Engineering concealment process compliance acceptance method based on unmanned aerial vehicle imaging

Abstract

The present application relates to an engineering concealment process compliance acceptance method based on unmanned aerial imaging, which collects multi-angle multi-spectral images through a drone, generates high-quality data using gimbal control and image correction, and realizes pixel-level segmentation of positions such as wire connection and anti-corrosion coating and automatic extraction of key process parameters combined with an AI model. After comparing the measured parameters with the construction standards, compliance determination results are generated, and are structured and packaged with image information, and the acceptance data is stored in a tamper-proof manner using a hash algorithm and a blockchain to ultimately generate a digital archive supporting full life cycle quality tracing. This method realizes automatic detection, intelligent judgment and reliable storage of the results of the concealed process, improving inspection efficiency and data security.

Description

Technical Field

[0001] This invention relates to the field of intelligent acceptance and blockchain data storage technology for engineering construction, and in particular to a method for compliant acceptance of concealed engineering processes based on UAV-borne imaging. Background Technology

[0002] In the current field of construction quality acceptance and data storage, traditional compliance acceptance of concealed works largely relies on manual on-site inspections, measuring instrument checks, and paper or photographic records, resulting in limited standardization and traceability of information collection. In recent years, with the development of drone imaging technology, multispectral / hyperspectral sensors, and AI models, automated drone inspections and intelligent image analysis have gradually become important means of construction quality acceptance, enhancing the digital visualization capabilities for high-altitude, hazardous, and hard-to-reach areas. Some emerging applications combine blockchain technology to store construction acceptance data summaries on the blockchain, improving data tamper-proofing and accountability.

[0003] Current mainstream technical solutions mainly include: using drones equipped with high-resolution visible light or infrared imaging equipment to conduct regular flight inspections of facilities such as power transmission lines, bridges, and utility tunnels; after collecting images, image stitching and manual defect annotation are performed via ground workstations; some systems integrate simple AI detection models or use assisted interpretation to identify anomalies such as structural deformation, corrosion, and foreign objects. For compliance determination, it is mostly based on human experience or quantitative analysis of some sensor parameters. Compliance determination results and evidence images are archived in separate documents or reports. Projects with the necessary resources may use blockchain for evidence storage, but often only the hash of the acceptance results or acceptance report is stored.

[0004] Current technologies still suffer from the following prominent problems: First, images, AI judgment parameters, and spatiotemporal information of hidden engineering process quality are recorded in a scattered manner from multiple sources and in multiple formats, lacking a unified structured encapsulation mechanism. This leads to fragmented information during subsequent data tracing and compliance assessment, making it impossible to efficiently trace the chain of responsibility. Second, current blockchain evidence storage mechanisms mainly target hash digest storage at the file or report level, making it difficult to support chained records of multidimensional data such as multispectral images, AI inference parameters, and spatial positioning metadata, resulting in insufficient evidence integrity. Third, during the acceptance assessment process, AI parameter inference and standard comparison lack a strong correlation with the original collected data, which can easily cause breakpoints during subsequent archiving, making it difficult to guarantee legal validity and technical credibility. Summary of the Invention

[0005] This application provides a compliance acceptance method for concealed engineering processes based on UAV-borne imaging, aiming to solve the problem of difficulty in judging and tracing the quality of concealed engineering processes mentioned in the background art.

[0006] The compliance acceptance method for concealed engineering processes based on UAV-borne imaging provided in this application specifically includes:

[0007] S1: Obtain the preset precise waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system, obtain the gimbal servo control signal based on the preset precise waypoint coordinate sequence and real-time attitude data, and drive the multispectral imager to perform multi-angle hovering shooting operation according to the gimbal servo control signal to obtain multispectral image sequence data packet.

[0008] S2: Input the multispectral image sequence data packet into the neural network model, and use the target region segmentation method to extract the hidden process parts in the image at the pixel level to generate a mask of the target region to be tested.

[0009] S3: Project the target area mask to obtain a set of image blocks of the region of interest, and process the set of image blocks of the region of interest to obtain a set of measured values ​​of key process parameters;

[0010] S4: Call the locally stored construction acceptance standard library, compare the set of measured values ​​of the key process parameters with the standard thresholds in the construction acceptance standard library item by item, and generate a compliance judgment result of qualified or unqualified.

[0011] S5: Based on the compliance judgment result, the set of measured values ​​of key process parameters and the unique code of the UAV equipment, perform a structured encapsulation operation to generate a structured acceptance record with complete traceability information, and complete the compliance acceptance of the hidden process of the project.

[0012] The engineering concealed process compliance acceptance method based on UAV-borne imaging provided in this application has the following beneficial effects:

[0013] This solution integrates UAV precision waypoint control, multispectral imaging, and neural network models to construct an automated, high-precision quality parameter perception system for complex power construction scenarios. Utilizing a pre-set waypoint sequence and real-time attitude data for collaborative guidance, it achieves multi-angle, multi-band precise imaging of key concealed areas such as conductor crimping pipes, hardware fastening points, and anti-corrosion coatings. This significantly improves the coverage completeness and temporal consistency of image acquisition, greatly enhancing the accuracy and robustness of parameter acquisition, and realizing a paradigm shift from "experience-based judgment" to "data-driven" construction quality assessment. Furthermore, the introduction of a localized real-time compliance verification mechanism significantly enhances the credibility and timeliness of acceptance results. After parameter calculation, the system immediately calls upon the embedded storage of the construction acceptance standard library for multi-dimensional threshold comparison, generating a qualified / unqualified judgment result with high temporal resolution. This result, along with the original image, measured values, high-precision spatiotemporal tags, and equipment identification information, is encapsulated into a structured record, ensuring the complete traceability of data elements throughout the entire process. Attached Figure Description

[0014] Figure 1This is the main flowchart of the compliance acceptance method for engineering concealed processes based on UAV-borne imaging.

[0015] Figure 2 This is a sub-flowchart of the compliance acceptance method for engineering concealed processes based on UAV-borne imaging.

[0016] Figure 3 This is another sub-flowchart of the compliance acceptance method for engineering concealed processes based on UAV-borne imaging. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0019] like Figure 1 As shown, this application provides a compliance acceptance method for engineering concealed processes based on UAV-borne imaging, specifically including:

[0020] S1: Obtain the preset precise waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system, obtain the gimbal servo control signal based on the preset precise waypoint coordinate sequence and real-time attitude data, and drive the multispectral imager to perform multi-angle hovering shooting operation according to the gimbal servo control signal to obtain multispectral image sequence data packet.

[0021] S2: Input the multispectral image sequence data packet into the neural network model, and use the target region segmentation method to extract the hidden process parts in the image at the pixel level to generate a mask of the target region to be tested.

[0022] S3: Project the target area mask to obtain a set of image blocks of the region of interest, and process the set of image blocks of the region of interest to obtain a set of measured values ​​of key process parameters;

[0023] S4: Call the locally stored construction acceptance standard library, compare the set of measured values ​​of the key process parameters with the standard thresholds in the construction acceptance standard library item by item, and generate a compliance judgment result of qualified or unqualified.

[0024] S5: Based on the compliance judgment result, the set of measured values ​​of key process parameters and the unique code of the UAV equipment, perform a structured encapsulation operation to generate a structured acceptance record with complete traceability information, and complete the compliance acceptance of the hidden process of the project.

[0025] Step S1: Obtain the preset precise waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system; obtain the gimbal servo control signal based on the preset precise waypoint coordinate sequence and real-time attitude data; and drive the multispectral imager to perform multi-angle hovering shooting operations according to the gimbal servo control signal to obtain a multispectral image sequence data packet. Specifically, this includes:

[0026] S1.1: Obtain the preset precision waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system, and perform spatial path planning processing based on the preset precision waypoint coordinate sequence to generate a multi-dimensional flight trajectory instruction set including the wire crimping pipe, the fastening part of the hardware and the target area of ​​the anti-corrosion coating.

[0027] The system receives a preset precision waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system. The data includes the three-dimensional spatial coordinates of the target area and the current roll, pitch, and yaw angular velocity information of the UAV.

[0028] Spatial topology analysis is performed on the preset precise waypoint coordinate sequence to identify the geometric distribution characteristics of key hidden process targets such as wire crimping pipes, hardware fastening parts and anti-corrosion coatings, and a weighted priority queue based on the importance of the targets is constructed.

