Unmanned aerial vehicle inspection report intelligent generation method based on data feature extraction

Abstract

The invention discloses an unmanned aerial vehicle inspection report intelligent generation method based on data feature extraction, and the method comprises the steps: collecting multi-voyage unmanned aerial vehicle inspection data, and carrying out the preprocessing to generate a standardized data set; the improved VMama coding is utilized, and cross-voyage-number feature flow alignment and residual error correction of the same object are carried out; analyzing multi-temporal change, extracting state change, deducing a normal trajectory, and calculating deviation; constructing a causal evidence graph by taking an inspection feature structure as an evidence node, and extracting a convergence sub-graph; inputting the convergence subgraph into a report inversion generator, and carrying out structure inverse solution, time sequence inverse deduction and reconstruction; and outputting a final inspection report, and performing association storage on the report and the evidence node. According to the invention, by introducing the improved feature extraction and intelligent report inversion technology, the automatic analysis of the unmanned aerial vehicle inspection data, the accurate abnormal identification and the verifiable automatic generation of the inspection report are realized.

Description

Technical Field

[0001] This invention relates to the field of drone inspection data processing technology, and in particular to a method for intelligent generation of drone inspection reports based on data feature extraction. Background Technology

[0002] With the development of drone and sensor technologies, drone inspections have been widely applied in scenarios such as power lines, bridge structures, industrial facilities, and energy pipelines. Drones can acquire high-resolution images and video data in complex or hazardous environments, effectively reducing the cost of manual inspections and improving operational safety. In existing technologies, the inspection process typically relies on drones equipped with visible light or infrared imaging devices to periodically photograph the inspected objects. The collected data is then analyzed manually or semi-automatically to generate corresponding inspection reports. While these methods improve inspection efficiency to some extent, they still largely depend on human experience in data processing and result presentation, resulting in limited levels of automation and intelligence.

[0003] Current UAV inspection technologies generally focus on target detection or defect identification based on data from a single flight. The processing of inspection images or videos often employs models such as convolutional neural networks or visual Transformers for feature extraction. While these methods are effective in identifying anomalies in single frames, they often lack effective spatiotemporal alignment and consistency modeling techniques for inspection data spanning multiple flights and time phases, making it difficult to accurately reflect the true changes in the inspected object over time. Data collected from different flights is easily affected by changes in shooting angle, lighting conditions, and flight attitude, leading to unstable change analysis results and insufficient accuracy in identifying the root causes of anomalies.

[0004] Current inspection reports are typically generated based on rule templates or manual editing. The generated report content lacks a clear, structured relationship with the underlying inspection data, making it difficult to trace the report's conclusions back to corresponding inspection evidence and lacking an effective verification mechanism. Once the inspection data changes or the identification results are biased, the report content often requires repeated manual revisions, resulting in low efficiency and high subjectivity. Current technology cannot yet achieve structured representation of multi-flight UAV inspection data, stable analysis of cross-temporal changes, or automatic generation of inspection reports with traceable and verifiable evidence.

[0005] Therefore, how to provide an intelligent method for generating drone inspection reports based on data feature extraction is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent generation method for UAV inspection reports based on data feature extraction. This invention performs unified preprocessing on multi-flight UAV inspection data, combines an improved VMamba network to achieve cross-flight feature alignment and multi-temporal change analysis, and constructs a causal inspection evidence graph to structure and organize anomaly information. Based on this, an inspection report inversion generator with structural deconstruction and temporal inverse verification capabilities is introduced to achieve automatic generation and consistency verification of inspection reports. This invention fully utilizes computer vision, spatiotemporal data modeling, and intelligent information generation technologies to accurately depict the state changes of inspected objects, ensuring that the inspection report content corresponds one-to-one with actual inspection evidence. It possesses advantages such as high automation, strong anomaly identification stability, and verifiable and traceable report results.

[0007] The intelligent generation method for UAV inspection reports based on data feature extraction according to embodiments of the present invention includes: The drone inspection platform collects drone inspection data from multiple flights, performs preprocessing on the drone inspection data, and generates a standardized inspection dataset. On the standardized inspection dataset, dual-scale selective scanning feature encoding is performed using an improved VMamba network. Spatiotemporal alignment and residual correction are performed on the feature streams of different voyages of the same inspection object, generating cross-voyage consistent feature maps and converging them into a cross-voyage associated feature set of the inspection object. Multi-temporal change analysis is performed on the feature set of cross-flight inspection objects to extract the state change information of the inspection objects, and the normal evolution trajectory is deduced by combining historical inspection data. The deviation between the actual change results and the normal evolution trajectory is calculated to generate the inspection feature structure. A causal inspection evidence graph is constructed using the inspection feature structure as evidence nodes. Causal consistency, evolutionary reversal and counterfactual convergence relationships are established between the evidence nodes, and causal convergence subgraphs are extracted. The causal convergence subgraph is input into the inspection report inversion generator to generate an initial inspection report. The structural inverse solution processing and time-series back-inference verification are then executed sequentially. The anomalies, locations and changes in the initial inspection report are mapped back to the corresponding evidence nodes. Content that does not meet the forward derivation and backtracking conditions is reconstructed. After the inspection report inversion generator completes the structural inverse solution processing and time sequence inverse verification, it outputs the final inspection report and associates and stores the final UAV inspection report with the corresponding inspection evidence nodes.

[0008] Optionally, the UAV inspection data includes visible light image data or video data, infrared thermal imaging data, corresponding timestamp information, flight identification information, UAV pose parameter information, and geospatial location information of the inspected object collected by the UAV during the inspection process.

[0009] Optionally, the preprocessing of the UAV inspection data includes data format unification processing, abnormal data removal processing, timestamp correction processing, time synchronization processing of multi-flight data, spatial registration processing based on UAV pose parameters, and spatial alignment processing based on a unified coordinate system to generate a standardized inspection dataset.

