A Method and System for Identifying Abnormal Behavior in Wind Farms Based on Multi-Model Concatenation

Abstract

This invention discloses a method and system for identifying abnormal behavior in wind farms based on multi-model cascaded calls, belonging to the field of abnormal behavior identification technology. The method includes: acquiring real-time video streams and determining scene semantic information; constructing an anomaly identification task graph; adaptively filtering and arranging a wind farm model library; obtaining a model cascaded call chain; performing multi-level progressive detection on the real-time video stream according to the model cascaded call chain; obtaining the output sequences of each model; performing error assessment and source tracing; and establishing a wind farm identification interference tree; performing multi-dimensional redundancy verification on the output sequences of each model; and establishing a wind farm identification redundancy verification space; and performing credibility enhancement processing on the output sequences of each model based on the wind farm identification interference tree and the wind farm identification redundancy verification space; and establishing a wind farm abnormal behavior heatmap. This invention solves the technical problem of insufficient accuracy in wind farm abnormal behavior identification in existing technologies, achieving the technical effect of improving the accuracy of wind farm abnormal behavior identification.

Description

Technical Field

[0001] This invention relates to the field of abnormal behavior recognition technology, specifically to a method and system for identifying abnormal behavior in wind farms based on multi-model serial invocation. Background Technology

[0002] Wind farm operating environments are typically characterized by large monitoring areas, dispersed equipment distribution, significant variations in sunlight and weather, large differences in target scale, and complex types of abnormal behavior. Monitoring videos are easily affected by factors such as rain, fog, strong sunlight, shadows, blade rotation, and obstructions from personnel and vehicles. Traditional abnormal behavior identification methods often rely on fixed detection models or single identification processes, making it difficult to flexibly adjust identification strategies according to different monitoring scenarios and anomaly types. This leads to insufficient coordination between target anchoring, behavior judgment, and progressive anomaly analysis, resulting in missed detections, false detections, or biased anomaly severity assessments, thus affecting the accuracy of abnormal behavior identification in wind farms. Summary of the Invention

[0003] This application provides a method and system for identifying abnormal behavior in wind farms based on multi-model serial invocation, which is intended to address the technical problem of insufficient accuracy in identifying abnormal behavior in wind farms in the prior art.

[0004] In view of the above problems, this application provides a method and system for identifying abnormal behavior in wind farms based on multi-model serial invocation.

[0005] The first aspect of this application provides a method for identifying abnormal behavior in wind farms based on multi-model serial invocation, the method comprising: The system acquires real-time video streams from a wind farm monitoring platform and determines the scene semantic information of these streams. Based on this semantic information, it constructs an anomaly identification task graph and adaptively filters and arranges the wind farm model library according to the graph, obtaining a model chain. It then performs multi-level progressive detection on the real-time video streams according to the model chain, obtaining the output sequences of each model. Each output sequence includes a target anchoring sequence, an anomaly behavior identification sequence, and an anomaly progression characteristic sequence. Error assessment and source tracing are performed on the output sequences to establish a wind farm identification interference tree. Multi-dimensional redundancy verification is performed on the output sequences to establish a wind farm identification redundancy verification space. Finally, based on the wind farm identification interference tree and the redundancy verification space, the system performs credibility enhancement processing on the output sequences to establish a wind farm anomaly behavior heatmap.

[0006] A second aspect of this application provides a wind farm abnormal behavior identification system based on multi-model serial invocation, the system comprising: The system includes a data acquisition module for acquiring real-time video streams from a wind farm monitoring platform and determining the scene semantic information of the video streams; a filtering and arrangement module for constructing an anomaly recognition task graph based on the scene semantic information and adaptively filtering and arranging the wind farm model library based on the anomaly recognition task graph to obtain the model chain; a progressive detection module for performing multi-level progressive detection on the real-time video streams according to the model chain to obtain the output sequences of each model, including target anchoring sequences, anomaly behavior recognition sequences, and anomaly progressive characteristic sequences; an evaluation and source tracing module for performing error evaluation and source tracing on the output sequences of each model to establish a wind farm recognition interference tree; a redundancy verification module for performing multi-dimensional redundancy verification on the output sequences of each model to establish a wind farm recognition redundancy verification space; and a heatmap creation module for performing credibility enhancement processing on the output sequences of each model based on the wind farm recognition interference tree and the wind farm recognition redundancy verification space to establish a wind farm anomaly behavior heatmap.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application acquires real-time video streams from a wind farm monitoring platform and determines the scene semantic information of the real-time video streams; constructs an anomaly recognition task graph based on the scene semantic information, and adaptively filters and arranges the wind farm model library according to the anomaly recognition task graph to obtain a model chain; performs multi-level progressive detection on the real-time video streams according to the model chain to obtain the output sequences of each model, wherein each model output sequence includes a target anchoring sequence, an anomaly behavior recognition sequence, and an anomaly progressive characteristic sequence; performs error assessment and source tracing on each model output sequence to establish a wind farm recognition interference tree; performs multi-dimensional redundancy verification on each model output sequence to establish a wind farm recognition redundancy verification space; and performs credibility enhancement processing on each model output sequence based on the wind farm recognition interference tree and the wind farm recognition redundancy verification space to establish a wind farm anomaly behavior heatmap. This invention addresses the technical problem of insufficient accuracy in identifying abnormal behavior in wind farms in existing technologies. By adaptively constructing a model chain based on scene semantic information and performing multi-level progressive detection and error interference correction on real-time acquired video streams, it achieves the technical effect of improving the accuracy of identifying abnormal behavior in wind farms. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic flowchart of a wind farm abnormal behavior identification method based on multi-model serial invocation provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of a wind farm abnormal behavior recognition system based on multi-model serial invocation provided in an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, filtering and arrangement module 12, progressive detection module 13, evaluation and tracing module 14, redundancy verification module 15, heat map creation module 16. Detailed Implementation

[0011] This application provides a method and system for identifying abnormal behavior in wind farms based on multi-model serial invocation. It addresses the technical problem of insufficient accuracy in identifying abnormal behavior in wind farms in existing technologies by adaptively constructing a model serial invocation chain based on scene semantic information and performing multi-level progressive detection and error interference correction on real-time acquired video streams, thereby improving the technical effect of identifying abnormal behavior in wind farms.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides a method for identifying abnormal behavior in wind farms based on multi-model serial invocation, the method comprising: Step S100: Obtain the real-time video stream from the wind farm monitoring platform and determine the scene semantic information of the real-time video stream.

[0015] Furthermore, the method provided in the application embodiments also includes: The scene semantic information includes the monitoring area type, target detection area, and identifiable anomaly type corresponding to the real-time acquired video stream.

[0016] In this embodiment of the application, in order to obtain the real-time video stream from the wind farm monitoring platform, the video acquisition device in the wind farm is accessed through the wind farm monitoring platform, and the video image data continuously acquired and uploaded by the video acquisition device during the current monitoring period is read to obtain the real-time video stream; at the same time, the video source identifier, camera number, acquisition timestamp and camera location information corresponding to the real-time video stream are read.

[0017] Next, the semantic information of the scene for real-time video streaming is determined. Specifically, firstly, based on the video source identifier and camera number, a match is made in the pre-stored camera location file on the wind farm monitoring platform to determine the camera installation location and monitoring range corresponding to the real-time video stream; then, based on the camera installation location and monitoring range, the monitoring area type corresponding to the real-time video stream is determined. The monitoring area type includes wind turbine tower area, blade rotation area, transformer substation area, booster station area, maintenance road area, perimeter fence area, or personnel operation area.

[0018] After determining the type of monitoring area, monitoring frames are extracted from the real-time video stream according to the preset frame extraction frequency. Based on the pre-marked monitoring boundaries, equipment boundaries, road boundaries, fence boundaries, or work boundaries in the camera location file, the monitoring frames are divided into regions to obtain the target detection area. The target detection area is the image area in the real-time video stream that needs to be identified and analyzed for target recognition, including the tower entrance area, blade sweep area, transformer substation perimeter area, substation equipment area, road passage area, fence inner and outer areas, or maintenance work area.

