model Training Results and Storage

[0270] Perform long-term model training on the constructed model and store the model training results to ensure its deployability and long-term maintainability. At the same time, as the data continues to be enriched and accumulated, the model training needs to be streamlined and automated, including triggering the update of the sample library to train the latest model results, accuracy calibration, parameter adjustment and optimization, result storage and recording, and cleaning of useless and discarded files at regular intervals every day.

[0271] Step ①: Training Process Automation

[0272] Saving Strategy:

[0273] 1. Function to save checkpoint (save_checkpoint)

[0274] Function: Package and save the training states such as the model, optimizer, and learning rate scheduler as files, and print the save path.

[0275] Key Parameters:

[0276] checkpoint: A dictionary containing model parameters, optimizer states, etc.

[0277] filename: The target file name (e.g., best_checkpoint_5.pt).

[0278] 2. Function to clean old checkpoints (clean_old_checkpoints)

[0279] Function: Regularly clean old checkpoint files and keep the specified number of the most recent files.

[0280] Key Parameters:

[0281] pattern: File matching pattern (e.g., "checkpoint_epoch_*.pt").

[0282] num_keep: Number of files to keep (e.g., keep the last 10 files).

[0283] Resume training script:

[0284] 1. Load checkpoint file:

[0285] Function: Load the saved model checkpoint from the file checkpoint_epoch_50.pt.

[0286] Details:

[0287] torch.load() deserializes the file into a dictionary containing model parameters, optimizer state, learning rate scheduler state, etc.

[0288] Checkpoints usually save the key states of training (e.g., model parameters at the 50th epoch).

[0289] 2. Resume model parameters:

[0290] Function: Load the saved model parameters into the current model.

[0291] Details:

[0292] checkpoint["model_state"] is the state dictionary of the model (containing weights and biases).

[0293] Ensure that the model structure is the same as when it was saved, otherwise an error will be reported.

[0294] 3. Resume optimizer state:

[0295] Function: Resume the internal state of the optimizer (such as momentum, gradient history, etc.).

[0296] Details:

[0297] Optimizers (such as Adam, SGD) rely on historical gradient information to update parameters and must be resumed to ensure the continuity of training.

[0298] 4. Resume learning rate scheduler state:

[0299] Function: Resume the state of the learning rate scheduler (such as StepLR, ReduceLROnPlateau).

[0300] Details:

[0301] The scheduler may record the current learning rate, adjustment steps, etc., and seamless connection of training can be achieved after resuming.

[0302] Step ②: Deployment Optimization and Monitoring

[0303] TensorRT Engine Building:

[0304] 1. Specify the file path of the input ONNX model.

[0305] 2. Specify the file path of the output TensorRT engine (the engine after INT8 quantization).

[0306] 3. Enable the INT8 quantization mode (calibration data is required).

[0307] 4. Specify the path of the calibration data (used to determine quantization parameters).

[0308] Performance Monitoring:

[0309] Monitor the CPU / GPU / memory usage rate and output it every 1 second.

[0310] Step ③: Model Version and Update Management

[0311] a. Version Control Strategy:

[0312] Training Phase:

[0313] The training script reads the parameters in config.yaml to initialize the model and trainer.

[0314] After training, export the model to the ONNX format and generate an evaluation report simultaneously.

[0315] Deployment Phase:

[0316] The inference service loads model.onnx to perform predictions.

[0317] Verify whether it meets the online standards (e.g., accuracy > 90%) through eval_results.txt.

[0318] Version Traceability:

[0319] The corresponding model + configuration + evaluation result combination can be accurately located through the version number (e.g., v1.1.0).

[0320] Upgrade the minor version number after modifying the configuration, and upgrade the major version number when the model structure is adjusted.

[0321] b. Automated Update Pipeline (GitLabCI Example):

[0322] Training Task (train_job)

[0323] Logic:

[0324] 1. Executed in the test phase, generates the model file best_model.pth and saves it to the output directory.

[0325] 2. Pass the output directory to subsequent tasks (such as deployment tasks) through artifacts.

[0326] Purpose: Automate the training and evaluation process and save model files for deployment.