[0029] Based on a weighted priority queue, an improved A-path planning algorithm combined with a UAV dynamics constraint model is used to calculate the optimal flight trajectory between adjacent waypoints, generating an initial flight path point set containing position, velocity, and acceleration vectors. For this initial flight path point set, a B-spline curve smoothing mechanism is introduced to continuously optimize the path curvature, eliminating the risk of attitude abrupt changes caused by sharp turns and ensuring the optical axis stability of the multispectral imager during hovering. By incorporating wind speed disturbance components in real-time attitude data, the hovering time and shooting trigger threshold in the flight trajectory command set are dynamically adjusted to compensate for the impact of environmental airflow on image quality. The optimized flight path point set and dynamically adjusted parameters are encapsulated into a standard MAVLink protocol command frame, generating a multi-dimensional flight trajectory command set including wire guide pipes, hardware fastening parts, and anti-corrosion coating target areas. Through the above spatial path planning and dynamic compensation processing methods, discrete target coordinates are transformed into smooth flight commands that conform to the kinematic characteristics of the UAV, achieving high-precision and stable coverage acquisition of concealed process areas, laying a spatiotemporal benchmark for subsequent high-quality acquisition of multispectral images.

[0030] For example, in the acceptance scenario of a 500kV transmission line conductor crimping pipe, the preset waypoint includes the coordinates of three crimping points. After the system parses the coordinates, a priority queue is constructed, setting the weight of the crimping points to 1.0 and the weight of ordinary towers to 0.5. An A-planning path planning algorithm is used, setting the maximum acceleration to 2 m / s² and the maximum angular velocity to 30° / s. After B-spline smoothing, the radius of curvature of the path is greater than 5 meters, meeting the stability requirements for hovering shooting. When a crosswind speed of 3 m / s is detected, the hovering time is automatically extended from 2 seconds to 3.5 seconds, and the gimbal compensation angle is adjusted. The final generated MAVLink command set contains 120 waypoints, with a command issuance delay of less than 10 ms, ensuring that the UAV reaches the designated shooting position with a positioning accuracy of ±0.1 meters, significantly improving the spatial alignment and signal-to-noise ratio of multispectral images.

[0031] S1.2: Receive the multi-dimensional flight trajectory instruction set and real-time attitude data, and use the multi-axis gimbal stabilization control algorithm to calculate the compensation for the drone's body shaking, so as to generate a gimbal servo control signal for stabilizing the optical axis pointing of the multispectral imager.

[0032] It receives multi-dimensional flight trajectory instruction sets and real-time attitude data, the attitude telemetry data including the angular velocity, acceleration and Euler angle information of the UAV body in a three-axis coordinate system.

[0033] Low-pass filtering is applied to the real-time attitude data to remove high-frequency mechanical vibration noise and extract low-frequency attitude components that reflect the macroscopic motion trend of the UAV.

[0034] The angle deviation vector of the multispectral imager's optical axis relative to the target area is calculated based on the low-frequency attitude component, and the reverse rotation angle that the gimbal needs to compensate is determined.

[0035] A Kalman filter algorithm is introduced to estimate the state of the viewpoint deviation vector. Combined with the response characteristics of the gimbal servo motor at historical moments, the optimal compensation amount for the next moment is predicted.

[0036] Based on the predicted optimal compensation amount, a pulse width modulation (PWM) control signal is generated to drive the pitch, roll, and yaw motors of the three-axis gimbal to perform reverse motion.

[0037] The residual between the actual attitude and the desired attitude of the gimbal is monitored in real time through a closed-loop feedback mechanism, and the PID control parameters are dynamically adjusted to eliminate steady-state error.

[0038] The multi-axis gimbal stabilization control algorithm described above converts the flight trajectory command from the previous step into a high-precision gimbal servo control signal, enabling the multispectral imager's optical axis to maintain stable pointing at the micro-arc level during hovering and shooting, thus ensuring the spatial geometric consistency of subsequent image acquisition.

[0039] S1.3: Based on the gimbal servo control signal, drive the multispectral imager to perform multi-angle hovering shooting operation, and collect multi-band spectral data for the wire crimping pipe, hardware fastening parts and anti-corrosion coating target areas to generate multispectral raw image frames containing visible light to near-infrared bands.

[0040] It receives the gimbal servo control signal generated in the previous steps, analyzes the pitch angle, yaw angle and roll angle correction values ​​contained therein, and converts them into driving voltage commands for the microelectromechanical system (MEMS) lens group inside the multispectral imager.

[0041] The deflection angle of the MEMS mirror group is adjusted according to the driving voltage command to accurately compensate for the high-frequency micro-vibrations caused by airflow disturbance during the drone's hovering and shooting process, ensuring that the optical axis of the multispectral imager is always perpendicular to or points at the center of the target area such as the wire crimping pipe, hardware fastening parts, and anti-corrosion coating at a preset tilt angle.

[0042] After the optical axis is stably locked onto the target area, the electronic shutter of the multispectral imager is triggered, and multiple narrowband filter channels from visible light to near-infrared bands are opened simultaneously to perform time-division exposure acquisition operation, thereby acquiring high dynamic range original grayscale images of different spectral bands under the same field of view.

[0043] Dark current noise subtraction is performed on the acquired multi-band raw grayscale images. The influence of sensor thermal noise is eliminated by using a pre-calibrated black level reference value, thereby improving the image signal-to-noise ratio and ensuring the discernibility of subtle spectral reflectance differences.

[0044] The denoised image data of each band are spatially registered and stacked according to wavelength order to construct a three-dimensional spectral data cube, in which two dimensions are spatial pixel coordinates and the third dimension is the spectral band index, forming an original data frame containing complete spectral information.

[0045] Through the above-mentioned optical axis stabilization compensation, multi-band synchronous acquisition and spectral data cube construction processing methods, the gimbal control signal of the previous step is transformed into a multispectral raw image frame with high spectral resolution and geometric stability, realizing the accurate capture of multidimensional spectral features of the hidden process parts, and providing a high-quality data foundation for subsequent radiometric correction and AI parameter calculation.

[0046] For example, for the acceptance inspection scenario of a wire crimping tube with a diameter of 30mm, 5 characteristic wavebands within the range of 450nm to 900nm covered by the multispectral camera are set, with each waveband having a bandwidth of 10nm. When the UAV generates an attitude jitter of ±2° due to the impact of gusts, the gimbal servo system responds at a frequency of 500Hz to control the optical axis deviation within 0.05°. The camera exposure time is set to 5ms, and the gain is set to 10dB, so that the resolution of the collected single-waveband image is 1280×1024 pixels. After dark current deduction, the standard deviation of background noise is reduced from 15 gray levels to 3 gray levels. The size of the finally generated 5-waveband spectral data cube is 1280×1024×5, which completely records the spectral reflection characteristics of the surface of the crimping tube, ensuring that the accuracy of the input data for the subsequent coating thickness inversion model is significantly improved.

[0047] S1.4: Perform radiation correction and geometric distortion correction processing on the original multispectral image frames to eliminate the influence of ambient illumination differences and optical deformation, so as to generate a standardized corrected multispectral image sequence.

[0048] Acquire the original multispectral image frames generated in S1.3, read the corresponding dark current reference frame and whiteboard reflectance calibration data, and construct a radiation correction reference model.

[0049] Perform dark current subtraction operation on the original image frames to eliminate the base offset caused by the thermal noise of the sensor, and normalize the pixel values of each waveband by using the linear gain coefficient to compensate for the response difference between different wavebands.

[0050] Construct a perspective projection transformation model based on the real-time attitude angle of the UAV and the intrinsic parameter matrix of the camera, and calculate the mapping relationship between the lens distortion parameter and the image point coordinates.

[0051] Adopt the polynomial distortion correction method to reversely compensate the radial distortion and tangential distortion in the edge area of the image, so as to restore the real geometric shape of the target object.

[0052] Register and fuse the spectral data after radiation correction with the spatial coordinates after geometric correction, so as to generate a standardized corrected multispectral image sequence with a unified physical dimension and geometric accuracy.

[0053] Through the above joint radiation and geometric correction processing method, the original multispectral image frames of the previous step are converted into a standardized corrected multispectral image sequence that eliminates the influence of ambient illumination differences and optical deformation, realizing high-fidelity restoration of the spectral features of hidden process parts and accurate mapping of geometric dimensions, and significantly improving the accuracy of subsequent AI model parameter extraction.

[0054] S1.5: Integrate the standardized corrected multispectral image sequence and the corresponding spatiotemporal positioning metadata, and perform data encapsulation and cache writing operations, so as to generate a multispectral image sequence data packet to be input into the neural network model.

