Photovoltaic power station construction quality defect ai identification system based on digital twinning

Abstract

The present application relates to the technical field of photovoltaic power station, and particularly relates to a photovoltaic power station construction quality defect AI identification system based on digital twinning, which comprises a model training module, a model adjustment module, a model updating module and a model iteration module. Through the cooperative operation of the four modules, an iterative cycle of "virtual pre-training, on-site self-adaptation, global knowledge aggregation and continuous closed-loop evolution" is constructed, the problem that a single module cannot independently solve model generalization and continuous adaptation is avoided, the dependence on a large amount of annotation data of any single project is significantly reduced, the model can realize high-performance migration and application among different photovoltaic power station projects, and the identification ability of the model is continuously enhanced with the expansion of the system deployment range and the accumulation of the running time, so that the inherent problems of the prior art, such as insufficient generalization ability, difficulty in direct cross-project reuse and the need for repeated investment of resources for data collection and model training due to the high dependence of model training on specific project data, are solved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant technology, and specifically to an AI-based system for identifying construction quality defects in photovoltaic power plants based on digital twins. Background Technology

[0002] In the construction of photovoltaic power plants, timely and accurate identification of construction quality defects is crucial. It not only directly affects the construction efficiency and safety of the power plant, but also determines the long-term operating performance and investment returns of the power plant.

[0003] Currently, the industry is exploring identification methods that combine digital twins and artificial intelligence technologies. This involves constructing virtual models of photovoltaic power plants and training AI recognition algorithms to automate and intelligently inspect quality. However, these methods typically rely on data collected and labeled within a specific project environment for model training and optimization. This results in models that are severely limited by the specific construction conditions, component types, and technological characteristics of each project. When the deployed model is applied to another photovoltaic power plant project, its recognition performance significantly declines, making direct reuse difficult. This forces each new project to invest substantial resources in data re-collection and model retraining, severely restricting the large-scale application and economic benefits of the model. Summary of the Invention

[0004] To address the technical problem that existing models cannot achieve high-performance applications across photovoltaic power plant projects, this application provides an AI-based photovoltaic power plant construction quality defect identification system based on digital twins.

[0005] The technical solution of the photovoltaic power plant construction quality defect identification system based on digital twin provided in this application is as follows:

[0006] A photovoltaic power plant construction quality defect identification system based on digital twins includes:

[0007] The model training module is used to obtain virtual construction scenarios under various quality defect types and the corresponding labeled data for each virtual construction scenario through the digital twin platform, and to pre-train the initial recognition model to obtain the pre-trained model.

[0008] The model adjustment module is used to load pre-trained models on each edge AI node of the collaborative AI node network after deploying the collaborative AI node network on the photovoltaic power station, and adjust the corresponding pre-trained models based on the construction images collected on the edge AI nodes to obtain the node optimization model.

[0009] The model update module is used to upload the model parameters of the optimized model at each node to the digital twin platform, so that the model parameters can be aggregated through the digital twin platform to obtain a global model;

[0010] The model iteration module is used to update each edge AI node based on the global model, and then generate corresponding adversarial operations in the virtual construction scenario corresponding to the digital twin platform based on the new quality defect types identified during the operation of the photovoltaic power station, so as to iteratively update the pre-trained model based on the adversarial operations.

[0011] Furthermore, the steps for obtaining virtual construction scenarios for each quality defect type and the corresponding annotation data for each virtual construction scenario through the digital twin platform include:

[0012] By integrating the parameterized component library and construction process rules into the digital twin platform, an initial virtual construction scenario is generated, and the various quality defect types defined by historical quality defect data and the corresponding defect generation parameters for each quality defect type are determined.

[0013] The parameters for generating each defect are substituted into the initial virtual construction scenario for simulation to obtain the virtual construction scenarios corresponding to each type of quality defect, and the labeled data on each virtual construction scenario is obtained.

[0014] Furthermore, the steps for pre-training the initial recognition model to obtain a pre-trained model include:

[0015] The initial recognition model is pre-trained in the first stage based on each virtual construction scenario and the corresponding labeled data to obtain the basic model. The quality defects in each virtual construction scenario are then identified based on the basic model to obtain the recognition results.

[0016] Based on each identification result, corresponding adversarial quality defect samples and defect annotations are generated. The adversarial quality defect samples and defect annotations are then integrated into the corresponding virtual construction scenario to obtain the training dataset.

[0017] The base model is pre-trained in the second stage based on the training dataset to obtain the pre-trained model.

[0018] Furthermore, the steps for deploying a collaborative AI node network on photovoltaic power plants include:

[0019] Based on the real-time construction accuracy and quality risk prediction synchronized by the digital twin platform, the deployment priority and node function configuration strategy for each construction area are determined.

[0020] Based on deployment priority and node function configuration strategy, corresponding edge AI nodes are deployed in each construction area, and each edge AI node is initialized based on the initial configuration parameters set on the digital twin platform to form an initial AI node network.

[0021] Based on the real-time operating data and load status reported by each edge AI node, resource scheduling analysis is performed through the digital twin platform to generate resource scheduling strategies;

[0022] In the initial AI node network, based on the resource scheduling strategy, nearby idle computing resources are allocated to each edge AI node and / or the task priority of the edge AI nodes is adjusted to form a collaborative AI node network.

[0023] Furthermore, based on the construction images collected from the edge AI nodes, the steps for adjusting the corresponding pre-trained model to obtain the node optimization model include:

[0024] After collecting construction images of the construction area to which the edge AI node belongs, the image features are obtained by performing augmented representation learning on each construction image.

[0025] Based on image features and virtual scene features of various virtual construction scenarios stored on the digital twin platform, the pre-trained model is adjusted for feature alignment to obtain an adaptive model;

[0026] Based on the labeled defect data stored on each edge AI node, the adaptive model is sample-adapted to obtain the initial optimized model;

[0027] Based on the initial optimization model, the construction images are simulated and inferred to obtain inferred quality defect data. Then, the initial optimization model is iteratively updated using the inferred quality defect data and construction images to obtain the node optimization model.

[0028] Furthermore, the steps to aggregate the parameters of each model through the digital twin platform to obtain the global model include:

[0029] Based on the historical recognition records and quality confirmation feedback on each edge AI node, the performance of the corresponding node optimization model is evaluated, and the quality confidence coefficient of each node optimization model is generated.

[0030] Based on each quality confidence coefficient, the model parameters on the corresponding node optimization model are weighted and aggregated to obtain the initial model;

[0031] The virtual construction scenarios stored on the digital twin platform are substituted into the initial model for reasoning. After obtaining the reasoning results and reasoning logic, the model is trained based on the reasoning results and reasoning logic to obtain the global model.

[0032] Furthermore, the steps for updating each edge AI node based on the global model include:

[0033] By analyzing the twin construction environment data and historical quality defect distribution characteristics corresponding to each edge AI node through the digital twin platform, the adaptation requirements of each edge AI node to the global model are evaluated, and the model update strategy is obtained.

[0034] Based on the model update strategy, inference model parameters adapted to each edge AI node are extracted from the global model, and virtual scene adjustment data of each virtual construction scenario are combined to generate targeted model update parameters.

[0035] The update parameters of each orientation model are sent to the corresponding edge AI nodes so that each edge AI node can update the corresponding node optimization model according to the update parameters, thus obtaining the updated node optimization model.

