Photovoltaic power station construction safety intelligent monitoring method and system based on AI video identification

Abstract

The invention relates to the technical field of intelligent monitoring, in particular to a photovoltaic power station construction safety intelligent monitoring method and system based on AI video recognition, and the method comprises the steps: employing an input image frame sequence, carrying out the feature extraction of the image frame sequence through a convolutional network, generating an image vector sequence, inputting the image vector sequence into a bidirectional gating circulation unit, and carrying out the recognition of the image vector sequence; outputting a stage representation vector, and outputting a construction stage number corresponding to the stage representation vector through a stage classifier; on the basis of the construction stage number and the segment representation vector, identifying the specific behavior type of each constructor in the current operation window and the station number of the constructor; constructing a job graph structure, and identifying risk scores of nodes in the job graph structure through a node risk scoring function; and on the basis of the constructed operation graph structure, final judgment on the violation behavior is completed.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and more particularly to an intelligent monitoring method and system for construction safety of photovoltaic power plants based on AI video recognition. Background Technology

[0002] With the large-scale construction of photovoltaic power plants, construction sites exhibit typical characteristics such as diverse work phases, dispersed worker distribution, large geographical span, and frequent equipment collaboration. During construction phases such as bracket installation, module laying, and electrical connections, there are often various high-risk scenarios, including working at heights, working near edges, overlapping operations, and multi-person collaborative work. The safety regulations for personnel behavior differ significantly at different stages, making real-time monitoring of the construction site a crucial means of ensuring safety. Currently, the industry commonly uses video detection-based violation identification methods, such as detecting not wearing a safety helmet, not wearing a safety belt, or entering a restricted area. However, existing systems typically rely on fixed rules or independent target recognition models, lacking the ability to understand the construction phases and dynamically adjust the identification focus according to the task flow. This leads to frequent instances of the same action being misjudged or missed at different stages.

[0003] Furthermore, such systems often rely solely on the behavior of individual workers, failing to analyze the relationship between a worker's workstation and its task attributes at the current stage, or to model structural risks arising from multi-person collaboration or workstation conflicts. For example, during component installation, different workstations should maintain consistent and coordinated behavior; however, during cable laying, a seemingly normal spatial location might be within an unfinished work area from the previous step—such structural errors cannot be detected through single-point behavior identification. Moreover, photovoltaic power plants commonly experience strong glare, obstructed views, and large spatial spans, making it difficult for simple detection methods to reliably reconstruct the contextual relationships of personnel behavior. Existing systems lack a multi-dimensional joint understanding mechanism for personnel behavior, workstation location, and construction stages, making it difficult to accurately identify hidden violations and collaborative risks in complex scenarios. Therefore, an intelligent monitoring method is needed that comprehensively utilizes construction stage information, personnel behavior patterns, and their spatial relationships to address the shortcomings of existing video detection systems in photovoltaic construction scenarios, including insufficient intelligence, poor contextual understanding, and lack of collaborative relationship judgment. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method and system for intelligent monitoring of construction safety in photovoltaic power plants based on AI video recognition.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A smart monitoring method for construction safety of photovoltaic power plants based on AI video recognition includes:

[0007] The input image frame sequence is taken, and features are extracted from the image frame sequence through a convolutional network to generate an image vector sequence. The image vector sequence is input into a bidirectional gated recurrent unit and outputs a stage representation vector. The construction stage number corresponding to the stage representation vector is output through a stage classifier.

[0008] The image processing window corresponding to the construction stage number is obtained. Each frame of the image in the image processing window is input into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction personnel in each frame. The position coordinates are correlated across frames by Kalman filtering and Hungarian matching algorithm to construct continuous image segments. Key point estimation is performed on the continuous image segments to generate key point temporal tensors. Action recognition is performed on the key point temporal tensors to generate behavioral feature vectors. The center point of the bounding box is used to perform planar coordinate back projection to generate position points. The position points are matched with the known work station division map to generate work station numbers.

