Metal roof construction quality detection system based on AI
Through the AI-based metal roof construction quality detection system, the problem of difficult to identify dynamic abnormalities in metal roofs in the existing technology is solved by using the included angle mutation, tangential anomalies and curvature stability recognition technology, combined with the bidirectional long and short-term memory network, and the problem of difficult to identify dynamic abnormalities in metal roofs is realized, and accurate detection and classification of metal roof defects is achieved.
Patent Information
- Application Number
- CN202510743233.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-05
- Publication Date
- 2025-08-15
- Estimated Expiration
- 2045-06-05
AI Technical Summary
In the prior art, metal roof construction quality detection systems based on image acquisition and recognition methods are difficult to identify abnormal areas in dynamic evolution, especially boundary misalignment and node deformation, resulting in insufficient accuracy and coverage breadth of laying quality detection, and it is easy to misjudgment or miss detection of key defects.
Using the AI-based metal roof construction quality detection system, through the recognition of included angle mutations, tangential anomaly extraction, curvature stability recognition and path fitting offset analysis, combined with the bidirectional long and short-term memory network, the angle change rate, tangential displacement sequence and curvature distribution abnormality of boundary nodes are constructed to realize dynamic detection of metal roof defects.
It realizes accurate identification of metal roof splicing parts, reveals local structural changes trends, captures imbalanced areas of direction continuity, strengthens the extraction of geometric performance of abnormal laying accuracy, and can identify multiple types of defects and make classified judgments, which improves the accuracy of detection and coverage breadth.
Smart Images

Figure CN120495279A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of roof image recognition, and in particular to an AI-based metal roof construction quality detection system. Background Art
[0002] The technical field of metal roof construction quality inspection systems includes applications related to roof image recognition. This technology, primarily based on computer vision and image processing, captures and analyzes images of roof structures to identify defects, cracks, splicing errors, and quality issues encountered during construction.
[0003] Among them, the metal roof construction quality inspection system refers to a system designed for quality inspection during the metal roof construction process. It mainly uses an image acquisition device to perform multi-angle and multi-time image acquisition of the roof structure at the construction site, and uses image recognition methods to detect the splicing gaps, plate misalignment, abnormal connection boundary nodes, surface cracks and laying conditions of the roof panels.
[0004] In existing technologies, static inspections based on image acquisition and recognition rely excessively on the boundary clarity and material characteristics of a single-frame image. There is a lack of continuity analysis methods for structural defects, especially boundary dislocations and node deformations, making it difficult to identify abnormal areas in dynamic evolution. Inspections are performed solely based on surface manifestations such as gaps and cracks, ignoring changes in structural geometry and path continuity. This makes it difficult to detect deep-seated laying deviations such as angular fractures and curvature disturbances in actual applications. The image recognition process lacks temporal modeling capabilities, and there is no systematic extraction strategy for outlier states of changes in node distribution density. This makes it very easy to misjudge or miss key defect fragments in dense structures. For example, in the connection transition area, if local node instability does not manifest as obvious cracks, but only a loss of tangential continuity, traditional methods cannot detect such potential risks, severely restricting the accuracy and coverage of full-cycle monitoring of laying quality, resulting in increased maintenance costs and increased safety hazards. Summary of the Invention
[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose an AI-based metal roof construction quality detection system.
[0006] To achieve the above objectives, the present invention adopts the following technical solutions: an AI-based metal roof construction quality inspection system includes: The angle mutation identification module obtains the entire boundary node sequence of the metal roof target location, extracts the angle change rate of adjacent boundary nodes in the node sequence, and screens the angle change mutation points to form the first boundary node anomaly preliminary screening set; The tangential anomaly extraction module obtains the tangential displacement sequence between adjacent boundary nodes at the target position of the metal roof through the first boundary node anomaly preliminary screening set, performs continuity analysis on the tangential gradient, and selects boundary nodes with continuous anomalies to form a first boundary node set with continuous directional anomalies; The curvature stability identification module obtains all boundary node sequences of the target position of the metal roof, determines the stability of the curvature change state of the node sequence, and obtains the second boundary node set with abnormal curvature distribution; The path fitting offset analysis module inputs the first boundary node set with continuous direction anomalies and the second boundary node set with curvature distribution anomalies into a bidirectional long short-term memory network to perform path fitting offset analysis and construct a path fitting offset sequence; The defect type determination module performs abnormal type detection on the paving path of the metal roof target position based on the path fitting offset sequence to obtain a metal roof quality detection result.
[0007] As a further solution of the present invention, the first boundary node anomaly preliminary screening set includes angle change sudden increase points, normal vector mutation boundary nodes, and boundary direction break points; the first boundary node set of continuous direction anomalies is specifically a multi-segment continuous mutation boundary node sequence, a tangential angle discontinuous area, and a local path disturbance zone; the second boundary node set of curvature distribution anomalies includes curvature extreme points, curvature fluctuation concentration areas, and boundary curvature instability sections; the path fitting offset sequence specifically refers to the fitting residual value sequence, the fitting trend mutation area, and the fitting error time series distribution; the metal roof quality inspection results include path tearing areas, angle torsion areas, and curvature instability areas.