[0055] For example, in the acceptance scenario of conductor crimping pipes, the multispectral camera resolution is set to 1280x1024, acquiring images in 5 bands, with the system generating a data packet for each frame. Assuming the GNSS coordinates of the current acquisition point are longitude 113.264, latitude 23.129, and elevation 50.5, after normalization, they are converted into a 64-bit integer spatial index key (0x7A3B...C1) using Z-order encoding, with a timestamp accurate to the microsecond level (1715623400123456). Protocol Buffers are used to define the message structure, and the serialized single-frame data packet size is approximately 2.5MB. When writing to the buffer, the buffer size is set to 512MB, which can hold approximately 200 frames of data. A CRC32 checksum ensures that there are no bit-flipping errors during data transmission from the sensor to the AI ​​processing module. This mechanism enables the AI ​​model to stably receive spatiotemporally labeled image data at a speed of 30 FPS. When the drone flies at a speed of 5 m / s, the spatial displacement of each frame of image is less than 0.2 meters, meeting the sub-meter level accuracy acceptance requirements and significantly improving the real-time performance and traceability accuracy of hidden process parameter extraction.

[0056] Step S2: Input the multispectral image sequence data packet into the neural network model, and use the target region segmentation method to extract the hidden process areas in the image at the pixel level to generate a mask for the target region to be tested. Specifically, this includes:

[0057] S2.1: Obtain multispectral image sequence data packets and input them into the neural network model. Based on the pre-trained lightweight convolutional neural network backbone architecture, perform feature pyramid construction processing on the multispectral image sequence data packets to generate a multi-scale feature map set containing deep semantic information and shallow spatial details.

[0058] The system reads multispectral image sequence data packets from the onboard high-speed solid-state buffer and parses the Protocol Buffers binary stream to extract the normalized multispectral corrected image matrix and associated spatiotemporal metadata. The extracted multispectral image matrix is ​​then stacked according to band dimensions to construct a four-dimensional tensor input structure with C spectral channels, H rows of pixels, and W columns of pixels, where C corresponds to the number of visible to near-infrared bands. A pre-trained lightweight convolutional neural network backbone architecture is loaded. This architecture uses depthwise separable convolutions instead of standard convolutions to reduce computational complexity and introduces residual connections to alleviate the gradient vanishing problem. The four-dimensional tensor is input into the first layer of the backbone network, and a convolutional downsampling operation with a stride of 2 is performed, halving the spatial resolution while increasing the number of feature channels to generate a primary shallow feature map. This primary shallow feature map is then passed sequentially through multiple stacked inverse residual modules, utilizing a linear bottleneck structure to compress intermediate feature dimensions and extract low-level visual features containing texture details and edge information. In the middle layers of the backbone network, dilated convolutional modules are introduced to expand the receptive field, capturing a wider range of contextual information without reducing spatial resolution, and generating intermediate semantic feature maps. In the deeper layers of the backbone network, further pooling and convolution operations are performed to extract high-level feature maps with highly abstract semantic information. At this point, spatial resolution is significantly reduced, but semantic discriminability is enhanced. 1x1 convolutional kernels are applied to the shallow, middle, and deep feature maps to unify the number of channels, ensuring that each level of feature map has the same channel dimension for subsequent fusion. The shallow high-resolution feature maps with unified channel dimensions are then aligned with the deep low-resolution feature maps using bilinear interpolation upsampling, restoring all feature maps to the same spatial size. The aligned multi-level feature maps are then concatenated along the channel dimension to construct an initial feature pyramid structure containing rich scale information. Layer normalization is performed on each layer of the initial feature pyramid to eliminate the influence of batch statistics fluctuations on feature distribution and improve model convergence stability. Through the above processing method, the multispectral image data of the previous step is transformed into a set of multi-scale feature maps containing deep semantic information and shallow spatial details, realizing a comprehensive representation of the multi-scale features of the hidden process parts, and providing a high-quality feature foundation for subsequent attention mechanism fusion.

[0059] For example, in the case of wire crimping pipe acceptance, the input image resolution is 512x512, and the number of spectral bands is 5. The backbone network adopts the MobileNetV3-Small architecture. The first layer outputs a 64-channel 128x128 feature map. After passing through 3 inverse residual blocks, a 96-channel 64x64 intermediate feature map is generated. Then, after passing through 5 inverse residual blocks, a 160-channel 32x32 high-level feature map is generated. The 160-channel high-level feature map is upsampled to 64x64 using nearest neighbor interpolation, concatenated with the 96-channel intermediate feature map, and then fused with a 1x1 convolution to obtain a 128-channel map. It is then upsampled again to 128x128 and concatenated with the 64-channel shallow feature map to obtain a 96-channel map. Finally, feature maps at three scales are generated: 96-channel 128x128, 128-channel 64x64, and 160-channel 32x32. The process takes about 15ms on the airborne Jetson Nano platform. The feature pyramid effectively preserves the subpixel-level details of the pipe edge and the global position information of the bolt, significantly improving the robustness of small target detection.

[0060] S2.2: Receive a set of multi-scale feature maps, and use the attention mechanism fusion module to perform channel weighting and spatial focusing processing on the set of multi-scale feature maps to generate an enhanced feature representation vector for the hidden process parts.

[0061] Receive a set of multi-scale feature maps generated by step S2.1, which contains deep semantic information and shallow spatial details. The set covers different levels of data, from low-level high-resolution texture features to high-level abstract semantic features.

[0062] A global average pooling operation is performed on the multi-scale feature map set along the channel dimension to compress spatial dimensional information to generate channel descriptors for features at each level, and to quantify the response intensity of each channel in the global context.

[0063] A fully connected neural network layer is constructed based on channel descriptors. The dependencies between channels are learned through a nonlinear activation function, and a weight vector representing the importance of each channel is generated to achieve adaptive channel recalibration.

[0064] The generated channel weight vector is multiplied element-wise with the original multi-scale feature map to suppress the response of the background noise channel and enhance the weight of the key feature channel of the hidden process, thus completing the attention focus of the channel dimension.

[0065] In the spatial dimension, a convolution operation is performed on the channel-weighted feature map to generate a spatial attention mask, which identifies high-confidence process areas in the image, such as the edge of the crimping pipe or the head of the bolt.

[0066] By using the Sigmoid activation function to normalize the spatial attention mask to the 0 to 1 range, a spatial attention map in the form of a probability distribution is formed, highlighting the pixel contribution of the target region.

[0067] The spatial attention map is multiplied element-wise with the channel-weighted feature map to further suppress interference signals in non-target areas and enhance the feature representation of the target area to be tested.

[0068] Bilinear interpolation upsampling is performed on the multi-scale feature maps after spatial focusing to uniformly map all levels of feature maps to the original resolution size of the input image.

[0069] A pixel-by-pixel stitching strategy is adopted to fuse upsampled feature maps of different scales along the channel dimension, thereby constructing a comprehensive feature tensor containing multi-scale contextual information.

[0070] The dimensionality reduction of the fused comprehensive feature tensor is performed by a 1x1 convolutional layer to eliminate redundant channel information and generate a compact and discriminative enhanced feature representation vector for hidden process parts.

[0071] By employing the dual attention mechanism of channel weighting and spatial focusing, the multi-scale feature map set from the previous step is transformed into an enhanced hidden process feature representation vector with high signal-to-noise ratio and strong semantic consistency. This enables the accurate extraction and enhancement of minute process defect features in complex backgrounds, significantly improving the boundary localization accuracy of subsequent segmentation tasks.

[0072] For example, in the acceptance scenario of crimped pipes, the input multi-scale feature map set includes three levels: C3, C4, and C5, with resolutions of 64x64, 32x32, and 16x16, respectively, and 256 channels for each level. Global average pooling is performed on the C5 level feature map to generate a 256-dimensional channel descriptor. Channel weights are calculated through a two-layer fully connected network (reduction ratio of 16), where the channel weights representing metal texture features are amplified by 1.8 times, while the channel weights representing the sky background are suppressed to 0.2 times. The spatial attention module generates a 64x64 single-channel mask, with pixel values ​​close to 1.0 in the crimped pipe region and close to 0.0 in the background region. After dual attention processing, the signal-to-noise ratio of the feature map is significantly improved, and the final generated enhanced feature representation vector has a size of 64x64x256, effectively preserving the edge details of sub-millimeter-level crimping defects and providing a highly discriminative feature foundation for subsequent pixel-level segmentation.