[0010] Optionally, generating a cross-ferry consistency feature map and aggregating it into a cross-ferry association feature set of the inspection object includes: The image or video frames corresponding to the same inspection object in each voyage in the standardized inspection dataset are input into the improved VMamba network. The improved VMamba network adds a voyage embedding layer and a spatial absolute position encoding layer at the input end. The voyage embedding layer maps the voyage identifier into an embedding vector and concatenates it with the input features. The spatial absolute position encoding layer generates a position code based on the pixel row and column coordinates and adds it to the input features by channel to obtain an initial feature stream containing voyage information and spatial information. The initial feature stream is fed into the VMamba main structure. Frequency domain-spatial domain coupling enhancement modules are inserted after the first selective scan coding block and the third selective scan coding block, respectively. The frequency domain-spatial domain coupling enhancement module first performs fast discrete Fourier transform on the local feature block to extract the frequency domain texture component. The defect texture with periodic or semi-periodic features is enhanced by a learnable bandpass filter layer and then fused with the corresponding spatial features through residual method. During the encoding process, the cross-ferry spatiotemporal alignment module is called. The cross-ferry spatiotemporal alignment module consists of an affine alignment layer based on pose parameters and a dynamic offset correction layer based on depth features, which performs spatial registration and displacement residual correction on feature streams from different ferry sources. The multi-ferry feature stream, processed by the cross-ferry spatiotemporal alignment module, is input into the consistency aggregation module. The consistency aggregation module adopts a credibility-weighted fusion strategy to perform channel-by-channel weighted accumulation of the feature maps of each voyage for the same inspection object, and outputs a cross-ferry consistent feature map. The instance re-identification clustering unit is invoked on the cross-voyage consistency feature map. The instance re-identification clustering unit calculates the similarity of the feature vectors of the same inspection object based on the global average convergent feature and the cosine distance metric. The cross-voyage associated feature set of the inspection object is generated with the similarity threshold as the aggregation criterion.

[0011] Optionally, the generated inspection feature structure includes: Object-level trajectories are established in chronological order for cross-flight associated feature sets. Missing time phases are interpolated and aligned based on flight identifiers and pose parameters. Low-quality frames are removed based on frame quality scores to obtain the time series. The multi-temporal change analysis module is called on the time series. The multi-temporal change analysis module consists of appearance change branch, geometric change branch and position change branch in parallel. It calculates the texture and edge difference, instance outline and area difference, and projection position and orientation difference for adjacent flight pairs within the sliding time window. It uses pose parameters to generate a disparity suppression map to cancel the pseudo changes caused by the viewpoint change and outputs the change records between each pair of flights. The change records are subjected to a persistence determination. They are filtered according to the preset time window length, minimum effective area threshold, position drift threshold and duration threshold. The persistent change set and the instantaneous change set are obtained by cross-flight consistency weight aggregation. The forward cumulative change and the backward cumulative change are compared by closed-loop consistency check. Records that do not meet the closed-loop consistency are removed. Baseline samples corresponding to the target object or object category are extracted from historical inspection data to generate phased normal evolution paths and rate boundaries. A dual-anchor alignment difference strategy is used to compare the current time series with the previous time phase state and the normal evolution path simultaneously to generate deviation information consisting of deviation magnitude, deviation direction and deviation duration. By binding appearance information, spatial information, evolution information and deviation information at the field level, an inspection feature structure is formed.

[0012] Optionally, the extraction of the causal convergence subgraph includes: The causal inspection evidence graph is initialized with the inspection feature structure as the inspection evidence node. For each inspection evidence node, the object identifier, time index, flight index, pose index, appearance information field, spatial information field, evolution information field and deviation information field are written, and a unique index key is generated for each inspection evidence node. Under the constraints of preset time window length, spatial distance threshold, and similarity threshold, the candidate inspection evidence node pairs are subjected to relationship determination, generating three types of relationship edges: The causal consistency relationship is established based on the consistency of direction and type of deviation information. Evolutionary reversal relationships are established when the trend direction of evolutionary information reverses and the magnitude exceeds a threshold. The counterfactual convergence relation edges are established when the gap is below the threshold and the deviation converges to the same root cause entry after alignment according to the normal evolution path; The established relationship edges are weighted and processed. The weights are obtained by weighting and fusing the comprehensive correlation of the inspection evidence nodes in terms of temporal continuity, spatial consistency, feature similarity and change stability. The weights are then normalized according to a unified scale, and relationship edges below a preset threshold are removed. Perform causal propagation and path convergence analysis on the preserved relation edges, iteratively update the path score and coverage according to the maximum propagation rounds or convergence halting conditions, prune the coverage or broken paths below the threshold, and aggregate the path set that meets the coverage threshold and path connectivity to obtain the candidate convergence path cluster. Select a set of closure nodes with candidate convergence path clusters as the core, extract the causal convergence subgraph, record the inspection evidence node list, relation edge list and edge weight list in the causal convergence subgraph, and generate a mapping index from the subgraph to the original inspection feature structure.

[0013] Optionally, the step of generating an initial inspection report and sequentially performing structural reverse engineering and temporal inverse verification includes: Within the inspection report inversion generator, an evidence planning layer, a bidirectional consistency layer, and a reconstructed audit layer are established sequentially, with the layers connected through status markers and index mappings. Input the causal convergence subgraph into the evidence planning layer, call the chapter skeleton planning unit to generate chapter, paragraph and sentence sequence, call the evidence pointer table generation unit to build the evidence pointer table based on the unique index identifier of the inspection evidence node, synthesize the initial inspection report and output the report-evidence mapping index; The initial inspection report is input into the bidirectional consistency layer. The structure of the report content is reversed to form an element table of abnormal items, location items and change items. Each element item is mapped back to the corresponding inspection evidence node according to the report-evidence mapping index. Forward derivable verification and reverse traceable verification are completed in sequence. Verification marks are added to sentences and paragraphs that fail the verification and missing coverage dimensions are recorded. The report with verification marks is input into the reconstructed audit layer. Based on the appearance, space, evolution and deviation information of the corresponding evidence node, the marked segments are rewritten. The quota control length is described according to the anomaly type and the missing dimensions are filled in. After writing the evidence citation number, the revised draft is generated iteratively according to the set rounds. The revised draft is returned to the bidirectional consistency layer for review. For content that fails the forward derivation check or the reverse backtracking check, reconstruction continues until all sentences and paragraphs pass the check or reach the preset reconstruction round.

