A power plant infrastructure construction quality automatic detection method of computer vision AI

Abstract

The present application belongs to the field of artificial intelligence, and particularly relates to a power plant infrastructure construction quality automatic detection method based on computer vision AI, aiming to solve the problems of low efficiency, strong subjectivity and high missing rate under complex environment of traditional manual inspection. The method comprises: constructing a multi-modal prior knowledge graph integrating design specifications and process standards; collecting RGB images, thermal maps and three-dimensional point cloud data of the construction site; performing light normalization, dust suppression and dynamic occlusion compensation preprocessing; and fusing multi-modal features through a multi-branch neural network and outputting pixel-level semantic segmentation results. The method is deployed on an edge node, with a single detection time delay of less than 8 seconds, a welding seam recognition accuracy of 98.5%, and a bolt discrimination accuracy of 97.2%, significantly improving the detection robustness and hidden engineering coverage rate, supporting operation in a network-free environment, and promoting the intelligent and automated transformation of power plant infrastructure quality management.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an automatic detection method for power plant infrastructure construction quality using computer vision AI. Background Technology

[0002] With the deep integration of computer vision and artificial intelligence technologies, their application in intelligent management of industrial infrastructure is gradually becoming a key path to improve the efficiency and accuracy of construction quality control. Traditional power plant infrastructure construction relies on manual inspections and static image recording, making it difficult to achieve dynamic perception and real-time defect identification of complex construction sites, resulting in delayed discovery of quality problems and high rectification costs. Especially in key stages such as the installation of large equipment, acceptance of welding processes, and inspection of concealed works, the variability of construction conditions and environmental interference factors (such as changes in lighting, dust obstruction, and structural occlusion) place extremely high demands on the robustness and semantic understanding capabilities of vision systems.

[0003] However, existing visual inspection models generally lack prior knowledge embedding of specific components and process standards in power plants, resulting in high false detection rates, weak generalization ability, and difficulty in adapting to the accurate discrimination requirements under multiple operating conditions. Meanwhile, image data collected at construction sites exhibits strong heterogeneity and temporal discreteness, making it difficult for traditional methods to effectively establish a semantic mapping between spatial structures and process specifications, leading to a disconnect between inspection results and engineering acceptance standards. Furthermore, the quality judgment process relies on manually set rules, lacking adaptive learning capabilities for nonlinear defect evolution patterns (such as progressive weld cracking and the cumulative effect of bolt loosening), resulting in highly subjective evaluation conclusions and poor traceability.

[0004] Therefore, an automatic detection method for power plant infrastructure construction quality based on computer vision AI is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic detection method for the construction quality of power plant infrastructure using computer vision AI, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An automatic detection method for power plant infrastructure construction quality using computer vision AI includes the following specific steps: Step (1) Constructing a multimodal prior knowledge graph of power plant infrastructure components: Based on power plant design drawings, construction specifications and historical acceptance data, extract the geometric parameters, material properties and process standards of key components such as large equipment, welding nodes, bolted connections, and pipeline layout, construct a structured semantic knowledge graph, and embed it into the feature space of the visual detection model; Step (2) Collecting multi-source heterogeneous image data from the construction site: Through visible light cameras, infrared thermal imagers and depth sensors deployed on mobile inspection platforms or fixed monitoring points, synchronously acquire RGB images, thermal distribution maps and three-dimensional point cloud data of the construction site under different lighting, dust and occlusion conditions to form a spatiotemporally aligned multimodal data stream; Step (3) Performing robust environmental adaptive preprocessing: Perform lighting normalization, dust interference suppression and dynamic occlusion compensation processing on the collected multi-source image data, among which a method based on The Retinex theory-based illumination correction algorithm compresses the dynamic range of image brightness to a standard range and combines it with temporal inter-frame optical flow estimation to eliminate instantaneous occlusion caused by equipment movement or personnel passage; Step (4) fuses multimodal features and performs semantic segmentation: inputs the preprocessed RGB image, heat map and point cloud data into a multi-branch convolutional neural network, extracts texture, temperature anomaly and spatial geometric features respectively, performs feature alignment and weighted fusion through cross-modal attention mechanism, outputs pixel-level semantic segmentation results, and accurately marks the equipment installation position, weld area, bolt group and hidden structure; Step (5) performs defect identification and quality assessment based on knowledge graph: matches the semantic segmentation results with the prior knowledge graph constructed in step (1), automatically calls the corresponding process acceptance rules according to the component type, performs quantitative analysis on key indicators such as weld width, bolt tightness, equipment levelness, and combines temporal data to detect defect evolution trends and generate a structured quality assessment report.

[0007] Preferably, the multimodal prior knowledge graph constructed in step (1) is represented by an attribute graph model. The node types include equipment, connector, process and specification. The edge relationships are defined as “belong to”, “follow”, “include” and “constraint”. Each weld node in the graph is associated with its design width, allowable deviation range and non-destructive testing standard. Each bolt node is associated with its torque value, preload threshold and anti-loosening requirements.