[0073] S2.3: Input the enhanced hidden process feature representation vector, and perform pixel-by-pixel class probability mapping on the enhanced hidden process feature representation vector based on the pixel classifier constrained by the metric learning loss function to generate the initial segmentation probability map of the hidden process.

[0074] It receives an enhanced hidden process feature representation vector output by the preceding attention mechanism fusion module. This vector integrates deep semantic information and shallow spatial details of the multispectral image and serves as the input basis for pixel-level classification.

[0075] A fully convolutional pixel classifier network head based on metric learning constraints is constructed, which maps the enhanced feature representation vector to a preset process category space. The category space includes four categories of labels: press-fit metal area, bolt fastener, anti-corrosion coating, and background noise.

[0076] The logits output of the classifier are normalized using the Softmax activation function. The probability distribution of each pixel belonging to each process category is calculated to generate an initial segmentation probability map. The calculation formula is as follows:

[0077]

[0078] in, This indicates that the pixel at coordinates (i,j) belongs to the category. The probability, For this pixel in category The logits output value on the screen. This represents the total number of categories.

[0079] A triplet loss function is introduced as a metric learning constraint to optimize the minimization of distances between pixels of the same process and the maximization of distances between pixels of different processes in the feature space. The loss function... Defined as:

[0080]

[0081] in, For feature mapping function, For anchor point samples, As a positive sample, For negative samples, For boundary margin, This indicates the operation of retrieving the maximum value. This represents the maximum number of iterations.

[0082] By jointly optimizing the cross-entropy classification loss and triplet metric loss through backpropagation, and adjusting the classifier weight parameters, the generated initial segmentation probability map has higher confidence and discriminative power at the class boundaries.

[0083] The output contains an initial segmentation probability map of the hidden process area with four channels. Each channel corresponds to the pixel belonging probability of a process category, providing a highly discriminative probability basis for subsequent conditional random field post-processing.

[0084] By using pixel classification processing under the aforementioned metric learning constraints, the abstract feature vector is transformed into a concrete category probability distribution, achieving preliminary semantic segmentation of hidden process parts and significantly improving the recognition accuracy of small defect areas in complex backgrounds.

[0085] S2.4: Call the initial segmentation probability map of the hidden process area, and use the conditional random field post-processing method to perform boundary refinement and topology consistency optimization on the initial segmentation probability map of the hidden process area to generate a high-precision binary mask of the target area to be tested.

[0086] S2.5: Based on the high-precision binary mask of the target region to be tested, perform morphological connected component analysis to remove invalid noise points and mark target instances in the high-precision binary mask of the target region to be tested, so as to generate the target region mask and output it to the geometric parameter calculation module.

[0087] A high-precision binary mask of the target area to be tested generated by the previous steps is obtained. The binary mask is a single-channel grayscale image, where a pixel value of 255 represents the target area of ​​the hidden process and a pixel value of 0 represents the background area.

[0088] The binary mask is subjected to an eight-neighbor connected component labeling method. All pixels in the image are traversed, and spatially adjacent foreground pixels are aggregated into independent connected components based on the pixel connectivity in the top, bottom, left, right and diagonal directions. A unique temporary label ID is assigned to each connected component.

[0089] Calculate the geometric properties of each independent connected component, including the total pixel area, the aspect ratio of the bounding rectangle, and the centroid coordinates. Based on prior knowledge of the physical dimensions of the engineering concealment process, set the lower limit of the area threshold and the aspect ratio constraint range.

[0090] Connected components whose total pixel area is less than the lower limit of the area threshold or whose aspect ratio exceeds the constraint range are identified as invalid noise or false detection fragments. The values ​​of their corresponding pixel regions are forcibly set to 0, and background cleaning is performed to remove isolated noise interference.

[0091] Perform renumbering mapping on the retained valid connected components, and reallocate the continuous target instance indexes according to the order of pixel area from largest to smallest or spatial scan line order to ensure that each valid process part has a unique and ordered identity.

[0092] Construct structured mask data containing target instance index, centroid coordinates, and bounding box information, write the renumbered pixel values ​​into the output matrix, and generate the mask of the target region to be tested.

[0093] By using morphological connected component analysis and instance indexing, the initial segmentation results from the previous step are transformed into a mask of the target region to be tested that has denoising properties and unique identity. This enables accurate extraction and instance separation of concealed process targets, providing unambiguous region of interest input for subsequent geometric parameter calculation.

[0094] For example, in the scenario of inspecting crimped pipes, a binary mask with a resolution of 1280x1024 is used as the input. The area threshold is set to 500 pixels, and the aspect ratio is constrained to [0.8, 1.2]. Three connected components are identified, labeled 1, 2, and 3. Component 1 has an area of ​​200 pixels, component 2 has an area of ​​5000 pixels with an aspect ratio of 1.0, and component 3 has an area of ​​4800 pixels with an aspect ratio of 1.1. Component 1 is determined to be noise and removed, while components 2 and 3 are retained and renumbered as 1 and 2. The final output mask contains only two effective crimped pipe regions, eliminating background clutter interference and ensuring that step S3 only performs parameter calculations on the actual process areas, significantly improving measurement robustness.

[0095] like Figure 2 As shown, step S3 involves projecting the image block set of the region of interest (ROI) onto the mask of the target region to be tested, and then processing the ROI image block set to obtain a set of measured values ​​for key process parameters. Specifically, this includes:

[0096] S3.1: Obtain the mask of the target area to be tested and the standardized multispectral correction image sequence. Use the pixel coordinate mapping method to project the mask of the target area to be tested onto the geometric space of the correction image sequence to generate a set of region of interest image blocks containing the edge features of the compression fitting, the bolt head features, and the anti-corrosion coating area features.

[0097] Obtain the final binary mask matrix of the target region output by S2.5 and the normalized multispectral corrected image sequence data stream generated by S1.4.

[0098] The set of non-zero pixel coordinates in the mask matrix is ​​analyzed to extract the boundary extreme points of the pressure pipe, bolt and coating areas, and the minimum bounding box of each independent connected domain is calculated.

[0099] Based on the spatial index range of the region of interest defined by the bounding box coordinates, a slice cropping operation is performed on the standardized multispectral corrected image sequence to separate the local image subsets of the corresponding bands.

[0100] By calling the camera intrinsic parameter matrix and the real-time attitude angle data of the UAV, an inverse projection mapping relationship from the pixel coordinate system to the world coordinate system is constructed to correct the geometric deformation caused by the tilt of the shooting angle.

[0101] Bilinear interpolation is used to resample the cropped local image subset to eliminate subpixel-level displacement errors and ensure strict alignment between image texture and mask boundary.

[0102] The corrected local image subset is subjected to an element-wise logical AND operation with the corresponding binary mask to shield background noise interference and generate a set of image patches of region of interest with high signal-to-noise ratio.

[0103] By using pixel coordinate mapping and geometric correction, the segmentation results from the previous step are transformed into a set of region-of-interest image blocks containing precise spatial location information, enabling accurate localization and isolation of hidden process features and providing a clean data foundation for subsequent parameter calculation.

[0104] S3.2: Receive the set of image blocks of the region of interest, call the sub-pixel level edge detection operator to extract the parallel contour lines on both sides based on the edge features of the press-fit pipe, and perform perspective transformation correction processing based on the camera intrinsic parameter matrix and the real-time attitude data of the UAV to generate an orthophoto geometric model of the press-fit pipe that eliminates the influence of perspective distortion.

[0105] The system receives a set of image patches representing the region of interest and extracts the gradient information of the manhole cover edge. It then applies a subpixel-level Canny edge detection operator to perform non-maximum suppression on the image grayscale gradient, pinpointing the precise location of edge pixels. Finally, it uses quadratic polynomial interpolation to fit the grayscale distribution curve along the edge normal direction and calculates the subpixel coordinate offsets corresponding to gradient extrema. This generates a set of parallel contour line points on both sides of the manhole cover, containing floating-point precision coordinates.

[0106] Acquire the intrinsic parameter matrix and real-time attitude data of the UAV's onboard camera. Construct a mapping relationship from the pixel coordinate system to the camera coordinate system based on the camera intrinsic parameter matrix. Combine the UAV's real-time attitude angle data to construct a rotation matrix, transforming 3D points in the camera coordinate system to the world coordinate system. Establish a camera imaging geometric model including focal length, principal point coordinates, and distortion coefficients.

[0107] Based on the mapping relationship between the parallel contour point sets on both sides of the pressure tube and the world coordinate system, perspective transformation correction is performed. The geometric transformation parameters from the image plane to the orthographic projection plane are calculated using the homography matrix solution method. Inverse perspective mapping is applied to the sub-pixel edge coordinates to eliminate trapezoidal distortion and length contraction effects caused by the tilted shooting angle. The orthographic projection geometric model of the pressure tube in the direction perpendicular to the axis is reconstructed.