[0014] Optionally, the step of associating and storing the final drone inspection report with the corresponding inspection evidence nodes includes: After the inspection report inversion generator completes the structural inverse solution processing and time-series reverse verification, the completeness of the inspection report content that has passed the verification is checked to confirm that the inspection report includes anomaly description, spatial location description and change trend description for each abnormal object. The inspection report that has completed the integrity check is indexed and bound to the corresponding causal convergence subgraph, a one-to-one correspondence between the inspection report paragraphs and the inspection evidence nodes is established, and a report-evidence association index table is generated. Based on the report-evidence association index table, the final inspection report content, the corresponding inspection evidence node identifier, time index, flight index, and pose index are jointly encapsulated to form a traceable inspection report data package. The inspection report data packet is written into the storage system, and the report-evidence association index table is stored synchronously to complete the association storage and traceability query between the final UAV inspection report and the inspection evidence nodes.

[0015] The beneficial effects of this invention are: This invention performs unified preprocessing and feature modeling on UAV multi-flight inspection data, and utilizes an improved feature extraction network to achieve spatiotemporal alignment and consistent representation between inspection data from different flights. This enables the appearance, spatial location, and time-varying changes of the inspected objects to be uniformly represented in a structured form, thereby effectively overcoming the problems of scattered, difficult-to-associate, and unstable change characteristics of inspection data in existing technologies, and significantly improving the overall usability and analytical reliability of UAV inspection data.

[0016] This invention introduces a causal inspection evidence graph to organize and converge the causal relationships of inspection feature structures. This allows multi-source anomaly information generated during the inspection process to aggregate around potential root causes, reducing the risk of misjudgment from single-frame or single-flight assessments and improving the stability and consistency of anomaly identification. Through the extraction of the causal convergence subgraph, the formation path and correlation of inspection anomalies are clearly expressed, providing reliable structured evidence support for inspection result analysis and decision-making.

[0017] This invention further establishes an inspection report inversion generator that performs structural inversion and temporal reverse verification on the report content while automatically generating the inspection report. This enables the anomaly descriptions, location descriptions, and change descriptions in the inspection report to automatically align with the underlying inspection evidence and possess traceability. This effectively avoids the problems of existing technologies where inspection reports rely on manual editing, are highly subjective, and are difficult to verify. It automates, standardizes, and verifies the inspection report generation process, thereby improving the intelligence level and application value of the UAV inspection system. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the intelligent generation method for UAV inspection reports based on data feature extraction proposed in this invention; Figure 2This is a schematic diagram of the cross-flight feature extraction and consistency feature construction process of the improved VMamba network for the intelligent generation method of UAV inspection reports based on data feature extraction proposed in this invention. Figure 3 This is a schematic diagram of the workflow of the inspection report inversion generator of the UAV inspection report intelligent generation method based on data feature extraction proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A method for intelligent generation of drone inspection reports based on data feature extraction includes: The drone inspection platform collects drone inspection data from multiple flights, performs preprocessing on the drone inspection data, and generates a standardized inspection dataset. On the standardized inspection dataset, dual-scale selective scanning feature encoding is performed using an improved VMamba network. Spatiotemporal alignment and residual correction are performed on the feature streams of different voyages of the same inspection object, generating cross-voyage consistent feature maps and converging them into a cross-voyage associated feature set of the inspection object. Multi-temporal change analysis is performed on the feature set of cross-flight inspection objects to extract the state change information of the inspection objects, and the normal evolution trajectory is deduced by combining historical inspection data. The deviation between the actual change results and the normal evolution trajectory is calculated to generate the inspection feature structure. A causal inspection evidence graph is constructed using the inspection feature structure as evidence nodes. Causal consistency, evolutionary reversal and counterfactual convergence relationships are established between the evidence nodes, and causal convergence subgraphs are extracted. The causal convergence subgraph is input into the inspection report inversion generator to generate an initial inspection report. The structural inverse solution processing and time-series back-inference verification are then executed sequentially. The anomalies, locations and changes in the initial inspection report are mapped back to the corresponding evidence nodes. Content that does not meet the forward derivation and backtracking conditions is reconstructed. After the inspection report inversion generator completes the structural inverse solution processing and time sequence inverse verification, it outputs the final inspection report and associates and stores the final UAV inspection report with the corresponding inspection evidence nodes.

[0021] In this embodiment, the UAV inspection data includes visible light image data or video data, infrared thermal imaging data, corresponding timestamp information, flight identification information, UAV pose parameter information, and geospatial location information of the inspected object collected by the UAV during the inspection process.

[0022] In this embodiment, the preprocessing of UAV inspection data includes data format unification processing, abnormal data removal processing, timestamp correction processing, time synchronization processing of multi-flight data, spatial registration processing based on UAV pose parameters, and spatial alignment processing based on a unified coordinate system to generate a standardized inspection dataset.