[0009] A task graph structure is constructed based on behavioral feature vectors, stage representation vectors, and workstation numbers. A graph embedding update function is introduced to update the task graph structure. The risk score of nodes in the updated task graph structure is identified through a node risk scoring function.

[0010] Based on the preset judgment logic and combined with risk scoring, a violation label is generated through a violation judgment function, and a control response intensity is output through a response intensity calculation function. Based on the violation label and the control response intensity, linked response content is generated.

[0011] Preferably, after generating location points by back-projecting the center point of the bounding box into planar coordinates and matching them with the known workstation division diagram to generate workstation numbers, the method further includes:

[0012] The consistency loss function is used to determine the consistency between workstation number, stage representation vector, and behavioral feature vector. The consistency loss function includes location coordinates, the center of the main construction area corresponding to different stages, the cosine similarity between the behavioral feature vector and the stage representation vector, the average position of the construction personnel in the previous stage, weight coefficients, and small constants.

[0013] Preferably, the stage classifier includes a weight matrix, intermediate layer dimension, total number of construction stage categories, bias vector, activation function, and normalization mapping function.

[0014] Preferably, the work graph structure includes a node set and a boundary set. The node set is generated by splicing together behavioral feature vectors, stage representation vectors, and workstation numbers. The boundary set is constructed based on behavioral relationship rules and spatial layout logic.

[0015] Preferably, the parameters of the graph embedding update function include the node embedding result, the weight matrix of the graph convolutional layer, the node's neighborhood set, node degree, activation function, behavioral feature similarity, modulation coefficient, and behavioral conflict correction term.

[0016] Preferably, the parameters of the node risk scoring function include the degree of deviation between the structural reasoning result and the original behavioral characteristics, the set of workstation numbers allowed in the current stage, the indicator sub-function, and the penalty intensity.

[0017] Preferably, the parameters of the violation determination function include behavior label, behavior whitelist, current workstation number, allowed workstation set, risk score, and soft determination weight coefficient.

[0018] Preferably, the response intensity calculation function includes the workstation risk level, the severity of the violation of the behavior in the current stage, the risk score, and the balance coefficient. The workstation risk level is determined by the workstation deployment icon, the severity of the violation of the behavior in the current stage is obtained by looking up a rule table, and the balance coefficient is used to balance structural risk, rule intensity, and model warning.

[0019] Preferably, the preset judgment logic includes behavior legality judgment, workstation location legality judgment, and model-assisted judgment. The model-assisted judgment is used to judge the violation of personnel who are not clearly in violation but are in a boundary situation.

[0020] The intelligent monitoring system for construction safety of photovoltaic power plants based on AI video recognition includes:

[0021] The image recognition module is used to take the input image frame sequence, extract features from the image frame sequence through a convolutional network, generate an image vector sequence, input the image vector sequence into a bidirectional gated recurrent unit, and output a stage representation vector. The stage classifier outputs the construction stage number corresponding to the stage representation vector.

[0022] The behavior and workstation recognition module is used to acquire the image processing window numbered according to the corresponding construction stage. Each frame of the image in the image processing window is input into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction personnel in each frame. The position coordinates are correlated across frames by Kalman filtering and Hungarian matching algorithm to construct continuous image segments. Key point estimation is performed on the continuous image segments to generate key point temporal tensors. Action recognition is performed on the key point temporal tensors to generate behavior feature vectors. The center point of the bounding box is used to perform planar coordinate back projection to generate position points. The position points are matched with the known workstation division map to generate workstation numbers.

[0023] The task graph construction module is used to construct the task graph structure based on behavioral feature vectors, stage representation vectors, and workstation numbers. It introduces a graph embedding update function to update the task graph structure and identifies the risk scores of nodes in the updated task graph structure through a node risk scoring function.

[0024] The linkage response generation module is used to generate violation tags based on preset judgment logic and risk scoring through a violation judgment function, and output control response intensity through a response intensity calculation function. Based on the violation tags and control response intensity, linkage response content is generated.