[0008] As a further solution of the present invention, the angle mutation recognition module includes: The normal vector extraction submodule collects the target position image data of the metal roof, obtains the arrangement order of the boundary nodes in the three-dimensional coordinate space, calculates the normal vector directions constructed by adjacent boundary nodes, and establishes the boundary node normal direction sequence based on the spatial coordinate data between adjacent normal vectors; The angle increment calculation submodule selects two segments of normal vectors formed by each group of three adjacent boundary nodes in the boundary node normal direction sequence, calls the direction angle value between each group of vectors and constructs an angle difference sequence, performs a comparative analysis of the boundary node angle change amplitude, and obtains a boundary node angle change rate sequence; The mutation point screening submodule calls the angle change rate amplitude of each boundary node in the boundary node angle change rate sequence and compares the boundary node change trends in sequence, identifies boundary nodes with inconsistent continuous change directions and sudden changes in amplitude as feature points, and generates a first boundary node anomaly preliminary screening set.
[0009] As a further solution of the present invention, the tangential anomaly extraction module includes: The tangent vector construction submodule obtains the spatial coordinate information of the adjacent boundary nodes of the metal roof target position based on the first boundary node anomaly preliminary screening set, sequentially connects the adjacent boundary nodes to form a boundary line segment direction vector sequence, and uniformly normalizes each group of boundary node direction vectors to establish a boundary tangent direction sequence; The gradient sequence extraction submodule calls adjacent vector pairs in the boundary tangential direction sequence, obtains the degree of directional mutation between consecutive boundary nodes by comparing the direction angles, and constructs a continuous sequence reflecting the overall path change trend to obtain a tangential gradient sequence; The abnormal segment identification submodule sequentially extracts the boundary node direction change amplitude according to the tangential gradient sequence, and performs continuity judgment on the change interval. When the boundary node direction mutation continues to occur and the gradient change trend is unstable, the corresponding boundary node segment is delineated to obtain the first boundary node set with continuous direction anomalies.
[0010] As a further solution of the present invention, the curvature stability identification module includes: The coordinate sequence acquisition submodule obtains the boundary image of the target position of the metal roof, extracts the positions of all boundary nodes on the boundary line segment, and forms a spatial point sequence according to the boundary node numbering order. The coordinates of each group of three adjacent boundary nodes in the sequence are constructed into a boundary node combined coordinate sequence; The curvature trend extraction submodule constructs the geometric relationship between three consecutive points according to the boundary node combined coordinate sequence, extracts the curvature value of the geometric relationship arc, and calculates the continuous trend of the center difference value between adjacent curvatures to obtain the curvature change trend sequence; The unstable point identification submodule compares the amplitude of each boundary node in the curvature change trend sequence, determines the fluctuation state of the continuous change segment formed by multiple adjacent boundary nodes, and selects the second boundary node set with abnormal curvature distribution.
[0011] As a further solution of the present invention, the path fitting offset analysis module includes: The sequence construction submodule extracts the rate of change of the normal vector angle of the curvature abnormal boundary nodes in the corresponding image frame based on the first boundary node set with continuous direction abnormalities, and extracts the curvature change gradient between adjacent coordinate boundary nodes based on the second boundary node set with curvature distribution abnormalities. The two types of boundary node sets are integrated into an independent time series structure to generate an input feature sequence set; The feature combination submodule calls the time window segment of each type of sequence in the input feature sequence set, performs temporal feature splicing in the order of boundary node numbers, and generates a path sequence feature group; The fitting training submodule inputs the path sequence feature group into a bidirectional long short-term memory network, carries out spatial change trend learning of boundary nodes based on the time dimension, extracts the residual values of boundary node outputs, and obtains a path fitting offset sequence.
[0012] As a further solution of the present invention, the defect type determination module includes: The abnormal segment extraction submodule extracts the time series segments with prominent continuous offset amplitudes based on the residual values output by the boundary nodes in the path fitting offset sequence, matches the corresponding boundary node numbers with the sequence positions, and obtains a set of continuous abnormal offset segments; The outlier region identification submodule calls the boundary node offset data in the continuous abnormal offset segment set, performs density ratio judgment through the local anomaly factor algorithm, identifies the time series segments with sparse local distribution and lower density than the surrounding boundary nodes in the offset sequence, and obtains the outlier boundary node segment; The defect type determination submodule performs classification judgment on the boundary node morphology based on the spatial structural position and angle change trend corresponding to the boundary node in the outlier boundary node segment, combined with the boundary direction and curvature fluctuation characteristics of the boundary node, and generates the metal roof defect type result.
[0013] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by constructing a sequence of boundary node angle change rates, the mutation points of the metal roof splicing parts are accurately identified, effectively revealing the local structural change trend; combining the tangential displacement sequence and angle mutation characteristics between adjacent boundary nodes, the directional continuity imbalance section is captured, so that the abnormal directional variation presents a dynamic evolution trajectory; on this basis, the stability of the curvature distribution state of the boundary nodes is further judged, and the curvature fluctuation concentration area and extreme points are identified through differential analysis, thereby strengthening the extraction of the geometric expression of the paving accuracy anomaly; the path fitting process embeds the change sequence of the structural space into the time domain learning model, combines the normal angle change and the curvature gradient to construct a feature sequence, and obtains the residual extraction capability of the abnormal trend in the time series evolution, making the abnormal trend more three-dimensional and traceable; further integrating the abnormal offset fragment and the density factor algorithm, the identification of the outlier area in the paving path is realized, and the classification judgment is completed according to the geometric position and direction state of the boundary node, effectively delineating the structural range of multiple types of defects such as tearing, torsion and instability. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 is a system flow chart of the present invention; Figure 2 This is a flow chart of the angle mutation identification module of the present invention; Figure 3 This is a flow chart of the tangential anomaly extraction module of the present invention; Figure 4This is a flow chart of the curvature stability identification module of the present invention; Figure 5 This is a flow chart of the path fitting offset analysis module of the present invention; Figure 6 This is a flow chart of the defect type determination module of the present invention. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0016] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.