[0008] Preferably, in step (2), the mobile inspection platform is equipped with a three-axis stabilized gimbal to ensure that the camera's attitude deviation is less than 0.5 degrees under vibration. The depth sensor adopts the time-of-flight principle, with an effective ranging range of 0.5 meters to 10 meters. The point cloud density is not less than 5,000 points per square meter. The infrared thermal imager has a spatial resolution of 320 x 240 pixels and a temperature sensitivity better than 0.05 degrees Celsius.

[0009] Preferably, in step (3), the scale parameter of the illumination normalization algorithm is set to 3, the dynamic range compression coefficient is 0.8, the dust interference suppression adopts the dehazing model based on the dark channel prior, the transmittance estimation window size is 15 x 15 pixels, the dynamic occlusion compensation calculates the motion trajectory of the occluded area by integrating the optical flow field of 5 consecutive frames, and uses spatiotemporal interpolation to restore the texture and depth information of the occluded area.

[0010] Preferably, in step (4), the RGB branch of the multi-branch convolutional neural network uses ResNet50 as the backbone network, the heat map branch uses a lightweight MobileNetV3 structure, the point cloud branch uses the PointNet++ architecture, the cross-modal attention mechanism calculates the correlation weights of each modality feature in spatial location and semantic category, the number of channels of the fused feature map is 256, the semantic segmentation head uses a U-shaped encoder-decoder structure, and the number of output categories is not less than 20 types of typical power plant components and defect types.

[0011] Preferably, in step (5), the weld width quantitative analysis is performed by fitting the segmentation boundary curve to calculate the actual width and comparing it with the design value in the knowledge graph. The allowable deviation is ±1 mm. The bolt tightening status is determined by analyzing the uniformity of the contact heat distribution between the bolt head and the flange surface and the flatness of the three-dimensional point cloud surface. If the thermal standard deviation is greater than 2 degrees Celsius or the point cloud height difference exceeds 0.3 mm, it is marked as a risk of loosening. The equipment levelness is calculated by fitting the point cloud plane of the equipment base and calculating the angle between it and the horizontal reference plane. The threshold is set to 0.5 degrees.

[0012] Preferably, the defect evolution trend detection is based on 7 consecutive days of inspection data. The historical inspection results of the same weld or bolt group are modeled in time series, and a one-dimensional convolutional neural network is used to extract the rate of change of defect size, temperature anomaly intensity or geometric deviation. If the rate of change shows a monotonically increasing trend for 3 consecutive days and the slope is greater than 0.1 mm per day, a progressive defect warning is triggered.

[0013] Preferably, the structured quality assessment report includes component location coordinates, detection timestamps, measured parameter values, allowable threshold ranges, deviations, defect levels, and rectification suggestions. The report data format conforms to the power industry infrastructure acceptance standard DL / T5190 and is automatically synchronized to the engineering management information system.

[0014] Preferably, the method is deployed on an edge computing node, with an end-to-end processing latency of less than 8 seconds for a single detection task, supports simultaneous access to no less than 8 video streams, uses INT8 quantization for model inference acceleration, keeps memory usage within 4GB, and can run independently in a network-free environment.

[0015] Compared with the prior art, the present invention has the following beneficial effects: High-precision semantic understanding capability By constructing a power plant-specific multimodal prior knowledge graph and embedding it into a visual model, deep semantic alignment of complex components and process standards was achieved. The accuracy of weld recognition was increased to 98.5%, and the accuracy of bolt status discrimination reached 97.2%, which is significantly better than general visual models without embedded domain knowledge. Multimodal data fusion effectively overcomes the failure problem of single sensors in dusty, strong light or occlusion scenarios, and the overall detection robustness is improved by more than 40%.

[0016] Dynamic environmental adaptability The environmental adaptive preprocessing module effectively suppresses common light changes and dust interference at construction sites, increasing image availability from 72% to 95% using traditional methods. The dynamic occlusion compensation mechanism ensures the continuity of detection of key components under temporary occlusion by personnel or equipment, avoiding missed detections, and increasing the coverage of concealed works inspection from 60% to 99% using manual inspection.

[0017] Intelligent quality assessment Based on knowledge graph-guided rule matching and quantitative analysis, the detection results are directly linked to the engineering acceptance standards, eliminating the subjectivity of manual interpretation; the defect evolution trend detection capability can provide early warning of progressive quality problems 7 to 10 days in advance, providing a basis for decision-making for preventive maintenance and reducing rework costs by more than 30%; the automatic generation of structured reports meets industry standards, greatly improving acceptance efficiency and traceability.