[0036] Furthermore, the steps to obtain the updated node optimization model include:

[0037] The updated node optimization model is used to identify quality defects in real-time construction images on edge AI nodes, and the performance gain of the current model update is evaluated based on the consistency between the identification results and the feedback from manual verification.

[0038] Based on the performance gains, the quality confidence coefficients of the updated node optimization models are updated.

[0039] Beneficial effects achieved:

[0040] This application provides an AI-based photovoltaic power plant construction quality defect recognition system based on digital twins, comprising: a model training module, used to acquire virtual construction scenarios for each quality defect type and corresponding labeled data for each virtual construction scenario through a digital twin platform, and pre-train an initial recognition model based on the labeled data to obtain a pre-trained model; a model adjustment module, used to load the pre-trained model onto each edge AI node of the collaborative AI node network after deploying it on the photovoltaic power plant, and adjust the corresponding pre-trained model based on construction images collected on the edge AI nodes to obtain a node optimized model; a model update module, used to upload the model parameters of each node optimized model to the digital twin platform so that the model parameters can be aggregated through the digital twin platform to obtain a global model; and a model iteration module, used to update each edge AI node based on the global model, and generate corresponding adversarial operations in the virtual construction scenario corresponding to the digital twin platform based on new quality defect types identified during the operation of the photovoltaic power plant, so as to iteratively update the pre-trained model based on the adversarial operations.

[0041] In this application, the initial identification model is pre-trained using virtual construction scenarios and labeled data under various quality defect types generated by a digital twin platform through a model training module. This enables the generated pre-trained model to acquire basic identification capabilities covering a wide range of quality defect types, thereby reducing reliance on real data from a single project and laying a generalization foundation for cross-project applications. Next, the model adjustment module deploys a collaborative AI node network on the photovoltaic power station. After loading the pre-trained model, each edge AI node on the collaborative AI node network can perform targeted adjustments to the pre-trained model based on locally acquired construction images, resulting in a node-optimized model. This process allows the node-optimized model to adapt to the construction conditions and environmental differences of specific projects, effectively mitigating performance degradation caused by changes in scenarios between projects. Finally, the model update module updates the node-optimized models... The model parameters are uploaded to the digital twin platform for aggregation to obtain a global model. This aggregation operation integrates optimization experience from multiple projects, enabling the global model to learn common features across projects and enhance generalization knowledge. Finally, the model iteration module updates each edge AI node based on the global model and generates adversarial operations in the virtual construction scenario corresponding to the digital twin platform based on new quality defect types identified during the operation of the photovoltaic power station to iteratively update the pre-trained model. This closed-loop process enables continuous learning of new knowledge from actual projects and feedback to the virtual training environment, thereby continuously strengthening the model's adaptability and robustness. Ultimately, the system can gradually improve the model's recognition performance in different photovoltaic power station projects without relying on a large amount of new project labeled data, solving the problem that existing models cannot achieve high-performance cross-project applications. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the module of the photovoltaic power plant construction quality defect AI recognition system based on digital twins, which is the subject of this application.

[0043] Explanation of icon numbers:

[0044] 10. Model training module; 20. Model adjustment module; 30. Model update module; 40. Model iteration module. Detailed Implementation

[0045] The following combination Figure 1 This application will be described in further detail.

[0046] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0047] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0048] This application discloses an AI-based system for identifying construction quality defects in photovoltaic power plants based on digital twins.

[0049] Please refer to Figure 1 The photovoltaic power plant construction quality defect identification system based on digital twin proposed in this embodiment includes:

[0050] The model training module 10 is used to acquire virtual construction scenarios for each quality defect type and the corresponding labeled data for each virtual construction scenario through the digital twin platform, and to pre-train the initial recognition model based on the labeled data to obtain a pre-trained model. The model adjustment module 20 is used to load the pre-trained model on each edge AI node of the collaborative AI node network after deploying the collaborative AI node network on the photovoltaic power station, and to adjust the corresponding pre-trained model based on the construction images collected on the edge AI nodes to obtain the node optimized model. The model update module 30 is used to upload the model parameters of each node optimized model to the digital twin platform so that the model parameters can be aggregated through the digital twin platform to obtain a global model. The model iteration module 40 is used to update each edge AI node based on the global model, and then generate corresponding adversarial operations in the virtual construction scenario corresponding to the digital twin platform based on the new quality defect types identified during the operation of the photovoltaic power station, so as to iteratively update the pre-trained model based on the adversarial operations.

[0051] This embodiment proposes a complete closed-loop optimization process constructed through the collaborative operation of four modules. The model training module 10 utilizes a digital twin platform to generate virtual construction scenarios and labeled data covering various quality defect types, pre-training the initial recognition model to endow it with recognition capabilities and generalization independent of any specific real-world project. The model adjustment module 20 deploys a collaborative AI node network at the specific photovoltaic power station project site. It standardizes, compresses, and encrypts the pre-trained model files to form a securely transmittable model package. This package is then broadcast or distributed point-to-point to each edge AI node via a stable network communication link (such as wired broadband or a 5G private network). Each edge node receives the package... Upon receiving the model package, data integrity verification and decryption are performed first. Then, the model loading interface of the locally deployed AI framework (such as PyTorch or TensorFlow) is called to parse and load the model parameters and structure into memory, completing the loading of the pre-trained model onto the edge AI nodes. This allows each edge AI node with the pre-trained model to make targeted adjustments to the pre-trained model based on locally collected real construction images, resulting in a node-optimized model. The purpose is to enable the model to quickly adapt to the unique environment and subtle differences of specific projects, achieving a smooth migration from virtual to physical scenes. The model update module 30 collects the model parameters of each node's optimized model and uploads them to the digital twin platform for aggregation, resulting in a model that integrates multiple... The global model, based on the project's on-site experience, enhances the model's comprehensive judgment and robustness through knowledge aggregation. The model iteration module 40 distributes the global model to each edge AI node to improve its performance. Based on new quality defect types identified during actual operation, it generates corresponding adversarial operations in the digital twin platform to iteratively update the initial pre-trained model. This forms the core loop from the physical world feeding back new knowledge to the virtual world. Specifically, when an edge AI node detects a previously unrecorded new quality defect type through sensors or image recognition models during real-time operation, it synchronously uploads the characteristic data of this new quality defect type (such as defect morphology, location, environmental conditions, etc.) to the digital twin platform. In its corresponding virtual construction scenario, the platform uses scene editing and parametric modeling tools to simulate and generate virtual instances corresponding to the new quality defect type. For example, it adjusts the texture, structure, or lighting parameters of the virtual photovoltaic module to accurately reproduce the visual or physical features of the defect, thereby generating virtual scene data containing the new quality defect type as adversarial operations. Subsequently, using the adversarial virtual data in the adversarial operations, combined with the original training dataset, the pre-trained model is incrementally trained or fine-tuned. The model parameters of the pre-trained model are optimized through backpropagation, enabling the pre-trained model to learn the ability to identify the new quality defect type. For example, adversarial training algorithms can be used to alternate between normal data and adversarial data during the training process to enhance the robustness of the model.

[0052] The combination of the above four modules forms an iterative cycle of "virtual pre-training → on-site adaptation → global knowledge aggregation → continuous closed-loop evolution," which avoids the problem that a single module cannot independently solve the problem of model generalization and continuous adaptation. It significantly reduces the dependence on a large amount of labeled data for any single project, enabling the model to achieve high-performance transfer and application between different photovoltaic power station projects. As the system deployment scope expands and the running time accumulates, its recognition ability will continuously improve. This solves the inherent problems of insufficient generalization ability, difficulty in direct cross-project reuse, and the need for repeated investment of resources for data collection and model training caused by the high dependence of model training on specific project data.