[0108] By using subpixel edge extraction and perspective correction processing, the region of interest from the previous step is transformed into an orthophoto geometric model of the press-fit pipe that eliminates the influence of viewing angle distortion, thus achieving the expected technical effect of benchmark construction for high-precision geometric parameter measurement.

[0109] For example, for a multispectral image with a resolution of 1920x1080, the gradient magnitude is calculated using a 3x3 Sobel operator. Dual threshold detection is performed with high and low thresholds of 50 and 150. A quadratic polynomial fit is applied to the neighborhood of the edge pixels to determine a sub-pixel offset of 0.35 pixels. The camera focal length is f=8mm, and the pixel size is 3.45μm. The UAV pitch angle is 15 degrees, and the yaw angle is 5 degrees. A rotation matrix R is constructed, and the homography matrix H is calculated. The inverse matrix of H is applied to the extracted edge point set for coordinate transformation. After correction, the parallelism error of the press-fit pipe edge is reduced to within 0.02 pixels, effectively eliminating the influence of perspective distortion on the edge distance measurement.

[0110] S3.3: Based on the orthophoto geometric model of the press-fit pipe, perform Euclidean distance metric calculation to calculate the measured value of the edge distance. At the same time, use the Hough circle transformation method to locate the bolt reference plane and the end of the thread for the bolt head features. Generate the measured value of the bolt protrusion length by converting the pixel spacing. Form an intermediate result set of hidden process geometric parameters based on the measured values ​​of the edge distance and the measured values ​​of the bolt protrusion length.

[0111] The orthophoto geometric model of the pressure pipe is received, and the pixel coordinate sequence of the parallel contour lines on both sides of the model is extracted. A two-dimensional planar distance mapping function is constructed based on the camera imaging principle to calculate the pixel mean of the vertical Euclidean distance between the contour lines. The pixel mean is multiplied by the spatial resolution coefficient to obtain the measured physical value of the edge distance of the pressure pipe.

[0112] For the characteristic region of the bolt head, the Hough circle transform method is applied to detect the gradient distribution of the circular edge. The center coordinates and radius parameters of the bolt reference surface are located by the peak value of the accumulator. Gray-scale gradient scanning is performed along the axial extension direction of the bolt to identify the boundary transition point between the thread end and the nut contact surface.

[0113] Calculate the axial pixel displacement from the center of the reference plane to the boundary point of the thread end. Correct projection distortion using the UAV attitude angle and convert the actual axial length using trigonometric relationships. Generate the measured value of the bolt protrusion length.

[0114] By integrating the measured values ​​of edge distance and buckle length, an intermediate result set of hidden process geometric parameters is constructed.

[0115] Through the above geometric measurement and projection correction processing, the image pixel features are transformed into physically meaningful process size indicators, realizing non-contact high-precision parameter quantification.

[0116] For example, for a crimped pipe image with a resolution of 0.5 mm / pixel, the contour coordinates are extracted, the average pixel value is calculated to be 40 px, and multiplied by a coefficient to obtain an offset of 20.0 mm. For an M16 bolt, the Hough transform is used to locate the center (100, 100), and the end boundary y=125 is scanned, with an axial displacement of 25 px. After conversion with an attitude correction coefficient of 1.02, the thread length is calculated to be 25.5 mm. This process significantly improves the robustness of micro-dimensional measurements.

[0117] S3.4: Based on the regional features of the anti-corrosion coating in the image patch set of the region of interest, nonlinear inversion processing is performed using a pre-calibrated regression analysis model of spectral reflectance and dry film thickness to estimate the measured value of the dry film thickness of the anti-corrosion coating.

[0118] Receive spectral data of the anti-corrosion coating region from the set of image blocks of the region of interest, extract the reflectance values ​​of pixels in each band, and construct a multidimensional spectral feature vector.

[0119] The pre-calibrated regression analysis model of spectral reflectance and dry film thickness is invoked. This model establishes a nonlinear mapping relationship between light absorption and medium thickness based on the Lambert-Beer law.

[0120] The least squares method is used to iteratively optimize the model parameters and determine the absorption coefficient and scattering coefficient under a specific waveband in order to eliminate the reflection interference of the substrate material.

[0121] The measured dry film thickness of the anti-corrosion coating is calculated using the following formula:

[0122]

[0123] in, For the dry film thickness of the coating, The absorption coefficient is... The scattering coefficient is... As the substrate reference reflectivity, This represents the measured reflectivity for the current band.

[0124] The calculated thickness values ​​are correlated with spatial coordinates to generate a pixel-level thickness distribution matrix, and outliers are removed to improve measurement robustness.

[0125] By using nonlinear inversion processing, the spectral reflectance data from the previous step is converted into the measured value of the dry film thickness of the anti-corrosion coating, thus achieving non-contact, high-precision quantification of the physical parameters of the concealed process.

[0126] S3.5: Integrate the measured values ​​of the edge distance, bolt protrusion length, and dry film thickness of the anti-corrosion coating obtained from the intermediate results set of the hidden process geometric parameters, and perform a multi-source data fusion and encapsulation operation to generate a set of measured values ​​of key process parameters containing information on all key process dimensions.

[0127] like Figure 3 As shown, step S4 involves: calling the locally stored construction acceptance standard library, comparing the measured values ​​of the key process parameters with the standard thresholds in the library item by item to generate a compliance judgment result of "qualified" or "unqualified". Specifically, this includes:

[0128] S4.1: Based on the set of measured values ​​of key process parameters generated in the previous steps, read the standard threshold configuration information in the locally stored construction acceptance standard library, and parse the standard threshold configuration information to construct a standardized comparison benchmark set that includes the upper limit value of the edge distance between the press-fit pipe, the lower limit value of the bolt protrusion length, and the range of coating dry film thickness values.

[0129] The system receives a set of measured values ​​for key process parameters generated in the previous steps. This set includes quantitative indicators such as the edge distance of the crimped pipe, the bolt protrusion length, and the dry film thickness of the coating. It reads the locally stored construction acceptance standard library, which stores compliance threshold configuration information corresponding to different voltage levels and hardware models in the form of structured data tables. The read standard threshold configuration information is parsed to extract the engineering type identifier matching the current acceptance task. Based on the engineering type identifier, the corresponding crimping process specification is retrieved to obtain the maximum allowable deviation upper limit value for the edge distance of the crimped pipe. Simultaneously, the fastening process specification is retrieved to obtain the minimum retention length lower limit value that the bolt protrusion length must meet. Further, the anti-corrosion process specification is retrieved to obtain the lower and upper limits of the acceptable range for the dry film thickness of the coating. The extracted upper limit value for edge distance, lower limit value for protrusion length, and range value for coating thickness are standardized and converted to ensure that the dimensions are consistent with the measured values. A standardized comparison benchmark set containing the above three key indicators is constructed. This benchmark set serves as the logical reference for subsequent item-by-item consistency comparisons. By parsing the standard library and extracting threshold parameters under specific working conditions, discrete industry standards are transformed into calculable digital comparison benchmarks, enabling automated loading and precise matching of acceptance criteria.

[0130] For example, in the acceptance scenario of conductor crimping for a 500kV transmission line, the system reads the configuration of hydraulic splice pipe model JYD-400 / 35 from the standard library. The parsing reveals an upper limit of 2.5mm for the distance between the crimped pipe edges, a lower limit of 3.0mm for the M16 bolt extension length, and a dry film thickness range of [80μm, 120μm] for the anti-corrosion coating. The system aligns the measured distance between the edges (2.1mm), extension length (3.5mm), and coating thickness (95μm) with the aforementioned benchmark set. If the measured distance between the edges exceeds 2.5mm or the extension length is less than 3.0mm, it is marked as unqualified. This process eliminates the lag of manual standard review, ensuring that the comparison benchmark and measured data are matched within milliseconds, significantly improving acceptance efficiency.

[0131] S4.2: Based on the upper limit value of the edge distance of the crimped pipe in the standardized comparison benchmark set, perform numerical boundary verification processing on the measured edge distance values ​​in the set of measured values ​​of key process parameters to generate a single compliance mark for the crimping process.

[0132] Extract the measured values ​​of the edge distance between the press-fit pipes from the set of measured values ​​of key process parameters, and read the upper limit threshold of the edge distance between the press-fit pipes corresponding to the standardized comparison benchmark set.

