A method for monitoring the formation of a particle board based on machine vision

Abstract

The application discloses a kind of based on machine vision's shaving board forming monitoring method, comprising the following steps: step one: obtaining transmission image sequence;Step two: execute energy spectrum response decoupling and fusion processing, construct fusion transmission image sequence;Step three: form three-dimensional data body;Step four: construct interlayer structure continuity topological field;Step five: form layer sequence unfolding tensor group;Step six: layer sequence unfolding tensor group is input to improved RepLKNet model, including input embedding module, layer sequence enhancement feature construction module, layer sequence jump compression module and output coding module, embedding layer sequence penetration mechanism, output interlayer structure feature tensor;Step seven: generate interlayer structure change trajectory set;Step eight: identify abnormal change position and execute structure intrusion path unfolding analysis, generate mixed layer space distribution result.The application adopts layer sequence topological modeling and improved RepLKNet model, realizes mixed layer intelligent monitoring in the process of shaving board forming.

Description

Technical Field

[0001] This invention relates to the field of online monitoring technology for particleboard production, and in particular to a method for monitoring particleboard forming based on machine vision. Background Technology

[0002] With the continuous improvement of automation and intelligent manufacturing in the engineered wood products industry, the demand for online monitoring of the internal structural uniformity and interlayer quality during particleboard forming is increasing. In current particleboard production processes, the structural state of the board blank before laying and hot pressing directly affects the mechanical properties and stability of the final product. Therefore, real-time, non-destructive testing of the board blank's internal structure is of great significance. Currently, commonly used testing methods mainly rely on manual experience or simple surface image inspection techniques. Some systems introduce X-ray imaging for internal structural observation, but in practical applications, the following problems are commonly encountered:

[0003] The acquired transmission images are significantly affected by variations in material thickness and uneven density distribution. Single-energy imaging struggles to simultaneously capture the imaging effects of both high-density and low-density regions, resulting in insufficient contrast of the internal structure and difficulty in accurately reflecting subtle interlayer changes. During continuous scanning, the temporal and spatial correspondence of image data is easily affected by fluctuations in transport speed and mechanical vibrations. Existing methods struggle to guarantee spatial consistency during 3D reconstruction, thus impacting the reliability of structural analysis. Regarding the continuity of the layered structure within the slab, traditional image processing methods are mostly based on 2D feature extraction or local thresholding, lacking the ability to comprehensively model the structural continuity in the thickness direction, making it difficult to effectively identify interlayer transition regions and mixed-layer phenomena. In terms of feature analysis, existing deep learning models are mostly designed for planar images, lacking the ability to express cross-layer structural relationships. They struggle to effectively characterize the dynamic changes in the structure as it expands from one layer to another, leading to unstable mixed-layer identification results and a high false detection rate.

[0004] Therefore, how to provide a machine vision-based method for monitoring particleboard forming is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine vision-based method for monitoring particleboard forming. This invention employs a continuous topological field of interlayer structure and an improved RepLKNet model to achieve intelligent monitoring of mixed layers during the particleboard forming process. By decoupling and fusing the energy spectrum response of the transmission image sequence, a fused transmission image sequence is constructed, and further, a three-dimensional data volume inside the slab is formed. Based on this, a continuous topological field of interlayer structure is established and a layer sequence unfolding is performed. Combined with the layer sequence penetration mechanism, the features of the unfolded tensor group are extracted and enhanced to obtain the interlayer structure feature tensor. By peeling off layers one by one and continuously tracking structural changes, a set of interlayer structure change trajectories is constructed. Based on the structural intrusion path unfolding analysis, the accurate generation of mixed layer spatial distribution results is achieved, thereby realizing continuous monitoring and spatial positioning of the internal structural change process of particleboard.

[0006] A method for monitoring particleboard forming based on machine vision according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Install an online X-ray flaw detection device along the particleboard laying and conveying path to continuously scan the slab in the forming stage and obtain a sequence of transmission images.

[0008] Step 2: Perform energy spectrum response decoupling and fusion processing sequentially on the transmission image sequence to construct a fused transmission image sequence;

[0009] Step 3: Stack the fused transmission image sequence according to spatial correspondence to form a three-dimensional data volume inside the slab;

[0010] Step 4: In the three-dimensional data volume inside the slab, establish structural nodes based on voxels as basic units, establish connection relationships, and construct the interlayer structural continuity topological field;

[0011] Step 5: Perform a sequence unfolding of the interlayer structural continuity topological field along the slab thickness direction to generate a structural slice group that maintains the interlayer correspondence, forming a sequence unfolding tensor group;

[0012] Step 6: Input the layer-order unfolded tensor group into the improved RepLKNet model. The improved RepLKNet model includes an input embedding module, a layer-order enhanced feature construction module, a layer-order skip compression module, and an output encoding module. The layer-order enhanced feature construction module embeds a layer-order penetration mechanism and outputs a layer-by-layer structure feature tensor.