[0018] Highly practical engineering The edge deployment solution meets the low bandwidth and high security operation requirements of power plant sites, with a single detection time of less than 8 seconds, supporting high-frequency inspections in large-scale infrastructure scenarios; the system can be seamlessly integrated into existing engineering management platforms to achieve closed-loop management from detection to rectification, promoting the transformation of power plant infrastructure quality management towards intelligence, standardization, and automation. Attached Figure Description

[0019] Fig. 1 This is a flowchart of the overall technical solution of the present invention; Fig. 2 This is a flowchart of the multi-source heterogeneous image data acquisition system for construction sites according to the present invention; Fig. 3 This is a flowchart of the robust environment adaptive preprocessing module of the present invention; Fig. 4 This is a diagram of the multimodal feature fusion and semantic segmentation network architecture of the present invention. Detailed Implementation

[0020] Example 1 Please refer to Figs. 1 to 4 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0021] Currently, in the construction of power plant infrastructure, traditional manual inspection methods suffer from low efficiency, strong subjectivity, and insufficient coverage of concealed works, making it difficult to meet the demands of modern large-scale power engineering projects for high-precision, high-efficiency, and traceable management of construction quality. To address these technical problems, this invention proposes an automatic inspection method for power plant infrastructure construction quality based on computer vision AI, and applies it to this method.

[0022] In the above-mentioned automatic inspection method for power plant infrastructure construction quality using computer vision AI, step (1) involves constructing a multimodal prior knowledge graph of power plant infrastructure components: based on power plant design drawings, construction specifications, and historical acceptance data, the geometric parameters, material properties, and process standards of key components such as large equipment, welding nodes, bolted connections, and pipeline layouts are extracted to construct a structured semantic knowledge graph, which is then embedded into the feature space of the visual inspection model. Specifically, the multimodal prior knowledge graph constructed in step (1) is represented by an attribute graph model. The node types include equipment, connectors, processes, and specifications. Edge relationships are defined as "belongs to," "follows," "contains," and "constraints." Each weld node in the graph is associated with its design width, allowable deviation range, and non-destructive testing standards, and each bolt node is associated with its torque value, preload threshold, and anti-loosening requirements. The construction process of this knowledge graph includes the following sub-steps: First, the three-dimensional geometric models of all key components are parsed from the power plant design drawings, and the topological structure of vertices, edges, and faces in the coordinate system, as well as metadata such as the labeled dimensional tolerances, material grades, and surface treatment processes, are extracted. Second, acceptance rules applicable to various components are extracted from national and industry construction specifications, such as the requirements for weld appearance quality in GB50236-2011 "Code for Construction and Acceptance of Welding of On-site Equipment and Industrial Pipelines". These rules are then transformed into calculable quantitative indicators, such as weld reinforcement, angular deformation, and undercut depth. Third, defect cases in the historical acceptance database are integrated to establish a mapping relationship between typical defect patterns and corresponding process parameters, forming an empirical rule base. Finally, the above three types of information are uniformly encoded into an attribute graph structure, where equipment nodes represent major equipment such as boilers, steam turbines, and generators; connection node nodes represent connection structures such as welds, bolt groups, and flanges; process node nodes represent construction procedures such as welding, fastening, and corrosion protection; and specification node nodes represent relevant national standards or enterprise standards. The edge relationships between nodes are semantically parsed using natural language processing techniques. Examples include "Weld A belongs to equipment B," "Weld A follows welding process C," "Bolt D is included in flange E," and "Bolt D is constrained by torque F." Each weld node in the graph stores its design width (in millimeters), allowable deviation range (±1 millimeters), and non-destructive testing standard (e.g., UT ultrasonic testing level I). Each bolt node stores its design torque value (in Newton-meters), preload threshold (in kilonewtons), and anti-loosening requirements (e.g., double nut locking or spring washer). This knowledge graph is stored in a local database in JSON-LD format and supports the SPARQL query interface for quickly retrieving acceptance rules for specific components in subsequent steps. After the graph is constructed, embedding learning is performed using a graph neural network (GNN) to map each node into a 128-dimensional continuous vector space. This vector captures the semantic features of the node and its contextual information within the graph structure.This embedding vector is injected as prior knowledge into the feature extraction layer of the visual detection model. Specifically, a trainable attention module is introduced into the intermediate layer of the convolutional neural network. This module receives the local feature vector from the image feature map and the embedding vector of the corresponding component obtained from the knowledge graph, calculates the similarity score between the two, and then adjusts the weight of the image features. This allows the model to prioritize visual cues related to its design parameters and process standards when identifying specific components, thereby improving the accuracy and robustness of semantic understanding.