[0133] A numerical boundary verification logic is constructed to compare the extracted measured value of the pipe joint distance with the upper limit threshold of the joint distance.

[0134] If the measured distance between the edges of the press-fit pipe is less than or equal to the upper limit threshold of the edge distance, then the single process is determined to meet the geometric constraint conditions.

[0135] If the measured distance between the edges of the crimped pipe is greater than the upper limit threshold, then the process is deemed to have a dimensional defect that does not meet compliance requirements.

[0136] Based on the judgment result, a Boolean-type compliance flag is generated for each crimping process item, where a true value represents compliance and a false value represents non-compliance.

[0137] By using numerical boundary verification, the geometric measurement data from the previous step is transformed into standardized compliance logic tags, enabling automated initial screening and quantitative judgment of the pressing process quality.

[0138] For example, the upper limit threshold for the edge distance of the crimped pipe is set to 10.5mm. When the measured edge distance calculated by the airborne AI is 10.2mm, a logical judgment of 10.2 ≤ 10.5 is executed, the result is true, and a compliance mark of "qualified" is generated. When the measured edge distance is 10.8mm, a logical judgment of 10.8 ≤ 10.5 is executed, the result is false, and a compliance mark of "unqualified" is generated. This process eliminates the subjective error of manual visual inspection and significantly improves the objectivity and consistency of acceptance judgment.

[0139] S4.3: Based on the lower limit of bolt protrusion length in the standardized comparison benchmark set, perform minimum threshold logic judgment processing on the measured bolt protrusion length in the set of measured values ​​of key process parameters to generate a single compliance mark for fastening process.

[0140] The lower limit threshold parameter of bolt protrusion length is extracted from the standardized comparison benchmark. This parameter is set according to the minimum number of exposed thread turns or length requirements of the bolt fastening process in the power fittings installation and acceptance specifications, and serves as a rigid constraint boundary for compliance judgment.

[0141] Read the measured bolt protrusion length from the set of measured values ​​of key process parameters. This value is calculated by positioning the bolt reference plane and the thread end through Hough circle transformation in the previous steps, and then performing perspective correction in combination with camera intrinsic parameters and UAV attitude data. It represents the actual exposed physical length of the bolt at the current detection point.

[0142] A minimum threshold logic judgment model is constructed, which compares the measured value of the bolt protrusion length with the lower threshold value, and uses the following inequality logic to determine the compliance status:

[0143]

[0144] in, For each fastening process item, 1 represents compliance and 0 represents non-compliance. This is the measured value of the bolt protrusion length; This is the lower limit threshold for the bolt protrusion length.

[0145] Based on the logical judgment result, a corresponding Boolean compliance mark is generated. If the measured value is greater than or equal to the lower limit threshold, it is marked as compliant, indicating that the bolt tightness meets the requirements for preventing loosening and electrical contact. If the measured value is less than the lower limit threshold, it is marked as non-compliant, indicating that there is a risk of under-tightening or stripping of the bolt threads.

[0146] The generated fastening process compliance tags are encapsulated in a temporary judgment result cache area, waiting to be aggregated with the individual tags of the pressing process and corrosion prevention process to form the final overall acceptance conclusion.

[0147] Through the above minimum threshold logic judgment process, continuous geometric measurement data is transformed into discrete compliance status indicators, realizing automated, standardized and rapid screening of bolt fastening process quality and eliminating subjective errors of manual visual inspection.

[0148] For example, in the acceptance test of M20 high-strength bolts at a node on a transmission line tower, the standardized comparison benchmark sets a lower limit threshold of 2.5mm for bolt protrusion length (corresponding to approximately 2-3 threads). The onboard AI module calculates the actual bolt protrusion length to be 3.2mm. The logic checks: 3.2mm ≥ 2.5mm, the condition is met, and a compliance mark M=1 (qualified) is generated for the fastening process. If the actual bolt protrusion length at another location is 1.8mm, since 1.8mm < 2.5mm, the condition is not met, a mark M=0 (unqualified) is generated, and a local re-inspection instruction is triggered. This process is completed within milliseconds, significantly improving the efficiency and consistency of on-site acceptance.

[0149] S4.4: Based on the coating dry film thickness range values ​​in the standardized comparison benchmark set, perform range inclusion detection processing on the measured values ​​of anti-corrosion coating dry film thickness in the set of measured values ​​of key process parameters to generate a single compliance mark for the anti-corrosion process.

[0150] Obtain the coating dry film thickness range values ​​from the standardized comparison benchmark set, analyze the minimum and maximum thresholds for compliance of the anti-corrosion process, and establish the judgment boundary conditions.

[0151] Read the measured value of the dry film thickness of the anti-corrosion coating from the set of measured values ​​of key process parameters, and input it as the target physical quantity to be verified into the interval detection logic unit.

[0152] A logic for determining the inclusion of a range is constructed. The measured value of the dry film thickness of the anti-corrosion coating is compared with the minimum and maximum threshold values ​​to assess whether the measured value falls within the valid and compliant range.

[0153] If the measured dry film thickness of the anti-corrosion coating is greater than or equal to the minimum threshold and less than or equal to the maximum threshold, then the measuring point is determined to meet the construction specifications for the anti-corrosion coating thickness, and a Boolean truth value marker representing the compliance status is generated.

[0154] If the measured dry film thickness of the anti-corrosion coating is less than the minimum threshold or greater than the maximum threshold, it is determined that the coating is too thin or too thick at the measuring point, and a Boolean false value mark representing the violation is generated.

[0155] The generated Boolean truth or false value tags are encapsulated as individual compliance tags for corrosion prevention processes, serving as the basic input data for subsequent multi-source logical aggregation.

[0156] By using interval inclusion detection processing, the spectral inversion thickness data from the previous step is transformed into standardized binary compliance status markers, thereby achieving automated quantitative judgment of anti-corrosion process quality and expected technical effects of anomaly early warning.

[0157] For example, the standard range for the dry film thickness of the anti-corrosion coating is set to [80μm, 120μm]. When the measured value obtained from spectral inversion is 95μm, the logic checks whether 95≥80 and 95≤120 are true, and outputs a compliance flag (True). When the measured value is 70μm, the logic checks whether 70≥80 is true, and outputs a violation flag (False). When the measured value is 130μm, the logic checks whether 130≤120 is true, and outputs a violation flag (False). This process eliminates the subjective error of manual visual inspection and significantly improves the objectivity and consistency of the acceptance of concealed works.

[0158] S4.5: Combine the compliance marks of the integrated crimping process, fastening process, and corrosion prevention process, perform multi-source logical AND operation aggregation processing to generate the final compliance judgment result of qualified or unqualified.

[0159] The system receives compliance tags for each of the crimping, fastening, and corrosion prevention processes, and constructs a multi-dimensional process compliance state vector. It then performs logical truth mapping on this multi-dimensional process compliance state vector, converting each tag into a binary Boolean variable, where a pass tag is mapped to logic 1 and a fail tag to logic 0.

[0160] Based on a set of binary Boolean variables, a multi-source logical AND operation aggregation engine is invoked to perform an item-by-item logical AND operation. This operation follows a unanimous pass principle; the aggregation result is only considered to be logical 1 if all individual process markers are logical 1.

[0161] If any process step is marked as logic 0, the result of the logical AND operation is immediately set to logic 0, triggering a veto mechanism. This process ensures the rigor of hidden process acceptance and prevents local defects from being masked by the appearance of overall compliance.

[0162] The binary aggregation result generated by the logical AND operation is mapped back to the business semantic space to generate the final compliance judgment result string. If the result is logical 1, a "qualified" status flag is output; if the result is logical 0, a "unqualified" status flag is output and an exception item index is triggered.

[0163] By using a multi-source logic and operation aggregation processing method, the discrete single compliance markers of the previous step are transformed into a unified final compliance judgment result with legal effect, thereby achieving the expected technical effect of rigor and non-repudiation of the acceptance conclusion of concealed processes.

[0164] For example, in the acceptance scenario of a transmission line conductor crimping point, the system identifies the following compliance indicators: crimping process compliance is marked as 1 (measured edge distance 32.5mm, less than the upper limit 33.0mm); fastening process compliance is marked as 0 (measured bolt protrusion length 18.0mm, less than the lower limit 20.0mm); and anti-corrosion process compliance is marked as 1 (measured coating thickness 120μm, within the 100-150μm range). A state vector [1, 0, 1] is constructed, and a logical AND operation is performed: 1 AND 0 AND 1. Since the intermediate term is 0, according to Boolean algebra laws, the result is forced to be 0. The system marks the final compliance judgment result as "unqualified" and automatically identifies the fastening process as the defect source. This judgment result is directly used as input to step S5, ensuring that the evidence data accurately reflects the on-site quality status, avoiding shirking responsibility due to subjective human judgment, and significantly improving the objectivity and traceability credibility of the acceptance data.