[0053] In one feasible implementation, the specific execution steps of the model training module include steps S11 to S15:

[0054] Step S11: Using the parameterized component library and construction process rules integrated in the digital twin platform, an initial virtual construction scene is generated, and the various quality defect types defined by historical quality defect data and the defect generation parameters corresponding to each quality defect type are determined.

[0055] It should be noted that the parametric component library contains three-dimensional digital models of various components of photovoltaic power plants, such as photovoltaic panels, brackets, combiner boxes, and cables, as well as their adjustable parameters, such as dimensions, tilt angle, orientation, and spacing. At the same time, the construction process rules define the assembly logic, connection methods, and installation sequence constraints of these components.

[0056] Based on project design drawings or configuration parameters, the digital twin platform automatically selects and instantiates corresponding component models from the parametric component library. Then, according to the logic and constraints stipulated by the construction process rules, these components are assembled and laid out in the digital twin platform to generate a complete three-dimensional virtual model of a photovoltaic power station that conforms to design specifications and has accurate geometric and topological relationships. This model is the initial virtual construction scene.

[0057] Based on this, historical quality defect data collected from past engineering cases are summarized and abstracted to define discrete quality defect types such as "hidden cracks in battery cells," "tilted frame installation," and "incomplete welds on the welding strips." Defect generation parameters are analyzed for each type of quality defect. For example, for "hidden cracks," the defect generation parameters may include crack length, width, direction, and position coordinates on the battery cell. These predefined quality defect types and their corresponding defect generation parameters provide a blueprint for determining virtual construction scenarios in the correct component locations and in a form that conforms to physical laws.

[0058] Step S12: Substitute the generation parameters of each defect into the initial virtual construction scene for simulation to obtain the virtual construction scene corresponding to each quality defect type, and obtain the annotation data on each virtual construction scene.

[0059] The simulation engine on the digital twin platform locates the corresponding parametric components in the initial virtual construction scene based on the defect generation parameters defined for each quality defect type in the previous step. Then, it drives the physical simulation module or geometric deformation algorithm to simulate the physical form and visual appearance of the quality defects on the components according to these defect generation parameters. For example, it cuts and renders the grid of the battery cell model according to the crack parameters (defect generation parameters) to simulate the crack effect, or adjusts the Euler angle of the support model according to the installation tilt parameters (defect generation parameters), thereby generating a new virtual construction scene that embeds a specific type and shape of defect. By traversing all quality defect types and performing the above process for each set of corresponding defect generation parameter combinations, a massive collection of virtual construction scenes covering all quality defect types and their various morphological variations can be generated in batches.

[0060] At the same time, since each quality defect type is generated programmatically according to preset parameters, its category, geometric location, shape and size, and the corresponding defect generation parameters are automatically recorded and structured during the simulation process. This information constitutes the annotation data corresponding to the virtual construction scene.

[0061] Step S13: Based on each virtual construction scenario and the corresponding labeled data, the initial recognition model is pre-trained in the first stage to obtain a basic model. Based on the basic model, the quality defects in each virtual construction scenario are identified to obtain the recognition results.

[0062] These virtual construction scenes and their corresponding labeled data are used as training samples and input into an initial recognition model, such as a deep learning network or convolutional neural network. The predicted output is calculated through forward propagation and then compared with the labeled data to calculate the loss function (e.g., cross-entropy loss for classification tasks or mean squared error loss for localization tasks). Then, using the chain rule, the gradient of the loss function with respect to each parameter of the initial recognition model is calculated layer by layer from the output layer to determine the contribution of each parameter to the total error. Finally, for each parameter in the initial recognition model, its new value is equal to the current value minus the learning rate multiplied by the gradient of that parameter. All parameters of the model are updated synchronously to reduce the loss function and ensure that they move simultaneously along their respective paths. The gradient is adjusted in the opposite direction. Through repeated iterations of this update process, the model parameters are gradually optimized to a region with a smaller loss function, thereby driving the model's predicted output to be closer to the real label, and ultimately improving the overall recognition performance of the model. After this process is completed in one iteration cycle (i.e., traversing the entire training dataset once), multiple iteration cycles are repeated. After each cycle, the model's recognition accuracy, recall, and other performance indicators are evaluated on an independent validation dataset, and the trend of the loss function is monitored. When the loss function tends to stabilize and converges to a low level for several consecutive cycles, it is considered that the recognition performance of the initial recognition model has reached a stable state. At this point, the iteration is stopped, and the resulting model is the basic model that has learned the basic features of various defects.

[0063] Next, in order to evaluate the initial capabilities of the basic model and obtain its recognition characteristics, the virtual construction scenarios previously used for training are input into the basic model again. The basic model will perform reasoning analysis on each virtual construction scenario and output prediction information on whether there are quality defects, as well as the specific type, location, and confidence level of the quality defects. This prediction information constitutes the recognition result.

[0064] The process of re-identifying the trained virtual construction scene using the base model is not simply a repetition of training, but rather aims to obtain the base model's specific predictive performance on the current data to serve the subsequent model enhancement process. Although the base model is trained on virtual construction scenes and labeled data, its training process aims to learn generalized feature representations. After training, inference evaluation is part of the standard machine learning process. Here, the identification results are directly used in the next step to generate adversarial quality defect samples and defect labels. That is, by analyzing the identification results of the base model on the virtual construction scene, more challenging adversarial samples can be constructed in a targeted manner and then integrated into the training dataset. This further improves the model's ability to identify complex or marginal defects in the second stage of pre-training, ensuring that the model can learn from its own predictions and continuously improve.

[0065] Step S14: Based on each recognition result, generate corresponding adversarial quality defect samples and defect annotations, and then integrate the adversarial quality defect samples and defect annotations into the corresponding virtual construction scene to obtain the training dataset.

[0066] By analyzing the recognition results of the basic model on the virtual construction scene, samples with low model prediction confidence, misclassification, or positioning deviations are identified. These samples reflect the weaknesses in the current cognition of the basic model. To address these weaknesses, a batch of new virtual defect samples, namely adversarial quality defect samples, are generated programmatically in the digital twin platform by adjusting the defect generation parameters in the corresponding virtual construction scene (e.g., slightly changing the length, angle, or contrast of the crack in the background) or combining multiple defects to generate new complex forms. These samples aim to deceive or challenge the current basic model. Since the adversarial quality defect samples are generated based on parameterization, their corresponding defect categories and location information are also automatically generated, forming defect annotations.

[0067] Then, the adversarial quality defect samples and defect annotations are merged with the virtual scene samples and their annotation data generated in step S12 to form a training dataset. This constructs difficult samples that the model cannot accurately identify and adds them to the training dataset. In the subsequent second-stage pre-training, the basic model is forced to deepen its learning, break through its original capability boundaries, and significantly improve the model's robustness and generalization ability to identify fuzzy, rare and complex morphological defects.

[0068] Step S15: Perform a second-stage pre-training on the base model based on the training dataset to obtain a pre-trained model.