[0019] After obtaining the monitoring area type and target detection area, these are matched against a preset anomaly identification rule base to determine the identifiable anomaly types corresponding to the real-time video stream. Identifiable anomaly types include personnel intrusion, abnormal vehicle parking, maintenance personnel working beyond designated boundaries, foreign object accumulation, perimeter intrusion, abnormal approach to blade areas, abnormal smoke or fire, abnormal equipment appearance, or road obstruction. The monitoring area type, target detection area, and identifiable anomaly types are then associated and recorded to obtain the scene semantic information of the real-time video stream.

[0020] Step S200: Construct an anomaly recognition task graph based on the scene semantic information, and adaptively filter and arrange the wind farm model library based on the anomaly recognition task graph to obtain the model serial call chain.

[0021] In this embodiment, when constructing an anomaly recognition task graph based on scene semantic information, the monitoring area type, target detection area, and identifiable anomaly type in the scene semantic information are first read. The monitoring area type is used as the scene node of the task graph, the target detection area is used as the detection area node of the task graph, and the identifiable anomaly type is used as the anomaly task node of the task graph. The anomaly recognition task graph is used to represent the relationship between the scene, detection area, and anomaly recognition task corresponding to the current real-time video stream. Subsequently, based on the monitoring area type and target detection area, the target categories that need to be prioritized for detection are determined, such as personnel, vehicles, foreign objects, smoke, or equipment parts, and the target categories are added to the anomaly recognition task graph as target recognition nodes. Then, based on the identifiable anomaly type, the judgment conditions required for each abnormal behavior are determined, such as whether the target has entered the target detection area, whether the target's dwell time exceeds a preset time, whether the target's movement trajectory crosses a preset boundary, and whether the target's appearance is abnormal. The judgment conditions are added to the anomaly recognition task graph as anomaly judgment nodes. Finally, the scene nodes, detection area nodes, target recognition nodes, anomaly task nodes, and anomaly judgment nodes are connected in the order of recognition to obtain the anomaly recognition task graph.

[0022] Next, the wind farm model library is adaptively filtered and arranged based on the anomaly identification task map. This process involves first matching and filtering the wind farm model library according to the anomaly identification task map to obtain multiple candidate models suitable for the current anomaly identification task; then, anomaly behavior identification is decomposed based on the anomaly identification task map to obtain anomaly pre-identification conditions and anomaly post-judgment conditions; subsequently, based on the anomaly pre-identification conditions and anomaly post-judgment conditions, multiple candidate models are sequentially arranged and their input / output connections are configured to obtain the initial model call chain; finally, the initial model call chain is configured with backup and anomaly self-recovery features to generate a model chain that can be called sequentially according to the identification process.

[0023] Furthermore, in the method provided in the application embodiments, the adaptive filtering and arrangement of the wind farm model library based on the anomaly identification task map to obtain the model serial call chain also includes: The wind farm model library is matched and filtered according to the anomaly identification task graph to obtain multiple candidate models; anomaly behavior identification is decomposed according to the anomaly identification task graph to obtain anomaly pre-identification conditions and anomaly post-judgment conditions; the multiple candidate models are sequentially arranged and input / output connection configured according to the anomaly pre-identification conditions and the anomaly post-judgment conditions to obtain the model call initial chain; the model call initial chain is configured with backup and anomaly self-recovery to generate the model serial call chain.

[0024] In this embodiment, when matching and filtering the wind farm model library based on the anomaly identification task map, the monitoring area type, target detection area, identifiable anomaly type, target identification node, and anomaly judgment node in the anomaly identification task map are read first. Then, the applicable area, model identification object, model input format, model output format, and applicable anomaly type of each model in the wind farm model library are read. Subsequently, the monitoring area type is compared with the applicable area of ​​the model, the target detection area and target identification node are compared with the model identification object, and the identifiable anomaly type and anomaly judgment node are compared with the applicable anomaly type of the model. It is also determined whether the model input format and model output format meet the requirements for serial calling. Models that match the comparison and meet the format requirements are determined as candidate models, resulting in multiple candidate models.

[0025] Next, when decomposing abnormal behavior identification based on the anomaly identification task graph, we first take the identifiable anomaly type as the object and find its associated target identification node, detection area node, and anomaly judgment node in the anomaly identification task graph; then, we decompose it in the order of target identification, target localization, target tracking, and anomaly confirmation. The target category identification, target location identification, target area attribution identification, and target motion state identification that need to be completed before anomaly confirmation are determined as the anomaly pre-identification conditions; the anomaly post-judgment conditions are determined as whether the target enters the target detection area, whether the target stays for more than a preset time, whether the target trajectory crosses a preset boundary, whether the target behavior state meets the anomaly rules, and whether the anomaly behavior is continuous.

[0026] Then, based on the anomaly pre-identification conditions and anomaly post-judgment conditions, multiple candidate models are sequentially arranged and configured with input-output connections. In this process, firstly, a model satisfying the anomaly pre-identification conditions is selected from multiple candidate models and configured at the pre-call position; then, a model satisfying the anomaly post-judgment conditions is selected from multiple candidate models and configured at the post-call position; subsequently, multiple candidate models are arranged in the order of target detection, target localization, target tracking, behavior recognition, and anomaly judgment. After the sequential arrangement is completed, the model output format and model input format of adjacent candidate models are read. When the target category, target location, target bounding box, target trajectory, or target state output by the previous-level candidate model can be used as input for the next-level candidate model, a direct transmission relationship is established; when the formats are inconsistent, the output results of the previous-level candidate model are processed by field correspondence, coordinate transformation, timestamp alignment, or data normalization before being passed to the next-level candidate model, thus obtaining the initial model call chain.

[0027] Finally, when configuring backup and anomaly self-recovery for the initial model call chain, backup models are first searched in the wind farm model library for key candidate models in the initial model call chain. Backup models and corresponding candidate models have the same applicable region, model identification object, and applicable anomaly type, and their input and output formats are the same or can be converted and connected. Then, a replacement relationship is established between candidate and backup models, and switching conditions are set for model execution failure, empty output, output confidence below a preset threshold, model call timeout, or incompatible output formats. Simultaneously, anomaly self-recovery processing is configured for the initial model call chain. When input data is missing, input / output formats do not match, or adjacent model connections fail, re-calling, field matching, coordinate transformation, timestamp alignment, backup model switching, or previous node output rollback processing are performed. After completion, a model serial call chain is generated.

[0028] Step S300: Perform multi-level progressive detection on the real-time acquired video stream according to the model serial call chain, and obtain the output sequence of each model. The output sequence of each model includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence.

[0029] In this embodiment, when performing multi-level progressive detection on the real-time acquired video stream according to the model concatenation call chain, the real-time acquired video stream is first enhanced and frame extracted to obtain each enhanced monitoring frame of the wind farm; then, the enhanced monitoring frames are anchored to the detection target according to the model concatenation call chain to obtain the target anchoring sequence; subsequently, abnormal behavior identification is performed on the target anchoring sequence according to the model concatenation call chain to obtain the abnormal behavior identification sequence; finally, the abnormal behavior identification sequence is subjected to temporal progressive analysis according to the model concatenation call chain to obtain the abnormal progressive characteristic sequence. The resulting model output sequences include the target anchoring sequence, the abnormal behavior identification sequence, and the abnormal progressive characteristic sequence.

[0030] Furthermore, in the method provided in the application embodiment, multi-level progressive detection is performed on the real-time acquired video stream according to the model serial call chain to obtain the output sequence of each model. The output sequence of each model includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence, and further includes: The real-time acquired video stream is enhanced and frames are extracted to obtain enhanced monitoring frames of the wind farm; target anchoring is performed on each enhanced monitoring frame according to the model's serial call chain to obtain the target anchoring sequence; abnormal behavior identification is performed on the target anchoring sequence according to the model's serial call chain to obtain the abnormal behavior identification sequence; and time-series progressive analysis is performed on the abnormal behavior identification sequence according to the model's serial call chain to obtain the abnormal progressive characteristic sequence.