[0327] Deployment Task (deploy_job)

[0328] logic:

[0329] 1. Pre-preparation (before_script):

[0330] Configure the SSH private key (read from the GitLab secret variable SSH_PRIVATE_KEY).

[0331] Restrict private key file permissions to prevent security risks.

[0332] Automatically trust the target device (avoiding manual confirmation on first connection).

[0333] 2 core operations (script):

[0334] Use scp to transfer the training product best_model.pth to the specified path of the edge device.

[0335] Purpose: Safely deploy the model to the production environment, triggering it only from the main branch to ensure process reliability.

[0336] Step 3: Construction of business functions of embankment inspection system:

[0337] The business functions of the levee inspection system are designed and constructed based on the inspection business logic process, mainly including video surveillance AI robot inspection functions and manual inspection functions. The specific steps are as follows:

[0338] Step 1: Build the video surveillance AI robot patrol function:

[0339] This function includes inspection tasks, inspection analysis, alarm handling and other sub-items. The specific contents are as follows:

[0340] 1. Inspection tasks: Build inspection task functions, including the formulation of normal inspection tasks (plans) and special inspection tasks (plans).

[0341] ① Normal inspection task (plan) formulation: According to the requirements of dike inspection management, the management personnel compile the inspection plan for the video AI robot and file it for implementation. During the implementation process, the system will call the dike video AI recognition algorithm developed in this patent for analysis and judgment, and the relevant results will be automatically synchronized to the inspection analysis function as records related to the normal inspection task (plan).

[0342] ② Special inspection task (plan) formulation: According to the early warning information including water regime, video, displacement, seepage pressure, etc. constructed by the dike management unit and the requirements of dike inspection management, the management personnel compile the special inspection plan for the video AI robot and file it for implementation. During the implementation process, the system will call the dike video AI recognition algorithm developed in this patent for analysis and judgment, and the relevant results will be automatically synchronized to the inspection analysis function as records related to the special inspection task (plan).

[0343] 2. Inspection analysis: Users can query the normal inspection tasks (plans) and special inspection tasks (plans) according to the inspection task name. The query content includes the plan name, construction time, execution time, problem content found during execution, etc. At the same time, for the problems that have been found, the video AI algorithm can be manually called to intercept video clips of the problems found during the dike inspection process, and then manually analyze and judge again, and edit the early warning content and disposal suggestions, or close the alarm.

[0344] 3. Alarm disposal function, including three functions: alarm query, general disposal, and professional disposal.

[0345] ① Alarm query: The alarm query function provides various alarm information found in the normal inspection tasks (plans) and special inspection tasks (plans). The system can automatically assign the alarms to the inspection personnel for on-site disposal, and at the same time supports the administrator to manually dispatch the alarm disposal and close the alarm in this function.

[0346] ② General disposal: The general disposal function mainly enables the inspection personnel to dispose of the alarms found in the inspection by the video AI robot dispatched by the system. The inspection personnel can view various alarm disposal tasks (including various basic information of the alarms) found in the normal inspection tasks (plans) and special inspection tasks (plans) and upload and edit the content after disposal, including but not limited to documents, texts, videos, pictures, etc. After the disposal is completed, the alarm will be automatically ended; if the inspection personnel cannot dispose of the alarm due to reasons such as the alarm involving the authority of the professional department, the information that cannot be disposed of will be uploaded to the administrator and the administrator will reassign the task.

[0347] ③ Professional Disposal: The professional disposal function provides the functions of querying and disposing of alarms that ordinary inspection personnel cannot handle. For alarms transferred from the ordinary disposal function to this professional disposal function, the administrator fills in the disposal opinions and disposal content for archiving, and contacts the professional disposal team for disposal. The disposal results feedback by the professional disposal team are filled in by the administrator or the professional disposal team. At this time, the alarm is terminated.

[0348] Step 2: The manual inspection function includes sub-items such as inspection system, manual inspection, and inspection statistics.

[0349] 1. Inspection System: This function provides the administrator with the function of uploading inspection-related system documents for users of all levels of units to query and preview the inspection management system online.