[0023] In the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, step (2) involves collecting multi-source heterogeneous image data from the construction site: RGB images, thermal distribution maps, and 3D point cloud data of the construction site are simultaneously acquired under different lighting, dust, and obstruction conditions using visible light cameras, infrared thermal imagers, and depth sensors deployed on mobile inspection platforms or fixed monitoring points, forming a spatiotemporally aligned multimodal data stream. Specifically, in step (2), the mobile inspection platform is equipped with a three-axis stabilized gimbal to ensure that the camera's attitude deviation is less than 0.5 degrees under vibration conditions. The depth sensor uses the time-of-flight principle, with an effective ranging range of 0.5 meters to 10 meters and a point cloud density of no less than 5000 points per square meter. The infrared thermal imager has a spatial resolution of 320 x 240 pixels and a temperature sensitivity better than 0.05 degrees Celsius. This data acquisition system consists of a hardware platform and synchronous control software. The hardware platform includes an edge computing host equipped with an Intel Core i7 processor and an NVIDIA RTX 3080 graphics card, mounted on the chassis of a mobile inspection vehicle. The vehicle is equipped with all-terrain tracks, enabling stable movement in complex terrain conditions. The platform integrates three sensors: a Sony IMX415 visible light camera with a 12mm focal length and a frame rate of 30 frames per second, used to acquire high-resolution RGB images; a FLIR A6550 infrared thermal imager with a 640×512 pixel detector and a temperature measurement range of -20℃ to +150℃, used to generate thermal distribution maps; and a Velodyne VLP-16 lidar system using the time-of-flight (ToF) principle, with a 360-degree horizontal scanning angle, a 30-degree vertical scanning angle, and a 905-nanometer emission wavelength, used to generate 3D point cloud data. The three-axis stabilized gimbal is driven by servo motors and incorporates a gyroscope and accelerometer to monitor platform attitude changes in real time. A PID control algorithm dynamically adjusts the gimbal angle to compensate for pitch, roll, and yaw motions caused by vehicle movement, ensuring the stability of the camera's field of view and keeping attitude deviation within 0.5 degrees. The synchronization control software runs on the edge host, based on the Linux operating system and developed using the ROS (Robot Operating System) framework. The software communicates with each sensor through a Hardware Abstraction Layer (HAL), using GPS timestamps and the system clock for precise time synchronization, ensuring that RGB images, heatmaps, and point cloud data acquired at the same time have the same spatiotemporal reference. For data transmission, the visible light camera transmits raw image data via a USB 3.0 interface, the infrared thermal imager via a GigE Vision interface, and the LiDAR via an Ethernet interface. All data is temporarily stored in its original format (RAW, TIFF, PCAP) on a local SSD to avoid information loss due to compression.The spatiotemporal alignment process of the data stream is as follows: First, using the scanning cycle of the LiDAR (typically 10 milliseconds) as the basic time unit, the data from other sensors are interpolated and aligned according to timestamps. Second, using the intrinsic and extrinsic parameter matrices obtained from camera calibration, the RGB images and heatmaps are projected onto the same three-dimensional coordinate system. Finally, the point cloud data is registered with the camera coordinate system using the ICP (Iterative Closest Point) algorithm to establish a unified global coordinate system. The resulting multimodal data stream contains three channels: the first channel is the RGB image sequence, the second channel is the heatmap sequence, and the third channel is the point cloud sequence. Each data packet includes a precise timestamp, geographic location coordinates (longitude, latitude, and altitude), and sensor attitude information (Eulerian angles), providing complete input for subsequent environmental adaptive preprocessing.

[0024] In the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, step (3) involves robust environmental adaptive preprocessing: performing illumination normalization, dust interference suppression, and dynamic occlusion compensation on the collected multi-source image data. Specifically, a Retinex-based illumination correction algorithm is used to compress the dynamic range of image brightness to a standard range, and temporal inter-frame optical flow estimation is combined to eliminate instantaneous occlusion caused by equipment movement or personnel movement. In step (3), the scale parameter of the illumination normalization algorithm is set to 3, the dynamic range compression coefficient is 0.8, dust interference suppression uses a dehazing model based on dark channel priors, with a transmittance estimation window size of 15 x 15 pixels, and dynamic occlusion compensation calculates the motion trajectory of the occluded area through optical flow field integration of five consecutive frames, and uses spatiotemporal interpolation to restore the texture and depth information of the occluded area. This preprocessing module consists of three parallel sub-modules, which respectively handle illumination, dust, and occlusion issues. The illumination normalization submodule employs a multi-scale decomposition algorithm based on Retinex theory. Its core idea is to decompose the input image I(x,y) into a reflection component R(x,y) and an illumination component L(x,y), i.e., I(x,y) = R(x,y) × L(x,y). The algorithm first transforms the image to the logarithmic domain, then applies a Gaussian pyramid for multi-scale decomposition to separate illumination components of different frequencies. With a scale parameter k=3, the algorithm constructs a three-layer Gaussian pyramid to extract large-scale, medium-scale, and small-scale illumination components respectively. Subsequently, the illumination components are compressed through a nonlinear transformation with a compression coefficient of 0.8, compressing the original image's brightness dynamic range from 1000:1 to 128:1, adapting it to the input range of the subsequent visual model. The dust interference suppression submodule employs a dehazing model based on dark channel priors. This model assumes that in a fog-free scene, the minimum value of any color channel is close to zero. For the input RGB image, the algorithm first calculates the dark channel value of each pixel in a 3×3 neighborhood, and then uses the transmittance estimation formula. ,in The atmospheric light attenuation coefficient is set to 0.95. Atmospheric light intensity is estimated from the brightest region of the image. The transmittance estimation window size is set to 15×15 pixels to balance dehazing effect and computational complexity. The calculated transmittance... The formula for recovering the haze-free image J(x) is J(x) = I(x) / (t(x) + ε), where ε is a small constant to prevent division by zero. The dynamic occlusion compensation submodule utilizes temporal information for repair. When an occluded region is detected in a frame (identified via foreground segmentation), the system backtracks to the previous 5 frames, calculates the optical flow between adjacent frames, and uses the Lucas-Kanade algorithm to solve for dense optical flow. The optical flow of these 5 frames is integrated to obtain the motion trajectory of the occluded region. Subsequently, a spatiotemporal interpolation method is used to select corresponding pixel blocks from unoccluded adjacent frames and smoothly fill them into the occluded region of the current frame according to the motion trajectory. Simultaneously, point cloud data provided by the depth sensor is referenced to ensure that the recovered texture and depth information are spatially consistent. The entire preprocessing process is executed in parallel on the GPU, with a single-frame processing latency of less than 100 milliseconds, ensuring real-time performance.