[0165] Step S5: Based on the compliance judgment result, the set of measured values ​​of key process parameters, and the unique code of the UAV equipment, perform a structured encapsulation operation to generate a structured acceptance record with complete traceability information, thus completing the compliance acceptance of the hidden engineering processes. Specifically, this includes:

[0166] S5.1: Based on the compliance judgment results, the set of measured values ​​of key process parameters, timestamps, spatial location information and the unique code of UAV equipment, construct the hidden process acceptance metadata object using the structured data mapping protocol to generate a standardized acceptance information body containing all discrete traceability elements.

[0167] The system receives the compliance judgment results and the set of measured values ​​of key process parameters generated in the preceding steps, as well as the high-precision timestamps, GNSS spatial coordinates, and unique equipment codes synchronously pushed by the UAV flight control system. It parses the data types and semantic structures of these discrete data sources and establishes a mapping table between each data field and standard JSON key-value pairs based on the predefined hidden process acceptance metadata schema specification. It performs unit standardization on the set of measured values ​​of key process parameters, uniformly converting the pipe fitting edge distance to millimeters, bolt thread length to millimeters, and coating dry film thickness to micrometers to eliminate dimensional differences. It performs WGS-84 coordinate system verification on the GNSS spatial coordinates, supplementing the altitude field to form a three-dimensional position vector containing longitude, latitude, and elevation. It maps the compliance judgment results to Boolean status identifiers and the unique equipment codes to string identity identifiers, ensuring strict consistency of data types. Using structured data construction methods, it assembles the standardized parameter values, position vectors, timestamps, identity identifiers, and status identifiers according to a nested hierarchy to generate a standardized acceptance information body object with complete traceability elements. By using a structured data mapping protocol, multi-source heterogeneous acceptance elements are transformed into a standardized acceptance information body with a unified format and clear semantics, realizing the standardized encapsulation of acceptance data and providing a standardized metadata foundation for subsequent image payload fusion.

[0168] For example, the compliance judgment result "qualified" is received, with the following key process parameters: {edge distance: 35.2mm, protrusion length: 12.5mm, coating thickness: 120μm}, timestamp "2023-10-27T10:00:00Z", coordinates {lat: 23.12,lon: 113.25, alt: 50}, and device ID "UAV_001". According to the schema specification, the edge distance and protrusion length are mapped to double type, the coating thickness to integer type, and the coordinates to object type. Construct a JSON object: {"meta": {"timestamp": "2023-10-27T10:00:00Z", "location": {"lat": 23.12, "lon": 113.25, "alt": 50}, "device_id": "UAV_001"}, "params": {"crimp_distance": 35.2, "bolt_length": 12.5, "coating_thickness": 120}, "status":true}. This standardized acceptance information body eliminates data format ambiguity and significantly improves the efficiency and compatibility of subsequent data assembly.

[0169] S5.2: Perform lossless compression encoding on the current acceptance image and use Base64 encoding to convert the compressed binary image stream into a text-format image payload to generate binary-to-text image data blocks adapted for JSON structure transmission.

[0170] This step obtains the standardized acceptance information body generated in step S5.1 and the current acceptance raw image data output in step S3 as input objects. Lossless compression encoding is performed on the current acceptance raw image data using PNG or TIFF lossless compression methods. This preserves the complete pixel grayscale values ​​and spectral reflectance information of each band of the multispectral image, eliminating the interference of quantization noise introduced by lossy compression on subsequent hash verification consistency, and generating a binary image compressed stream. The byte sequence of the binary image compressed stream is read, and according to the Base64 encoding standard, every three 8-bit bytes are divided into groups, mapped to four 6-bit index values. Based on the Base64 character set mapping table, the index values ​​are converted into printable characters in the ASCII character set, including uppercase letters, lowercase letters, numbers, plus signs, and forward slashes. If the last three bytes are less than three, equal signs are used for padding. The converted ASCII character sequences are concatenated into a continuous text string, forming a binary-to-text image data block. This data block has plain text attributes and can be seamlessly embedded into JSON structure field values, avoiding data truncation or parsing errors caused by incompatible encoding formats during binary data serialization and transmission. By combining lossless compression and Base64 encoding, the binary image stream from the previous step is transformed into binary-to-text image data blocks adapted for JSON structure transmission, thus achieving the integrity preservation and standardization compatibility of visual evidence data in heterogeneous encapsulation environments.

[0171] For example, for a multispectral corrected image with a resolution of 1920x1080, a lossless PNG compression method is used with a compression ratio set to the highest level of 9, generating a binary compressed stream of approximately 2.5MB. This binary stream is then Base64 encoded, with each 3 bytes mapped to 4 characters, ultimately generating an ASCII text string of approximately 3.4MB. This string contains only the characters AZ, az, 0-9, +, / , and =, without any invisible control characters. This text string is assigned to the `image_data` field of a JSON object, ensuring that in the subsequent heterogeneous data assembly step S5.3, the image data can be serialized uniformly with compliance judgment results, spatiotemporal stamps, and other text metadata, significantly improving the tamper-resistance of the evidence data and cross-platform parsing consistency.

[0172] S5.3: Based on the standardized acceptance information body and binary-to-text image data block, the heterogeneous data assembly engine is called to perform field alignment and nested encapsulation operations to generate a structured acceptance record original message containing complete business logic and visual evidence.

[0173] Receive standardized acceptance information and binary-to-text image data blocks, initialize the JSON object builder of the heterogeneous data assembly engine, and set the root node as the hidden process acceptance record structure.

[0174] The compliance judgment results, key process parameter measured values, timestamps, spatial location information, and unique UAV equipment codes in the standardized acceptance information body are parsed into key-value pairs and mapped to the metadata level fields of the JSON object.

[0175] For numerical parameters such as the edge distance of the crimped pipe, the bolt protrusion length, and the dry film thickness of the coating, floating-point precision standardization is performed, retaining four significant digits after the decimal point to ensure the consistency of data storage format.

[0176] The binary-to-text image data block is embedded as an independent field in the multimedia payload layer of the JSON object, with the field identifier set as image_payload and the data type marked as Base64 encoded string.

[0177] Establish a logical association index between the metadata level and the multimedia payload level, and use a unique transaction ID to strongly bind image evidence with the corresponding geometric and physical parameters and judgment results to prevent data misalignment.

[0178] Calling the serialization interface converts the constructed JSON object into a UTF-8 encoded byte stream, performs syntax validity checks, and ensures that the nested structure conforms to the RFC 8259 standard specification.

[0179] By using the field alignment and nested encapsulation operations of the heterogeneous data assembly engine, the discrete traceability elements and visual evidence from the previous step are transformed into a rigorously structured original message of the acceptance record, thus achieving the complete aggregation of multi-source heterogeneous data.

[0180] S5.4: Perform UTF-8 character set standardization conversion on the original structured acceptance record message and remove all non-printable control characters to eliminate encoding ambiguity, so as to generate a standardized acceptance record string with cross-platform parsing consistency.

[0181] Receive the original structured acceptance record message, call the UTF-8 character set encoder to scan the message character stream, identify and mark all non-standard ASCII control characters and illegal Unicode proxy pairs.

[0182] For the identified non-printable control characters, a filter mask is constructed according to the ISO / IEC 10646 standard to replace control codes whose hexadecimal values ​​are in the range of 0x00 to 0x1F and are not newline, carriage return, or tab characters with an empty string, thereby eliminating invisible interference characters introduced by serial communication of embedded devices.

[0183] The remaining valid text data is subjected to normalized form C (NFC) transformation, which merges the combined character sequences of the decomposed form into a single pre-assembled character to ensure the consistency of byte sequence during cross-platform parsing of multilingual metadata and prevent hash value drift caused by differences in encoding equivalence.

[0184] The acceptance record string is protected by a regular expression validation mechanism that detects and removes implicit formatting characters such as zero-width spaces and byte order mark (BOM) to ensure the structural purity of the JSON serialization process.

[0185] Through the above character cleaning and standardization processing, the heterogeneous original message generated in the previous step is transformed into a standardized acceptance record string with a definite byte sequence, thereby achieving the expected technical effect of eliminating encoding ambiguity and ensuring the uniqueness of subsequent hash digest operations.