[0025] The beneficial effects of this invention are as follows:

[0026] This invention establishes a construction phase identification mechanism, enabling the system to automatically identify the current construction phase based on the current work environment and pace. This dynamically links subsequent violation identification rules with the current construction status, solving the problem of existing systems having static rules that are difficult to adapt to changes in the construction process. The system further extracts personnel behavioral characteristics and workstation spatial coordinates, fusing behavioral semantics, spatial location, and phase semantics to construct a work relationship graph. Graph nodes represent construction personnel, and graph edges depict the work collaboration and spatial relationships between personnel. Through graph embedding reasoning, it identifies structural implicit risks such as behavioral deviations, workstation conflicts, and phase inconsistencies, thus overcoming the limitations of traditional single-point action identification in detecting errors in multi-person collaboration and cross-workstation violations. Based on the risk nodes output by the graph structure, the system combines a phase rule base to comprehensively determine the legality of behavior and spatial location, and automatically generates corresponding linkage response strategies based on the risk level, including audible and visual alarms, broadcast prompts, and area restrictions, giving the system a practically executable proactive intervention capability. This invention significantly improves the adaptability and accuracy of photovoltaic power plant construction monitoring systems in dynamic scenarios by introducing semantic modeling of the construction phase, structured expression of work behavior, and multi-dimensional consistency judgment mechanism. It can effectively identify potential violations and collaborative risks, reduce construction safety hazards, and improve on-site supervision efficiency. Attached Figure Description

[0027] Figure 1 This is a flowchart of an intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition, as described in a specific embodiment of the present invention.

[0028] Figure 2 This is a block diagram of an intelligent monitoring system for photovoltaic power plant construction safety based on AI video recognition, as described in a specific embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Step 1: Take the input image frame sequence, extract features from the image frame sequence through a convolutional network, and generate an image vector sequence. Input the image vector sequence into a bidirectional gated recurrent unit and output a stage representation vector. Then, output the construction stage number corresponding to the stage representation vector through a stage classifier. Specifically, this includes:

[0031] This step identifies the current stage of the photovoltaic power station construction site. It does not rely on a static timetable, but rather dynamically determines the actual work progress by combining image sequences captured by cameras in the construction area with task plan information to guide subsequent behavioral decisions. Input includes image frame sequences. Each frame For a moment The image frames below were captured by a fixed high-definition camera installed near the main construction work surface. The image acquisition frequency was 15 frames per second, and the resolution was 1920×1080. Continuous images were taken. The frame serves as the processing window. Task schedule information. This data comes from the construction project management system and is structured scheduling data. It includes stage number, expected start and end time, textual description of work type, and key equipment type. It can be used to provide rule references, but does not directly determine stage labels.

[0032] The image frame is first input to a feature extraction structure consisting of five convolutional layers, which extracts the spatial representation of each frame. The number of channels in each convolutional layer is... The kernel size is uniformly set to Each layer is followed by batch normalization and max pooling operations, and the output frame features are represented as a vector sequence. Each of them Indicates the first Spatial features of a frame image, dimensions .

[0033] Temporal modeling employs a bidirectional gated recurrent unit (Bi-GRU), consisting of two stacked layers, with each layer having a hidden state dimension of 1. Input is The output is a time-weighted average pooling vector representing the stages:

[0034] ;

[0035] Where the function This represents a bidirectional gated cyclic unit sequence modeling structure, and the output is... It is a stage semantic representation vector for the entire time window, used to represent the distribution characteristics of the job status within that time period.

[0036] The stage classifier is a two-layer fully connected structure, and its output is the construction stage number. :

[0037] ;

[0038] in and These are the weight matrices for the first and second fully connected layers, respectively. For intermediate layer dimensions, This represents the total number of construction phase categories. , These are the bias vectors for each layer; This represents an element-wise activation function using a modified linear unit. This represents the normalization mapping function, which transforms the output into a classification probability vector. Finally, through... The operation yields the current stage number. .