[0069] A supervised learning process similar to the first-stage pre-training is adopted. The training dataset is input into the base model for forward propagation and loss calculation. Then, the model parameters are updated through backpropagation and optimization algorithms. However, the focus and objectives of this training stage are different. Because the training dataset contains a large number of adversarial quality defect samples generated to target the weaknesses of the base model, the base model is forced to learn to distinguish the subtle differences between these highly similar or more deceptive quality defect patterns and their labeled data during training. This adjusts its internal feature representation and decision boundaries, enabling it to have a stronger ability to discriminate various defects, especially marginal cases and complex variants. After multiple rounds of iterative training until the performance of the base model converges on this training dataset, the final model is the pre-trained model that has completed the two-stage training. By building upon the base model's mastery of common defect features, targeted adversarial training significantly improves the model's robustness and generalization ability in identifying uncommon, ambiguous, and highly confusing defects. This ensures that the pre-trained model has the potential to handle complex and unknown defect scenarios before being deployed in subsequent real and varied photovoltaic power plant projects, laying a solid model foundation for the entire system to achieve high-performance recognition across projects.

[0070] In one feasible implementation, the specific execution steps of the model adjustment module include steps S21 to S28:

[0071] Step S21: Based on the real-time construction accuracy and quality risk prediction synchronized by the digital twin platform, determine the deployment priority and node function configuration strategy for each construction area.

[0072] The digital twin platform receives and integrates progress data, sensor monitoring data, and historical quality database information from the actual construction site in real time. It synchronizes multi-source data from the actual construction site. For real-time construction accuracy calculations, such as in the component installation area, it processes high-precision GPS or vision sensor data installed on construction machinery to obtain the position coordinates of installed components. This coordinates are then compared in real time with the precise design coordinates of the component in the digital twin to calculate its positional deviation. Simultaneously, it collects real-time bolt tightening torque values ​​through IoT bolt sensors or smart wrenches, and statistically analyzes the proportion of torque values ​​within the design standard range in the construction area to determine the torque qualification rate. For quality risk level prediction, it simulates… The True and Analytical Model comprehensively analyzes inputs from multiple dimensions, including environmental conditions such as current wind speed, temperature, and humidity obtained in real time through an environmental sensor network; the inherent complexity score of the current operation process retrieved from the construction process knowledge base; and the frequency and distribution patterns of past defects (such as hidden cracks and incomplete welds) under similar processes and environments queried from a historical quality database. Then, it uses a pre-trained risk assessment model (such as an ensemble machine learning model that integrates logistic regression, decision trees, or neural networks) to comprehensively analyze these multi-dimensional features, output the probability value of various specific quality defects occurring in the construction area during a specific construction phase in the future, and quantify it into a quality risk level.

[0073] Next, construction areas with real-time construction accuracy below a set threshold (i.e., current construction quality is already weak) and construction areas with high quality risk prediction levels are identified as critical construction areas, and their deployment priority is set to the highest to ensure that monitoring resources prioritize coverage of the most controllable aspects. Simultaneously, relying on in-depth mining of historical photovoltaic power plant project big data, the digital twin platform constructs a defect knowledge base, which records the multi-dimensional characteristic conditions corresponding to various quality defects (such as microcracks in photovoltaic modules, poor soldering of solar cells, scratches on frames, and poor sealing of junction boxes) when they occur. These multi-dimensional characteristic conditions include the type of construction area where the quality defect occurs, the specific construction technology used in that area at the time, environmental parameters during operation (such as temperature, humidity, and wind speed), batch information of materials used, and the operating team. When making predictions for a new construction project, the digital twin platform acquires the construction data stream of each construction area in real time, including environmental conditions collected by IoT sensors, process steps parsed from the construction plan, and material tracking information. These real-time features are then matched and similarity calculated with historical patterns in the defect knowledge base. Specifically, the platform uses classification models (such as gradient boosting decision trees or deep neural network classifiers) for analysis. This model takes the feature vector of the construction area as input, performs complex internal nonlinear calculations, and outputs the probability distribution of various known defect types occurring in the construction area. The top one or several defect types with the highest probability are identified as the main quality risk types of the construction area at the current stage. For example, the system may predict that the main risk of the construction area is "poor sealing of junction boxes" based on the current high humidity in the wiring area, the historical poor records of a certain batch of sealant, and the ongoing sealing process. Then, the node function configuration strategy is determined based on the predicted main quality risk types of different construction areas. For example, for construction areas predicted to be prone to appearance defects, high-definition visual recognition functions and corresponding equipment are configured for the nodes there; for construction areas predicted to be prone to electrical connection problems, infrared temperature measurement or electrical parameter monitoring interfaces and data analysis equipment may be added to the nodes. In this way, the deployment of edge AI nodes is transformed from a uniform or experience-based arrangement to a dynamic and targeted configuration driven by real-time data and prediction. This ensures that monitoring capabilities can be prioritized and most effectively focused on weak links in construction quality and potentially high-risk construction areas, thereby maximizing the overall efficiency of quality defect identification and preventive intervention capabilities at the system level, even with limited resources.

[0074] Step S22: Based on the deployment priority and node function configuration strategy, deploy corresponding edge AI nodes in each construction area, and initialize each edge AI node based on the initial configuration parameters set on the digital twin platform to form an initial AI node network.

[0075] Based on the instructions issued by the digital twin platform, the robots or construction personnel first go to the construction area marked as the highest priority, and then install the corresponding physical hardware devices on each edge AI node according to the model type specified by the node function configuration strategy corresponding to the construction area.

[0076] After the physical hardware devices are powered on and connected to the network, each edge AI node automatically downloads its pre-configured initial configuration parameters from the digital twin platform. These parameters include the node's assigned network identifier, communication protocol and key, associated digital twin sub-model identifier, model weights tailored and allocated according to its functions, and an initial detection task list and threshold parameters customized based on its regional risk. After loading the corresponding initial configuration parameters, the edge AI node completes the configuration and startup of its software environment and registers with the digital twin platform, announcing its joining the network.

[0077] With all planned edge AI nodes having completed registration and interconnection, a distributed sensing and computing network covering key construction areas, possessing differentiated detection capabilities, and having established bidirectional data and control channels with the digital twin platform has been constructed. This is the initial AI node network, which transforms the node function configuration strategy based on predictive analysis into a physically ready, functionally customized, and logically unified initial AI node network. This lays a plug-and-play hardware and system foundation for subsequent real-time acquisition of real-time working condition data, execution of local quality defect identification, and realization of collaborative computing within the network, thereby ensuring that the quality monitoring system can operate efficiently and purposefully from the very beginning of its construction.

[0078] Step S23: Based on the real-time operating data and load status reported by each edge AI node, resource scheduling analysis is performed through the digital twin platform to generate a resource scheduling strategy.

[0079] Each edge AI node periodically reports its real-time operating data to the digital twin platform, including its construction image acquisition frequency, current task queue, number and type of identified quality defects, and its own load status, such as CPU and GPU utilization, memory usage, power consumption, and network bandwidth usage.

[0080] After aggregating real-time operational data from all edge AI nodes, the digital twin platform constructs corresponding optimization functions with the optimization goals of balancing the overall network computing load and minimizing the total processing latency of tasks.