[0023] In this embodiment, generating a cross-ferry consistency feature map and aggregating it into a cross-ferry association feature set of the inspection object includes: The image or video frames corresponding to the same inspection object in each voyage from the standardized inspection dataset are input into an improved VMamba network. This improved VMamba network adds a voyage embedding layer and a spatial absolute position encoding layer at the input. The voyage embedding layer maps the voyage identifier to an embedding vector and concatenates it with the input features. The spatial absolute position encoding layer generates position codes based on pixel row and column coordinates and adds them to the input features by channel, resulting in an initial feature stream containing voyage information and spatial information. The voyage embedding layer maps voyage identifiers to embedding vectors and concatenates them with the input features, specifically: Each flight identifier in the UAV inspection mission is converted into a unique discrete index, and a trainable embedding matrix is ​​constructed for all flight indices during VMamba network initialization. Before feature encoding, the cruise index corresponding to the current input image is retrieved into the embedding matrix to obtain the cruise embedding vector with the same dimension as the backbone feature channel. The obtained cruise embedding vector is concatenated with the original feature tensor of the same image along the channel dimension; The spatial absolute position encoding layer generates position codes based on pixel row and column coordinates and adds them to the input features by channel, specifically: The row coordinate encoding vector and column coordinate encoding vector are constructed for the input image, and then combined into a two-dimensional position encoding matrix with the same size as the image by broadcasting. Copy or linearly map the position encoding matrix along the channel dimension to an encoding tensor with the same number of channels as the input feature tensor; The copied position encoding tensor and the corresponding input feature tensor are added element by element at the same spatial location, incorporating the absolute spatial location information into the input feature; The initial feature stream is fed into the VMamba main structure. Frequency-spatial coupling enhancement modules are inserted after the first and third selective scan coding blocks, respectively. These modules first perform a fast discrete Fourier transform on the local feature blocks to extract frequency-domain texture components. These components are then enhanced with a learnable bandpass filter layer to improve the defect texture with periodic or semi-periodic features. Finally, they are fused with the corresponding spatial features using a residual method. Specifically, the enhancement of the defect texture with periodic or semi-periodic features through the learnable bandpass filter layer is as follows: For each local frequency domain feature block, a trainable center frequency and bandwidth parameter are set, and the bandpass filter layer determines the frequency band to be retained based on the parameter combination, suppressing irrelevant high-frequency noise and low-frequency background components. The amplitude spectrum within the retained frequency band is amplified according to a learnable gain coefficient to highlight texture signals such as crack stripes and corrosion particles that are distributed periodically or semi-periodically. The amplified frequency domain features are restored to spatial domain features by inverse discrete Fourier transform, and then added to the original spatial domain features in the form of residuals to achieve explicit enhancement of the defect texture without destroying the overall feature distribution. During the encoding process, a cross-ferry spatiotemporal alignment module is invoked. This module consists of a pose-parameter-based affine alignment layer and a depth-feature-based dynamic offset correction layer connected in series. It performs spatial registration and displacement residual correction on feature streams from different voyages. Specifically, the spatial registration and displacement residual correction on feature streams from different voyages are performed as follows: The affine alignment layer calculates the inter-flight transformation matrix based on the pose parameters recorded by the UAV in each flight, and performs affine transformation on the feature flow to be aligned according to the transformation matrix so that the same inspection object obtains a consistent spatial projection in the reference coordinate system. The dynamic offset correction layer takes the affine-aligned feature stream and the reference flight feature stream as inputs, generates a pixel-by-pixel displacement deviation field through a trainable offset prediction sub-network, and performs multi-channel sampling correction on the feature stream based on the deviation field to eliminate residual displacement errors caused by local parallax, minor flight jitter and optical distortion. The feature stream after dynamic offset correction is compared with the feature stream of the reference voyage. For areas where there is still local misalignment, iterative refinement correction is triggered until the overall displacement error is lower than the preset threshold, thus completing the spatial registration and displacement residual correction of the cross-voyage feature stream. The multi-ferry feature streams processed by the cross-ferry spatiotemporal alignment module are input into the consistency aggregation module. The consistency aggregation module adopts a credibility-weighted fusion strategy to perform channel-by-channel weighted accumulation of the feature maps of each voyage for the same inspection object, and outputs a cross-ferry consistent feature map. The credibility-weighted fusion strategy is as follows: For the same inspection object, the quality score is estimated for the feature map corresponding to each voyage. The quality score integrates four indicators: image sharpness, exposure balance, pose stability and cross-voyage alignment residual. The four indicators are linearly merged to obtain a single confidence coefficient. The credibility coefficients of all voyages are normalized so that the sum of the coefficients of each voyage is equal to one, and the normalized coefficients are used as weights to be assigned to each channel of the corresponding feature map. The feature maps of the same inspection object in each voyage are weighted and accumulated according to the channel to obtain a consistent feature map across voyages, and the edges of the fusion result are smoothed to reduce local abrupt noise. The instance re-identification clustering unit is invoked on the cross-voyage consistency feature map. The instance re-identification clustering unit calculates the similarity of the feature vectors of the same inspection object based on the global average convergent feature and the cosine distance metric. The cross-voyage associated feature set of the inspection object is generated with the similarity threshold as the aggregation criterion.

[0024] This invention constructs a five-level architecture based on the original VMamba state-space visual backbone: cruise location enhancement, dual-scale encoding, frequency-space coupling enhancement, cross-cruise alignment, and consistency aggregation. At the input end, cruise identity and pixel coordinates are injected into the feature stream through a cruise embedding layer and a spatial absolute position encoding layer. The backbone maintains VMamba dual-scale selective scanning encoding blocks to capture long-range dependencies. A frequency-space coupling enhancement module is inserted after the first and third encoding blocks to enhance periodic and semi-periodic defect textures through learnable bandpass filtering. In the encoding stage, an affine alignment layer and a dynamic offset correction layer are connected in series to form a cross-cruise spatiotemporal alignment module, utilizing pose parameters and depth features to achieve spatial registration and residual correction. At the tail end, a credibility-weighted consistency aggregation module is set up to weight and accumulate multi-cruise features channel-by-channel, outputting a cross-cruise consistent feature map. Instance re-identification clustering forms a cross-cruise associated feature set of inspection objects, thus completing four functions—feature injection, fine-grained enhancement, cross-cruise alignment, and credibility aggregation—within the same network.