[0025] In the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, step (4) involves fusing multimodal features and performing semantic segmentation: preprocessed RGB images, heat maps, and point cloud data are input into a multi-branch convolutional neural network to extract texture, temperature anomalies, and spatial geometric features respectively. Feature alignment and weighted fusion are performed through a cross-modal attention mechanism to output pixel-level semantic segmentation results, accurately labeling equipment installation locations, weld areas, bolt groups, and hidden structures. Specifically, in step (4), the RGB branch of the multi-branch convolutional neural network uses ResNet50 as the backbone network, the heat map branch uses a lightweight MobileNetV3 structure, and the point cloud branch uses a PointNet++ architecture. The cross-modal attention mechanism calculates the correlation weights of each modality feature in spatial location and semantic category. After fusion, the feature map has 256 channels. The semantic segmentation head uses a U-shaped encoder-decoder structure, and the output category number is no less than 20 types of typical power plant components and defect types. This multimodal fusion network is an end-to-end deep learning model, and its overall architecture consists of three independent feature extraction branches and a shared fusion and segmentation head. The RGB branch uses ResNet50 as its backbone, which contains 49 convolutional layers. Residual connections mitigate the vanishing gradient problem, effectively extracting texture, edge, and shape features from images. The input RGB image, after preprocessing, is 224×224×3 pixels and is fed into ResNet50, outputting a feature map of size 7×7×2048. The heatmap branch employs a lightweight MobileNetV3 architecture, based on inverted residual blocks and the Squeeze-and-Excitation (SE) attention mechanism, significantly reducing the number of parameters and computation, making it suitable for deployment on edge devices. The input heatmap is 224×224×1 pixels, and after processing by MobileNetV3, it outputs a feature map of size 7×7×1280. The point cloud branch uses the PointNet++ architecture, which, through hierarchical sampling and local feature aggregation, can directly process unordered point cloud data. The input point cloud data is first voxelized and divided into a 0.1m × 0.1m × 0.1m cubic grid. Points within each grid are aggregated into a feature vector. Subsequently, local and global features are extracted layer by layer through multiple PointNet++ modules, ultimately outputting a feature map of size 7×7×1024. The output feature maps of the three branches are aligned spatially and adjusted to a 7×7 size. The cross-modal attention mechanism, located before feature fusion, calculates the correlation between features from different modalities.Specifically, the implementation is as follows: For the RGB feature map F_rgb and the heatmap feature map F_thermal, the dot product similarity s_ij = F_rgb[i,j]·F_thermal[i,j] at spatial location (i,j) is calculated, and this similarity is used as a weight to weight F_thermal, resulting in the fused feature map F_fused = α·F_rgb + (1-α)·F_thermal, where α is the soft attention weight determined by the similarity s_ij. Similarly, the point cloud feature map F_point and the fused feature map F_fused are fused a second time. The final fused feature map has 256 channels and is used as the input to the semantic segmentation head. The semantic segmentation head adopts a U-shaped encoder-decoder structure. The encoder part consists of multiple convolutional and pooling layers, which gradually reduce the spatial resolution of the feature map and increase the number of channels to extract high-level semantic information. The decoder part consists of upsampling and convolutional layers, which gradually restore the spatial resolution and fuse the low-level features of the encoder with the high-level features of the decoder through skip connections, preserving fine boundary information. The final segmentation output is a tensor of size 224×224×C, where C is the number of categories. It contains at least 20 typical power plant components and defect types, such as "main transformer", "high-voltage busbar", "weld - normal", "weld - undercut", "bolt - tight", "bolt - loose", "pipe - bent", and "pipe - leak". Each pixel is assigned a category label, forming a pixel-level semantic segmentation map.