[0039] For example, during the component installation phase, the image sequence will continuously show repetitive bending, panel handling, and panel placement actions. Combining this with keywords for this phase in the task scheduling description, the classifier will tend to identify this time window as the "component installation" phase and output the corresponding number. If behavioral characteristics that do not conform to the time plan appear, such as entering wiring work ahead of schedule, the model will automatically update the phase judgment based on image inference priority.

[0040] Step 2: Obtain the image processing window corresponding to the construction stage number. Input each frame of the image within the image processing window into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction workers in each frame. Kalman filtering and Hungarian matching algorithms are used to correlate the position coordinates across frames to construct continuous image segments. Keypoint estimation is performed on the continuous image segments to generate keypoint temporal tensors. Action recognition is performed on the keypoint temporal tensors to generate behavioral feature vectors. Planar coordinate back projection is performed using the center point of the bounding box to generate position points. The position points are then matched with the known workstation division map to generate workstation numbers. Specifically, this includes:

[0041] This step involves receiving the construction stage number identified in the previous stage. With stage representation vector The next step is to identify the specific behavior type of each construction worker within the current work window. and its workstation number This step plays a crucial role in the entire system by assigning semantic values ​​to tasks. Its output serves as the core node attribute in subsequent graph structure modeling, accurately expressing "who, where, and what they did," forming the foundation for structured violation identification. In the specific scenario of photovoltaic power plant construction, behavior type and spatial location not only determine the work status but also relate to whether the safety requirements of the current stage are met. For example, during the support installation stage, if a person is detected moving components in the component placement area, it may constitute a risk of process misalignment. Therefore, the task of this step is not only identification but also to tightly bind behavior, space, and stage semantics and achieve consistent modeling.

[0042] The input image sequence is the same as in step one, with each 16 frames forming a processing window, captured by fixed wide-angle cameras deployed in the component area, support area, and cable area. Each frame is first input into a pre-set human detection network, with a five-layer convolutional structure in the backbone, and each layer having [number of channels]. Anchor frame regression is used to detect bounding boxes and output the position of each construction worker in each frame. Kalman filtering and Hungarian matching algorithms are used to correlate the trajectories of workers across frames, thus obtaining the position of each worker. Continuous image segments. To further obtain details of the operation, for each... Keypoint estimation was performed on the image segments, and the two-dimensional coordinates of 17 standard keypoints were extracted to form a keypoint temporal tensor. .

[0043] The keypoint sequence is input into a temporal modeling structure for action recognition. Considering that actions at photovoltaic construction sites are typically small in amplitude and slow in pace (such as screw fixing and terminal connection), traditional fast-switching action recognition structures are not suitable. Therefore, a one-dimensional convolutional network structure is used instead of the RNN structure, with the number of channels set to [value missing]. With a kernel size of 3, output the behavioral feature vector for each person. ,in This serves as a semantic representation of the current task action.

[0044] The spatial location is calculated based on the center point of the detection box. By performing planar coordinate back projection using camera intrinsic parameters, the position points of personnel in the planar coordinate system of the construction site can be obtained. Then, match it with the known workstation layout diagram to obtain the workstation number. However, in real-world scenarios, the location of personnel often fluctuates due to factors such as equipment obstruction and terrain changes in the construction area, meaning not all locations constitute valid workstations. Therefore, we designed a workstation constraint based on stage-specific semantic guidance, through... Control whether the spatial location is within the target workstation area of ​​the current stage.

[0045] Define the behavior-space-stage consistency loss function:

[0046] ;

[0047] in: For personnel The current spatial coordinates; For the stage The corresponding main operational area center is given by the deployment map configuration file; Cosine similarity between behavioral semantics and stage semantics; The average position of each person in the previous phase, used to limit sudden position changes (especially during inter-component jump operations); coefficient Separately control the weights for semantic consistency and positional continuity. It is a small constant used for numerical stability.

[0048] The loss function has the following advantages: (1) it uses stage semantics to control the spatial region; (2) it uses behavior vectors and stage vectors to ensure that the action is consistent with the current task; (3) it adds a rare "inverse distance regularization term" as a position smoothing term to model the physical continuity of personnel behavior migration, which is especially suitable for continuous workstation transition modeling of "drag-and-drop" operations such as cable laying.