[0025] In this embodiment, the generation of the inspection feature structure includes: Object-level trajectories are established in chronological order for cross-flight associated feature sets. Missing time phases are interpolated and aligned based on flight identifiers and pose parameters. Low-quality frames are removed based on frame quality scores to obtain the time series. The multi-temporal change analysis module is invoked on the time series. This module consists of parallel branches for appearance change, geometric change, and position change. It calculates texture and edge differences, instance contour and area differences, and projection position and orientation differences for adjacent flight pairs within the sliding time window. It uses pose parameters to generate a disparity suppression map to cancel out pseudo-changes caused by viewpoint variations, and outputs the change records between each pair of flights. Specifically, the calculation of texture and edge differences, instance contour and area differences, and projection position and orientation differences for adjacent flight pairs within the sliding time window is as follows: In the appearance change branch, the corresponding regions of the feature maps of the two voyages are converted into grayscale texture vectors, the difference of the local binary mode histogram is calculated and the change of the number of Sobel edge pixels is measured to obtain the difference between texture and edge. In the geometric transformation branch, the equivalent contour is extracted based on the instance mask generated by the cross-flight consistency feature map. The percentage change of contour length and mask area is compared to obtain the difference between instance contour and area. In the position change branch, the instance centroid is projected onto a unified reference plane using the pose parameters of the two flights, the Euclidean distance and orientation angle difference of the projected coordinates are calculated, and the displacement caused by the difference in viewpoint is canceled out by the disparity suppression map, and the change record composed of three difference measures is output. Persistence determination is performed on change records, filtering them according to preset time window length, minimum effective area threshold, position drift threshold, and duration threshold. Persistent change sets and instantaneous change sets are obtained based on cross-flight consistency weight aggregation. A closed-loop consistency check is used to compare forward cumulative changes with backward cumulative changes; records that do not meet the closed-loop consistency requirement are removed. Specifically: The preset time window length is set to 3 months, the minimum effective area threshold is set to 80 pixels, the position drift threshold is set to 9 pixels, and the duration threshold is set to 6 weeks. The closed-loop consistency check compares the forward cumulative changes with the backward cumulative changes, specifically as follows: The changes in each voyage are accumulated from the starting point of the time series to the current voyage, forming a forward cumulative change curve; The backward cumulative change curve is generated by accumulating backward from the end of the time series to the current voyage, and the time order of this curve is reversed to align with the forward curve. The difference between the change amplitude and direction of the two curves at the same time node is compared. When the difference exceeds the preset consistency threshold or the directions are opposite, it is judged as closed-loop inconsistency. Remove inconsistent change records from the closed loop and mark the remaining records as consistent in the closed loop. Baseline samples corresponding to the target object or object category are extracted from historical inspection data to generate phased normal evolution paths and rate boundaries. A dual-anchor alignment differencing strategy is used to compare the current time series with both the previous time phase state and the normal evolution path, generating deviation information consisting of deviation magnitude, deviation direction, and deviation duration. Specifically, the dual-anchor alignment differencing strategy involves comparing the current time series with both the previous time phase state and the normal evolution path. Using the object characteristics of the current voyage as the alignment benchmark, and selecting the characteristic state of the same object in the previous voyage as the first anchor point, the two are subjected to first-order difference to obtain short-period changes, which are used to reflect recent fluctuations. Align the current cruise object features with the normal evolution path generated from historical baseline samples at the same time point, use it as the second anchor point, perform difference on the two to obtain the long-term deviation; The short-cycle changes and long-term deviations are accumulated by sliding over a time window. Three indicators are calculated: the change magnitude threshold, the consistency of direction, and the duration. The three indicators are then combined to generate the deviation magnitude, the deviation direction, and the deviation duration. By binding appearance information, spatial information, evolution information and deviation information at the field level, an inspection feature structure is formed.

[0026] In this embodiment, the extraction of the causal convergence subgraph includes: The causal inspection evidence graph is initialized with the inspection feature structure as the inspection evidence node. For each inspection evidence node, the object identifier, time index, flight index, pose index, appearance information field, spatial information field, evolution information field and deviation information field are written, and a unique index key is generated for each inspection evidence node. Under the constraints of preset time window length, spatial distance threshold, and similarity threshold, the candidate inspection evidence node pairs are subjected to relationship determination, generating three types of relationship edges: The causal consistency relationship is established based on the consistency of direction and type of deviation information. Evolutionary reversal relationships are established when the trend direction of evolutionary information reverses and the magnitude exceeds a threshold. The counterfactual convergence relation edges are established when the difference is lower than the similarity threshold and the deviation converges to the same root cause entry after alignment according to the normal evolution path; The relationship determination for candidate inspection evidence node pairs is as follows: In the time series, only node pairs with a time interval of no more than 90 days and a spatial distance of less than 0.4 meters are retained as comparable pairs. The similarity of appearance, location and evolution features is calculated for comparable node pairs. If the comprehensive similarity is higher than 0.75, the next step of comparison is carried out. The three rules of deviation information direction and type, evolution trend direction and magnitude, and residual after alignment with normal evolution path are checked in turn. When the direction and type are consistent, it is marked as a causal consistent relationship edge. When the evolution trend is reversed and the magnitude exceeds the threshold, it is marked as an evolution reversal relationship edge. When the residual is lower than the threshold and the deviation converges to the same root cause entry, it is marked as a counterfactual convergence relationship edge. The preset time window length is 90 days, the spatial distance threshold is 0.4 meters, and the similarity threshold is 0.75. The established relationship edges are weighted and assigned values. These weights are obtained by weighting and fusing the correlation between the inspection evidence nodes in terms of temporal continuity, spatial consistency, feature similarity, and stability of change. The weights are then normalized according to a unified standard, and relationship edges below a preset threshold are removed. Temporal continuity is obtained by calculating the ratio of the time interval between node pairs to the length of a preset time window; the shorter the interval, the higher the score. Spatial consistency is determined based on the ratio of the Euclidean distance between node pairs in a unified coordinate system to a spatial distance threshold; the smaller the distance, the higher the score. Feature similarity is measured by the combined cosine similarity of node pairs across three categories of features: appearance, location, and evolution. Higher similarity results in higher scores. Change stability is measured by the range of deviation fluctuations at statistical nodes within a sliding time window; the smaller the fluctuation, the higher the score. Perform causal propagation and path convergence analysis on the preserved relation edges, iteratively update the path score and coverage according to the maximum propagation rounds or convergence halting conditions, prune the coverage or broken paths below the threshold, and aggregate the path set that meets the coverage threshold and path connectivity to obtain the candidate convergence path cluster. Select a set of closure nodes with candidate convergence path clusters as the core, extract the causal convergence subgraph, record the inspection evidence node list, relation edge list and edge weight list in the causal convergence subgraph, and generate a mapping index from the subgraph to the original inspection feature structure.