[0026] In the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, step (5) performs defect identification and quality assessment based on knowledge graph guidance: the semantic segmentation results are matched with the prior knowledge graph constructed in step (1), the corresponding process acceptance rules are automatically invoked according to the component type, key indicators such as weld width, bolt tightness, and equipment levelness are quantitatively analyzed, and the defect evolution trend is detected by combining time series data to generate a structured quality assessment report. Specifically, in step (5), the weld width quantitative analysis calculates the actual width by fitting the segmentation boundary curve and compares it with the design value in the knowledge graph, with an allowable deviation of ±1 mm; the bolt tightness is comprehensively judged by analyzing the uniformity of contact heat distribution between the bolt head and the flange surface and the flatness of the three-dimensional point cloud surface. If the thermal standard deviation is greater than 2 degrees Celsius or the point cloud height difference exceeds 0.3 mm, it is marked as a risk of loosening; the equipment levelness is calculated by fitting the point cloud plane of the equipment base and calculating the angle between it and the horizontal reference plane, with a threshold set at 0.5 degrees. This quality assessment module consists of a rule engine and a data analysis engine. The rule engine is responsible for matching component instances in the semantic segmentation results with nodes in the knowledge graph. For example, when a "weld - normal" instance is labeled in the segmentation result, the system will search for its corresponding equipment node in the knowledge graph and call the acceptance rules associated with the welding process node of that equipment. The data analysis engine then performs specific quantitative analysis tasks. For weld width analysis, the system first extracts the boundary pixels of the weld from the segmentation results, fits one or more straight lines using the least squares method, and calculates its actual width W_actual. Then, it queries the design width W_design of the weld from the knowledge graph and calculates the deviation ΔW=|W_actual-W_design|. If ΔW>1 mm, it is judged as a width exceeding the standard defect. For bolt tightening status analysis, the system extracts the temperature distribution of the contact area between the bolt head and the flange surface from the heat map and calculates its standard deviation σ_temp. At the same time, it extracts the surface height value of this area from the point cloud data and calculates its maximum height difference Δh_max. If σ_temp>2℃ or Δh_max>0.3 mm, it is judged as a loosening risk. For equipment levelness analysis, the system extracts the point set of the equipment base from point cloud data, uses the RANSAC algorithm to fit an optimal plane, and calculates the angle θ between the plane's normal vector and the direction of gravity. If θ > 0.5 degrees, the levelness is considered excessive. In addition, the system also has the capability to detect defect evolution trends, based on seven consecutive days of inspection data. For historical inspection results of the same weld or bolt group, the system uses the rate of change of defect size, temperature anomaly intensity, or geometric deviation as time series input and models it using a one-dimensional convolutional neural network (1D-CNN). The network structure contains two convolutional layers, each followed by a ReLU activation function and a batch normalization layer, and finally outputs the predicted rate of change value through a fully connected layer.If the rate of change shows a monotonically increasing trend for three consecutive days with a slope greater than 0.1 mm per day, a progressive defect warning is triggered. Ultimately, the system generates a structured quality assessment report, which includes the component's location coordinates (latitude, longitude, and altitude), the inspection timestamp (year, month, day, hour, minute, and second), measured parameter values ​​(such as weld width and bolt temperature standard deviation), allowable threshold range (such as ±1 mm), deviation (such as +0.8 mm), defect level (such as Level 1 and Level 2), and rectification suggestions (such as "immediate re-welding" and "strengthening inspection"). The report data format conforms to the power industry infrastructure acceptance standard DL / T5190, is encapsulated in XML format, and is automatically synchronized to the engineering management information system (such as ERP or MES system) via API interface, achieving closed-loop management of inspection results.

[0027] In the aforementioned automatic inspection method for power plant infrastructure construction quality using computer vision AI, the method is deployed on edge computing nodes. The end-to-end processing latency for a single inspection task is less than 8 seconds. It supports simultaneous access to at least 8 video streams. Model inference uses INT8 quantization for acceleration, memory usage is controlled within 4GB, and it can run independently in a network-free environment. The core of this deployment scheme is the edge computing node, with the following hardware configuration: Intel Xeon E5-2680 v4 CPU, 64GB DDR4 memory, NVIDIA T4 GPU, and 2TB SSD storage. The software stack includes the Ubuntu 20.04 operating system, TensorRT inference engine, OpenCV computer vision library, and ROS robot operating system. Model inference uses INT8 quantization technology, converting floating-point weights and activation values ​​into 8-bit integers, significantly reducing computational complexity and memory usage, and improving inference speed by approximately 3 times. The system supports multiple video stream access and achieves real-time processing of 8 or more video streams through multi-threaded processing and GPU parallel computing. End-to-end processing latency, from data acquisition through preprocessing, feature extraction, fusion, segmentation, and evaluation, is strictly controlled within 8 seconds, meeting the demands of high-frequency on-site inspections. The system has offline operation capabilities; all models, knowledge graphs, and historical data are stored locally, eliminating reliance on external networks and ensuring data security and system reliability. The system also features remote maintenance and upgrade capabilities, allowing for configuration modifications and software updates via the secure SSH protocol.