[0031] In this embodiment, when enhancing and extracting frames from the real-time acquired video stream, the continuous video frames in the real-time acquired video stream are first sorted according to the acquisition timestamp, and monitoring frames are extracted from the sorted continuous video frames according to a preset frame extraction frequency. The preset frame extraction frequency is set based on the video frame rate, target movement speed, and abnormal behavior recognition requirements of the wind farm monitoring platform. Subsequently, the monitoring frames undergo brightness equalization, contrast enhancement, image denoising, motion blur correction, and resolution normalization. Brightness equalization uses histogram equalization to adjust the grayscale distribution of the monitoring frames; contrast enhancement uses adaptive histogram equalization to enhance the grayscale difference between the target and the background; image denoising uses median filtering and Gaussian filtering to remove noise from the monitoring frames; motion blur correction uses Wiener filtering to restore blurred areas; and resolution normalization uses bilinear interpolation to adjust the monitoring frames to the input size required by the preceding model in the model chain. After completing the above processing, the processed monitoring frames are numbered according to the acquisition timestamp to obtain the enhanced monitoring frames of the wind farm.

[0032] When detecting and anchoring targets in each enhanced monitoring frame according to the model chain, each enhanced monitoring frame is first input into the pre-level model used for target anchoring in the model chain. The pre-level model processes each enhanced monitoring frame and outputs the target category, target bounding box, and target confidence score. Subsequently, the target bounding box is input into the subsequent model used for location resolution in the model chain. The subsequent model outputs the target center coordinates based on the top-left and bottom-right corner coordinates of the target bounding box and the image size, and outputs the target region attribution result in conjunction with the target detection region. Then, the target category, target bounding box, target center coordinates, target confidence score, and target region attribution result in each enhanced monitoring frame are arranged according to the acquisition timestamp. The same target in consecutive enhanced monitoring frames is associated through the model used for target association in the model chain, and the continuous anchoring results of the same target in consecutive enhanced monitoring frames are output. The continuous anchoring results are combined in chronological order to obtain the target anchoring sequence.

[0033] When identifying abnormal behavior in a target anchoring sequence based on a model chain, the target anchoring sequence is first input into the model for abnormal behavior identification within the chain. This model then outputs the target trajectory, dwell time, direction of movement, and area change information for the same target in continuous enhanced monitoring frames, based on the target category, bounding box, center coordinates, confidence level, and region attribution result in the target anchoring sequence. Subsequently, the target trajectory, dwell time, direction of movement, and area change information are input into the model for anomaly determination within the chain. This model, combined with the anomaly post-determination conditions corresponding to the identifiable anomaly types, outputs the abnormal behavior identification results for personnel intrusion, abnormal vehicle dwelling, maintenance personnel working beyond designated boundaries, foreign object accumulation, perimeter intrusion, abnormal approach to blade areas, abnormal smoke / fire, abnormal equipment appearance, and road obstruction. Finally, the abnormal behavior identification sequence is obtained by arranging the abnormal behavior category, anomaly occurrence area, abnormal target, anomaly confidence level, and anomaly determination result according to the acquisition timestamp.

[0034] When performing time-series progressive analysis on anomaly behavior identification sequences based on the model's serial call chain, the abnormal behavior category, anomaly occurrence area, anomaly target, anomaly confidence level, and anomaly judgment result for consecutive moments in the anomaly behavior identification sequence are first read according to the acquisition timestamp. The anomaly behavior identification results corresponding to the same anomaly target and the same anomaly occurrence area are then merged. Subsequently, the merged anomaly behavior identification results are input into the model used for time-series progressive analysis in the model's serial call chain. This model outputs the duration of the anomaly behavior, the frequency of the anomaly behavior, the change in the trajectory of the anomaly target, the change in the anomaly occurrence area, and the change in the anomaly confidence level based on a preset time window. Specifically, the duration of the anomaly behavior is obtained from the acquisition timestamps of the first and last occurrences of the anomaly judgment result; the frequency of the anomaly behavior is obtained from the number of times the anomaly judgment result appears within the preset time window; the change in the trajectory of the anomaly target is obtained from the change in the center coordinates of consecutive targets; the change in the anomaly occurrence area is obtained from the change in the target area attribution result; and the change in the anomaly confidence level is obtained from the change in the anomaly confidence level for consecutive moments. Finally, the model used for time-series progressive analysis outputs the progressive status corresponding to each collection timestamp based on the duration of abnormal behavior, frequency of abnormal behavior, changes in the trajectory of abnormal targets, changes in the area where abnormal occurs, and changes in the confidence level of abnormality. The progressive status, duration of abnormal behavior, frequency of abnormal behavior, changes in the trajectory of abnormal targets, changes in the area where abnormal occurs, and changes in the confidence level of abnormality are arranged in chronological order to obtain the abnormal progressive characteristic sequence.

[0035] Step S400: Perform error assessment and source tracing on the output sequences of each model, and establish a wind farm identification interference tree.

[0036] In this embodiment, when performing error assessment and source tracing on the output sequences of each model, cross-model correlation analysis is first performed on the output sequences of each model to generate an error candidate set; then, based on the error candidate set, the temporal location, spatial location, and model hierarchical location of the error are jointly located to construct the error occurrence node; subsequently, based on the error occurrence node, the propagation relationship of the error in the model serial call chain is traced in reverse to obtain the error propagation path; then, based on the real-time acquired video stream, interference factors are attributed to the error propagation path to determine the interference triggering conditions; finally, based on the error occurrence node, the error propagation path, and the interference triggering conditions, a wind farm identification interference tree is constructed.

[0037] Furthermore, the method provided in the application embodiments, which performs error assessment and source tracing on the output sequences of each model and establishes a wind farm identification interference tree, further includes: Cross-model correlation analysis is performed on the output sequences of each model to generate an error candidate set; based on the error candidate set, the temporal location, spatial location, and model hierarchical location of the error are jointly located to construct the error occurrence node; the error propagation path is traced in reverse according to the error occurrence node; interference factors are attributed to the error propagation path based on the real-time acquired video stream to determine the interference triggering condition; the wind farm identification interference tree is constructed based on the error occurrence node, the error propagation path, and the interference triggering condition.

[0038] In this embodiment of the application, when performing cross-model correlation analysis on the output sequences of each model, the collection timestamp, target category, target location, target bounding box, target center coordinates, target region attribution result, abnormal behavior result, abnormal occurrence area, abnormal confidence level, and abnormal progression status of the target anchoring sequence, abnormal behavior identification sequence, and abnormal progression characteristic sequence are read first. Then, based on the collection timestamp, the target anchoring sequence, abnormal behavior identification sequence, and abnormal progression characteristic sequence within the same time range are aligned, and whether they belong to the same target is determined by the overlap range of the target center coordinates and the target bounding box. Specifically, the overlap ratio between the bounding boxes of the targets in the target anchoring sequence and the bounding boxes of the abnormal targets in the abnormal behavior recognition sequence is calculated. When the overlap ratio is lower than a preset overlap threshold, the target positions are determined to be inconsistent. The target region attribution result in the target anchoring sequence is compared with the abnormal occurrence region in the abnormal behavior recognition sequence. When the two are inconsistent, the region output is determined to be inconsistent. The abnormal behavior result in the abnormal behavior recognition sequence is compared with the abnormal progression state in the abnormal progression characteristic sequence. When the abnormal behavior result is no abnormality but the abnormal progression state shows continuous abnormality, or the abnormal behavior result is abnormality but the abnormal progression state shows no progression state, the abnormal output is determined to be inconsistent. The collection timestamp, target category, target location, abnormal occurrence region, abnormal behavior result, abnormal confidence, and model call level corresponding to the above inconsistent results are summarized to generate an error candidate set.