[0017] See also Figure 1 , the AI-based metal roof construction quality inspection system includes: The angle mutation identification module obtains the entire boundary node sequence of the metal roof target location, extracts the angle change rate of adjacent boundary nodes in the node sequence, and screens the angle change mutation points to form the first boundary node anomaly preliminary screening set; The tangential anomaly extraction module obtains the tangential displacement sequence between adjacent boundary nodes at the target position of the metal roof through the first boundary node anomaly preliminary screening set, and performs continuity analysis on the tangential gradient to select boundary nodes with continuous anomalies to form the first boundary node set with continuous directional anomalies. The curvature stability identification module obtains all boundary node sequences of the target position of the metal roof, determines the stability of the curvature change state of the node sequence, and obtains the second boundary node set with abnormal curvature distribution; The path fitting offset analysis module inputs the first boundary node set with continuous direction anomalies and the second boundary node set with curvature distribution anomalies into the bidirectional long short-term memory network to perform path fitting offset analysis and construct a path fitting offset sequence; The defect type determination module performs abnormal type detection on the paving path of the metal roof target position based on the path fitting offset sequence to obtain the metal roof quality detection result; The initial screening set of the first boundary node anomaly includes angle change sudden increase points, normal vector mutation boundary nodes, and boundary direction break points. The first boundary node set of continuous direction anomaly is specifically a multi-segment continuous mutation boundary node sequence, tangential angle discontinuity area, and local path disturbance zone. The second boundary node set of curvature distribution anomaly includes curvature extreme value points, curvature fluctuation concentration area, and boundary curvature instability section. The path fitting offset sequence specifically refers to the fitting residual value sequence, fitting trend mutation area, and fitting error time series distribution. The metal roof quality inspection results include path tearing area, angle torsion area, and curvature instability area.
[0018] See also Figure 2 , the angle mutation recognition module includes: The normal vector extraction submodule collects the target position image data of the metal roof, obtains the arrangement order of the boundary nodes in the three-dimensional coordinate space, calculates the normal vector directions constructed by adjacent boundary nodes, and establishes the boundary node normal direction sequence based on the spatial coordinate data between adjacent normal vectors; The image acquisition device is positioned directly above the target area of the metal roof to ensure that the captured image has a vertical viewing angle and eliminates perspective distortion interference. The target area is scanned and photographed at a frame rate of 0.5fps using a continuous shutter method to obtain at least 5 image frames covering the entire boundary. The edge node coordinates in each frame of the image are mapped from two-dimensional pixel coordinates to a three-dimensional rectangular coordinate system using an image coordinate system conversion algorithm. The corresponding spatial conversion coefficient is set according to the metal roof material parameters and roof configuration characteristics (such as a reflectivity of 0.85 and a thickness of 5mm), and the endpoint information of each boundary segment in the image is called to construct a boundary node array. ,in Indicates the Boundary nodes, are the horizontal, vertical and elevation coordinate components in three-dimensional space, is the index of the node in the boundary sequence, and its value range is , then for any two adjacent nodes such as and Find the difference vector ,in Indicates that from The node points to The vector of each node is constructed by combining the three points of the local patch based on the vector cross product operation , find the face normal , Represents the patch normal vector composed of adjacent boundary vectors, normalized to obtain the unit normal vector , After normalization Normal direction unit vectors, and finally construct the normal vector sequence of the boundary nodes according to the node order The order of arrangement is determined by the clockwise or counterclockwise direction of the boundary nodes in the image. For example, select any three points in the boundary node sequence 、 、 , then 、 , cross product , the normalized result is , repeat this process to complete the spatial arrangement of the normals of all adjacent boundary nodes.
[0019] The angle increment calculation submodule selects two segments of normal vectors formed by each group of three adjacent boundary nodes in the boundary node normal direction sequence, calls the direction angle value between each group of vectors and constructs an angle difference sequence, performs a comparative analysis of the boundary node angle change amplitude, and obtains the boundary node angle change rate sequence; In the established normal direction sequence, select each group of three consecutive normal vectors in turn, e.g. ,in Indicates the The unit normal vector at the node, , respectively for the vector pair and Call the arc cosine function to calculate the angle value and ,in For nodes and The angle between the unit vectors, For nodes and The angle between the unit vectors is the difference between them. Representation node The amplitude of the normal angle change constitutes the angle difference sequence ,in Indicates the The change in the angle formed by the three-point group, during the calculation process, it is necessary to determine whether the adjacent vector dot product values are in the interval Otherwise, the abnormal vector group should be eliminated, for example, , then 、 、 , in this way, the change value of the normal angle between each node is obtained in turn and recorded in the form of a series. Each difference must be stored in an array structure in a unit of measurement (unit: °).