[0081] Specifically, the digital twin platform formally models the real-time aggregated operating conditions and load data of each edge AI node, as well as the network topology, into an optimization function with the joint objective of minimizing the load variance of all network nodes (to achieve load balancing) and the weighted sum of the total processing latency of all tasks. The upper limit of the computing power of each edge AI node, the task constraint that tasks must be processed before their latest completion time, and the additional communication latency caused by the migration of tasks from one edge AI node to another are used as constraints. Faced with this typical NP-hard combinatorial optimization function, the digital twin platform adopts a hybrid solution strategy. First, it quickly handles simple scenarios (such as directly scheduling some low-priority tasks on obviously overloaded edge AI nodes to their neighboring idle edge AI nodes). For complex scenarios, a candidate resource scheduling scheme is encoded as a chromosome. Each gene position on the chromosome corresponds to a task to be scheduled, and the value of that gene position represents a unique identifier for the target edge AI node to which the task is assigned. Furthermore, to determine the execution order and start time, the chromosome structure can be expanded to include two parts: the first part is the allocation mapping from the tasks to the target edge AI nodes, and the second part is the encoding of the relative execution order of each task on the assigned node. Thus, a single chromosome completely represents a scheme that allocates all tasks to each target edge AI node. The target edge AI node was identified and a complete scheduling scheme for execution order was determined. Then, a population consisting of a large number of randomly generated chromosomes was initialized. In each generation of evolution, the fitness of each chromosome was evaluated by calculating the objective function value (i.e., the weighted sum of load balancing and total latency) of the scheduling scheme represented by each chromosome and checking whether it met constraints such as node load limit, task priority, and communication latency. The fitness function was designed to be higher when the objective function value is smaller and the constraints are less violated. Then, selection operations were performed based on fitness to retain excellent individuals. Crossover operations were used to randomly exchange some gene segments of two parent chromosomes to generate offspring, exploring new allocation combinations. Mutation operations were used to randomly change the values ​​of some genes in the offspring chromosomes with a low probability (i.e., change the allocation node of the corresponding task or locally adjust the execution order), introducing random perturbation to maintain population diversity.This iterative process of selection, crossover, and mutation is repeated, driving the overall fitness of the population to continuously improve. When the preset maximum number of generations is reached or the fitness no longer significantly improves over several consecutive generations, the algorithm terminates and outputs the chromosome with the highest fitness in the current generation. Finally, the digital twin platform decodes the mechanical energy of the selected chromosome to obtain the target edge AI nodes to which each scheduled task should be assigned and their execution order. Combining this with the node's computing power and task processing time estimation, the start time of each task can be calculated, forming a resource scheduling strategy. This resource scheduling strategy clearly specifies which edge AI nodes' computing tasks need to be allocated in the future. The system migrates tasks to nearby, less loaded, idle edge AI nodes, or dynamically adjusts the execution order and resource quotas of different priority tasks on various edge AI nodes. This transforms the initial AI node network from a statically deployed collection into a system with elastic computing capabilities. It can automatically rebalance tasks and resources based on real-time load fluctuations, thereby maximizing the overall utilization of computing resources. This prevents individual edge AI nodes from becoming bottlenecks due to overload, leading to recognition delays, while ensuring that high-priority quality defect recognition tasks are processed promptly. This guarantees the stable, efficient, and low-latency continuous operation of the distributed AI recognition system.

[0082] Step S24: In the initial AI node network, according to the resource scheduling strategy, allocate nearby idle computing resources to each edge AI node and / or adjust the task priority of the edge AI nodes to form a collaborative AI node network.

[0083] The digital twin platform's resource scheduling strategy is issued as a control command to the relevant edge AI nodes. For edge AI node pairs that require task migration as specified in the resource scheduling strategy, the digital twin platform coordinates the two nodes to establish a direct point-to-point communication link. Overloaded edge AI nodes package the specified computing tasks in their task queues (such as a batch of construction image data to be identified and its context) and transmit them to a nearby idle edge AI node that is allocated as a supplement to its computing resources. The idle edge AI node then uses its idle computing power to execute the actual computing tasks and returns the results, thereby achieving the sharing and rebalancing of computing resources. At the same time, for edge AI nodes that require task priority adjustment in the resource scheduling strategy, the edge AI node will dynamically adjust the order of its local task scheduling queue according to the control command. For example, it will promote the quality defect identification tasks involving high-risk areas or key processes to the head of the queue for priority execution, while postponing some background analysis or low real-time requirements. Through this dynamic adjustment, the initially independent and statically configured edge AI nodes in the initial AI node network are tightly organized into a cohesive whole that can work collaboratively and dynamically allocate load based on the global state. This collaborative AI node network enables the entire node network to flexibly cope with local load peaks and sudden computing demands without changing the total amount of hardware resources. This significantly improves the overall system throughput and response time for real-time AI identification of construction quality defects, ensuring the continuity of critical quality monitoring tasks.

[0084] Step S25: After acquiring construction images of the construction area to which the edge AI node belongs, perform enhanced representation learning on each construction image to obtain image features.

[0085] Edge AI nodes deployed in various construction areas use their integrated image acquisition devices (such as industrial cameras or high-definition cameras) to automatically capture images of key work surfaces (such as photovoltaic module installation, cable crimping, bolt tightening, etc.) within their respective construction areas at regular intervals or based on motion detection, according to their configured acquisition plans or trigger commands issued by the digital twin platform.

[0086] Next, on the edge AI node, a pre-trained self-supervised learning model (e.g., a visual representation model based on contrastive learning) is used to perform a series of preset data transformation operations (such as random cropping, color jittering, Gaussian noise addition, etc.) on the original image to generate different enhanced views of the same construction image. Then, an encoder network is trained to bring the representations of different enhanced views from the same construction image closer together in the feature space, while pushing the representations from different images further apart. In this way, image features sensitive to essential attributes such as object shape, texture, and relative positions between components are extracted from the construction image. This allows the edge AI node to learn autonomously and extract image features that can effectively represent the construction status and potential quality defect patterns without the need for expensive manual image annotation.

[0087] Step S26: Based on the image features and the virtual scene features of each virtual construction scene stored on the digital twin platform, the pre-trained model is adjusted for feature alignment to obtain an adaptive model.

[0088] It should be noted that virtual scene features refer to feature vectors extracted from scene images of virtual construction scenes stored on those digital twin platforms using the same pre-trained self-supervised learning model to characterize virtual quality defects and the environment. These virtual scene features correspond semantically to image features but have different distributions.

[0089] A domain discriminator is introduced after the feature extraction network of the pre-trained model, and a hybrid feature set containing image features and virtual scene features is constructed. During training, while maintaining the backbone parameters of the pre-trained model for forward propagation feature extraction, a gradient inversion layer is used to prevent the domain discriminator from accurately determining whether the feature source is real or virtual. This forces the feature extraction network of the pre-trained model to adjust its parameters, making the feature distribution of its output blur the boundary between the two domains (i.e., real and virtual) as much as possible. This makes the feature distribution of image features as aligned with the feature distribution of virtual scene features as much as possible, effectively reducing the domain difference or distribution gap between the knowledge learned by the pre-trained model on rich virtual data (virtual scene features) and real construction scene data (image features). This allows the generated adaptive model to more smoothly and accurately transfer and apply the powerful defect recognition capabilities learned in the virtual environment to the current real construction environment, significantly improving the model's adaptability and recognition accuracy to complex real-world situations, and laying a key foundation for subsequent rapid sample adaptation.

[0090] Step S27: Based on the labeled defect data stored on each edge AI node, the adaptive model is sample-adapted to obtain the initial optimized model.

[0091] It should be noted that the labeled defect data refers to the actual construction defect image data stored locally on the edge AI node and confirmed and labeled by the on-site supervisor or quality inspector. It includes information such as the type and location of the quality defect.