[0039] Next, based on the error candidate set, the temporal location, spatial location, and model hierarchical location of the error are jointly located. In this process, the acquisition timestamp, augmented monitoring frame number, target bounding box, target center coordinates, target region attribution result, anomaly occurrence region, and model call level are read item by item from the error candidate set. Then, the temporal location of the error is determined based on the acquisition timestamp and augmented monitoring frame number; the coordinate position of the error in the image is determined based on the target bounding box and target center coordinates; the region location corresponding to the error is determined based on the target region attribution result and the anomaly occurrence region; and the model hierarchical location of the error is determined based on the level of the model that produced the inconsistent output. Subsequently, error information under the same acquisition timestamp, spatial location, and model call level is merged, retaining the corresponding target category, abnormal behavior result, and anomaly confidence. When the same error information simultaneously possesses temporal location, spatial location, and model hierarchical location, it is constructed as an error occurrence node.

[0040] Then, based on the error occurrence node, the error propagation path is traced backward. In this process, the model hierarchy corresponding to the error occurrence node is first determined, and the model is traced level by level along the input-output connection relationship of the model's serial call chain. During each level of tracing, the output results of the preceding model are read under the same acquisition timestamp, the same target category, and the same spatial location, and compared with the input data corresponding to the current error occurrence node. When determining target position offset, the distance between the target center coordinates output by the preceding model and the target center coordinates received by the subsequent model is calculated. If this distance exceeds a preset position offset threshold, a target position offset is determined to exist. When determining target category error, the target category output by the preceding model is compared with the target category used for anomaly detection by the subsequent model. If they are inconsistent, a target category error is determined to exist. When determining target region assignment error, the target region assignment result output by the preceding model is compared with the target detection region where the target center coordinates actually fall. If the target center coordinates are located within a certain target detection region but the output target region assignment result is for another region, a target region assignment error is determined to exist.

[0041] When an abnormal behavior result is incorrectly determined, the abnormal behavior result output by the preceding model is compared with the subsequent abnormal judgment conditions. If the target enters the target detection area, the dwell time exceeds a preset time, the movement trajectory crosses a preset boundary, and the preceding model output does not show any abnormality, a missed judgment is determined. If the target does not enter the target detection area, the dwell time does not exceed a preset time, the movement trajectory does not cross a preset boundary, but the preceding model output shows an abnormality, a misjudgment is determined. When an abnormal confidence level is determined to be abnormal, the abnormal confidence level output by the preceding model is read and compared with a preset confidence threshold and the abnormal confidence level of adjacent acquisition timestamps. If the abnormal confidence level is lower than the preset confidence threshold but the subsequent model continues to generate abnormal results, or the change in abnormal confidence level of adjacent acquisition timestamps exceeds a preset change threshold, an abnormal confidence level is determined to be abnormal. When determining an abnormal progressive state error, the abnormal progressive state in the abnormal progressive characteristic sequence is compared with the continuous abnormal behavior results in the abnormal behavior identification sequence. If multiple consecutive acquisition timestamps show abnormal behavior results but the abnormal progressive state does not show continuity, or if only a single acquisition timestamp shows abnormal behavior results but the abnormal progressive state shows continuity, an abnormal progressive state error is determined. If there is an input-output connection between the aforementioned error in the preceding model and the current error occurrence node, the corresponding model level position of the preceding model is added to the error propagation path, and this process continues until the preceding model no longer has a corresponding error, thus obtaining the error propagation path.

[0042] Next, interference factors are attributed to the error propagation path based on the real-time acquired video stream. Specifically, the corresponding video frames and adjacent video frames are extracted from the real-time acquired video stream based on the acquisition timestamp, enhanced monitoring frame number, and spatial location in the error propagation path; then, the video frames and adjacent video frames are analyzed for brightness, occlusion, sharpness, target scale, background motion, and image continuity. When determining a sudden change in illumination, the average brightness difference between adjacent video frames is calculated. If the average brightness difference exceeds a preset brightness change threshold, a sudden change in illumination is determined. When determining target occlusion, the proportion of the visible target area within the target bounding box to the target bounding box is calculated. If this proportion is lower than a preset visible proportion threshold, target occlusion is determined. When determining target blur, the sharpness value of the image area within the target bounding box is calculated. If the sharpness value is lower than a preset sharpness threshold, target blur is determined. When determining target scale change, the proportion of change in the target bounding box area in adjacent video frames is calculated. If the proportion of change exceeds a preset scale change threshold, target scale change is determined. When determining background motion interference, the magnitude of change in background pixels within the target detection area in adjacent video frames is calculated. If the magnitude of change exceeds a preset background change threshold, background motion interference is determined. When determining consecutive frame loss, the acquisition timestamp interval of adjacent video frames is compared. If the timestamp interval exceeds a preset frame interval threshold, consecutive frame loss is determined. When determining image jitter, the position offset of the fixed device boundary in adjacent video frames is compared. If the position offset exceeds a preset jitter threshold, image jitter is determined. Finally, the above judgment results are matched with the error occurrence nodes in the error propagation path to determine the interference triggering conditions.

[0043] Finally, a wind farm identification interference tree is constructed based on the error occurrence node, error propagation path, and interference triggering conditions. In this process, the target anchoring sequence is evaluated for positioning deviation to obtain target positioning deviation characteristics; the abnormal behavior identification sequence is evaluated for classification deviation to obtain behavior classification deviation characteristics; and the abnormal progression characteristic sequence is evaluated for temporal continuity deviation to obtain abnormal progression deviation characteristics. Subsequently, the target positioning deviation characteristics are mapped to the spatial offset attribute of the error occurrence node, the behavior classification deviation characteristics are mapped to the category uncertainty attribute of the error occurrence node, and the abnormal progression deviation characteristics are mapped to the temporal oscillation attribute of the error occurrence node. Finally, using the error occurrence node as the tree node, the error propagation path as the tree edge, the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute as node interference features, and the interference triggering condition as the edge triggering feature, the wind farm identification interference tree is obtained.

[0044] Furthermore, in the method provided in the application embodiments, constructing the wind farm identification interference tree based on the error occurrence node, the error propagation path, and the interference triggering condition further includes: The target anchoring sequence is evaluated for positioning deviation to obtain target positioning deviation characteristics; the abnormal behavior identification sequence is evaluated for classification deviation to obtain behavior classification deviation characteristics; the abnormal progression characteristic sequence is evaluated for temporal continuity deviation to obtain abnormal progression deviation characteristics; the target positioning deviation characteristics, the behavior classification deviation characteristics, and the abnormal progression deviation characteristics are mapped to the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute of the error occurrence node, respectively; the error occurrence node is used as a tree node, the error propagation path is used as a tree edge, the spatial offset attribute, the category uncertainty attribute, and the temporal oscillation attribute are used as node interference features, and the interference triggering condition is used as edge triggering features to obtain the wind farm identification interference tree.

[0045] In this embodiment, when evaluating the positioning deviation of the target anchoring sequence, the target category, target bounding box, target center coordinates, target confidence, and target region attribution result corresponding to each acquisition time stamp in the target anchoring sequence are first read. Then, the difference between the target center coordinates of the same target at adjacent acquisition time stamps is calculated to obtain the target position change, and the overlap ratio between the target bounding box and the target detection area at the corresponding acquisition time stamp is calculated. If the change in the target center coordinates between adjacent acquisition time stamps exceeds a preset position change threshold, or the overlap ratio between the target bounding box and the target detection area is lower than a preset overlap threshold, or the target region attribution result is inconsistent with the target detection area actually falling into by the target center coordinates, then it is determined that the target anchoring sequence has a positioning deviation. Subsequently, the acquisition time stamp, target center coordinate offset, target bounding box offset, target region attribution deviation, and target confidence corresponding to the positioning deviation are recorded to obtain the target positioning deviation characteristics.