[0013] Step 7: Based on the layered structure feature tensor, perform layer-by-layer peeling process according to the thickness direction of the slab to generate a structural sequence from the surface layer to the core layer, and continuously track the structural changes in different layers at the same spatial location to generate a set of interlayer structural change trajectories.

[0014] Step 8: Identify the locations of abnormal changes based on the set of interlayer structural change trajectories, and perform structural intrusion path expansion analysis to generate mixed-layer spatial distribution results.

[0015] Optionally, step one specifically includes:

[0016] An online X-ray flaw detection device is arranged below the particleboard laying and conveying path along the conveying direction. The online X-ray flaw detection device includes an X-ray emission source and a linear array detector, with the linear array detector located on the opposite side of the X-ray emission source.

[0017] During the process of the slab moving along the conveying path in the forming stage, the X-ray emission source continuously emits a beam of X-rays, which is received by the linear array detector after passing through the slab.

[0018] The linear array detector samples the transmitted rays line by line to obtain transmission signal data corresponding to the internal structure of the slab.

[0019] A rotary encoder is installed on the conveyor belt drive shaft. Based on the displacement signal output by the rotary encoder and the sampling frequency of the linear array detector, the spatial positions corresponding to adjacent sampling rows are matched.

[0020] Subtract the reference transmission signal value at the corresponding position in the blank-free state from the transmission signal value at each sampling point, and divide the subtraction result by the reference transmission signal value to obtain the normalized transmission intensity value.

[0021] The normalized transmission intensity values ​​are arranged into a transmission image according to the sampling row order, and stored sequentially by number according to the transport direction to generate a transmission image sequence.

[0022] Optionally, step two specifically involves:

[0023] The transmitted image sequence is subjected to energy spectrum response decoupling processing. Using the linear array detector in the online X-ray flaw detection device, the transmitted rays are divided into three energy ranges according to their energy magnitude: low energy range, medium energy range, and high energy range.

[0024] The pixels of each sampling row in the transmission image sequence are processed point by point according to their spatial location. The normalized transmission intensity values ​​corresponding to the same spatial location in the low-energy region, medium-energy region and high-energy region are extracted respectively and classified according to the energy range.

[0025] The normalized transmission intensity values ​​belonging to the low-energy region are arranged according to the original sampling row order and pixel position to generate a low-energy region transmission image. The normalized transmission intensity values ​​belonging to the medium-energy region are arranged according to the original sampling row order and pixel position to generate a medium-energy region transmission image. The normalized transmission intensity values ​​belonging to the high-energy region are arranged according to the original sampling row order and pixel position to generate a high-energy region transmission image.

[0026] The low-energy region transmission image, the medium-energy region transmission image, and the high-energy region transmission image are fused. The normalized transmission intensity value of the low-energy region is added to the normalized transmission intensity value of the medium-energy region and then divided by two to obtain a first intermediate value. The first intermediate value is added to the normalized transmission intensity value of the high-energy region and then divided by two to obtain the fused transmission intensity value.

[0027] The fused transmission intensity values ​​are arranged according to the original sampling row order and pixel position to generate a fused transmission image. Each fused transmission image is then continuously numbered and stored sequentially according to the sampling order corresponding to the transport direction to construct a fused transmission image sequence.

[0028] Optionally, step three specifically includes:

[0029] The fused transmission image sequence is read in chronological order according to the transport direction, and the time interval between adjacent fused transmission images is converted into the corresponding spatial interval;

[0030] Based on the correspondence between the conveyor belt speed and the sampling frequency, each fused transmission image is assigned a corresponding spatial position coordinate in the conveying direction, and the fused transmission images are arranged in sequence according to the spatial position coordinates.

[0031] Using the pixels in each fused transmission image as two-dimensional plane coordinates and the spatial position coordinates of the corresponding fused transmission image as the third-dimensional coordinates, a three-dimensional coordinate mapping is performed on each pixel.

[0032] According to the three-dimensional coordinate mapping relationship, the fused transmission images are arranged sequentially according to their spatial positions, and the distance relationship between adjacent fused transmission images is determined based on the spatial interval.