[0350] 2. Manual Inspection

[0351] ① Daily Inspection: Daily inspection includes inspection plan, inspection record, and problem handling. The inspection plan mainly realizes the customization of the inspection plan, and displays the inspection plan and its implementation status in the form of a calendar on a monthly basis; the inspection record mainly realizes that the process of each complete inspection forms an inspection record, which is summarized and displayed, supporting query, viewing, and supplementary recording of inspections; problem handling mainly realizes the summary display of the problems reported during the inspection. During the inspection process, the executor can call the embankment video AI recognition algorithm developed by this patent for auxiliary analysis and judgment. For problems that cannot be handled on the spot, further assignment, handling, and review can be carried out to achieve a closed-loop of problem handling.

[0352] ② Regular Inspection: Regular inspection mainly realizes the comprehensive inspection and problem recording of all facilities and projects within the jurisdiction after the flood season. During the inspection process, the executor can call the embankment video AI recognition algorithm developed by this patent for auxiliary analysis and judgment.

[0353] ③ Special Inspection: Special inspection mainly realizes the targeted inspection of key parts after special or extreme weather or disasters occur, and records the problems found. During the inspection process, the executor can call the embankment video AI recognition algorithm developed by this patent for auxiliary analysis and judgment.

[0354] 3. Inspection Statistics: Summarize and statistically calculate the number of inspections participated by each inspector, as well as data such as inspection duration and mileage generated by inspections.

[0355] Based on the existing business requirements for levee inspection and management, this invention utilizes the existing relatively complete video surveillance system to construct various video recognition algorithms that meet the safety management requirements of levees. It realizes video robot inspections based on video AI algorithms and video AI algorithm-assisted manual inspections in a low-cost manner, and constructs a levee inspection and management system centered on video AI recognition algorithms, fully exploring the potential of the existing video surveillance equipment and facilities of levee management units, initially realizing the informatization and intelligentization of levee inspection and management, and improving the management level of levee management units.

[0356] Based on the video surveillance systems owned by most existing levee project management units, this invention constructs a levee inspection and management system that meets the inspection and management requirements of levee management units by developing inspection and management business applications based on video AI recognition algorithms for levee safety monitoring and analysis, and improves the intelligentization ability of levee inspection and management in a low-cost manner to ensure the timely discovery and handling of levee safety problems.

[0357] This invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0358] In the description of this invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of this invention, "a plurality" means two or more, unless otherwise specifically defined.

[0359] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalent technologies, the invention is also intended to include these modifications and variations.

Claims

1. A method for dike inspection and management based on video recognition AI algorithm, characterized in that, Including the following steps: Based on the historical and real-time video monitoring data obtained from the existing video monitoring system of the dike, learning and training are carried out through the convolutional network algorithm to construct a video recognition AI model and obtain a training sample library; Based on the inspection business logic process of the dike, by dynamically collecting the video data of the dike, using the video recognition AI model, and based on the training sample library, the dike is inspected and managed.

2. The method for dike inspection and management based on the video recognition AI algorithm according to claim 1, wherein: When inspecting and managing the dike, according to the inspection business logic, combined with the video recognition AI model, an inspection business logic process is constructed, including: the video monitoring AI robot inspection business flow and the manual inspection business flow. Among them, the video monitoring AI robot inspection business flow represents the process of performing unmanned automatic inspection by carrying a video acquisition device, and the manual inspection business flow represents the process of assisting manual dike inspection through the video recognition AI model.

3. The method for dike inspection and management based on the video recognition AI algorithm according to claim 1, wherein: When performing model training and learning, by configuring a distributed training network, the convolutional network algorithm is trained based on the basic data to construct the video recognition AI model.

4. The method for dike inspection and management based on the video recognition AI algorithm according to claim 3, wherein: When constructing the training sample library, after annotating and enhancing the basic data, the training sample library is generated through the video recognition AI model. Among them, annotation is performed through pixel-level segmentation and multi-modal alignment, and data enhancement processing is performed through infrared enhancement, environment simulation, and multi-spectral fusion.