[0046] When evaluating the classification bias of anomaly behavior recognition sequences, the following steps are taken: First, the anomalous behavior category, anomalous occurrence area, anomalous target, anomalous confidence level, and anomalous judgment result are read for each acquisition time stamp in the sequence. Then, the anomalous behavior category is compared with identifiable anomalous types to determine if it belongs to the current monitoring area type and the identifiable anomalous type corresponding to the target detection area. Next, the anomalous judgment result is compared with the subsequent anomalous judgment conditions. If the target has not entered the target detection area but the anomalous judgment result is "anomaly has occurred," or if the target enters the target detection area and meets the conditions of dwell time, boundary crossing trajectory, abnormal approach, continuous smoke and fire detection, abnormal equipment appearance, and road occupancy but the anomalous judgment result is "no anomaly has occurred," the anomalous behavior recognition sequence is considered to have classification bias. Then, the anomalous behavior category and anomalous confidence level of the same anomalous target at adjacent acquisition time stamps are compared. If the anomalous behavior category changes frequently, or the anomalous confidence level is lower than a preset confidence threshold, or the change in anomalous confidence level between adjacent acquisition time stamps exceeds a preset confidence change threshold, the corresponding anomalous behavior category bias, anomalous judgment bias, and anomalous confidence bias are recorded to obtain the behavior classification bias characteristics.

[0047] When evaluating the temporal continuity deviation of an abnormal progressive characteristic sequence, the following steps are taken: first, the progressive state, duration of abnormal behavior, frequency of abnormal behavior, changes in the trajectory of the abnormal target, changes in the area of ​​abnormal occurrence, and changes in the confidence level of the abnormality are read for each acquisition time stamp in the abnormal progressive characteristic sequence; then, the continuity of the progressive state of the same abnormal target and the same area of ​​abnormal occurrence is checked according to the acquisition time stamp. If multiple consecutive acquisition time stamps show abnormal behavior identification results, but no continuous progressive state is formed in the abnormal progressive characteristic sequence, a progressive interruption deviation is determined to exist; if only a single acquisition time stamp shows an abnormal behavior identification result, but a continuous progressive state is formed in the abnormal progressive characteristic sequence, a progressive false continuation deviation is determined to exist; if the duration of abnormal behavior, frequency of abnormal behavior, changes in the trajectory of the abnormal target, and changes in the area of ​​abnormal occurrence are inconsistent with the continuous abnormal behavior results in the abnormal behavior identification sequence, a temporal continuity deviation is determined to exist. Subsequently, the number of progressive state changes, progressive interruption positions, progressive false continuation positions, deviations in abnormal duration, and fluctuations in abnormal confidence levels are recorded to obtain the characteristics of the abnormal progressive deviation.

[0048] Subsequently, the target positioning deviation characteristics, behavior classification deviation characteristics, and abnormal progression deviation characteristics are mapped to the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute of the error occurrence node, respectively. Specifically, based on the temporal position, spatial position, and model hierarchical position corresponding to the error occurrence node, the target positioning deviation characteristics, behavior classification deviation characteristics, and abnormal progression deviation characteristics under the same acquisition timestamp, the same target position, and the same model calling level are first searched. Then, the target center coordinate offset, target bounding box offset, and target region attribution deviation are written into the error occurrence node to form the spatial offset attribute; the abnormal behavior category deviation, abnormal judgment deviation, and abnormal confidence deviation are written into the error occurrence node to form the category uncertainty attribute; and the deviations in the number of progression state changes, progression interruption position, progression error continuation position, and abnormal duration are written into the error occurrence node to form the temporal oscillation attribute.

[0049] Finally, using the error occurrence node as the tree node, the error propagation path as the tree edge, spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute as node interference features, and interference triggering condition as edge triggering feature, the wind farm identification interference tree is obtained. In this process, firstly, the connection order between each error occurrence node is determined according to the forward and backward propagation relationship of each error occurrence node in the error propagation path; then, each error occurrence node is set as a tree node, and the propagation relationship between adjacent error occurrence nodes is set as a tree edge. Subsequently, the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute corresponding to each tree node are recorded as node interference features, and the corresponding illumination abrupt change, target occlusion, target blurring, target scale change, background motion interference, continuous frame loss, and image jitter between adjacent tree nodes are recorded as edge triggering features. Finally, according to the model hierarchy position in the model serial call chain, the acquisition timestamp order, and the error propagation direction, the tree nodes and tree edges are organized to obtain the wind farm identification interference tree.

[0050] Step S500: Perform multi-dimensional redundancy verification on the output sequences of each model to establish a redundancy verification space for wind farm identification.

[0051] In this embodiment, when performing multi-dimensional redundancy verification on the output sequences of each model, firstly, target consistency verification is performed on the target anchoring sequence to obtain the target redundancy verification result; then, behavior consistency verification is performed on the abnormal behavior identification sequence to obtain the behavior redundancy verification result; and finally, temporal consistency verification is performed on the abnormal progressive characteristic sequence to obtain the progressive redundancy verification result. Subsequently, a joint mapping guide pointer is obtained, which includes the model call level, time series position, and spatial region position. Finally, multi-dimensional joint mapping is performed on the target redundancy verification result, behavior redundancy verification result, and progressive redundancy verification result based on the joint mapping guide pointer to obtain the wind farm identification redundancy verification space.

[0052] Furthermore, the method provided in the application embodiment, which performs multi-dimensional redundancy verification on the output sequences of each model to establish a wind farm identification redundancy verification space, further includes: The target anchoring sequence is subjected to target consistency verification to obtain target redundancy verification results; the abnormal behavior identification sequence is subjected to behavior consistency verification to obtain behavior redundancy verification results; the abnormal progressive characteristic sequence is subjected to temporal consistency verification to obtain progressive redundancy verification results; a joint mapping guide pointer is obtained, which includes model call level, time series position, and spatial region position; the target redundancy verification results, behavior redundancy verification results, and progressive redundancy verification results are subjected to multi-dimensional joint mapping based on the joint mapping guide pointer to obtain the wind farm identification redundancy verification space.

[0053] In this embodiment, when performing target consistency verification on the target anchoring sequence, the target category, target bounding box, target center coordinates, target confidence, and target region attribution result corresponding to each acquisition timestamp in the target anchoring sequence are first read. Then, the target category, target bounding box, and target center coordinates of the same target in continuous augmentation monitoring frames are compared according to the acquisition timestamp. If the target category of the same target remains consistent in continuous augmentation monitoring frames, the overlap ratio of the target bounding box reaches a preset overlap threshold, the change in the target center coordinates does not exceed a preset position change threshold, and the target region attribution result is consistent with the target detection area where the target center coordinates fall, then the target consistency verification is deemed to have passed. If the target category changes, the overlap ratio of the target bounding box is lower than a preset overlap threshold, the change in the target center coordinates exceeds a preset position change threshold, and the target region attribution result is inconsistent with the target detection area where the target center coordinates fall, then the target consistency verification is deemed to have failed. Subsequently, the acquisition timestamp, target category, target bounding box, target center coordinates, target region attribution result, target confidence, and verification result corresponding to the verification are recorded to obtain the target redundancy verification result.

[0054] When performing behavior consistency verification on the abnormal behavior identification sequence, the following steps are taken: First, the abnormal behavior category, abnormal occurrence area, abnormal target, abnormal confidence level, and abnormal judgment result corresponding to each collection timestamp in the abnormal behavior identification sequence are read. Then, the abnormal behavior category is compared with the current monitoring area type and the identifiable abnormal type corresponding to the target detection area to determine whether the abnormal behavior category belongs to an identifiable abnormal type. At the same time, the abnormal occurrence area is compared with the target area assignment result corresponding to the abnormal target to determine whether the abnormal occurrence area is consistent with the target's location area. Subsequently, the abnormal judgment result is compared with the abnormal post-judgment conditions. When the target enters the target detection area, the target stays for more than a preset time, the target trajectory crosses a preset boundary, the target state meets the abnormal rules, and the abnormal judgment result is "abnormality occurred," the behavior consistency verification is deemed to have passed. When the target does not meet the abnormal post-judgment conditions but the abnormal judgment result is "abnormality occurred," or the target meets the abnormal post-judgment conditions but the abnormal judgment result is "no abnormality occurred," the behavior consistency verification is deemed to have failed. Finally, the abnormal behavior category, abnormal occurrence area, abnormal target, abnormal confidence level, abnormal judgment result, and verification result are recorded to obtain the behavior redundancy verification result.