[0033] The superimposed three-dimensional coordinate points and their corresponding fused transmission intensity values ​​are organized in a unified manner. The fused transmission intensity value corresponding to each three-dimensional coordinate position is used as a voxel value to construct a three-dimensional data volume inside the slab composed of multiple three-dimensional coordinate points.

[0034] Optionally, step four specifically involves:

[0035] In the three-dimensional data volume inside the slab, each voxel corresponding to a three-dimensional coordinate position is used as a structural node, and the fused transmission intensity value is used as the attribute value of the structural node.

[0036] Within the 3D data body, for each structural node, spatially adjacent voxels are selected as adjacent nodes, and the adjacency range is limited to directly adjacent voxels in the X, Y and Z directions.

[0037] Subtract the attribute values ​​between each structural node and its corresponding adjacent node, and take the absolute value of the subtraction result to obtain the voxel intensity difference between adjacent voxels.

[0038] The voxel intensity differences between all adjacent voxels are summarized and statistically analyzed. The voxel intensity differences are sorted from smallest to largest, and the median of the sorting results is selected as the criterion for structural similarity judgment.

[0039] When the difference in voxel intensity between a structural node and its adjacent node is less than the structural similarity criterion, a connection relationship is established between the structural node and its adjacent node.

[0040] The process of establishing connection relationships for all structural nodes is repeated to form a spatial network structure composed of multiple structural nodes and connection relationships, forming a connection structure distributed along the thickness direction of the slab, and constructing a continuous topological field of interlayer structure.

[0041] Optionally, step five specifically includes:

[0042] The interlayer structural continuity topological field is processed in layers along the thickness direction of the slab. The set of structural nodes corresponding to each thickness position is extracted in ascending order of Z coordinate, and a structural slice is formed by the set of structural nodes at each thickness position.

[0043] Arrange the structural slices sequentially according to the thickness direction of the slab to generate a structural slice group;

[0044] For two adjacent structural slices, for structural nodes at the same two-dimensional plane coordinate position, it is determined whether there is a connection relationship in the Z direction. The position with a connection relationship is marked with a value of 1, and the position without a connection relationship is marked with a value of 0, forming an inter-layer connection mark matrix with the same size as the structural slice.

[0045] Arrange the structural node attribute values ​​in each layer of structural slice into a two-dimensional matrix according to the two-dimensional plane coordinate order, and arrange the corresponding inter-layer connection mark matrix according to the same two-dimensional plane coordinate order.

[0046] Following the order of the slab thickness, the structural slice matrix of each layer and the corresponding interlayer connection mark matrix are arranged alternately. After the structural slice matrix of the i-th layer, the interlayer connection mark matrix between the i-th layer and the (i+1)-th layer is arranged to form a layer sequence expansion tensor group.

[0047] Optionally, step six specifically includes:

[0048] The hierarchical unfolded tensor group is input into the input embedding module. The structural slice matrix and the inter-layer connection label matrix are mapped to the same spatial position. The value of each position in the structural slice matrix is ​​multiplied with the value of the corresponding position in the inter-layer connection label matrix. The product result is added to the value of the corresponding position in the structural slice matrix to obtain the connection modulation feature. The feature is then arranged according to the hierarchical index order to generate the hierarchical embedding feature tensor.

[0049] The hierarchical embedding feature tensor is input into the hierarchical enhancement feature construction module, which introduces a hierarchical penetration mechanism to obtain the hierarchical enhancement feature tensor.

[0050] The sequence augmentation feature tensor is input into the sequence skip compression module, which performs cross-sequence difference calculation on the sequence augmentation features at each sequence position along the sequence index direction.

[0051] For any sequence position i, select the sequence enhancement features corresponding to the i-th and i+2-th layers at the same spatial position. When the i+2-th sequence position does not exist, use the sequence enhancement features of the i-th sequence position as the substitute value of the i+2-th sequence position in the calculation.

[0052] Subtract the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer and take the absolute value to obtain the inter-layer difference features at the corresponding sequence position i.

[0053] Add the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer, and then divide by two to obtain the inter-layer average feature at the corresponding sequence position i.

[0054] The inter-layer difference features and the inter-layer average features are concatenated along the channel dimension to form the compressed stratum sequence feature at stratum position i. The compressed stratum sequence features at all stratum positions are arranged in stratum sequence index order to generate a compressed stratum sequence feature tensor.

[0055] The compressed stratum feature tensor is input into the output encoding module, and nonlinear transformation is performed point by point on the feature values ​​of the compressed stratum feature corresponding to each stratum position in the compressed stratum feature tensor.