[0092] In this process, the sample adaptation employs a few-shot learning algorithm. First, a subset is randomly selected from the stored labeled defect data. Stratified random sampling is used to ensure that each quality defect type is represented in the partitioned set. This subset is then further divided into two non-overlapping parts: a support set and a query set. During adaptation, starting with the parameter initialization of the current adaptive model, a copy of the model's parameters is used as temporary parameters in the inner loop. In each iteration of the inner loop, labeled defect data from the support set is input into the adaptive model for forward propagation. The loss between the predicted labeled missing data and labeled defect data is calculated. Then, the gradient of this loss relative to the temporary parameters is calculated using backpropagation. A pre-set, relatively high learning rate is used to update these temporary parameters according to the stochastic gradient descent update rule. This process is repeated a predetermined number of times. Thus, using only a small support set, the temporary parameters of the adaptive model are adjusted through several rapid gradient steps towards minimizing the loss on the specific support set of the current edge AI node, allowing it to quickly adapt to the specific distribution of the labeled defect data.

[0093] After several rapid gradient updates in the inner loop to obtain temporary model parameters adapted to the current support set, the labeled missing data from the query set is input into this temporary model for forward propagation to obtain its prediction results for the query set samples. These prediction results are then compared with the labeled missing data from the query set, and the loss value of the temporary model on this query set is calculated using a loss function (such as cross-entropy loss). This loss value is the meta-loss, which quantifies the generalization performance of the model after rapid adaptation in the outer loop on a relevant but unseen query set. Next, the gradient of this meta-loss with respect to the initial parameters of the adaptive model (i.e., the model before the inner loop) is calculated using the backpropagation algorithm. However, this gradient calculation passes through all update steps in the inner loop, which usually requires second-order optimization or first-order approximation. Finally, an outer loop learning rate (or meta-learning rate) and optimization parameters are used to calculate the meta-loss. An integrator (such as Adam) updates the initial parameters of the adaptive model (the model before the inner loop starts) based on the calculated gradient. This allows the model to start from this initial state and achieve better generalization performance (i.e., lower query set loss) on the query set in future new tasks. By iterating this process repeatedly, the initial parameters of the adaptive model are optimized. The adaptive model optimized in this process is then fine-tuned on all local labeled samples. The resulting model is the initial optimized model that is highly matched to the current construction environment of the edge AI node. By utilizing the labeled defect data stored on each edge AI node, the already feature-aligned adaptive model is adjusted so that it can capture and adapt to the characteristics and types of quality defects in the actual construction area where it is deployed. This further significantly improves the model's recognition accuracy and reliability at specific locations on the basis of general adaptability.

[0094] The update rule of stochastic gradient descent is as follows: In each parameter update iteration, firstly, a single data point or a mini-batch of data is randomly selected from the training dataset. The loss function value of the adaptive model on that data is calculated. Then, the gradient of the loss function with respect to the model parameters of the adaptive model is obtained. Finally, the model parameters are shifted one step in the opposite direction of the gradient. This step size is controlled by the learning rate, and its mathematical expression is: ,in Represents model parameters, Represents the learning rate. This represents the gradient of the loss function calculated based on a single data point or a small batch of data. By repeatedly executing this update rule, the model parameters will gradually adjust in the direction of minimizing the overall loss function.

[0095] Step S28: Based on the initial optimization model, perform simulation reasoning on the construction images to obtain reasoning quality defect data. Then, iteratively update the initial optimization model using the reasoning quality defect data and the construction images to obtain the node optimization model.

[0096] The construction images are input into the initial optimization model. The initial optimization model performs forward propagation calculations through its internal neural network, outputting predicted information such as the types, locations, and confidence levels of potential quality defects in the construction images. These predictions constitute the inference quality defect data. Next, the inference quality defect data and its corresponding construction image pair are treated as an augmented training set, input into the initial optimization model, and the loss between the inference quality defect data and the inference quality defect labels is calculated. Then, the gradient is calculated using the backpropagation algorithm, and the optimizer updates the model parameters of the initial optimization model. This process can be repeated; that is, the updated initial optimization model is used again to infer from the construction images, generating new inference quality defect data, which is then used for training, forming a self-learning loop until the performance of the initial optimization model stabilizes on the validation set. Finally, a node optimization model that better reflects the actual data distribution of the current edge AI nodes is obtained. The labeled defect samples mentioned here refer to training data with real labels, formed by professionals manually identifying and classifying defects in the construction images. The main effect of this approach is that, given the limited amount of manually labeled data, continuous self-training using pseudo-labels generated by the model can effectively utilize a large number of unlabeled on-site construction images. This drives the model to continuously iterate and optimize itself in specific construction scenarios, thereby significantly improving the detection accuracy and robustness of the model for complex and variable on-site defects.

[0097] In this context, inference defect labels refer to the supervisory information extracted from the inference quality defect data and used for model iteration updates. Specifically, after the initial optimization model performs simulation inference on the construction images, it outputs a complete set of inference quality defect data. This typically includes the types of quality defects identified by the initial optimization model in the construction images, their location bounding boxes, and confidence scores. In subsequent iterative update steps, the system filters out high-confidence inference quality defect data from this data (for example, only retaining inference quality defect data with a confidence score higher than a certain threshold) and uses these as inference defect labels.

[0098] In one feasible implementation, the specific execution steps of the model update module include steps S31 to S33:

[0099] Step S31: Based on the historical recognition records and quality confirmation feedback on each edge AI node, perform performance evaluation on the corresponding node optimization model and generate the quality confidence coefficient of each node optimization model.

[0100] The system retrieves historical recognition records from storage for each edge AI node, including past prediction data output by the node optimization model, such as quality defect types, bounding box locations, and confidence scores. It also retrieves corresponding quality confirmation feedback, i.e., the true labels and correctness judgments provided after external review. The system then matches and compares the predicted data with the true labels for the same construction image. Based on the comparison results, it calculates a series of standardized performance evaluation metrics. For example, it divides the number of data points predicted as "quality defect exists" and confirmed correctly by the review by the total number of data points predicted as "quality defect exists" by the node optimization model to calculate accuracy; or it divides the number of data points confirmed as "quality defect exists" and successfully predicted by the review by the total number of confirmed "quality defect exists" data points to calculate recall. The system further calculates recall using the formula: The system calculates the F1 score. During the calculation process, the system will properly handle false alarms and missed alarms based on the quality confirmation feedback to ensure the comprehensiveness of the evaluation, and directly use the calculated F1 score as the quality confidence coefficient of the optimization model of that node.

[0101] Step S32: Based on each quality confidence coefficient, the model parameters on the corresponding node optimization model are weighted and aggregated to obtain the initial model.

[0102] The digital twin platform first collects all model parameters of the node optimization models from each edge AI node, and simultaneously obtains the quality confidence coefficient calculated for each edge AI node. Then, it multiplies the modulus parameters of each network layer of each node optimization model by its corresponding quality confidence coefficient to obtain a parameter tensor. Next, it sums the parameter tensors of all edge AI nodes on the same network layer, and finally divides the sum by the sum of all quality confidence coefficients to obtain the aggregated parameters of that network layer. By repeating this operation for all network layers, a novel initial model incorporating the knowledge of all edge AI nodes is generated. This enables the secure and differentiated fusion of node optimization models distributed across various construction sites within the framework of federated learning, using quantified model parameters as trust weights. This allows node models that perform more reliably and provide better feedback in actual recognition tasks to contribute more to the final fused initial model, significantly improving the overall recognition accuracy, robustness, and generalization ability of the obtained initial model. This lays a solid and high-quality parameter foundation for subsequent refinement and compression in virtual scenes to generate a high-performance global model.