[0055] When performing time-series consistency verification on anomaly progression characteristic sequences, the following steps are first taken: the progression state, duration of abnormal behavior, frequency of abnormal behavior occurrence, changes in the trajectory of the abnormal target, changes in the area of ​​abnormal occurrence, and changes in the confidence level of the abnormality corresponding to each acquisition time stamp in the anomaly progression characteristic sequence. Then, the progression state of the same abnormal target and the same area of ​​abnormal occurrence is continuously compared according to the acquisition time stamp, and the anomaly progression characteristic sequence is compared with the continuous anomaly judgment results in the anomaly behavior identification sequence. If multiple consecutive acquisition time stamps show anomaly judgment results, and the anomaly progression characteristic sequence forms a corresponding continuous progression state, the time-series consistency verification is considered successful. If the anomaly behavior identification sequence shows continuous anomaly judgment results, but the progression state in the anomaly progression characteristic sequence is interrupted, or if only a single acquisition time stamp in the anomaly behavior identification sequence shows an anomaly judgment result, but the anomaly progression characteristic sequence forms a continuous progression state, the time-series consistency verification is considered unsuccessful. Subsequently, the progression state, duration of abnormal behavior, frequency of abnormal behavior occurrence, changes in the trajectory of the abnormal target, changes in the area of ​​abnormal occurrence, changes in the confidence level of the abnormality, and the verification results are recorded to obtain the progression redundancy verification result.

[0056] To obtain the joint mapping guide pointer, the model call level corresponding to each model output sequence is first determined based on the model serial call chain. The time series position is determined based on the acquisition timestamp and the enhanced monitoring frame number. The spatial region position is determined based on the target detection region, the anomaly occurrence region, and the target region attribution result. Subsequently, the same model call level, the same time series position, and the same spatial region position are combined to form the joint mapping guide pointer. The joint mapping guide pointer includes the model call level, the time series position, and the spatial region position, and is used to map the target redundancy verification results, the behavior redundancy verification results, and the progressive redundancy verification results to the same detection process, the same time range, and the same spatial range.

[0057] Next, multi-dimensional joint mapping is performed on the target redundancy verification results, behavioral redundancy verification results, and progressive redundancy verification results based on the joint mapping guide pointers. In this process, the model call level, time series position, and spatial region position corresponding to each joint mapping guide pointer are first read. Then, the corresponding target redundancy verification results, behavioral redundancy verification results, and progressive redundancy verification results are found according to the model call level. Verification results within the same acquisition timestamp range are matched according to the time series position, and verification results within the same target detection region and anomaly occurrence region are matched according to the spatial region position. Subsequently, the matched target redundancy verification results, behavioral redundancy verification results, and progressive redundancy verification results are written to the same joint mapping position, and the target consistency verification results, behavioral consistency verification results, and time series consistency verification results are recorded. After completing the mapping corresponding to all joint mapping guide pointers, the wind farm identification redundancy verification space is obtained.

[0058] Step S600: Based on the wind farm identification interference tree and the wind farm identification redundancy verification space, perform credibility enhancement processing on the output sequences of each model to establish a wind farm abnormal behavior heat map.

[0059] In this embodiment, when performing credibility enhancement processing on the output sequences of each model based on the wind farm identification interference tree and the wind farm identification redundancy verification space, firstly, interference propagation is performed on the target anchoring sequence, abnormal behavior identification sequence, and abnormal progressive characteristic sequence based on the wind farm identification interference tree to obtain the distribution of abnormal candidate interference effects; then, redundancy propagation is performed on the target anchoring sequence, abnormal behavior identification sequence, and abnormal progressive characteristic sequence based on the wind farm identification redundancy verification space to obtain the distribution of abnormal candidate redundancy effects; subsequently, credibility reconstruction and spatiotemporal diffusion mapping are performed on the output sequences of each model based on the distribution of abnormal candidate interference effects and the distribution of abnormal candidate redundancy effects to obtain a wind farm abnormal behavior heatmap.

[0060] Furthermore, in the method provided in the application embodiments, the reliability enhancement processing of the output sequences of each model is performed based on the wind farm identification interference tree and the wind farm identification redundancy verification space to establish a wind farm abnormal behavior heatmap, which further includes: The interference propagation of each model's output sequence is deduced based on the wind farm identification interference tree to obtain the distribution of abnormal candidate interference effects; the redundancy propagation of each model's output sequence is deduced based on the wind farm identification redundancy verification space to obtain the distribution of abnormal candidate redundancy effects; and the reliable reconstruction and spatiotemporal diffusion mapping of each model's output sequence are performed based on the distribution of abnormal candidate interference effects and the distribution of abnormal candidate redundancy effects to obtain the heat map of abnormal wind farm behavior.

[0061] In this embodiment, when performing interference propagation deduction on the output sequences of each model based on the wind farm identification interference tree, the error occurrence node, error propagation path, spatial offset attribute, category uncertainty attribute, temporal oscillation attribute, and interference triggering condition in the wind farm identification interference tree are first read. Then, according to the model call level, time series position, and spatial region position corresponding to the error occurrence node, the corresponding output is searched in the target anchoring sequence, abnormal behavior identification sequence, and abnormal progression characteristic sequence. For error occurrence nodes with spatial offset attributes, the horizontal offset and vertical offset are read. The horizontal coordinate of the target center coordinate is added to the horizontal offset, and the vertical coordinate of the target center coordinate is added to the vertical offset to obtain the offset target center coordinate. Then, the affected spatial region position is determined according to the spatial region position to which the offset target center coordinate belongs. For error occurrence nodes with category uncertainty attributes, the abnormal behavior category deviation in the category uncertainty attribute is compared with the abnormal behavior category in the abnormal behavior identification sequence. If the two are inconsistent, the corresponding abnormal candidate is determined as an affected abnormal candidate. For error occurrence nodes exhibiting temporal oscillations, the progressive state is read according to the time series position. If the progressive state in three consecutive time series positions is, in order, a continuous anomaly, a non-continuous anomaly, and a continuous anomaly, then the intermediate time series position is determined as the affected time series position. The affected anomaly candidates, the affected spatial region position, the affected time series position, the model call level, and the interference triggering conditions are correlated to obtain the distribution of the anomaly candidate interference impact.

[0062] When performing redundancy propagation and deduction on the output sequences of each model based on the redundancy verification space of the wind farm identification, the joint mapping guide pointer, target redundancy verification result, behavioral redundancy verification result, and progressive redundancy verification result in the wind farm identification redundancy verification space are first read. The joint mapping guide pointer includes the model call level, time series position, and spatial region position. Then, according to the same model call level, the same time series position, and the same spatial region position, the target redundancy verification result, behavioral redundancy verification result, and progressive redundancy verification result are mapped to the same anomaly candidate. If the target redundancy verification result passes, it is recorded as 1, and if it fails, it is recorded as 0; if the behavioral redundancy verification result passes, it is recorded as 1, and if it fails, it is recorded as 0; if the progressive redundancy verification result passes, it is recorded as 1, and if it fails, it is recorded as 0. The three values ​​are added together to obtain the number of verifications passed. The anomaly candidate, target redundancy verification result, behavioral redundancy verification result, progressive redundancy verification result, number of verifications passed, model call level, time series position, and spatial region position are correlated to obtain the redundancy impact distribution of the anomaly candidate.