[0056] Perform an exponential operation on the feature value of each compressed sequence feature, subtract 1 from the exponential operation result, and divide by the sum of the exponential operation result and 1 to obtain the corresponding layered structure feature.

[0057] Arrange all the inter-layer structure features corresponding to the layer sequence positions in order of the layer sequence index, and output the inter-layer structure feature tensor.

[0058] Optionally, the layer-by-layer penetration mechanism specifically includes:

[0059] For any layer sequence position i, select the connection modulation features corresponding to the i-th layer, the (i-1)-th layer and the (i+1)-th layer at the same spatial position;

[0060] When the (i-1)th or (i+1)th layer sequence position does not exist, the connection modulation feature of the i-th layer is used as the connection modulation feature of the (i-1)th or (i+1)th layer in the calculation. The feature values ​​of the connection modulation features of the three selected layer sequence positions are averaged to obtain the path average feature of the corresponding layer sequence position i.

[0061] Subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i-1)-th layer, and subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i+1)-th layer. Then, average the two results to obtain the path change features at the corresponding layer position i.

[0062] The path average feature and path variation feature are concatenated along the channel dimension to obtain the sequence enhancement feature at sequence position i. The sequence enhancement features of all sequence positions are arranged in sequence index order to obtain the sequence enhancement feature tensor.

[0063] Optionally, step seven specifically includes:

[0064] Based on the tensor of the interlayer structure features, a layer-by-layer peeling process is performed in the thickness direction of the slab. According to the order of the layer sequence index from the surface layer to the core layer, the interlayer structure features corresponding to each layer sequence position are extracted sequentially, and the interlayer structure features corresponding to each layer sequence position are arranged in the order of the layer sequence index to generate a structural sequence from the surface layer to the core layer.

[0065] For any spatial location, extract the feature values ​​of the inter-layer structure features corresponding to each layer sequence position in the structural sequence, and arrange them according to the layer sequence index order to form the layer sequence feature sequence corresponding to the spatial location.

[0066] To continuously track structural changes at the same spatial location in different layers, the feature values ​​corresponding to the previous layer sequence position in the corresponding sequence feature sequence are subtracted from the feature values ​​corresponding to the next layer sequence position in the sequence feature sequence, and the absolute value of the subtraction result is taken to obtain the sequence of interlayer change values.

[0067] The numerical sequence of inter-layer changes is arranged in stratum index order, and the numerical sequence of inter-layer changes corresponding to each spatial location is taken as the structural change trajectory of the corresponding spatial location.

[0068] By summarizing the structural change trajectories of all spatial locations, we obtain the set of inter-layer structural change trajectories.

[0069] Optionally, step eight specifically includes:

[0070] Based on the set of interlayer structural change trajectories, the structural change trajectory corresponding to each spatial location is read point by point, and the interlayer change values ​​corresponding to each layer sequence position in the structural change trajectory are extracted.

[0071] In the structural change trajectory, the inter-layer change value corresponding to the next sequence position is compared with the inter-layer change value corresponding to the previous sequence position. When the inter-layer change value corresponding to the next sequence position is greater than the inter-layer change value corresponding to the previous sequence position, the corresponding spatial position is marked as an abnormal change position.

[0072] For spatial locations marked as abnormal changes, a continuous search is performed in the stratum sequence direction according to the stratum index order. The locations of adjacent stratum sequence positions that satisfy the inter-stratum change value corresponding to the next stratum sequence position is greater than the inter-stratum change value corresponding to the previous stratum sequence position are connected to form a structural intrusion path.

[0073] All structural intrusion paths are summarized according to their spatial coordinates, and the spatial locations where structural intrusion paths exist are marked as mixed-layer locations;

[0074] Arrange all mixed-layer locations according to spatial coordinates to generate the mixed-layer spatial distribution result.