[0020] The mutation point screening submodule calls the angle change rate amplitude of each boundary node in the boundary node angle change rate sequence and compares the boundary node change trends in sequence, identifies boundary nodes with inconsistent continuous change directions and sudden changes in amplitude as feature points, and generates the first boundary node anomaly preliminary screening set; For each boundary node , read the corresponding angle change rate , Representation node The angle change at the node is compared with the absolute value of the difference between the node and the previous node. ,in Indicates the change in angle of the previous node. If this value changes sign twice in a row and the amplitude is greater than the set threshold ,in The threshold for determining angle mutation is in degrees, and the empirical setting range is , then the node is determined to have a mutation feature and is initially selected as a boundary feature point, such as 、 ,but , record the node, and further determine whether the direction of the change has changed. If the previous difference is an increase ( increases), the latter difference decreases ( Decreases), that is, there is a direction inconsistency condition, record this node as an abnormal mutation point, and construct an abnormal initial screening set , where the set It represents the set of normal mutation boundary feature points that have been initially screened out, and outputs a node index sequence for subsequent feature extraction operations.
[0021] See also Figure 3 ,The tangential anomaly extraction module includes: The tangent vector construction submodule obtains the spatial coordinate information of the adjacent boundary nodes of the metal roof target location based on the first boundary node anomaly preliminary screening set, connects the adjacent boundary nodes in sequence to form a boundary line segment direction vector sequence, and performs unified normalization processing on each group of boundary node direction vectors to establish a boundary tangent direction sequence; After obtaining the first boundary node anomaly screening set Then, traverse each abnormal boundary node in the set in turn ,in Indicates the first nodes, Indicates its index value in the original boundary node sequence, calling its adjacent boundary nodes before and after The spatial coordinate value of and ,in Respectively represent The horizontal, vertical and vertical coordinate values of each node in three-dimensional space are used to construct the boundary direction vector by the difference of spatial coordinates. 、 ,in Indicates that from The tangent direction vector from the first node to the next node, Represents the node index in the boundary segment sequence. Then for each Perform unified normalization processing, the normalization method is ,in Represents the normalized unit tangent vector, Represents a vector The Euclidean norm (modulus) of is calculated as: ; in and are the coordinate components of the starting point and the end point respectively. 、 For example, , the module length is , after normalization , and establish the tangential unit vector sequence in this way ,in Indicates the Normalized tangent unit vectors are used as the boundary tangent direction sequence for subsequent analysis. The order of the vector sequence is arranged in sequence according to the node space coordinate index to ensure path continuity.
[0022] The gradient sequence extraction submodule calls the adjacent vector pairs in the boundary tangential direction sequence, compares the direction angles, obtains the degree of directional mutation between consecutive boundary nodes, and constructs a continuous sequence reflecting the overall path change trend to obtain the tangential gradient sequence; For the tangential unit vector sequence ,in Indicates the The normalized unit vectors in the direction of the boundary segments are selected for every two adjacent vector pairs. , calculate the angle value to determine the degree of change in the tangential direction. The angle calculation method is: ,in Indicates the With the The angle between the unit vectors, in degrees. Before calculating the dot product, ensure that the vectors are normalized to maintain ; In actual sampling, if 、 , then the dot product is , corresponding to the angle In this way, the angles between all adjacent vectors are constructed into a sequence ,in Indicates the The angle between the directions of adjacent vectors reflects the degree of directional turning of the path. This sequence serves as an important reference for describing the continuity and broken line density of the path trend.
[0023] The abnormal segment identification submodule sequentially extracts the magnitude of boundary node direction changes based on the tangential gradient sequence and makes continuity judgments on the change intervals. When sudden changes in boundary node directions continue to occur and the gradient change trend is unstable, the corresponding boundary node segments are delineated to obtain the first boundary node set with continuous direction anomalies. In obtaining the tangential gradient sequence After that, the variation range of each angle value is extracted in turn and continuity analysis is performed. The judgment method is: for any continuous subsequence ,in Indicates that the sub-segment Item angle value, if there are at least three consecutive angle values that satisfy: 、 ,in Indicates the angle threshold difference used to identify sudden changes in direction, in degrees, with an empirical range of , then it is determined that the sequence has an abnormal directional trend; then the stability of the gradient sequence change trend is determined, and the gradient change rate is defined as ,in Indicates the and The absolute magnitude of the angle change between the vectors, if there is Shows oscillating changes and satisfies ,in Indicates the amplitude threshold of the gradient mutation trend, in degrees, and the recommended value is , then the direction of the section is determined to be unstable, and the two conditions are combined to define the continuous abnormal direction section. Add continuous abnormal boundary node sets ,in Indicates the first boundary node set of the continuous directional anomalies finally identified.
[0024] See also Figure 4 , the curvature stability identification module includes: The coordinate sequence acquisition submodule obtains the boundary image of the target position of the metal roof, extracts the positions of all boundary nodes on the boundary line segment, and forms a spatial point sequence according to the boundary node numbering order. The coordinates of each group of three adjacent boundary nodes in the sequence are constructed into a boundary node combined coordinate sequence; In the boundary extraction process of the target position of the metal roof, it is first necessary to extract the boundary image of the target area through a coordinate sequence. In practical applications, metal roofs usually have a specific geometric shape, so the image processing module first obtains an image containing the outline of the metal roof, which is a binary image in which the boundary line is the white part. Next, an edge detection algorithm (such as the Canny algorithm) is used to detect the connected areas of the edge in the image and mark the boundaries in the image. Through algorithm processing, the coordinate information of the boundary segments can be obtained, and then these coordinates are extracted to obtain the node position of each boundary segment. Further steps involve arranging these nodes in numerical order and forming a spatial point sequence. For example, suppose the first boundary node extracted is A (10, 15), the second node is B (20, 25), the third is C (30, 35), and so on. On this basis, every three consecutive adjacent nodes form a boundary node combination. For example, if A, B, and C are three consecutive nodes, the boundary node combination is This combination method allows for further adjustment of the number of nodes based on different needs. For example, four nodes can be considered for a combination, but in practice, a three-point combination is more common. This step ultimately forms a continuous sequence of boundary node combination coordinates, providing data support for the subsequent curvature trend extraction module.