[0049] The system ultimately outputs two results for each worker at the current moment: behavior label. Discrete values ​​such as "crouching," "raising hands," "pulling a line," and "stepping on high" are jointly determined by the aforementioned convolutional structure and stage-guided consistency constraints; workstation number. The spatial edge connections in the subsequent graph structure are determined by projection matching and region partitioning, ensuring that they have a positional basis.

[0050] Step 3: Construct a work graph structure based on behavioral feature vectors, stage representation vectors, and workstation numbers. Update the work graph structure using a graph embedding update function. Identify the risk scores of nodes in the updated work graph structure using a node risk scoring function. Specifically, this includes:

[0051] This step plays a crucial transitional role in the system, moving from individual job behavior information to structured job relationship reasoning, based on the behavioral semantic vectors of each person identified in the previous stage. Workstation Number and the global semantic vector of the construction phase Construct the work diagram structure of the work site at the current moment. Based on this graph structure, potential violation risk personnel nodes are identified. Unlike traditional methods that only judge whether individual actions are abnormal, the core innovation of this step lies in simultaneously modeling the three-dimensional consistency relationship of "behavior-location-stage" based on the graph structure, and using graph neural reasoning mechanisms to identify "whether the behavior is structurally reasonable", thereby forming a high-level risk perception capability. In photovoltaic power plant construction scenarios, especially in the stage of intensive multi-person collaboration such as component laying and cable connection, risks are often not caused by individual behaviors, but rather by "behavioral misalignment", "workstation crowding", or "out-of-stage collaboration" between personnel. Therefore, single-point identification methods cannot meet the requirements of accurate monitoring.

[0052] First, construct the node set in the task graph structure. ,in The total number of workers detected in the current time window, for each node. Characterization personnel Its feature vector is constructed by concatenating three elements:

[0053] ;

[0054] Among the symbols This represents a vector concatenation operation. The behavioral feature vector output from step two is obtained by encoding the keypoint sequence using a one-dimensional convolutional network. Number the workstation The embedding vector can be obtained by looking up a table or mapped using a positional coding network; This is the semantic representation for the construction phase, output by the Bi-GRU network from step one. The concatenated node vector... ,in , which serve as input features for graph neural networks.

[0055] edge set The construction is based on behavioral relationship rules and spatial arrangement logic: if the Euclidean spatial distance between two people does not exceed a specified threshold If the behavior combination is defined at the current stage as a type pair that may have a cooperative or conflicting relationship (e.g., "lifting board-fixing", "twisting-connecting"), then an undirected edge is established between its node pairs. Work diagram structure This demonstrates the structural relationship between personnel behavior and spatial relationships at the work site.

[0056] To enhance the sensitivity and interpretability of risk behavior identification, we design a novel graph embedding update mechanism. This mechanism introduces a "stage consistency modulation term" and a "behavioral mutation penalty term" on top of standard graph convolution. The update formula is as follows:

[0057] ;

[0058] in Represents a node The embedding representation result; Here is the weight matrix of the graph convolutional layer; Represents a node The set of neighbors; For node degree; For activation functions; This indicates the similarity between behavioral characteristics. A low similarity means that the two people have very different behavioral patterns. The modulation coefficient; This is a predefined behavioral conflict correction term used to enhance the system's ability to identify individuals with significantly different behaviors in the same space. For example, in a high-altitude work area, if one person is in a connected state while another is in a twisting state and they are very close to each other, It can be used to amplify such abnormal collaborations.