[0075] The beneficial effects of this invention are:

[0076] This invention addresses the challenges of accurately representing the internal structural continuity and stably identifying layer mixing phenomena during particleboard forming by collaboratively constructing a continuous topological field of interlayer structure and an improved RepLKNet model. In the data representation stage, a structure modeling method based on voxel connectivity is introduced to uniformly represent the spatial distribution and interlayer connectivity within the slab, transforming discrete information into a continuous topological structure. In the feature extraction stage, a sequence-enhanced feature construction module in the improved RepLKNet model incorporates a sequence penetration mechanism, extending the original large-kernel convolution calculation based solely on two-dimensional space to the sequence space, enabling cross-layer... The joint feature calculation of sequence positions enables the interlayer structure feature tensor to possess interlayer structure transfer capability while maintaining spatial resolution. Furthermore, in the sequence jump compression module, the combination of cross-sequence difference calculation and averaging calculation allows the model to simultaneously possess the dual response capability to structural abrupt changes and structural continuity regions. In the structural analysis stage, through layer-by-layer peeling processing and the construction of interlayer structural change trajectory sets, the complex three-dimensional structural change process is transformed into a traceable trajectory form. In the identification of abnormal change locations and the analysis of structural intrusion paths, discrete anomalies are transformed into continuous paths, achieving spatial localization and determination of the expansion direction of mixed-layer regions. Through the above technical means, this invention can significantly improve the ability to express the continuity and abrupt changes of interlayer structures while maintaining spatial resolution, improve the accuracy and stability of mixed-layer identification, and enhance the adaptability to structural changes in complex production environments. Attached Figure Description

[0077] 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:

[0078] Figure 1 This is an overall flowchart of a machine vision-based particleboard forming monitoring method proposed in this invention.

[0079] Figure 2 This is a flowchart illustrating the construction of the interlayer structural continuity topological field in a machine vision-based particleboard forming monitoring method proposed in this invention.

[0080] Figure 3 This is a schematic diagram of the improved RepLKNet model structure for a machine vision-based particleboard forming monitoring method proposed in this invention. Detailed Implementation

[0081] 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.

[0082] refer to Figures 1-3 A machine vision-based method for monitoring particleboard forming includes the following steps:

[0083] Step 1: Install an online X-ray flaw detection device along the particleboard laying and conveying path to continuously scan the slab in the forming stage and obtain a sequence of transmission images.

[0084] Step 2: Perform energy spectrum response decoupling and fusion processing sequentially on the transmission image sequence to construct a fused transmission image sequence;

[0085] Step 3: Stack the fused transmission image sequences according to spatial correspondence to form a three-dimensional data volume inside the slab;

[0086] Step 4: In the three-dimensional data volume inside the slab, establish structural nodes based on voxels as basic units, establish connection relationships, and construct the interlayer structural continuity topological field;

[0087] Step 5: Perform a sequence expansion of the interlayer structural continuity topological field along the thickness direction of the slab to generate a set of structural slices that maintain the interlayer correspondence, forming a sequence expansion tensor set;

[0088] Step 6: Input the layer-order unfolded tensor group into the improved RepLKNet model. The improved RepLKNet model includes an input embedding module, a layer-order enhanced feature construction module, a layer-order skip compression module, and an output encoding module. The layer-order enhanced feature construction module embeds a layer-order penetration mechanism and outputs a layer-by-layer structure feature tensor.

[0089] Step 7: Based on the feature tensor of the layered structure, perform layer-by-layer peeling process according to the thickness direction of the slab to generate a structural sequence from the surface layer to the core layer, and continuously track the structural changes in different layers at the same spatial location to generate a set of interlayer structural change trajectories.

[0090] Step 8: Identify the locations of abnormal changes based on the set of interlayer structural change trajectories, and perform structural intrusion path expansion analysis to generate mixed-layer spatial distribution results.

[0091] In this embodiment, step one specifically includes:

[0092] An online X-ray flaw detection device is arranged below the particleboard paving and conveying path along the conveying direction. The online X-ray flaw detection device includes an X-ray emission source and a linear array detector, with the linear array detector located on the opposite side of the X-ray emission source.

[0093] As the slab in the forming stage moves along the conveying path, the X-ray emission source continuously emits X-ray beams, which are then received by a linear array detector after passing through the slab.

[0094] The linear array detector samples the transmitted rays line by line to obtain transmission signal data corresponding to the internal structure of the slab.

[0095] A rotary encoder is installed on the conveyor belt drive shaft. Based on the displacement signal output by the rotary encoder and the sampling frequency of the linear array detector, the spatial positions corresponding to adjacent sampling rows are matched.

[0096] Subtract the reference transmission signal value at the corresponding position in the blank-free state from the transmission signal value at each sampling point, and divide the subtraction result by the reference transmission signal value to obtain the normalized transmission intensity value.

[0097] The normalized transmission intensity values ​​are arranged into transmission images according to the sampling row order, and stored sequentially according to the transport direction to generate a transmission image sequence.