[0025] The curvature trend extraction submodule constructs the geometric relationship between three consecutive points based on the combined coordinate sequence of the boundary nodes, extracts the curvature value of the geometric relationship arc, and calculates the continuous trend of the central difference value between adjacent curvatures to obtain the curvature change trend sequence; By analyzing the combined coordinate sequence of the boundary nodes, the geometric relationship between three consecutive nodes is first constructed. The geometric relationship is performed by calculating the curvature of the geometric figure formed by the three adjacent nodes. Assume that three consecutive boundary nodes 、 、 It forms a three-point combination. By calculating the degree of curvature between these nodes, the curvature value between the three points can be determined. When calculating, we first need to calculate the distance between each pair of adjacent points, such as arrive The distance is: ; Similarly, arrive The distance is: .
[0026] Then, the curvature values between the three points are calculated based on these distances and the geometric relationship between the three points. Each curvature value represents the degree of curvature of this boundary. Next, based on the curvature values calculated for all boundary node combinations, the differences between adjacent curvatures are extracted and a trend sequence is formed. For example, if the curvature values of the first two segments are and , then their central difference values are ,In this way, the changing trend of curvature can be extracted.
[0027] The unstable point identification submodule compares the amplitude of each boundary node in the curvature change trend sequence, determines the fluctuation state of the continuous change segment formed by multiple adjacent boundary nodes, and selects the second boundary node set with abnormal curvature distribution; Analyze the amplitude changes in the curvature change trend sequence to identify potential instability points. In specific implementation, the curvature change trend of each boundary node will be compared with the set amplitude threshold. If the difference between adjacent curvature values exceeds the preset threshold, it indicates that there may be a risk of instability. The set threshold is usually adjusted based on historical data or experimental results. Assuming that a threshold is set to 0.1, if the curvature change between two adjacent nodes is greater than 0.1, it is considered a potential instability point. For example, suppose that in a certain section of the boundary, the node arrive The curvature value increases from 0.1 to 0.3, and the amplitude difference is , exceeds the set threshold of 0.1, at this time The node is identified as an unstable point. This allows the system to identify nodes with abnormal curvature distribution, further identifying and screening areas that may require repair or monitoring. This step forms the second boundary node set through amplitude comparison, fluctuation state judgment, and anomaly screening.
[0028] See also Figure 5 , the path fitting offset analysis module includes: The sequence construction submodule extracts the rate of change of the normal vector angle of the curvature abnormal boundary nodes in the corresponding image frame based on the first boundary node set with continuous direction anomalies, and extracts the curvature change gradient between adjacent coordinate boundary nodes based on the second boundary node set with curvature distribution anomalies. The two types of boundary node sets are integrated into an independent time series structure to generate an input feature sequence set. The change rate of the normal vector angle of the curvature abnormal boundary node in the corresponding image frame is extracted based on the first boundary node set with continuous direction abnormality. In this process, the boundary node set is first extracted from the image frame. These nodes show obvious direction abnormality due to boundary continuity interruption, folding deformation or local concave-convex structure. For these nodes, the change rate of the normal vector angle between two adjacent nodes needs to be calculated. Assume that the first boundary node to be analyzed is The boundary nodes (denoted as node i with abnormal subscript direction) have normal vectors , its front and back adjacent nodes are and boundary nodes. Respectively represent the node normal vector in the three-dimensional coordinate system along 、 、 The component of the axis, the unit is the dimensionless direction ratio, representing a unit vector; and the The normal vector of the boundary node (i.e. the next node in the sequence, denoted as node i plus one) is ,in Similarly, its direction component in the three-dimensional coordinate axis. The calculation formula is: ; To further calculate the angle change rate, we can use the unit time or the distance between nodes as the denominator. If the unit node spacing is 1, the change rate is approximately equal to the difference between adjacent angle values. Example: If the node , the normal vector ,node , the normal vector , then the dot product is , the module lengths are , similarly the second modulus is also 1, so: , if the previous node angle is , then the angle change rate is , which is used to construct a time series of sudden changes in the direction of boundary nodes.
[0029] The feature combination submodule calls the time window segment of each type of sequence in the input feature sequence set, performs temporal feature splicing in the order of boundary node numbers, and generates a path sequence feature group; The curvature change gradient between adjacent coordinate boundary nodes is extracted based on the second boundary node set with abnormal curvature distribution. boundary nodes (where the subscript “curve” indicates that the node belongs to the curvature anomaly detection subset), and its corresponding curvature value is recorded as , which is derived from the curve curvature determined by the three-point fitting method and is expressed in degrees / meter ( ), describes the angle change of the unit length path segment at the node. The adjacent node is boundary nodes, denoted as nodes , whose curvature value is , which represents the curvature of the subsequent boundary of this segment. To calculate the curvature gradient between these two nodes, we need to obtain their spatial distance, and set their Euclidean distance to be , the unit is "meter". The calculation formula of the curvature gradient is: For example, let node The curvature value is ,node The curvature is , the actual distance between the two in the image coordinate space is , then the curvature gradient is: , which means that between the 5th and 6th nodes, the curvature increases at a rate of 0.001 per meter of path length. , which can be used to identify areas with sudden changes in curvature within a boundary segment. This calculation is performed sequentially on all pairs of adjacent nodes in the second node set of the curvature anomaly, gradually constructing a complete sequence of curvature change gradients.