5. The method for dike inspection and management based on the video recognition AI algorithm according to claim 4, wherein: When constructing the training sample library, the training sample library is constructed by obtaining the dam seepage video sample library, the dam damage sample library, the dam protection forest damage sample library, and the personnel behavior recognition library sample library.

6. The method for dike inspection and management based on the video recognition AI algorithm according to claim 5, wherein: When constructing the video recognition AI model, after optimizing the model through the definition of the hyperparameter search space, the gradient health check and the weight histogram are used to check the model training process.

7. The method for dike inspection and management based on the video recognition AI algorithm according to claim 6, wherein: When performing model training, after saving the function of the checkpoint and cleaning the function of the old checkpoint, the checkpoint file is loaded and the model parameters, optimizer state, and learning rate scheduler state are restored; By constructing a TensorRT engine, setting a version control policy and an automated update pipeline, the model is trained and the results of the model training are stored in the database for preservation.

8. The method for dike inspection and management based on the video recognition AI algorithm according to claim 2, wherein: When constructing the video monitoring AI robot inspection business flow, the video monitoring AI robot inspection business flow includes the following steps: Step 1.1: Develop video robot patrol tasks: Based on daily management work, formulate regular video robot patrol tasks and generate daily patrol routes. Based on daily monitoring and early warning information, formulate special video robot patrol tasks and generate special patrol routes. File the generated tasks separately to form corresponding patrol tasks. Step 1.2: Start the inspection task: According to the requirements of the regular inspection task, complete the inspection according to the regular inspection task inspection route and archive the inspection record information; according to the requirements of the special inspection task, complete the special inspection according to the special inspection task inspection route and archive the special inspection record information; Step 1.3: Automatically call the video AI recognition algorithm to analyze and judge the video image information collected during the video robot's patrol mission; Step 1.4: Complete the inspection and handling work according to the analysis and judgment results: If there is no problem, the inspection record will be directly archived and the inspection task is completed; if there is a problem, it will be reported to the administrator based on the problem identified by the video, and the administrator will assign personnel to handle it; Among them, general problems are assigned to the inspection team for inspection and handling, and professional problems are assigned to the professional handling team for professional problem handling.

9. The embankment inspection and management method based on video recognition AI algorithm according to claim 8 is characterized by: When constructing a manual inspection business flow, the manual inspection business flow includes the following steps: Step 2.1: Formulate manual inspection tasks: formulate manual inspection tasks in combination with daily management work and generate daily manual inspection plans; Step 2.2: The inspectors conduct inspections according to the manual inspection plan. The inspectors complete the inspection judgment based on the actual inspection situation. If the inspection is determined not to be carried out, the inspection process ends. If the inspection is determined to be carried out, the normal inspection process will be entered into the next step. Step 2.3: Patrol personnel conduct normal patrols: Patrol personnel conduct normal patrols as required. During the patrol process, the administrator delegates the use of the video surveillance system and video AI intelligent analysis algorithm to patrol personnel to assist in manual patrol work. Step 2.4: The inspectors make judgments on the problems found during the inspection: if a problem is found and the inspectors can handle it themselves, the inspectors will handle it according to the daily management regulations, and record and archive the handling situation, and then the manual inspection work ends; if a problem is found and the inspectors cannot handle it themselves, the inspectors will report it to the administrator according to the management regulations, and the administrator will assign a professional team to handle it. The handling situation will be summarized as required and collected by the reviewers. The reviewers will manually archive it according to the handling results of the professional team or automatically archive it according to the time regulations, and then the entire inspection process ends.

10. A dike inspection and management system based on a video recognition AI algorithm, characterized in that, The system is used to implement a dike inspection and management method based on a video recognition AI algorithm as described in any one of claims 1 to 9, and the system includes: The intelligent model training module is used to build a video recognition AI model using historical and real-time video surveillance data obtained from the existing video monitoring system of the dike through convolutional network algorithms for training and learning, and to obtain a training sample library. The inspection management module is used to conduct inspection management on the levee based on the inspection business logic process of the levee, by dynamically collecting video data of the levee, using the video recognition AI model, and based on the training sample library.

Images