[0098] In the specific implementation process, the X-ray emission source in the online X-ray flaw detection device is set to an operating voltage of 90kV and a tube current of 6mA. An energy spectrum-sensitive linear array detector is selected, with a pixel pitch of 0.2mm, capable of dividing the transmitted rays into energy ranges and outputting multi-channel signal data. The conveyor belt speed is set to 1.0m / s, and the rotary encoder resolution is set to 5000 pulses / second. Through synchronous matching of the encoder displacement signal and the sampling frequency, the spatial spacing between adjacent sampling rows is stabilized at 0.2mm, thus ensuring consistent longitudinal and transverse resolution. During transmission signal processing, reference transmission signal values ​​acquired in the absence of a slab are used as a unified benchmark, thereby preserving the relative intensity differences between different sampling rows and avoiding grayscale distortion caused by local normalization. The transmission image sequence generated through the above processing can realistically reflect the density changes and structural distribution within the slab, improving the accuracy and stability of topological field construction and mixed-layer identification.

[0099] In this embodiment, step two specifically includes:

[0100] The energy spectrum response decoupling process is performed on the transmission image sequence. Using the linear array detector in the online X-ray flaw detection device, the transmission rays are divided into three energy ranges according to their energy magnitude: low energy range, medium energy range, and high energy range. The low energy range corresponds to an energy range of 20keV to 40keV, the medium energy range corresponds to an energy range of 40keV to 60keV, and the high energy range corresponds to an energy range of 60keV to 90keV.

[0101] The pixels of each sampling row in the transmission image sequence are processed point by point according to their spatial location. The normalized transmission intensity values ​​corresponding to the same spatial location in the low-energy region, medium-energy region and high-energy region are extracted respectively and classified according to the energy range.

[0102] The normalized transmission intensity values ​​belonging to the low-energy region are arranged according to the original sampling row order and pixel position to generate a low-energy region transmission image. The normalized transmission intensity values ​​belonging to the medium-energy region are arranged according to the original sampling row order and pixel position to generate a medium-energy region transmission image. The normalized transmission intensity values ​​belonging to the high-energy region are arranged according to the original sampling row order and pixel position to generate a high-energy region transmission image.

[0103] The transmission images of the low-energy region, the medium-energy region, and the high-energy region are fused. The normalized transmission intensity value of the low-energy region is added to the normalized transmission intensity value of the medium-energy region and then divided by two to obtain the first intermediate value. The first intermediate value is then added to the normalized transmission intensity value of the high-energy region and then divided by two to obtain the fused transmission intensity value. The normalized transmission intensity value of the high-energy region is used to enhance the structural representation ability of the high-density region during the fusion process.

[0104] The fused transmission intensity values ​​are arranged according to the original sampling row order and pixel position to generate a fused transmission image. Each fused transmission image is then numbered consecutively and stored sequentially according to the sampling order corresponding to the transport direction to construct a fused transmission image sequence.

[0105] In the specific implementation process, the linear array detector is equipped with multiple independent energy spectrum channels, each corresponding to a fixed energy range, thus ensuring clear distinctions between different energy ranges. During energy spectrum data processing, data from multiple energy ranges at the same spatial location are combined using a point-to-point correspondence method, maintaining the spatial correspondence between energy ranges. During the generation of the fused transmission image, data is combined in a fixed order for each pixel position, ensuring that the fusion result maintains a consistent numerical trend across the entire image. Through these settings, the fused transmission image maintains the continuity of the overall structure while improving the contrast stability between regions of different densities and reducing local fluctuations caused by differences in energy response, thereby enhancing the stability of interlayer variation identification during structural analysis.

[0106] In this embodiment, step three specifically includes:

[0107] The fused transmission image sequence is read in chronological order according to the transport direction, and the time interval between adjacent fused transmission images is converted into the corresponding spatial interval;

[0108] Based on the correspondence between the conveyor belt speed and the sampling frequency, each fused transmission image is assigned a corresponding spatial position coordinate in the conveying direction, and the fused transmission images are arranged in sequence according to the spatial position coordinates.

[0109] Using the pixels in each fused transmission image as two-dimensional plane coordinates and the spatial position coordinates of the corresponding fused transmission image as the third-dimensional coordinates, a three-dimensional coordinate mapping is performed on each pixel.

[0110] According to the three-dimensional coordinate mapping relationship, the fused transmission images are arranged sequentially according to their spatial positions, and the distance relationship between adjacent fused transmission images is determined based on the spatial interval.

[0111] The superimposed three-dimensional coordinate points and their corresponding fused transmission intensity values ​​are uniformly organized, and the fused transmission intensity value corresponding to each three-dimensional coordinate position is used as a voxel value to construct a three-dimensional data volume inside the slab composed of multiple three-dimensional coordinate points.