[0030] The fitting training submodule inputs the path sequence feature group into the bidirectional long short-term memory network, expands the spatial change trend of boundary nodes based on the time dimension, extracts the residual value of the boundary node output, and obtains the path fitting offset sequence; The path sequence feature group is input into the Bidirectional Long Short-Term Memory Network (BiLSTM), and the spatial change trend of the boundary nodes is expanded based on the time dimension and the output residual value of the boundary nodes is extracted to obtain the path fitting offset sequence. The path sequence feature group is composed of the normal vector angle change rate sequence of the first boundary node set with direction anomaly and the curvature change gradient sequence of the second boundary node set with curvature anomaly, which constitutes the time step. The node input feature vector under ,in: : Indicates that the direction of abnormal concentration number is The boundary nodes of The rate of change of the normal vector angle at , in degrees / second ( ); : Indicates that the curvature anomaly is concentrated in The boundary nodes of The curvature gradient at , in degrees / m² ( The bidirectional LSTM neural network structure uses a combination of forward and reverse state learning to preserve the state dependencies of past and future nodes in the context of time series. Its state update formula is as follows: ; in: : Current time step The hidden state of the boundary node The changing trend in spatial location; 、 : forward (previous moment) and reverse (next moment) states; : before and after state adjustment factor (set to 0.8, based on the results of maximizing the average fitting accuracy of the training set cross validation); 、 : The weight matrices for the forward and reverse states, respectively (set to random values during the initial training phase, then adjusted by gradient descent to 0.7 and 0.6). : The weight matrix of the feature input (set to 0.8, determined by the convergence of the training error); : bias term (set to 0.1); : Hyperbolic tangent activation function, which maps the linear superposition state output to interval.
[0031] Assume that the current time step is , the previous state is , the state at the next moment is , the input features are: 、 , then the input features of node 5 are: Combined into scalar input: , substitute into the formula to calculate: ; This hidden state is used to generate the network's output prediction value. , and its result is used to calculate the residual with the actual node position value. Assume that the actual observation space position value of node 5 is , the predicted value is , then the output residual value is: , and finally, all the residual values of the time steps The chronological order of the paths is combined to form a path fitting offset sequence: this sequence measures the offset error of each node on the predicted path relative to the true path, and serves as data input for subsequent structural monitoring or correction steps. This result demonstrates a direct quantitative correspondence between the residual calculation and the spatial differences in the paths. The BiLSTM network output effectively reflects the trend of boundary node trajectory changes and the degree of divergence between the model prediction and the trajectory.
[0032] See also Figure 6 , the defect type determination module includes: The abnormal segment extraction submodule extracts the time series segments with prominent continuous offset amplitudes based on the residual values output by the boundary nodes in the path fitting offset sequence, matches the corresponding boundary node numbers with the sequence positions, and obtains the set of continuous abnormal offset segments; Based on the residual values output by the boundary nodes in the path fitting offset sequence, the residual value sequence of each time step is analyzed in sequence (for example, the residual at the first moment is 0.02 meters, the residual at the second moment is 0.03 meters... and so on to the Tth moment). Each residual value represents the absolute difference between the actual coordinate position of the boundary node at that time step and the predicted coordinate position by the bidirectional long short-term memory network (BiLSTM), in meters. This value reflects the fitting accuracy error of the boundary node. First, a linear traversal is performed on the entire time series. During the traversal process, the residual value of each time step is compared with the offset benchmark threshold set by the system. The threshold is recorded as , the value is 0.05 meters, which comes from the maximum acceptable local node displacement offset limit for metal roof structures in the industry standard. If the residual values of any three consecutive time steps (for example, the 8th, 9th, and 10th moments) are greater than , then mark the segment as a prominent offset phenomenon, and continue to determine whether the subsequent time steps meet the same conditions. If they do, they are incorporated into the current segment until the residual value of a certain time step falls back to no more than the threshold. This time point is the end point of the current abnormal segment. For example, the offset conditions are still met at the 11th and 12th moments, and the residual value is less than 0.05 meters at the 13th moment. Finally, the start and end times of the abnormal segment are determined to be 8 to 12, and the corresponding boundary nodes are numbered J8 to J12 (J is the node number prefix, and the subscript is the sequence index). Then, the segment time index position is matched with the spatial boundary node number one by one to form an abnormal offset segment set, such as segment {J8, J9, J10, J11, J12}. All time segments that meet the offset conditions are recorded and numbered in the above way.