[0063] Finally, based on the distribution of interference and redundancy effects of abnormal candidate models, the output sequences of each model are reliably reconstructed and subjected to spatiotemporal diffusion mapping. In this process, the distribution of interference and redundancy effects of abnormal candidate models is first mapped to the distribution of redundancy effects of abnormal candidate models according to the abnormal candidate, model call level, time series position, and spatial region position. If the spatial region position corresponding to the target center coordinates is affected by the spatial offset attribute and the target redundancy check result fails, the target center coordinates that passed the target redundancy check result in the previous and next time series positions are read. The two abscissas are added together and divided by 2 to obtain the reconstructed abscissa, and the two ordinates are added together and divided by 2 to obtain the reconstructed ordinate. The reconstructed target center coordinates are composed of the reconstructed abscissa and the reconstructed ordinate, and the target region attribution result is redefined. If the abnormal behavior category is affected by the category uncertainty attribute and the behavior redundancy check result fails, the abnormal behavior categories that passed the behavior redundancy check result in adjacent time series positions for the same abnormal candidate are read. The frequency of each abnormal behavior category is counted, and the abnormal behavior category with the highest frequency is taken as the reconstructed abnormal behavior category, and the abnormality judgment result is updated synchronously. If the progressive state is affected by temporal oscillation attributes and the progressive redundancy check result fails, then the anomaly judgment results in the continuous time series positions are read. The first time series position where consecutive anomaly judgment results appear is taken as the start position, and the last time series position where consecutive anomaly judgment results appear is taken as the end position. The collection timestamp corresponding to the end position is subtracted from the collection timestamp corresponding to the start position to obtain the reconstructed anomaly behavior duration, and the progressive state is updated. After the reliable reconstruction is completed, the reconstructed anomaly behavior category, anomaly occurrence area, anomaly confidence, anomaly behavior duration, and progressive state are written to the corresponding spatial region position. The anomaly behavior duration is divided by a preset duration threshold to obtain the duration coefficient. The anomaly confidence, duration coefficient, and preset value corresponding to the progressive state are multiplied to obtain the anomaly intensity. Finally, the anomaly behavior category, anomaly occurrence area, anomaly intensity, anomaly behavior duration, and progressive state are summarized according to the time series position and spatial region position to obtain the wind farm anomaly behavior heat map.

[0064] Furthermore, the method provided in the application embodiments, in establishing a heat map of abnormal behavior in a wind farm, further includes: A graded early warning system is implemented based on the heat map of abnormal behavior in the wind farm.

[0065] In this embodiment, when performing graded early warning based on the wind farm abnormal behavior heatmap, the following steps are taken: First, the abnormal behavior category, abnormal occurrence area, abnormal intensity, abnormal behavior duration, and progressive status are read from the wind farm abnormal behavior heatmap. Then, the abnormal intensity is compared with preset early warning thresholds, which include a first-level early warning threshold, a second-level early warning threshold, and a third-level early warning threshold. If the abnormal intensity is less than the first-level early warning threshold, it is recorded as a state of concern. If the abnormal intensity is greater than or equal to the first-level early warning threshold and less than the second-level early warning threshold, a first-level early warning is generated. If the abnormal intensity is greater than or equal to the second-level early warning threshold and less than the third-level early warning threshold, a second-level early warning is generated. If the abnormal intensity is greater than or equal to the third-level early warning threshold, a third-level early warning is generated. Subsequently, the abnormal behavior duration and progressive status are used as early warning correction conditions. When the abnormal behavior duration exceeds a preset duration threshold, or the progressive status indicates that the abnormal behavior has developed from a single abnormality to a continuous abnormality, the current early warning level is increased by one level. Finally, the early warning level, abnormal behavior category, abnormal occurrence area, abnormal intensity, and early warning time are sent to the wind farm monitoring platform to complete the graded early warning.

[0066] In summary, the embodiments of this application have at least the following technical effects: This application acquires real-time video streams from a wind farm monitoring platform and determines the scene semantic information of the real-time video streams; constructs an anomaly recognition task graph based on the scene semantic information, and adaptively filters and arranges the wind farm model library according to the anomaly recognition task graph to obtain a model chain; performs multi-level progressive detection on the real-time video streams according to the model chain to obtain the output sequences of each model, wherein each model output sequence includes a target anchoring sequence, an anomaly behavior recognition sequence, and an anomaly progressive characteristic sequence; performs error assessment and source tracing on each model output sequence to establish a wind farm recognition interference tree; performs multi-dimensional redundancy verification on each model output sequence to establish a wind farm recognition redundancy verification space; and performs credibility enhancement processing on each model output sequence based on the wind farm recognition interference tree and the wind farm recognition redundancy verification space to establish a wind farm anomaly behavior heatmap. This invention addresses the technical problem of insufficient accuracy in identifying abnormal behavior in wind farms in existing technologies. By adaptively constructing a model chain based on scene semantic information and performing multi-level progressive detection and error interference correction on real-time acquired video streams, it achieves the technical effect of improving the accuracy of identifying abnormal behavior in wind farms.

[0067] Example 2 is based on the same inventive concept as the wind farm abnormal behavior identification method based on multi-model serial calling in the previous examples, such as... Figure 2 As shown, this application provides a wind farm abnormal behavior identification system based on multi-model serial invocation. The system and method embodiments in this application are based on the same inventive concept. The system includes: The data acquisition module 11 is used to acquire real-time video streams from the wind farm monitoring platform and determine the scene semantic information of the real-time video streams; the filtering and arrangement module 12 is used to construct an anomaly recognition task map based on the scene semantic information and adaptively filter and arrange the wind farm model library based on the anomaly recognition task map to obtain the model serial call chain; the progressive detection module 13 is used to perform multi-level progressive detection on the real-time video stream according to the model serial call chain to obtain the output sequence of each model, wherein each model output sequence includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence; the evaluation and tracing module 14 is used to perform error evaluation and tracing on each model output sequence and establish a wind farm recognition interference tree; the redundancy verification module 15 is used to perform multi-dimensional redundancy verification on each model output sequence and establish a wind farm recognition redundancy verification space; the heat map establishment module 16 is used to perform credibility enhancement processing on each model output sequence based on the wind farm recognition interference tree and the wind farm recognition redundancy verification space to establish a wind farm abnormal behavior heat map.

[0068] Furthermore, the system is also used to implement the following functions: The wind farm model library is matched and filtered according to the anomaly identification task graph to obtain multiple candidate models; anomaly behavior identification is decomposed according to the anomaly identification task graph to obtain anomaly pre-identification conditions and anomaly post-judgment conditions; the multiple candidate models are sequentially arranged and input / output connection configured according to the anomaly pre-identification conditions and the anomaly post-judgment conditions to obtain the model call initial chain; the model call initial chain is configured with backup and anomaly self-recovery to generate the model serial call chain.

[0069] Furthermore, the system is also used to implement the following functions: The real-time acquired video stream is enhanced and frames are extracted to obtain enhanced monitoring frames of the wind farm; target anchoring is performed on each enhanced monitoring frame according to the model's serial call chain to obtain the target anchoring sequence; abnormal behavior identification is performed on the target anchoring sequence according to the model's serial call chain to obtain the abnormal behavior identification sequence; and time-series progressive analysis is performed on the abnormal behavior identification sequence according to the model's serial call chain to obtain the abnormal progressive characteristic sequence.

[0070] Furthermore, the system is also used to implement the following functions: Cross-model correlation analysis is performed on the output sequences of each model to generate an error candidate set; based on the error candidate set, the temporal location, spatial location, and model hierarchical location of the error are jointly located to construct the error occurrence node; the error propagation path is traced in reverse according to the error occurrence node; interference factors are attributed to the error propagation path based on the real-time acquired video stream to determine the interference triggering condition; the wind farm identification interference tree is constructed based on the error occurrence node, the error propagation path, and the interference triggering condition.

[0071] Furthermore, the system is also used to implement the following functions: The target anchoring sequence is evaluated for positioning deviation to obtain target positioning deviation characteristics; the abnormal behavior identification sequence is evaluated for classification deviation to obtain behavior classification deviation characteristics; the abnormal progression characteristic sequence is evaluated for temporal continuity deviation to obtain abnormal progression deviation characteristics; the target positioning deviation characteristics, the behavior classification deviation characteristics, and the abnormal progression deviation characteristics are mapped to the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute of the error occurrence node, respectively; the error occurrence node is used as a tree node, the error propagation path is used as a tree edge, the spatial offset attribute, the category uncertainty attribute, and the temporal oscillation attribute are used as node interference features, and the interference triggering condition is used as edge triggering features to obtain the wind farm identification interference tree.