[0112] In the specific implementation process, when spatially reconstructing the fused transmission image sequence, the continuously acquired fused transmission images are first cached in chronological order into an image queue of 2048 frames, with each frame corresponding to one linear array scan result. The spatial interval between two adjacent frames in the transport direction is determined to be 0.2 mm, and this interval is used as the interlayer distance in the 3D reconstruction. For each fused transmission image, a planar coordinate system is established according to a pixel resolution of 0.2 mm × 0.2 mm. The horizontal position of the pixel is defined as the X coordinate, the vertical position as the Y coordinate, and the spatial position coordinate of the corresponding frame is used as the Z coordinate, thereby constructing a 3D coordinate index. During the data organization process, the frames are arranged in ascending order of the Z coordinate, and the fused transmission intensity value of each pixel is written into the corresponding 3D coordinate position to form a regular 3D grid structure. The final 3D data volume has a resolution of 0.2 mm in the X, Y, and Z directions, achieving isotropic spatial sampling. This structure can continuously reflect the changes in density distribution inside the slab, providing a stable and consistent data foundation for the analysis of the continuity of the interlayer structure.

[0113] In this embodiment, step four specifically includes:

[0114] In the three-dimensional data volume inside the slab, each voxel corresponding to a three-dimensional coordinate position is used as a structural node, and the fused transmission intensity value is used as the attribute value of the structural node.

[0115] Within the 3D data body, for each structural node, spatially adjacent voxels are selected as adjacent nodes, and the adjacency range is limited to directly adjacent voxels in the X, Y and Z directions.

[0116] Subtract the attribute values ​​between each structural node and its corresponding adjacent node, and take the absolute value of the subtraction result to obtain the voxel intensity difference between adjacent voxels.

[0117] The voxel intensity differences between all adjacent voxels are summarized and statistically analyzed. The voxel intensity differences are sorted from smallest to largest, and the median of the sorting results is selected as the criterion for structural similarity judgment.

[0118] When the difference in voxel strength between a structural node and its adjacent node is less than the structural similarity criterion, a connection relationship is established between the structural node and its adjacent node.

[0119] The process of establishing connection relationships for all structural nodes is repeated to form a spatial network structure composed of multiple structural nodes and connection relationships, forming a connection structure distributed along the thickness direction of the slab, and constructing a continuous topological field of interlayer structure.

[0120] In the specific implementation process, the median of the voxel intensity difference distribution is used as the judgment benchmark, and the connection relationship is adaptively determined according to the overall change characteristics of the actual data, thereby avoiding the misjudgment problem caused by fixed thresholds. At the same time, the connection structure established by combining the three-dimensional spatial adjacency relationship effectively enhances the expressive ability of the continuous region inside the slab and makes the density abrupt change region clearly distinguishable, making the transition position of the interlayer structure clearer. By constructing the connection structure distributed along the thickness direction, it is beneficial to highlight the structural continuity and fracture characteristics between different layers, thereby improving the sensitivity and discrimination accuracy of abnormal structures in the mixed layer identification process.

[0121] In this embodiment, step five specifically includes:

[0122] The interlayer structural continuity topological field is processed in layers along the thickness direction of the slab. The set of structural nodes corresponding to each thickness position is extracted in ascending order of Z coordinate, and a structural slice is formed by the set of structural nodes at each thickness position.

[0123] Arrange the structural slices sequentially according to the thickness direction of the slab to generate a structural slice group;

[0124] For two adjacent structural slices, for structural nodes at the same two-dimensional plane coordinate position, it is determined whether there is a connection relationship in the Z direction. The position with a connection relationship is marked with a value of 1, and the position without a connection relationship is marked with a value of 0, forming an inter-layer connection mark matrix with the same size as the structural slice.

[0125] Arrange the structural node attribute values ​​in each layer of structural slice into a two-dimensional matrix according to the two-dimensional plane coordinate order, and arrange the corresponding inter-layer connection mark matrix according to the same two-dimensional plane coordinate order.

[0126] Following the order of the slab thickness direction, the structural slice matrix of each layer and the corresponding interlayer connection mark matrix are arranged alternately. After the structural slice matrix of the i-th layer, the interlayer connection mark matrix between the i-th layer and the (i+1)-th layer is arranged to form a layer sequence expansion tensor group.