[0033] The outlier region identification submodule calls the boundary node offset data in the continuous abnormal offset segment set, and uses the local anomaly factor algorithm to perform density ratio judgment to identify the time series segments with sparse local distribution and lower density than the surrounding boundary nodes in the offset sequence, thus obtaining the outlier boundary node segment; Obtain the residual value of each boundary node in the continuous abnormal offset segment set, process each boundary node one by one, construct a local neighborhood residual set centered on it, and perform density analysis and judgment on it. The density measurement method is to calculate the "reachable density" of the node in its residual neighborhood, which is defined as the residual value distance relationship between the node and its k nearest neighboring nodes, where k is the number of neighbors. In this embodiment, k=5 is set, which comes from the minimum support number used in statistical analysis to construct local distribution stability judgment to prevent misjudgment due to a single isolated value during the abnormality identification process. For the boundary node J10 (numbered J10, where J represents the boundary node and 10 is the time or space sequence subscript), assume that its residual value is 0.11 meters, and the residual values of its five adjacent nodes in the time series are {0.06 meters, 0.08 meters, 0.09 meters, 0.07 meters, 0.05 meters}. Then, the local point set constructed with J10 as the center is used to evaluate its distribution density difference in the offset dimension. The specific abnormality judgment indicator is the local anomaly factor, which is calculated as follows: ; In this formula: Indicates the current evaluation node (e.g. J10), and its subscript is the number of the node to be judged; Representation node The k-nearest neighbor set of the node The k closest nodes in the residual dimension, for example, the five nearest neighbors of node J10 are J6, J7, J8, J9, and J11; Indicates that in the collection Select any adjacent node , used with Perform density comparison; Indicates the number of elements in the set, that is, the value of k, which is 5 here; For nodes The local reachability density of nodes is defined as The reciprocal of the average reachable distance to all points in its k-neighborhood, in units of , represents density; Neighbor nodes The local reachable density of is the density ratio. If the average ratio is larger, it means that the node The sparser the neighborhood.
[0034] Assume that the reachability density of node J10 is =33.33 (the calculation method is that the average value of the 5 reachable distances is 0.03 meters, and the reciprocal is 33.33). Its average neighbor density is 25, so the calculation result is: , if the system sets the LOF threshold to 1.5 (this value is obtained by comparing abnormal and normal nodes in a large number of test samples), then nodes with LOF values less than 1.5 are not judged as outliers; if we change to node J13, its local density is 10 and the neighborhood average density is 25, then , is greater than the threshold, it is finally determined to be an outlier boundary node, and all nodes that meet the conditions are classified into the outlier boundary node segment.
[0035] The defect type determination submodule classifies the boundary node morphology based on the spatial structural position and angle change trend corresponding to the boundary node in the outlier boundary node segment, combined with the boundary direction and curvature fluctuation characteristics of the boundary node, and generates the metal roof defect type result; According to the spatial structure position of each node in the identified outlier boundary node segment (i.e. the two-dimensional coordinate value of the node in the image coordinate system ,in 、 is the physical position of node Jj on the horizontal and vertical axes, in meters), the trend of the normal vector angle change (expressed as , in degrees per second, representing the rate of fluctuation of the angle of the normal vector of the boundary node between consecutive time steps), the curvature change gradient (expressed as , in degrees / m², reflecting the intensity of the boundary's bending change), and a comprehensive evaluation is performed based on the distribution characteristics of the direction vector of each interval. A node sequence is constructed for each outlier boundary segment. The boundary's strike direction vector is obtained by calculating the difference in coordinates between the first and last nodes, and its strike change characteristics are determined by the angle (deflection angle) between the node and the main axis of the overall structure. For example, if nodes J13 to J18 constitute an outlier segment, its strike vector is , after normalization and comparison with the main axis direction, the included angle is 22°. Combined with the average node angle change rate of 6.1 degrees / second and the average curvature change gradient of 0.07 degrees / m², the system determines it as a "node concentrated convex defect" according to the classification rules. If the average angle change rate in a certain section, such as J21 to J25, reaches 8.4 degrees / second, the curvature change gradient is 0.11 degrees / m², the direction vector return direction fluctuation range is greater than 30 degrees, and the angle mutation points are concentrated in more than three nodes, it is classified as a "path tearing zone"; if the angle change trend between J30 and J35 shows a sudden change on one side while the other side remains continuous (for example , ), the direction vector has a bidirectional reversal feature, and the curvature change is asymmetric (for example, the average curvature on the left is 0.08 degrees / m², and on the right is 0.03 degrees / m²), then it is determined to be a "corner torsion zone"; if a section of the boundary such as J40 to J45 continuously experiences violent oscillations in curvature change ( If the fluctuation exceeds 0.05 degrees / m² at 5 consecutive nodes, but the direction vector remains smooth without obvious deflection, and the angle change trend is less than 3 degrees / second, it is judged to be a "curvature instability zone."
[0036] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.
Claims
1. AI-based metal roof construction quality inspection system, characterized by: The system comprises: The angle mutation identification module obtains the entire boundary node sequence of the metal roof target location, extracts the angle change rate of adjacent boundary nodes in the node sequence, and screens the angle change mutation points to form the first boundary node anomaly preliminary screening set; The tangential anomaly extraction module obtains the tangential displacement sequence between adjacent boundary nodes at the target position of the metal roof through the first boundary node anomaly preliminary screening set, performs continuity analysis on the tangential gradient, and selects boundary nodes with continuous anomalies to form a first boundary node set with continuous directional anomalies; The curvature stability identification module obtains all boundary node sequences of the target position of the metal roof, determines the stability of the curvature change state of the node sequence, and obtains the second boundary node set with abnormal curvature distribution; The path fitting offset analysis module inputs the first boundary node set with continuous direction anomalies and the second boundary node set with curvature distribution anomalies into a bidirectional long short-term memory network to perform path fitting offset analysis and construct a path fitting offset sequence; The defect type determination module performs abnormal type detection on the paving path of the metal roof target position based on the path fitting offset sequence to obtain a metal roof quality detection result.
2. The AI-based metal roof construction quality inspection system according to claim 1 is characterized in that: The first boundary node anomaly initial screening set includes angle change sudden increase points, normal vector mutation boundary nodes, and boundary direction break points. The first boundary node set of continuous direction anomalies is specifically a multi-segment continuous mutation boundary node sequence, tangential angle discontinuity area, and local path disturbance zone. The second boundary node set of curvature distribution anomalies includes curvature extreme value points, curvature fluctuation concentration area, and boundary curvature instability section. The path fitting offset sequence specifically refers to the fitting residual value sequence, fitting trend mutation area, and fitting error time series distribution. The metal roof quality inspection results include path tearing area, angle torsion area, and curvature instability area.