[0186] S5.5: Based on the normalized acceptance record string, the SHA-256 hash algorithm is used to perform a digest operation to generate a fixed-length data fingerprint, and this data fingerprint is used as an integrity verification field to fill back into the normalized acceptance record string, finally generating a structured acceptance record with self-verification capability.

[0187] Receive a normalized acceptance record string that has been standardized to UTF-8 and has had non-printable control characters removed, and use it as the input data stream for the SHA-256 hash algorithm.

[0188] The SHA-256 one-way hash function interface in the cryptographic hash library is called to perform byte-level traversal and block padding on the input normalized acceptance record string to ensure that the data length meets the requirement of being an integer multiple of 512 bits.

[0189] The core compression function based on the SHA-256 algorithm performs 64 rounds of logical operations on the padded data block, including Boolean function selection, modulo addition, and cyclic shift operations, and gradually updates the value of the internal hash status register.

[0190] After completing the iterative processing of all data blocks, the final 256-bit binary hash state value is extracted and converted into a 64-character hexadecimal string representation, forming a data fingerprint digest with anti-collision properties.

[0191] The data fingerprint summary field is embedded into the preset metadata node of the standardized acceptance record string to establish a strong binding relationship between the data content and the integrity verification value.

[0192] Perform secondary serialization validation on the complete string structure after embedding the data fingerprint to ensure the legality of the JSON syntax structure and the correctness of the field nesting relationship.

[0193] By using SHA-256 hash digest operation and backfilling mechanism, the standardized text from the previous step is transformed into a structured acceptance record with self-verification capability, achieving the expected technical effect of immediate detection of data tampering and traceability of information integrity.

[0194] For example, for a wire crimping acceptance task, the standardized acceptance record string includes the measured distance between the crimped pipe pairs (18.5 mm), the bolt protrusion length (12.0 mm), the coating thickness (120 μm), and the timestamp (2023-10-27T10:00:00Z). The system performs a SHA-256 operation on this string, compressing it through 64 rounds to generate a fixed-length hexadecimal hash value (a1b2c3...f9e8). This hash value is then filled into the "hash_digest" field of the record string. If any bit changes during subsequent transmission, such as increasing the timestamp by 1 second, the recalculated hash value will experience an avalanche effect, becoming completely inconsistent with the original value, significantly improving the credibility and tamper-proof capability of the stored data.

[0195] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0196] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0197] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for compliance acceptance of concealed engineering processes based on UAV-borne imaging, characterized in that, Specifically, it includes: S1: Obtain the preset precise waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system, obtain the gimbal servo control signal based on the preset precise waypoint coordinate sequence and real-time attitude data, and drive the multispectral imager to perform multi-angle hovering shooting operation according to the gimbal servo control signal to obtain multispectral image sequence data packets. S2: Input the multispectral image sequence data packet into the neural network model, and use the target region segmentation method to extract the hidden process parts in the image at the pixel level to generate a mask of the target region to be tested; S3: Project the target area mask to obtain a set of image blocks of the region of interest, and process the set of image blocks of the region of interest to obtain a set of measured values ​​of key process parameters; S4: Call the locally stored construction acceptance standard library, compare the set of measured values ​​of the key process parameters with the standard thresholds in the construction acceptance standard library item by item, and generate a compliance judgment result of qualified or unqualified. S5: Based on the compliance judgment result, the set of measured values ​​of key process parameters and the unique code of the UAV equipment, perform a structured encapsulation operation to generate a structured acceptance record with complete traceability information, and complete the compliance acceptance of the hidden process of the project.

2. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 1, characterized in that, The neural network model adopts a convolutional neural network backbone architecture and introduces an attention mechanism and a triplet loss function to extract pixel-level hidden process parts.

3. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 1, characterized in that, S1 specifically includes: The system acquires a preset precision waypoint coordinate sequence and real-time attitude data pushed by the UAV flight control system, and performs spatial path planning processing based on the preset precision waypoint coordinate sequence to generate a multi-dimensional flight trajectory instruction set. Receive the multi-dimensional flight trajectory instruction set and real-time attitude data, perform compensation calculations for UAV body vibration, and generate gimbal servo control signals; The multispectral imager is driven by the gimbal servo control signal to perform multi-angle hovering shooting operations to obtain multispectral raw image frames. Radiometric correction and geometric distortion correction are performed on the original multispectral image frames to generate a standardized multispectral corrected image sequence; Based on the standardized multispectral corrected image sequence, data encapsulation and cache writing operations are performed to generate a multispectral image sequence data packet.

4. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 3, characterized in that, The multi-dimensional flight trajectory instruction set includes wire crimping pipes, hardware fastening parts, and anti-corrosion coating target areas.

5. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 4, characterized in that, The multispectral imager driven by the gimbal servo control signal performs multi-angle hovering shooting operations to obtain multispectral raw image frames, including: Based on the gimbal servo control signal, the multispectral imager is driven to perform multi-angle hovering shooting operations, and multi-band spectral data is collected for the target areas of wire crimping pipes, hardware fastening parts and anti-corrosion coatings to generate multispectral raw image frames containing visible light to near-infrared bands.

6. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 1, characterized in that, S2 specifically includes: The multispectral image sequence data packet is input into a neural network model, and feature pyramid construction processing is performed on the multispectral image sequence data packet based on a pre-trained lightweight convolutional neural network backbone architecture to generate a multi-scale feature map set containing deep semantic information and shallow spatial details. The initial segmentation probability map of the hidden process area is obtained by processing the multi-scale feature map set; The initial segmentation probability map of the hidden process area is called, and boundary refinement and topology consistency optimization are performed on the initial segmentation probability map of the hidden process area to generate a high-precision binary mask of the target area to be measured. Based on the high-precision binary mask of the target region to be tested, morphological connected component analysis is performed to remove invalid noise points and mark target instances in the high-precision binary mask of the target region to be tested, so as to generate the target region mask.

7. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 6, characterized in that, The process of processing the multi-scale feature map set to obtain the initial segmentation probability map of the hidden process area includes: Channel weighting and spatial focusing are performed on the multi-scale feature map set to generate an enhanced hidden process feature representation vector. Based on the pixel classifier, pixel-wise category probability mapping is performed on the enhanced hidden process feature representation vector to generate an initial segmentation probability map of the hidden process.

8. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 1, characterized in that, S3 specifically includes: The mask of the target area to be tested and the standardized multispectral correction image sequence are obtained. The mask of the target area to be tested is projected onto the geometric space of the standardized multispectral correction image sequence using the pixel coordinate mapping method to generate a set of region of interest image blocks containing features of the pipe edge, bolt head and anti-corrosion coating area. The edge features of the pressure pipe in the image patch set of the region of interest are processed to obtain the orthophoto geometric model of the pressure pipe; Based on the orthophoto geometric model of the press-fit pipe, the Euclidean distance metric is used to calculate the measured value of the side distance. At the same time, the bolt head features in the image block set of the region of interest are processed to obtain the measured value of the bolt protrusion length. Based on the measured values ​​of the side distance and the measured values ​​of the bolt protrusion length, an intermediate result set of hidden process geometric parameters is formed. Nonlinear inversion processing is performed based on the region features of the anti-corrosion coating in the set of image blocks of the region of interest to estimate the measured value of the dry film thickness of the anti-corrosion coating. The measured values ​​of the edge distance, bolt protrusion length, and dry film thickness of the anti-corrosion coating obtained by spectral inversion are integrated from the intermediate results set of hidden process geometric parameters. A multi-source data fusion and encapsulation operation is performed to generate a set of measured values ​​of key process parameters containing information on all key process dimensions.

9. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 8, characterized in that, The process of processing the edge features of the bushing in the image patch set of the region of interest to obtain the orthophoto geometric model of the bushing includes: Based on the edge features of the pressure pipe in the image patch set of the region of interest, a sub-pixel-level edge detection operator is called to extract the parallel contour lines on both sides. Then, based on the camera intrinsic parameter matrix and the real-time attitude data of the UAV, perspective transformation correction processing is performed to generate the orthophoto geometric model of the pressure pipe.

10. The method for compliance acceptance of concealed engineering processes based on UAV-borne imaging according to claim 8, characterized in that, The process of processing the bolt head features in the image block set of the region of interest to obtain the measured bolt protrusion length includes: The Hough circle transform method is used to locate the bolt reference plane and the end of the thread based on the bolt head features, and the measured value of the bolt thread extension length is generated by converting the pixel spacing.

Images