[0059] Based on the graph embedding results, a node risk scoring function is defined to determine whether a node is a potential source of violation. The scoring function considers not only the deviation between its structural embedding and original behavioral features, but also introduces stage-workstation consistency constraints, as shown in the following formula:

[0060] ;

[0061] in For nodes The risk score, The first item indicates the degree of deviation between the structural reasoning result and the original behavioral characteristics; the second item is the legality check of the stage workstation. This is the set of workstation numbers allowed in the current stage. This is an indicator function; the third term is used to determine whether someone in the same workstation as a neighboring person exhibits inconsistent behavior. This indicates a mismatch between the behavioral labels. Controlling the intensity of penalties. This design is particularly suitable for scenarios in photovoltaic construction where "multiple people working at the same station" are prone to misoperation and mutual interference, effectively improving the system's ability to identify anomalies under the requirement of consistency in collective behavior.

[0062] The final output is: (1) Work diagram structure (2) Risk score for each node, for use in subsequent violation reasoning and graph backtracking; This serves as a prerequisite indicator for whether to trigger an alarm or control action. This step constructs a three-dimensional violation perception system of "task perception - behavior relationship - spatial collaboration" by introducing stage guidance, multi-source modulation, and anomaly amplification mechanisms. It breaks through the applicability limitations of traditional behavior recognition in complex multi-workstation collaborative environments and is one of the core mechanisms that support the entire system to maintain accuracy and on-site adaptability in actual deployment.

[0063] Step 4: Based on the preset judgment logic and combined with risk scoring, generate a violation label through the violation judgment function, and simultaneously output the control response intensity through the response intensity calculation function. Based on the violation label and the control response intensity, generate linked response content, specifically including:

[0064] This step is responsible for making the final judgment on violations based on the constructed work graph structure, and generating specific linkage response strategies accordingly. This step is a crucial link in the entire system's transition from "identification" to "intervention." Its judgment logic needs to combine graph structure reasoning results, behavioral semantic features, workstation spatial location, and stage task context to achieve high-confidence violation judgment and output a reasonable control response level. Because photovoltaic power plant construction often encounters problems such as "interleaved use of collaborative workstations," "unclear boundary demarcation during stage switching," and "chaotic regional operation permissions," the system, when performing the final judgment, not only needs to rely on model reasoning scores but must also consider whether the behavior / workstation restrictions defined in the construction specifications have been violated. To ensure reasonable system responses, reliable alarms, and the requirement of "implementable linkage execution," this step, during the modeling process, not only considers the accuracy of risk identification but also enhances the interpretability and operability of the decision-making.

[0065] The inputs for this step include the following items, all of which are outputs from the previous step or data from the preceding steps: Work diagram structure Represents the job relationship network within the current time period, with each node... Represents a construction worker; graph neural network pairs nodes The embedded feature output is , is calculated from the graph convolution structure; The original behavioral semantic vector is obtained by temporal convolutional encoding of the keypoint sequence in the second step; The workstation number is obtained from the spatial mapping process in the second step; The construction stage number is derived from the classification results of step one; Node risk scores are calculated based on the degree of semantic deviation from the graph structure, behavioral consistency, and workstation compliance. Additionally, a construction phase rule table is loaded during system deployment. Provided by safety management personnel, this record contains the set of permitted behaviors for each stage. Allowed workstation set And its risk level.

[0066] To achieve actual control response, the system needs to clearly determine whether each person constitutes a violation. The judgment logic is divided into three levels:

[0067] First is the judgment of behavior legality. The system matches the current behavior tags of each person with the behavior whitelist of the current construction stage ; if it does not belong to the whitelist range, it is initially marked as a behavior non - compliance. Secondly is the judgment of the legality of the work position. Query whether the current work position number belongs to the set of work positions allowed in the stage ; if someone is active in a non - permitted area (such as entering the cable laying area during the bracket installation stage), it is marked as a position violation. The third layer is the model - assisted judgment mechanism: for those personnel who are not clearly violated in the above two items but are in boundary situations, the system introduces a risk score as the basis for soft judgment. If the score exceeds the threshold (usually set to 0.75), the system can also mark it as a potential violation. The combination of the three constitutes the final violation judgment function:

[0068] ;

[0069] where is the violation label of person , with values of 0 or 1, is the logical judgment function, and if violated, it is 1; is the soft judgment weight, which can be dynamically set in the system and usually takes a value of 0.5 to balance the priorities of model perception and rule control. For example, during the component laying stage, a person is identified as performing the "bending and fixing" action, but his work position number is in the "inverter room wiring area" and not in , and at the same time , higher than the threshold, then , and is finally determined to be a violation.