[0103] Step S33: Substitute the virtual construction scenarios stored on the digital twin platform into the initial model for reasoning, obtain the reasoning results and reasoning logic, and then train the model based on the reasoning results and reasoning logic to obtain the global model.

[0104] The digital twin platform batches programmatically generated virtual construction scenarios, encompassing various types of quality defects and construction conditions, into an initial model. Forward propagation inference is performed on the virtual construction scene images corresponding to each scenario, outputting inference results reflecting the type and location of quality defects. Simultaneously, deep information such as activation feature maps and attention weight distributions from the intermediate layers of the initial model are recorded and extracted as inference logic. Then, a more lightweight learning model is constructed as a global model prototype to be trained. During training, this learning model is required not only to approximate the inference results of the initial model in predicting quality defects on virtual construction scene images (constrained by the inference result loss function), but also to ensure that the feature representations of the intermediate layers of the learning model correspond to those of the initial model. The reasoning logic is made as similar as possible, and the loss function is constrained by feature imitation. Through this dual supervision signal, the model to be learned is fully trained so that it can imitate and inherit the comprehensive decision-making ability and internal knowledge representation of the initial model. After the training converges, a global model is obtained. By taking the initial model, which is knowledge-rich but may have a large number of parameters and a complex structure obtained through weighted aggregation, its recognition ability and internal logic are extracted and transferred to a more concise and computationally efficient model to be learned in a safe, controllable and diverse virtual construction scenario. Thus, a global model that is easier to distribute, deploy and run efficiently on various edge AI nodes is obtained with almost no loss of recognition performance. This solves the redundancy and bloat problems that may occur after model fusion.

[0105] Among these, constraining through the inference result loss function refers to establishing a specific loss function during the training of the model to be learned. The core function of this loss function is to measure the difference between the inference results of the model to be learned on virtual construction scenarios and the inference results of the initial model on the same batch of virtual construction scenarios used as supervision signals. During training, the value of the loss function is continuously calculated, and all model parameters of the model to be learned are adjusted through the backpropagation algorithm. The adjustment direction is to minimize this loss value, thereby forcing the prediction behavior of the model to be learned to become closer and closer to the statistical distribution of the initial model. Constraining through feature imitation loss function refers to the initial model... When processing input data, the intermediate layers of the network generate a series of feature maps or feature vectors that can represent the abstract features of the data, i.e., the inference logic. During training, the same batch of input data is fed into the initial model and the model to be learned, respectively, and the output features of the two at the corresponding intermediate layers (such as a convolutional layer or after a Transformer block) are extracted. Then, the distance between the two feature sets is calculated by a metric function (such as mean squared error MSE, cosine similarity, or KL divergence of the attention matrix). This distance value is the feature imitation loss. The loss value is continuously reduced through this optimization algorithm, so that the inference logic of the model to be learned is aligned with the initial model.

[0106] In one feasible implementation, the specific execution steps of the model iteration module include steps S41 to S43:

[0107] Step S41: Analyze the twin construction environment data and historical quality defect distribution characteristics corresponding to each edge AI node through the digital twin platform, evaluate the adaptation requirements of each edge AI node to the global model, and obtain the model update strategy.

[0108] It should be noted that the construction twin data refers to the virtual mirror data that is constructed in the digital twin platform through IoT sensors, BIM models and on-site monitoring, and is synchronized with the actual construction environment in real time. It includes dynamic environmental information such as the three-dimensional geometry, material status, machinery deployment, and lighting and weather of the construction area where the edge AI node is located. The historical quality defect distribution characteristics refer to the statistical regularity of the types of quality defects (such as the frequency of occurrence of cracks, leaks and dimensional deviations) presented in the historical identification records of the edge AI node and their correlation with the construction location and process stage.

[0109] The digital twin platform compares these two types of data from each edge AI node with the data corresponding to the virtual construction scenario on which the global model was trained. For example, if the twin data of an edge AI node shows that it is in the high-altitude curtain wall installation stage with complex lighting conditions, and the false alarm rate of "glass scratches" in its historical defect features is high, the digital twin platform can assess that the current construction environment and task of the edge AI node have a strong need to adapt to the global model's capability modules such as "identification of curtain wall surface defects under strong light" and "distinguishing between scratches and reflections". Based on this adaptation need, the platform determines the fine-tuning training parameters, such as the learning rate, number of iterations, and loss function weights, to ensure that the model update focuses on specific weak links. The fine-tuning scheme will explicitly instruct the extraction of relevant parameters from the global model as initialization and use these training parameters for local training to generate the corresponding model update strategy. Its content may explicitly indicate that a subset of neural network parameters related to specific quality defect types and construction environment conditions needs to be extracted from the global model.

[0110] Step S42: Based on the model update strategy, extract the inference model parameters that are compatible with each edge AI node from the global model, and combine them with the virtual scene adjustment data of each virtual construction scenario to generate targeted model update parameters.

[0111] The digital twin platform analyzes the model update strategy, which clarifies the model capability dimensions that corresponding edge AI nodes need to strengthen to meet their adaptation requirements. For example, it improves the sensitivity to identify "hidden cracks in components" or optimizes the detection robustness under "low-light conditions." Next, based on the model update strategy, the digital twin platform locates and extracts inference model parameters directly related to the model capability dimensions from the global model's inference model parameter set. For example, it extracts the feature extraction layer parameters responsible for texture analysis and the weight vector corresponding to the "hidden crack" category in the final classification layer. Then, the digital twin platform calls upon virtual scene adjustment data specifically generated for the corresponding edge AI node. This data consists of a series of virtual images and annotations that simulate the real construction environment and target quality defect challenges of the corresponding edge AI node. Using this data, the extracted inference model parameters are fine-tuned and trained within the digital twin platform. By calculating loss and gradient descent, these parameters are made to better adapt to the specific scenarios faced by the edge AI node while retaining global knowledge. After fine-tuning, the specifically adjusted inference model parameters, along with the necessary model structure configuration information, are packaged to generate a targeted model update parameter specific to the corresponding edge AI node. This achieves personalized and precise model updates, ensuring that what is sent to each edge AI node is not a complete and bloated global model, but a targeted model update parameter closely aligned with its current actual needs. This allows for efficient and direct improvement of the edge AI node's ability to identify specific quality defects in its unique construction environment with minimal communication and computational overhead.

[0112] Step S43: Send the update parameters of each orientation model to the corresponding edge AI node so that each edge AI node can update the corresponding node optimization model according to the update parameters of the orientation model and obtain the updated node optimization model.

[0113] The digital twin platform transmits targeted model update parameters to corresponding edge AI nodes via network communication links. Upon receiving the targeted model update parameters, each edge AI node replaces or integrates them into the corresponding network layer parameters in the node optimization model (e.g., updating the weights of specific convolutional kernels or the parameters of the classification head). This process retains most of the original general knowledge and local adaptation experience of the node optimization model while injecting enhancement capabilities specifically targeting the weaknesses of the current edge AI node. After this operation, an updated node optimization model with targeted performance improvement is obtained. This enables the secure, rapid, and personalized distribution and deployment of global model knowledge to a massive number of edge AI nodes. Each edge AI node can obtain an enhanced AI model that integrates collective experience, has undergone specialized training in a virtual environment, and is highly matched to its own real-world work scenario almost in real time. This continuously and accurately improves the overall performance and adaptability of the entire distributed quality defect detection system at the system level.