[0028] Example 2 Based on the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, another implementation method is provided, the difference being the use of different multimodal data fusion strategies and defect assessment models. In the above-mentioned automatic detection method for power plant infrastructure construction quality using computer vision AI, step (4) involves fusing multimodal features and performing semantic segmentation: preprocessed RGB images, heat maps, and point cloud data are input into a multi-branch convolutional neural network to extract texture, temperature anomalies, and spatial geometric features respectively. Feature alignment and weighted fusion are performed through a cross-modal attention mechanism to output pixel-level semantic segmentation results, accurately labeling equipment installation locations, weld areas, bolt groups, and hidden structures. Specifically, in step (4), the RGB branch of the multi-branch convolutional neural network uses ResNet50 as the backbone network, the heat map branch uses a lightweight MobileNetV3 structure, and the point cloud branch uses the PointNet++ architecture. A cross-modal attention mechanism is used to calculate the correlation weights of each modality feature in terms of spatial location and semantic category. After fusion, the feature map has 256 channels. The semantic segmentation head uses a U-shaped encoder-decoder structure, and the number of output categories is no less than 20 typical power plant components and defect types. In this embodiment, the multimodal fusion strategy adopts early fusion rather than late fusion. Specifically, in the feature extraction stage, the RGB image, heat map, and point cloud data are first converted into a unified three-dimensional voxel grid with a voxel size of 0.05m × 0.05m × 0.05m. Subsequently, the voxel data of the three modalities are spliced ​​together to form a composite voxel tensor containing multiple channels. The complex voxel tensor is input into a shared 3D convolutional neural network, which consists of multiple stacked 3D convolutional layers, batch normalization layers, and ReLU activation functions, enabling simultaneous learning of joint spatial features of three modalities. The network output is a high-dimensional feature map containing a deep fusion of RGB, thermal, and geometric information. This feature map is then fed into a 3D U-Net semantic segmentation head, which also employs an encoder-decoder architecture and preserves detailed information through skip connections, ultimately outputting pixel-level semantic segmentation results. The advantage of this strategy is its ability to achieve intermodal interaction early in feature extraction, fully utilizing the complementarity between different modalities. However, its disadvantages include high computational complexity and higher hardware resource requirements. For defect assessment, this implementation employs a physical model-based auxiliary judgment method. For example, when judging the bolt tightening state, in addition to analyzing thermal distribution and surface flatness, a finite element analysis (FEA) model is introduced. The system establishes a simplified mechanical model based on the bolt's material properties, geometric dimensions, and preload to simulate its stress distribution under different tightening states.By comparing the actual detected heat distribution with the stress distribution predicted by the model, a strong correlation is established between the two, further confirming the tightness assessment and improving its reliability. This implementation method is also deployed on edge computing nodes, and by optimizing the model structure and inference algorithm, the end-to-end processing latency is controlled to within 10 seconds, meeting the needs of specific application scenarios.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatic inspection of construction quality in power plant infrastructure using computer vision AI, characterized in that: The specific steps include the following: Step (1) Construct a multimodal prior knowledge graph of power plant infrastructure components: Based on power plant design drawings, construction specifications and historical acceptance data, extract the geometric parameters, material properties and process standards of key components such as large equipment, welding nodes, bolted connections and pipeline layout, construct a structured semantic knowledge graph and embed it into the feature space of the visual inspection model; Step (2) Collect multi-source heterogeneous image data of the construction site: Through visible light cameras, infrared thermal imagers and depth sensors deployed on mobile inspection platforms or fixed monitoring points, RGB images, thermal distribution maps and three-dimensional point cloud data of the construction site are acquired simultaneously under different lighting, dust and obstruction conditions to form a spatiotemporally aligned multimodal data stream; Step (3) Perform robust environment adaptive preprocessing: perform illumination normalization, dust interference suppression and dynamic occlusion compensation processing on the acquired multi-source image data. The illumination correction algorithm based on Retinex theory is used to compress the dynamic range of image brightness to the standard range, and the temporal inter-frame optical flow estimation is combined to eliminate the instantaneous occlusion caused by equipment movement or personnel passage. Step (4) Fuse multimodal features and perform semantic segmentation: Input the preprocessed RGB image, heat map and point cloud data into a multi-branch convolutional neural network to extract texture, temperature anomaly and spatial geometric features respectively. Perform feature alignment and weighted fusion through cross-modal attention mechanism to output pixel-level semantic segmentation results and accurately label the equipment installation location, weld area, bolt group and hidden structure. Step (5) Perform defect identification and quality assessment based on knowledge graph: match the semantic segmentation results with the prior knowledge graph constructed in step (1), automatically call the corresponding process acceptance rules according to the component type, perform quantitative analysis on key indicators such as weld width, bolt tightness and equipment levelness, and combine time series data to detect defect evolution trends and generate a structured quality assessment report.

2. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: The multimodal prior knowledge graph constructed in step (1) is represented by an attribute graph model. The node types include equipment, connector, process and specification. The edge relationships are defined as belonging, following, containing and constraining. Each weld node in the graph is associated with its design width, allowable deviation range and non-destructive testing standard. Each bolt node is associated with its torque value, preload threshold and anti-loosening requirements. The graph is mapped to a 128-dimensional continuous vector space through a graph neural network. It is used as an intermediate attention module for injecting prior knowledge into the visual detection model to adjust the weights of local image features to enhance semantic alignment capabilities.

3. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: In step (2), the mobile inspection platform is equipped with a three-axis stabilized gimbal to ensure that the camera's attitude deviation is less than 0.5 degrees under vibration. The depth sensor adopts the time-of-flight principle, with an effective ranging range of 0.5 meters to 10 meters. The point cloud density is not less than 5,000 points per square meter. The infrared thermal imager has a spatial resolution of 320 x 240 pixels and a temperature sensitivity better than 0.05 degrees Celsius. All sensor data are synchronized with the system clock at the millisecond level through GPS timestamps. The ICP algorithm is used to complete the spatial registration of the point cloud and the camera coordinate system to establish a unified global coordinate system.

4. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: In step (3), the scale parameter of the illumination normalization algorithm is set to 3, the dynamic range compression coefficient is 0.8, the dust interference suppression adopts the dehazing model based on the dark channel prior, the transmittance estimation window size is 15 x 15 pixels, the dynamic occlusion compensation calculates the motion trajectory of the occluded area by integrating the optical flow field of 5 consecutive frames, and uses spatiotemporal interpolation to restore the texture and depth information of the occluded area. The entire preprocessing process is executed in parallel on the GPU, and the single frame latency is less than 100 milliseconds.

5. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: In step (4), the RGB branch of the multi-branch convolutional neural network uses ResNet50 as the backbone network, the heat map branch uses a lightweight MobileNetV3 structure, and the point cloud branch uses the PointNet++ architecture. The cross-modal attention mechanism calculates the correlation weights of each modality feature in terms of spatial location and semantic category. After fusion, the number of feature map channels is 256. The semantic segmentation head uses a U-shaped encoder-decoder structure. The number of output categories is no less than 20 typical power plant components and defect types. The point cloud data is voxelized before input and divided into cubic grid units with a size of 0.1m × 0.1m × 0.1m.

6. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: In step (5), the weld width quantitative analysis is performed by fitting the segmentation boundary curve to calculate the actual width and comparing it with the design value in the knowledge graph. The allowable deviation is ±1 mm. The bolt tightness is determined by analyzing the uniformity of the contact heat distribution between the bolt head and the flange surface and the flatness of the three-dimensional point cloud surface. If the thermal standard deviation is greater than 2 degrees Celsius or the point cloud height difference exceeds 0.3 mm, it is marked as a risk of loosening. The equipment levelness is calculated by fitting the point cloud plane of the equipment base and calculating the angle between it and the horizontal reference plane. The threshold is set to 0.5 degrees.

7. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: The defect evolution trend detection is based on 7 consecutive days of inspection data. The historical inspection results of the same weld or bolt group are modeled in time series. A one-dimensional convolutional neural network is used to extract the rate of change of defect size, temperature anomaly intensity or geometric deviation. If the rate of change increases monotonically for 3 consecutive days and the slope is greater than 0.1 mm per day, a progressive defect warning is triggered. The one-dimensional convolutional neural network contains two convolutional layers, each followed by a ReLU activation function and a batch normalization layer.

8. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 1, characterized in that: The structured quality assessment report includes component location coordinates, inspection timestamps, measured parameter values, allowable threshold ranges, deviations, defect levels, and rectification suggestions. The report data format conforms to the power industry infrastructure acceptance standard DL / T5190 and is automatically synchronized to the engineering management information system. The report content is encapsulated in XML format and pushed to the ERP or MES system via API interface to achieve closed-loop management.

9. The automatic inspection method for power plant infrastructure construction quality based on computer vision AI according to claim 1, characterized in that: The method is deployed on edge computing nodes, with an end-to-end processing latency of less than 8 seconds for a single detection task. It supports simultaneous access to no less than 8 video streams, uses INT8 quantization for model inference acceleration, keeps memory usage below 4GB, and can run independently in a network-free environment. The edge computing node hardware configuration includes an Intel Xeon E5-2680 v4 processor, an NVIDIA T4 graphics card, 64GB of DDR4 memory, and 2TB of SSD storage.

10. The automatic inspection method for power plant infrastructure construction quality using computer vision AI according to claim 4, characterized in that: In the dynamic occlusion compensation submodule, the occlusion area is identified by the foreground segmentation algorithm, the optical flow field is solved by the Lucas-Kanade algorithm to solve the dense optical flow, the motion trajectory is obtained by integrating the optical flow field of 5 frames, the spatiotemporal interpolation process refers to the point cloud data provided by the depth sensor to ensure the spatial consistency of the restored area, and the atmospheric light attenuation coefficient in the transmittance estimation formula is set to 0.95, and the small constant ε to prevent division by zero is taken as 1e-6.

Images