[0072] Furthermore, the system is also used to implement the following functions: The target anchoring sequence is subjected to target consistency verification to obtain target redundancy verification results; the abnormal behavior identification sequence is subjected to behavior consistency verification to obtain behavior redundancy verification results; the abnormal progressive characteristic sequence is subjected to temporal consistency verification to obtain progressive redundancy verification results; a joint mapping guide pointer is obtained, which includes model call level, time series position, and spatial region position; the target redundancy verification results, behavior redundancy verification results, and progressive redundancy verification results are subjected to multi-dimensional joint mapping based on the joint mapping guide pointer to obtain the wind farm identification redundancy verification space.

[0073] Furthermore, the system is also used to implement the following functions: The interference propagation of each model's output sequence is deduced based on the wind farm identification interference tree to obtain the distribution of abnormal candidate interference effects; the redundancy propagation of each model's output sequence is deduced based on the wind farm identification redundancy verification space to obtain the distribution of abnormal candidate redundancy effects; and the reliable reconstruction and spatiotemporal diffusion mapping of each model's output sequence are performed based on the distribution of abnormal candidate interference effects and the distribution of abnormal candidate redundancy effects to obtain the heat map of abnormal wind farm behavior.

[0074] Furthermore, the system is also used to implement the following functions: The scene semantic information includes the monitoring area type, target detection area, and identifiable anomaly type corresponding to the real-time acquired video stream.

[0075] Furthermore, the system is also used to implement the following functions: A graded early warning system is implemented based on the heat map of abnormal behavior in the wind farm.

[0076] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for identifying abnormal behavior in wind farms based on multi-model serial invocation, characterized in that, The method includes: Acquire real-time video streams from the wind farm monitoring platform and determine the scene semantic information of the real-time video streams; An anomaly identification task graph is constructed based on the scene semantic information, and the wind farm model library is adaptively filtered and arranged based on the anomaly identification task graph to obtain the model serial call chain. The real-time acquired video stream is subjected to multi-level progressive detection according to the model serial call chain to obtain the output sequence of each model. The output sequence of each model includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence. Error assessment and source tracing are performed on the output sequences of each model to establish a wind farm interference identification tree; Multidimensional redundancy verification is performed on the output sequences of each model to establish a redundancy verification space for wind farm identification; Based on the wind farm identification interference tree and the wind farm identification redundancy verification space, the output sequences of each model are subjected to credibility enhancement processing to establish a wind farm abnormal behavior heatmap.

2. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, Based on the anomaly identification task map, the wind farm model library is adaptively filtered and arranged to obtain the model serial call chain, including: The wind farm model library is matched and filtered according to the anomaly identification task map to obtain multiple candidate models; Based on the anomaly recognition task map, the abnormal behavior recognition is decomposed to obtain the anomaly pre-recognition conditions and anomaly post-judgment conditions. Based on the anomaly pre-identification conditions and the anomaly post-judgment conditions, the multiple candidate models are sequentially arranged and their input-output connections are configured to obtain the initial chain of model calls; The model is configured with backup and exception self-recovery configurations for the initial call chain, and the model is generated by the serial call chain.

3. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, The real-time acquired video stream is subjected to multi-level progressive detection according to the model's serial call chain to obtain the output sequences of each model. Each model output sequence includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence, including: The real-time video stream is enhanced and frames are extracted to obtain the enhanced monitoring frames of the wind farm. Based on the model's serial call chain, target anchoring is performed on each enhanced monitoring frame to obtain the target anchoring sequence; Based on the model's serial call chain, abnormal behavior identification is performed on the target anchoring sequence to obtain the abnormal behavior identification sequence; Based on the model's serial call chain, a time-series progressive analysis is performed on the abnormal behavior identification sequence to obtain the abnormal progressive characteristic sequence.

4. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, Error assessment and source tracing are performed on the output sequences of each model to establish a wind farm interference identification tree, including: Cross-model correlation analysis is performed on the output sequences of each model to generate an error candidate set; Based on the error candidate set, the temporal location, spatial location, and model hierarchical location of the error are jointly located to construct the error occurrence node; Based on the error occurrence node, trace the error propagation path in reverse; Based on the real-time acquired video stream, interference factors are attributed to the error propagation path to determine the interference triggering conditions; Based on the error occurrence node, the error propagation path, and the interference triggering condition, the wind farm identification interference tree is constructed.

5. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 4, characterized in that, Based on the error occurrence node, the error propagation path, and the interference triggering condition, the wind farm interference identification tree is constructed, including: The positioning deviation of the target anchoring sequence is evaluated to obtain the target positioning deviation characteristics; The abnormal behavior identification sequence is subjected to classification bias evaluation to obtain behavior classification bias characteristics; The abnormal progression characteristic sequence is evaluated for temporal continuity deviation to obtain abnormal progression deviation characteristics; The target positioning deviation characteristic, the behavior classification deviation characteristic, and the abnormal progression deviation characteristic are respectively mapped to the spatial offset attribute, category uncertainty attribute, and temporal oscillation attribute of the error occurrence node; Using the error occurrence node as the tree node, the error propagation path as the tree edge, the spatial offset attribute, the category uncertainty attribute, and the temporal oscillation attribute as node interference features, and the interference triggering condition as edge triggering features, the wind farm identification interference tree is obtained.

6. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, Multidimensional redundancy checks are performed on the output sequences of each model to establish a redundancy check space for wind farm identification, including: Perform target consistency verification on the target anchoring sequence and obtain the target redundancy verification result; The abnormal behavior identification sequence is subjected to behavior consistency verification to obtain behavior redundancy verification results; Perform a time-series consistency check on the abnormal progressive characteristic sequence and obtain the progressive redundancy check result; Obtain the joint mapping pointer, which includes the model call level, time series position, and spatial region position; Based on the joint mapping guide pointer, a multi-dimensional joint mapping is performed on the target redundancy verification result, the behavioral redundancy verification result, and the progressive redundancy verification result to obtain the wind farm identification redundancy verification space.

7. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, Based on the wind farm identification interference tree and the wind farm identification redundancy check space, the output sequences of each model are subjected to credibility enhancement processing to establish a wind farm abnormal behavior heatmap, including: Based on the wind farm identification interference tree, the interference propagation of each model output sequence is deduced to obtain the distribution of abnormal candidate interference effects. Based on the redundancy verification space for wind farm identification, redundancy propagation and deduction are performed on the output sequences of each model to obtain the distribution of abnormal candidate redundancy effects. Based on the distribution of abnormal candidate interference and the distribution of abnormal candidate redundancy, the output sequences of each model are reliably reconstructed and spatiotemporally diffused to obtain a heat map of the abnormal behavior of the wind farm.

8. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, The scene semantic information includes the monitoring area type, target detection area, and identifiable anomaly type corresponding to the real-time acquired video stream.

9. The wind farm abnormal behavior identification method based on multi-model serial invocation as described in claim 1, characterized in that, Establish a heat map of abnormal behavior in the wind farm, including: A graded early warning system is implemented based on the heat map of abnormal behavior in the wind farm.

10. A wind farm abnormal behavior identification system based on multi-model serial invocation, characterized in that, The system is used to execute the wind farm abnormal behavior identification method based on multi-model serial calling as described in any one of claims 1-9, and the system includes: The data acquisition module is used to acquire real-time video streams from the wind farm monitoring platform and determine the scene semantic information of the real-time video streams. The filtering and arrangement module is used to construct an anomaly recognition task graph based on the scene semantic information, and to adaptively filter and arrange the wind farm model library based on the anomaly recognition task graph to obtain the model serial call chain. The progressive detection module is used to perform multi-level progressive detection on the real-time acquired video stream according to the model serial call chain, and obtain the output sequence of each model. The output sequence of each model includes a target anchoring sequence, an abnormal behavior recognition sequence, and an abnormal progressive characteristic sequence. The evaluation and source tracing module is used to evaluate and trace the source of errors in the output sequences of each model and establish a wind farm identification interference tree. A redundancy check module is used to perform multi-dimensional redundancy check on the output sequences of each model and establish a redundancy check space for wind farm identification. The heatmap creation module is used to perform credibility enhancement processing on the output sequences of each model based on the wind farm identification interference tree and the wind farm identification redundancy verification space, and to create a heatmap of abnormal wind farm behavior.

Images