[0114] Furthermore, step S43 may be followed by steps S431 to S432:

[0115] Step S431: The updated node optimization model is used to identify quality defects in real-time construction images on edge AI nodes, and the performance gain of the current model update is evaluated based on the consistency between the identification results and the feedback from manual verification.

[0116] The edge AI node uses the updated node optimization model to perform forward inference on the newly acquired real-time construction images. The updated node optimization model outputs the prediction and identification results of the type, location and confidence level of quality defects in the real-time construction images to complete the quality defect identification. Then, the obtained prediction and identification results are submitted to the quality management personnel for review, and manual verification feedback including the correctness judgment (such as correct, false alarm, omission) and possible corrected labels is obtained.

[0117] When evaluating performance gains, the predicted recognition results are compared with their corresponding manual verification feedback. The performance metrics (such as accuracy, recall, or F1 score) of the updated node optimization model on the current data are calculated. At the same time, the performance metrics of the edge AI node before this update based on historical data of the same type are retrieved and calculated. By comparing the improvement of key performance metrics (such as F1 score) before and after the update, the performance gain of this model update is quantitatively evaluated. For example, if the F1 score increases from 0.85 to 0.92 after the update, the performance gain can be quantified as 0.07.

[0118] Step S432: Update the quality confidence coefficient of each updated node optimization model based on each performance gain.

[0119] Retrieve the historical quality confidence coefficients currently stored in the optimization model of this node using a predefined formula: ,in It is a decay factor between 0 and 1, used to balance the weight of historical performance and recent update effects, thereby ensuring the stability and adaptability of the quality confidence coefficient. Through this calculation, the quality confidence coefficient will be dynamically adjusted up or down according to the actual effect of this update, realizing real-time correction of the reliability assessment of the updated node optimization model. This makes the node trust assessment system of the entire system a system that can dynamically evolve with the model performance. It not only provides the latest basis for more reasonable weighted fusion in the next round of global model aggregation, but also incentivizes the continuous optimization of the model update strategy.

[0120] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A photovoltaic power plant construction quality defect identification system based on digital twins, characterized in that, include: The model training module is used to obtain virtual construction scenarios under each quality defect type and the labeled data corresponding to each virtual construction scenario through the digital twin platform, and to pre-train the initial recognition model to obtain the pre-trained model. The model adjustment module is used to load the pre-trained model onto each edge AI node of the collaborative AI node network after deploying the collaborative AI node network on the photovoltaic power station, and adjust the corresponding pre-trained model based on the construction images collected on the edge AI nodes to obtain the node optimization model. The model update module is used to upload the model parameters of the optimized models of each node to the digital twin platform, so that the model parameters can be aggregated through the digital twin platform to obtain a global model; The model iteration module is used to update each edge AI node based on the global model, and generate corresponding adversarial operations in the virtual construction scenario corresponding to the digital twin platform based on the new quality defect types identified during the operation of the photovoltaic power station, so as to iteratively update the pre-trained model based on the adversarial operations. The step of adjusting the corresponding pre-trained model based on the construction images collected on the edge AI nodes to obtain the node optimization model includes: After acquiring construction images of the construction area to which the edge AI node belongs, enhanced representation learning is performed on each construction image to obtain image features; Based on the image features and the virtual scene features of each virtual construction scene stored on the digital twin platform, the pre-trained model is adjusted for feature alignment to obtain an adaptive model; Based on the labeled defect data stored on each edge AI node, the adaptive model is sample-adapted to obtain an initial optimized model; Based on the initial optimization model, the construction image is simulated and inferred to obtain inferred quality defect data. Then, the initial optimization model is iteratively updated using the inferred quality defect data and the construction image to obtain the node optimization model. The step of aggregating the model parameters through the digital twin platform to obtain the global model includes: Based on the historical identification records and quality confirmation feedback on each edge AI node, the performance of the corresponding node optimization model is evaluated, and the quality confidence coefficient of each node optimization model is generated. Based on the quality confidence coefficients, the model parameters on the corresponding node optimization model are weighted and aggregated to obtain the initial model. The virtual construction scenarios stored on the digital twin platform are substituted into the initial model for reasoning. After obtaining the reasoning results and reasoning logic, the model is trained based on the reasoning results and reasoning logic to obtain the global model. The step of updating each edge AI node based on the global model includes: By analyzing the twin construction environment data and historical quality defect distribution characteristics corresponding to each edge AI node through the digital twin platform, the adaptation requirements of each edge AI node to the global model are evaluated, and a model update strategy is obtained. Based on the model update strategy, inference model parameters adapted to each edge AI node are extracted from the global model, and virtual scene adjustment data of each virtual construction scene are combined to generate targeted model update parameters. The directional model update parameters are sent to the corresponding edge AI nodes so that each edge AI node can update its corresponding node optimization model according to the directional model update parameters, thereby obtaining the updated node optimization model.

2. The AI-based photovoltaic power plant construction quality defect identification system according to claim 1, characterized in that, The steps of obtaining virtual construction scenarios for each quality defect type and the corresponding labeled data for each virtual construction scenario through a digital twin platform include: By using the parameterized component library and construction process rules integrated in the digital twin platform, an initial virtual construction scene is generated, and various quality defect types defined by historical quality defect data and the defect generation parameters corresponding to each quality defect type are determined. The defect generation parameters are substituted into the initial virtual construction scene for simulation to obtain the virtual construction scene corresponding to each of the quality defect types, and the annotation data on each virtual construction scene is obtained.

3. The AI-based photovoltaic power station construction quality defect identification system according to claim 2, characterized in that, The step of pre-training the initial recognition model to obtain the pre-trained model includes: The initial recognition model is pre-trained in the first stage based on each virtual construction scenario and the corresponding labeled data to obtain a basic model. The quality defects in each virtual construction scenario are then identified based on the basic model to obtain the recognition results. Based on the identification results, corresponding adversarial quality defect samples and defect labels are generated respectively. The adversarial quality defect samples and defect labels are then integrated into the corresponding virtual construction scene to obtain the training dataset. The base model is pre-trained in the second stage based on the training dataset to obtain the pre-trained model.

4. The AI-based photovoltaic power plant construction quality defect identification system according to claim 3, characterized in that, The steps for deploying a collaborative AI node network on a photovoltaic power plant include: Based on the real-time construction accuracy and quality risk prediction synchronized by the digital twin platform, the deployment priority and node function configuration strategy for each construction area are determined. Based on the deployment priority and the node function configuration strategy, corresponding edge AI nodes are deployed in each of the construction areas, and each edge AI node is initialized based on the initial configuration parameters set on the digital twin platform to form an initial AI node network. Based on the real-time operating data and load status reported by each edge AI node, resource scheduling analysis is performed through the digital twin platform to generate resource scheduling strategies; In the initial AI node network, according to the resource scheduling strategy, nearby idle computing resources are allocated to each edge AI node and / or the task priority of the edge AI node is adjusted to form the collaborative AI node network.

5. The AI-based photovoltaic power station construction quality defect identification system according to claim 4, characterized in that, The step of obtaining the updated node optimization model includes: The updated node optimization model is used to identify quality defects in real-time construction images on the edge AI nodes, and the performance gain of the current model update is evaluated based on the consistency between the identification results and the feedback from manual verification. Based on the performance gains described, the quality confidence coefficients of the updated node optimization models are updated.

Images