3. The AI-based metal roof construction quality inspection system according to claim 1 is characterized in that: The angle mutation recognition module includes: The normal vector extraction submodule collects the target position image data of the metal roof, obtains the arrangement order of the boundary nodes in the three-dimensional coordinate space, calculates the normal vector directions constructed by adjacent boundary nodes, and establishes the boundary node normal direction sequence based on the spatial coordinate data between adjacent normal vectors; The angle increment calculation submodule selects two segments of normal vectors formed by each group of three adjacent boundary nodes in the boundary node normal direction sequence, calls the direction angle value between each group of vectors and constructs an angle difference sequence, performs a comparative analysis of the boundary node angle change amplitude, and obtains a boundary node angle change rate sequence; The mutation point screening submodule calls the angle change rate amplitude of each boundary node in the boundary node angle change rate sequence and compares the boundary node change trends in sequence, identifies boundary nodes with inconsistent continuous change directions and sudden changes in amplitude as feature points, and generates a first boundary node anomaly preliminary screening set.
4. The AI-based metal roof construction quality inspection system according to claim 3 is characterized in that: The tangential anomaly extraction module includes: The tangent vector construction submodule obtains the spatial coordinate information of the adjacent boundary nodes of the metal roof target position based on the first boundary node anomaly preliminary screening set, sequentially connects the adjacent boundary nodes to form a boundary line segment direction vector sequence, and uniformly normalizes each group of boundary node direction vectors to establish a boundary tangent direction sequence; The gradient sequence extraction submodule calls adjacent vector pairs in the boundary tangential direction sequence, obtains the degree of directional mutation between consecutive boundary nodes by comparing the direction angles, and constructs a continuous sequence reflecting the overall path change trend to obtain a tangential gradient sequence; The abnormal segment identification submodule sequentially extracts the boundary node direction change amplitude according to the tangential gradient sequence, and performs continuity judgment on the change interval. When the boundary node direction mutation continues to occur and the gradient change trend is unstable, the corresponding boundary node segment is delineated to obtain the first boundary node set with continuous direction anomalies.
5. The AI-based metal roof construction quality inspection system according to claim 4 is characterized in that: The curvature stability identification module includes: The coordinate sequence acquisition submodule obtains the boundary image of the target position of the metal roof, extracts the positions of all boundary nodes on the boundary line segment, and forms a spatial point sequence according to the boundary node numbering order. The coordinates of each group of three adjacent boundary nodes in the sequence are constructed into a boundary node combined coordinate sequence; The curvature trend extraction submodule constructs the geometric relationship between three consecutive points according to the boundary node combined coordinate sequence, extracts the curvature value of the geometric relationship arc, and calculates the continuous trend of the center difference value between adjacent curvatures to obtain the curvature change trend sequence; The unstable point identification submodule compares the amplitude of each boundary node in the curvature change trend sequence, determines the fluctuation state of the continuous change segment formed by multiple adjacent boundary nodes, and selects the second boundary node set with abnormal curvature distribution.
6. The AI-based metal roof construction quality inspection system according to claim 5 is characterized in that: The path fitting offset analysis module includes: The sequence construction submodule extracts the rate of change of the normal vector angle of the curvature abnormal boundary nodes in the corresponding image frame based on the first boundary node set with continuous direction abnormalities, and extracts the curvature change gradient between adjacent coordinate boundary nodes based on the second boundary node set with curvature distribution abnormalities. The two types of boundary node sets are integrated into an independent time series structure to generate an input feature sequence set; The feature combination submodule calls the time window segment of each type of sequence in the input feature sequence set, performs temporal feature splicing in the order of boundary node numbers, and generates a path sequence feature group; The fitting training submodule inputs the path sequence feature group into a bidirectional long short-term memory network, carries out spatial change trend learning of boundary nodes based on the time dimension, extracts the residual values of boundary node outputs, and obtains a path fitting offset sequence.
7. The AI-based metal roof construction quality inspection system according to claim 6 is characterized in that: The defect type determination module includes: The abnormal segment extraction submodule extracts the time series segments with prominent continuous offset amplitudes based on the residual values output by the boundary nodes in the path fitting offset sequence, matches the corresponding boundary node numbers with the sequence positions, and obtains a set of continuous abnormal offset segments; The outlier region identification submodule calls the boundary node offset data in the continuous abnormal offset segment set, performs density ratio judgment through the local anomaly factor algorithm, identifies the time series segments with sparse local distribution and lower density than the surrounding boundary nodes in the offset sequence, and obtains the outlier boundary node segment; The defect type determination submodule performs classification judgment on the boundary node morphology based on the spatial structural position and angle change trend corresponding to the boundary node in the outlier boundary node segment, combined with the boundary direction and curvature fluctuation characteristics of the boundary node, and generates the metal roof defect type result.
Citation Information
Patent Citations
Movable metal roof detection equipment
CN104142253A
An intelligent roof panel deformation detection management system and detection management method
CN105318857A
Roof panel defect detection method and system
CN114529546A
Metal roof deformation analysis method based on linear interpolation
CN119494149A
Cited By
Strain key feature missing repair method and system for traffic infrastructure monitoring
CN122045630A