[0027] In this embodiment, the step of generating an initial inspection report and sequentially performing structural reverse engineering and temporal inverse verification includes: Within the inspection report inversion generator, an evidence planning layer, a bidirectional consistency layer, and a reconstructed audit layer are established sequentially, with the layers connected through status markers and index mappings. The causal convergence subgraph is input into the evidence planning layer. The chapter skeleton planning unit is called to generate chapter, paragraph, and sentence sequence. The evidence pointer table generation unit is called to build the evidence pointer table based on the unique index identifier of the inspection evidence node. The initial inspection report is synthesized and the report-evidence mapping index is output. Specifically, the chapter skeleton planning unit generates chapter, paragraph, and sentence sequence as follows: Based on the anomaly type, spatial location level, and temporal evolution stage of the inspection evidence nodes in the causal convergence subgraph, the corresponding chapter title and paragraph structure templates are obtained by searching the pre-set template library. Following the order of general overview—anomaly summary—object-by-object details—risk assessment—handling recommendations, the chapter and paragraph templates are sorted and placeholders for binding evidence nodes are inserted within the paragraphs to form a hierarchical skeleton framework; Match a set of applicable sentence templates for each placeholder in the skeleton framework, and generate a sentence sequence index table based on the field labels of the evidence nodes; The initial inspection report is input into the bidirectional consistency layer. A structural reverse analysis is performed on the report content to generate an element table of anomaly entries, location entries, and change entries. Each element entry is mapped back to its corresponding inspection evidence node according to the report-evidence mapping index. Forward derivability verification and reverse traceability verification are performed sequentially. Verification tags are added to sentences and paragraphs that fail verification, and missing coverage dimensions are recorded. Specifically, the structural reverse analysis of the report content involves: The initial inspection report is segmented into sentences and paragraphs. Based on the placeholders in the chapter skeleton, abnormal location words, component name words, change quantifiers and time words are extracted and the words are labeled as abnormal tags, location coordinate tags, change range tags and time tags. Based on the principle of merging tags of the same sentence, tags belonging to the same abnormal object are aggregated into one abnormal element record. The correspondence between location description and change description is analyzed by combining the inter-sentence dependency relationship, and location element record and change element record are generated respectively. The object number, spatial level, time index, abnormal type, change amount and change direction fields are written into the element table. Using the report-evidence mapping index, each record in the feature table is mapped one by one, and associated with the corresponding inspection evidence node in the causal convergence subgraph. The sentence position index is retained for subsequent consistency verification. The report with verification marks is input into the reconstructed audit layer. Based on the appearance, space, evolution and deviation information of the corresponding evidence node, the marked segments are rewritten. The quota control length is described according to the anomaly type and the missing dimensions are filled in. After writing the evidence citation number, the revised draft is generated iteratively according to the set rounds. The revised draft is returned to the bidirectional consistency layer for review. For content that fails the forward derivation check or the reverse backtracking check, reconstruction continues until all sentences and paragraphs pass the check or reach the preset reconstruction round.

[0028] In this embodiment, the step of associating and storing the final UAV inspection report with the corresponding inspection evidence nodes includes: After the inspection report inversion generator completes the structural inverse solution processing and time-series reverse verification, the completeness of the inspection report content that has passed the verification is checked to confirm that the inspection report includes anomaly description, spatial location description and change trend description for each abnormal object. The inspection report that has completed the integrity check is indexed and bound to the corresponding causal convergence subgraph, a one-to-one correspondence between the inspection report paragraphs and the inspection evidence nodes is established, and a report-evidence association index table is generated. Based on the report-evidence association index table, the final inspection report content, the corresponding inspection evidence node identifier, time index, flight index, and pose index are jointly encapsulated to form a traceable inspection report data package. The inspection report data packet is written into the storage system, and the report-evidence association index table is stored synchronously to complete the association storage and traceability query between the final UAV inspection report and the inspection evidence nodes.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the unmanned aerial vehicle (UAV) inspection of a 110kV overhead transmission line in a certain region. This line, which has been in operation for over twelve years, is approximately 24 kilometers long and has 45 towers. Some of these towers are located in an area bordering hills and farmland, where seasonal wind and rain, diurnal temperature variations, and prolonged sunlight have led to the accumulation of deposits on the insulator surfaces, corrosion of hardware, and localized aging issues year after year. The maintenance unit typically uses UAVs to conduct periodic inspections of this line every six months, and the inspection images serve as an important basis for subsequent maintenance decisions.

[0030] In traditional inspection processes, after a drone completes its flight mission, inspection personnel manually review the images captured that day and compare them with historical images to determine if any abnormal changes exist. Due to differences in flight altitude, shooting angle, and lighting conditions between different flights, the appearance of the same component can vary significantly at different times, making it difficult for manual comparison to reliably distinguish between actual structural changes and changes in shooting conditions. Inspection reports rely heavily on human experience, lacking a clear correspondence between report content and specific inspection images. Anomaly descriptions are difficult to trace back to concrete evidence, hindering review and accountability.

[0031] In this embodiment, the operation and maintenance unit incorporated drone inspection data from September 2024 and March 2025 into the intelligent drone inspection report generation method based on data feature extraction proposed in this invention for processing. Both inspections utilized the same model of drone and visible light imaging equipment, with an average of approximately 4,200 images collected per inspection, along with corresponding timestamps, flight identifiers, and drone pose parameters. All inspection data first entered a preprocessing flow, where the image data underwent format standardization, time synchronization, and spatial alignment based on pose parameters to generate a standardized inspection dataset.

[0032] Based on a standardized inspection dataset, the system employs an improved VMamba network for feature encoding of inspection images. By introducing cruise embedding and spatial absolute position encoding at the input stage, and performing dual-scale selective scanning and frequency-spatial coupling enhancement during the backbone encoding process, features of the same inspection object across different cruises can be aligned under a unified spatial semantics. Subsequently, the system performs spatiotemporal alignment and residual correction on the feature streams from multiple cruises belonging to the same inspection object, generating a cross-cruise consistent feature map and forming a cross-cruise associated feature set for the inspection object.