[0070] After completing the determination of the violation label, the system needs to further output the control response intensity to drive different levels of linkage control. For example, high - risk violation behaviors should trigger voice broadcasts and audible and visual warnings; medium - risk behaviors only record logs and prompt on - site management personnel; minor behaviors are only observed and recorded. To achieve hierarchical response, the following response intensity calculation function is designed:

[0071] ;

[0072] where is the control response score (normalized to [0, 1]), The risk level of a workstation is indicated by the workstation deployment map. For example, a risk level of 1.0 is given for areas near slopes, while a risk level of 0.2 is given for regular ground work areas. This indicates the severity of the violation in the current stage, obtained by looking up the rule table. For example, the "twisting" action has a score of 0.8 in the electrical connection stage and 0.1 in the bracket installation stage. Scoring the risk model; three coefficients The values ​​are set to [0.3, 0.3, 0.4] respectively to balance structural risk, rule strength, and model warning.

[0073] Ultimately, according to and Jointly decide on the content of the coordinated response, for example when and When [the event occurs], an audible and visual alarm is triggered, and the access control system is activated to restrict entry into the area; when [the event occurs], [the system will]... and At that time, a broadcast voice reminder is given, and the system notification module is simultaneously managed; when but At that time, violations are only recorded in the violation log for later review. This creates a clear, executable, and highly adaptable response logic that avoids both excessive system alerts and the omission of risky behaviors.

[0074] refer to Figure 2 As shown, in another aspect of this application, an intelligent monitoring system for construction safety of photovoltaic power plants based on AI video recognition is proposed, including:

[0075] The image recognition module is used to take the input image frame sequence, extract features from the image frame sequence through a convolutional network, generate an image vector sequence, input the image vector sequence into a bidirectional gated recurrent unit, and output a stage representation vector. The stage classifier outputs the construction stage number corresponding to the stage representation vector.

[0076] The behavior and workstation recognition module is used to acquire the image processing window numbered according to the corresponding construction stage. Each frame of the image in the image processing window is input into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction personnel in each frame. The position coordinates are correlated across frames by Kalman filtering and Hungarian matching algorithm to construct continuous image segments. Key point estimation is performed on the continuous image segments to generate key point temporal tensors. Action recognition is performed on the key point temporal tensors to generate behavior feature vectors. The center point of the bounding box is used to perform planar coordinate back projection to generate position points. The position points are matched with the known workstation division map to generate workstation numbers.

[0077] The task graph construction module is used to construct the task graph structure based on behavioral feature vectors, stage representation vectors, and workstation numbers. It introduces a graph embedding update function to update the task graph structure and identifies the risk scores of nodes in the updated task graph structure through a node risk scoring function.

[0078] The linkage response generation module is used to generate violation tags based on preset judgment logic and risk scoring through a violation judgment function, and output control response intensity through a response intensity calculation function. Based on the violation tags and control response intensity, linkage response content is generated.

[0079] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent monitoring of construction safety in photovoltaic power plants based on AI video recognition, characterized in that, include: The input image frame sequence is taken, and features are extracted from the image frame sequence through a convolutional network to generate an image vector sequence. The image vector sequence is input into a bidirectional gated recurrent unit and outputs a stage representation vector. The construction stage number corresponding to the stage representation vector is output through a stage classifier. The image processing window corresponding to the construction stage number is obtained. Each frame of the image in the image processing window is input into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction personnel in each frame. The position coordinates are correlated across frames by Kalman filtering and Hungarian matching algorithm to construct continuous image segments. Key point estimation is performed on the continuous image segments to generate key point temporal tensors. Action recognition is performed on the key point temporal tensors to generate behavioral feature vectors. The center point of the bounding box is used to perform planar coordinate back projection to generate position points. The position points are matched with the known work station division map to generate work station numbers. A task graph structure is constructed based on behavioral feature vectors, stage representation vectors, and workstation numbers. A graph embedding update function is introduced to update the task graph structure. The risk score of nodes in the updated task graph structure is identified through a node risk scoring function. Based on the preset judgment logic and combined with risk scoring, a violation label is generated through a violation judgment function, and a control response intensity is output through a response intensity calculation function. Based on the violation label and the control response intensity, linked response content is generated.

2. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, After generating location points by back-projecting the center point of the bounding box into planar coordinates and matching them with the known workstation division diagram to generate workstation numbers, the process further includes: The consistency loss function is used to determine the consistency between workstation number, stage representation vector and behavior feature vector. The consistency loss function includes location coordinates, the center of the main construction area corresponding to different stages, the cosine similarity between the behavior feature vector and the stage representation vector, the average position of the construction personnel in the previous stage, weight coefficients and small constants.

3. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The stage classifier includes a weight matrix, intermediate layer dimensions, total number of construction stage categories, bias vector, activation function, and normalization mapping function.

4. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The work graph structure includes a set of nodes and a set of edges. The set of nodes is generated by splicing together behavioral feature vectors, stage representation vectors, and workstation numbers. The set of edges is constructed based on behavioral relationship rules and spatial layout logic.

5. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The parameters of the graph embedding update function include the node embedding result, the weight matrix of the graph convolutional layer, the node's neighborhood set, node degree, activation function, behavioral feature similarity, modulation coefficient, and behavioral conflict correction term.

6. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The parameters of the node risk scoring function include the degree of deviation between the structural reasoning result and the original behavioral characteristics, the set of workstation numbers allowed in the current stage, the indicator sub-function, and the penalty intensity.

7. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The parameters of the violation determination function include behavior label, behavior whitelist, current workstation number, allowed workstation set, risk score, and soft determination weight coefficient.

8. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The response intensity calculation function includes the workstation risk level, the severity of the violation in the current stage, the risk score, and the balance coefficient. The workstation risk level is determined by the workstation deployment icon. The severity of the violation in the current stage is obtained by looking up a rule table. The balance coefficient is used to balance structural risk, rule intensity, and model warning.

9. The intelligent monitoring method for construction safety of photovoltaic power plants based on AI video recognition according to claim 1, characterized in that, The preset judgment logic includes behavior legality judgment, workstation location legality judgment, and model-assisted judgment. The model-assisted judgment is used to judge the violation of personnel who are not clearly in violation but are in a boundary situation.

10. An intelligent monitoring system for construction safety of photovoltaic power plants based on AI video recognition, characterized in that, include: The image recognition module is used to take the input image frame sequence, extract features from the image frame sequence through a convolutional network, generate an image vector sequence, input the image vector sequence into a bidirectional gated recurrent unit, and output a stage representation vector. The stage classifier outputs the construction stage number corresponding to the stage representation vector. The behavior and workstation recognition module is used to acquire the image processing window numbered according to the corresponding construction stage. Each frame of the image in the image processing window is input into a preset human detection network. The human detection network detects the bounding box of each frame and outputs the position coordinates of the construction personnel in each frame. The position coordinates are correlated across frames by Kalman filtering and Hungarian matching algorithm to construct continuous image segments. Key point estimation is performed on the continuous image segments to generate key point temporal tensors. Action recognition is performed on the key point temporal tensors to generate behavior feature vectors. The center point of the bounding box is used to perform planar coordinate back projection to generate position points. The position points are matched with the known workstation division map to generate workstation numbers. The task graph construction module is used to construct the task graph structure based on behavioral feature vectors, stage representation vectors, and workstation numbers. It introduces a graph embedding update function to update the task graph structure and identifies the risk scores of nodes in the updated task graph structure through a node risk scoring function. The linkage response generation module is used to generate violation tags based on preset judgment logic and risk scoring through a violation judgment function, and output control response intensity through a response intensity calculation function. Based on the violation tags and control response intensity, linkage response content is generated.

Images