[0033] The system performs multi-temporal change analysis on the feature set of cross-flight inspection objects. By jointly analyzing changes in component surface texture, geometric contour, and spatial stability, and combining historical inspection data to deduce the evolution trend of corresponding components under normal operating conditions, the system can identify persistent abnormal changes and write the deviation information between these changes and the normal evolution trajectory into the inspection feature structure. Using the inspection feature structure as evidence nodes, the system constructs a causal inspection evidence graph. Through the establishment of causal consistency relationships, evolutionary reversal relationships, and counterfactual convergence relationships, it extracts a causal convergence subgraph, which serves as the core evidence set for inspection anomalies along the route.

[0034] During the inspection report generation phase, a causal convergence subgraph is input into the inspection report inversion generator. The system first generates an initial inspection report and performs structural inverse decomposition and temporal inverse verification on the report content, mapping the anomaly descriptions, location descriptions, and change descriptions in the report back to the corresponding inspection evidence nodes. For descriptions that cannot be verified through forward derivation or backward backtracking, the system automatically reconstructs them based on the appearance information, spatial information, evolution information, and deviation information contained in the corresponding inspection evidence nodes until all descriptions pass the consistency verification. Finally, the system outputs the inspection report and establishes an association between the report content and the inspection evidence nodes, achieving a traceable correspondence between the inspection report and the inspection data.

[0035] Table 1. Comparison of UAV Inspection Analysis and Report Generation Results As shown in Table 1, under the same number of inspection flights and the same number of inspection images, different inspection analysis methods exhibit significant differences in anomaly identification accuracy and report generation efficiency. The traditional manual method identified the most suspected anomalies in the initial inspection stage, reaching nine, but only seven of these were confirmed as genuine anomalies during subsequent on-site verification. This indicates that manual interpretation is more prone to misjudgment due to the influence of subjective experience. The semi-automatic analysis method reduced the number of suspected anomalies to some extent, but still resulted in one misjudgment. The method of this invention, while maintaining the same number of genuine anomalies, further reduced the number of suspected anomalies to seven, and the number of anomalies confirmed during verification was completely consistent with the initial judgment, demonstrating stability and consistency in anomaly identification.

[0036] The differences in inspection report generation efficiency among the three methods are also significant. Traditional manual methods take an average of about 160 minutes to complete a full inspection report. Semi-automatic analysis methods, through template-based and partially automated processing, reduce this time to about 70 minutes, but still require considerable manual intervention. In contrast, the method of this invention can output a complete inspection report in only about 22 minutes after the inspection data processing is complete, reducing manual intervention time, improving the overall efficiency of inspection operations, and avoiding the time wasted on repetitive manual verification and organization.

[0037] From the perspective of the traceability of report results, traditional manual inspection reports struggle to establish a clear correspondence between anomaly descriptions and specific inspection images or flights. Subsequent verification and tracking rely on manual retrieval of the original data. Semi-automatic analysis methods improve this issue to some extent, but the correlation remains incomplete. The method of this invention links the inspection report content with inspection evidence nodes, enabling each anomaly in the report to be traced back to the corresponding inspection data and flight, thus improving the reliability and verifiability of inspection results and providing more credible data support for subsequent operation and maintenance decisions.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent generation of UAV inspection reports based on data feature extraction, characterized in that, include: The drone inspection platform collects drone inspection data from multiple flights, performs preprocessing on the drone inspection data, and generates a standardized inspection dataset. On the standardized inspection dataset, dual-scale selective scanning feature encoding is performed using an improved VMamba network. Spatiotemporal alignment and residual correction are performed on the feature streams of different voyages of the same inspection object, generating cross-voyage consistent feature maps and converging them into a cross-voyage associated feature set of the inspection object. Multi-temporal change analysis is performed on the feature set of cross-flight inspection objects to extract the state change information of the inspection objects, and the normal evolution trajectory is deduced by combining historical inspection data. The deviation between the actual change results and the normal evolution trajectory is calculated to generate the inspection feature structure. A causal inspection evidence graph is constructed using the inspection feature structure as evidence nodes. Causal consistency, evolutionary reversal and counterfactual convergence relationships are established between the evidence nodes, and causal convergence subgraphs are extracted. The causal convergence subgraph is input into the inspection report inversion generator to generate an initial inspection report. The structural inverse solution processing and time-series back-inference verification are then executed sequentially. The anomalies, locations and changes in the initial inspection report are mapped back to the corresponding evidence nodes. Content that does not meet the forward derivation and backtracking conditions is reconstructed. After the inspection report inversion generator completes the structural inverse solution processing and time sequence inverse verification, it outputs the final inspection report and associates and stores the final UAV inspection report with the corresponding inspection evidence nodes.

2. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The UAV inspection data includes visible light image data or video data, infrared thermal imaging data, corresponding timestamp information, flight identification information, UAV pose parameter information, and geospatial location information of the inspected object collected by the UAV during the inspection process.

3. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The preprocessing of UAV inspection data includes standardizing the data format, removing abnormal data, correcting timestamps, synchronizing the time of data from multiple flights, spatial registration based on UAV pose parameters, and spatial alignment based on a unified coordinate system, to generate a standardized inspection dataset.

4. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The process of generating cross-voyage consistency feature maps and aggregating them into a cross-voyage association feature set for the inspection objects includes: The image or video frames corresponding to the same inspection object in each voyage in the standardized inspection dataset are input into the improved VMamba network. The improved VMamba network adds a voyage embedding layer and a spatial absolute position encoding layer at the input end. The voyage embedding layer maps the voyage identifier into an embedding vector and concatenates it with the input features. The spatial absolute position encoding layer generates a position code based on the pixel row and column coordinates and adds it to the input features by channel to obtain an initial feature stream containing voyage information and spatial information. The initial feature stream is fed into the VMamba main structure. Frequency domain-spatial domain coupling enhancement modules are inserted after the first selective scan coding block and the third selective scan coding block, respectively. The frequency domain-spatial domain coupling enhancement module first performs fast discrete Fourier transform on the local feature block to extract the frequency domain texture component. The defect texture with periodic or semi-periodic features is enhanced by a learnable bandpass filter layer and then fused with the corresponding spatial features through residual method. During the encoding process, the cross-ferry spatiotemporal alignment module is invoked. The cross-ferry spatiotemporal alignment module consists of an affine alignment layer based on pose parameters and a dynamic offset correction layer based on depth features, which performs spatial registration and displacement residual correction on feature streams from different ferry sources. The multi-ferry feature streams processed by the cross-ferry spatiotemporal alignment module are input into the consistency aggregation module. The consistency aggregation module adopts a credibility-weighted fusion strategy to perform channel-by-channel weighted accumulation of the feature maps of each voyage for the same inspection object and outputs a cross-ferry consistent feature map. The instance re-identification clustering unit is invoked on the cross-voyage consistency feature map. The instance re-identification clustering unit calculates the similarity of the feature vectors of the same inspection object based on the global average convergent feature and the cosine distance metric. The cross-voyage associated feature set of the inspection object is generated with the similarity threshold as the aggregation criterion.

5. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The generated inspection feature structure includes: Object-level trajectories are established in chronological order for cross-flight associated feature sets. Missing time phases are interpolated and aligned based on flight identifiers and pose parameters. Low-quality frames are removed based on frame quality scores to obtain the time series. The multi-temporal change analysis module is called on the time series. The multi-temporal change analysis module consists of appearance change branch, geometric change branch and position change branch in parallel. It calculates the texture and edge difference, instance outline and area difference, and projection position and orientation difference for adjacent flight pairs within the sliding time window. It uses pose parameters to generate a disparity suppression map to cancel the pseudo changes caused by the viewpoint change and outputs the change records between each pair of flights. The change records are subjected to a persistence determination. They are filtered according to the preset time window length, minimum effective area threshold, position drift threshold and duration threshold. The persistent change set and the instantaneous change set are obtained by cross-flight consistency weight aggregation. The forward cumulative change and the backward cumulative change are compared by closed-loop consistency check. Records that do not meet the closed-loop consistency are removed. Baseline samples corresponding to the target object or object category are extracted from historical inspection data to generate phased normal evolution paths and rate boundaries. A dual-anchor alignment difference strategy is used to compare the current time series with the previous time phase state and the normal evolution path simultaneously to generate deviation information consisting of deviation magnitude, deviation direction and deviation duration. By binding appearance information, spatial information, evolution information and deviation information at the field level, an inspection feature structure is formed.

6. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The extraction of the causal convergence subgraph includes: The causal inspection evidence graph is initialized with the inspection feature structure as the inspection evidence node. For each inspection evidence node, the object identifier, time index, flight index, pose index, appearance information field, spatial information field, evolution information field and deviation information field are written, and a unique index key is generated for each inspection evidence node. Under the constraints of preset time window length, spatial distance threshold, and similarity threshold, the candidate inspection evidence node pairs are subjected to relationship determination, generating three types of relationship edges: The causal consistency relationship is established based on the consistency of direction and type of deviation information. Evolutionary reversal relationships are established when the trend direction of evolutionary information reverses and the magnitude exceeds a threshold. The counterfactual convergence relation edges are established when the gap is below the threshold and the deviation converges to the same root cause entry after alignment according to the normal evolution path; The established relationship edges are weighted and processed. The weights are obtained by weighting and fusing the comprehensive correlation of the inspection evidence nodes in terms of temporal continuity, spatial consistency, feature similarity and change stability. The weights are then normalized according to a unified scale, and relationship edges below a preset threshold are removed. Perform causal propagation and path convergence analysis on the preserved relation edges, iteratively update the path score and coverage according to the maximum propagation rounds or convergence halting conditions, prune the coverage or broken paths below the threshold, and aggregate the path set that meets the coverage threshold and path connectivity to obtain the candidate convergence path cluster. Select a set of closure nodes with candidate convergence path clusters as the core, extract the causal convergence subgraph, record the inspection evidence node list, relation edge list and edge weight list in the causal convergence subgraph, and generate a mapping index from the subgraph to the original inspection feature structure.

7. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The process of generating an initial inspection report and sequentially performing structural reverse engineering and temporal inverse verification includes: Within the inspection report inversion generator, an evidence planning layer, a bidirectional consistency layer, and a reconstructed audit layer are established sequentially, with the layers connected through status markers and index mappings. Input the causal convergence subgraph into the evidence planning layer, call the chapter skeleton planning unit to generate chapter, paragraph and sentence sequence, call the evidence pointer table generation unit to build the evidence pointer table based on the unique index identifier of the inspection evidence node, synthesize the initial inspection report and output the report-evidence mapping index; The initial inspection report is input into the bidirectional consistency layer. The structure of the report content is reversed to form an element table of abnormal items, location items and change items. Each element item is mapped back to the corresponding inspection evidence node according to the report-evidence mapping index. Forward derivable verification and reverse traceable verification are completed in sequence. Verification marks are added to sentences and paragraphs that fail the verification and the missing coverage dimension is recorded. The report with verification marks is input into the reconstructed audit layer. Based on the appearance, space, evolution and deviation information of the corresponding evidence node, the marked segments are rewritten. The quota control length is described according to the anomaly type and the missing dimensions are filled in. After writing the evidence citation number, the revised draft is generated iteratively according to the set rounds. The revised draft is returned to the bidirectional consistency layer for review. For content that fails the forward derivation check or the reverse backtracking check, reconstruction continues until all sentences and paragraphs pass the check or reach the preset reconstruction round.

8. The intelligent generation method for UAV inspection reports based on data feature extraction according to claim 1, characterized in that, The step of associating and storing the final drone inspection report with the corresponding inspection evidence nodes includes: After the inspection report inversion generator completes the structural inverse solution processing and time-series reverse verification, the completeness of the inspection report content that has passed the verification is checked to confirm that the inspection report contains anomaly description, spatial location description and change trend description for each abnormal object. The inspection report that has completed the integrity check is indexed and bound to the corresponding causal convergence subgraph, a one-to-one correspondence between the inspection report paragraphs and the inspection evidence nodes is established, and a report-evidence association index table is generated. Based on the report-evidence association index table, the final inspection report content, the corresponding inspection evidence node identifier, time index, flight index, and pose index are jointly encapsulated to form a traceable inspection report data package. The inspection report data packet is written into the storage system, and the report-evidence association index table is stored synchronously to complete the association storage and traceability query between the final UAV inspection report and the inspection evidence nodes.

Images