[0127] In the specific implementation, the sequence unfolding tensor group is stored using a regular three-dimensional array structure. The first dimension represents the sequence index along the thickness direction of the slab, and its length is the sum of the number of structural slices and the number of interlayer connection marker matrices. The second and third dimensions correspond to the horizontal and vertical pixel coordinates of the fused transmission image, respectively, with a resolution of 0.2 mm. For any i-th sequence position, when the index is even, the fused transmission intensity value at the corresponding position in the i-th layer structural slice matrix is ​​stored; when the index is odd, the 0 or 1 marker value at the corresponding position in the interlayer connection marker matrix between the i-th and (i+1)-th layers is stored, thus forming an alternating arrangement of structural layer-connection layer-structural layer. Through this organization, the node attribute information and interlayer connection relationships in the original three-dimensional topology are explicitly expressed in a unified data structure, allowing the structural continuity in the thickness direction to be directly read and calculated in sequence form. This is beneficial for the model to continuously model the changes in interlayer structure while maintaining spatial position information, improving the recognition accuracy and stability of mixed-layer regions.

[0128] In this embodiment, step six specifically includes:

[0129] The hierarchical unfolded tensor group is input into the input embedding module. The structural slice matrix and the inter-layer connection label matrix are mapped to the same spatial position. The value of each position in the structural slice matrix is ​​multiplied with the value of the corresponding position in the inter-layer connection label matrix. The product result is added to the value of the corresponding position in the structural slice matrix to obtain the connection modulation feature. The feature is then arranged according to the hierarchical index order to generate the hierarchical embedding feature tensor.

[0130] The hierarchical embedding feature tensor is input into the hierarchical enhancement feature construction module, which introduces a hierarchical penetration mechanism to obtain the hierarchical enhancement feature tensor.

[0131] The sequence augmentation feature tensor is input into the sequence skip compression module, which performs cross-sequence difference calculation on the sequence augmentation features at each sequence position along the sequence index direction.

[0132] For any sequence position i, select the sequence enhancement features corresponding to the i-th and i+2-th layers at the same spatial position. When the i+2-th sequence position does not exist, use the sequence enhancement features of the i-th sequence position as the substitute value of the i+2-th sequence position in the calculation.

[0133] Subtract the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer and take the absolute value to obtain the inter-layer difference features at the corresponding sequence position i.

[0134] Add the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer, and then divide by two to obtain the inter-layer average feature at the corresponding sequence position i.

[0135] The inter-layer difference features and the inter-layer average features are concatenated along the channel dimension to form the compressed stratum sequence feature at stratum position i. The compressed stratum sequence features at all stratum positions are arranged in stratum sequence index order to generate a compressed stratum sequence feature tensor.

[0136] The compressed stratum feature tensor is input into the output encoding module, and nonlinear transformation is performed point by point on the feature values ​​of the compressed stratum feature corresponding to each stratum position in the compressed stratum feature tensor.

[0137] Perform an exponential operation on the feature value of each compressed sequence feature, subtract 1 from the exponential operation result, and divide by the sum of the exponential operation result and 1 to obtain the corresponding layered structure feature.

[0138] Arrange all the inter-layer structure features corresponding to the layer sequence positions in order of layer sequence index, and output the inter-layer structure feature tensor.

[0139] In this invention, the improved RepLKNet model is obtained by expanding the layer-order dimension and reconstructing the structure based on the original RepLKNet large-kernel convolutional structure. The traditional RepLKNet model mainly performs large-kernel convolution calculations on two-dimensional images. Its convolutional kernel has a large receptive field in the spatial dimension and achieves feature extraction by fusing local and global information within a single layer plane, without involving cross-layer information modeling along the thickness direction. While maintaining the advantages of the original large-kernel convolution spatial receptive field, this invention introduces a layer-order penetration mechanism, which expands the feature calculation from two-dimensional space to a three-dimensional structure in layer-order space. The original feature extraction process, which only occurs within the same layer, is transformed into a joint calculation process across layer-order positions. The sampling unit in the convolution operation is expanded from a single-layer plane to a multi-layer combination unit that includes adjacent layer-order positions. By constructing path average features and path change features, the model simultaneously possesses the ability to express structural continuity and the ability to perceive structural abrupt changes.

[0140] Through the aforementioned structural improvements, the model, when processing data from the particleboard forming process, not only maintains the sensitivity of the original large-kernel convolution to large-scale structures but also explicitly characterizes the transition relationship and anomalous change regions between the core and surface layers, thereby enhancing its response to layer mixing phenomena. Compared to the traditional RepLKNet model, the improved RepLKNet model establishes a stable feature transfer path in the sequence direction, making continuous structural regions exhibit smooth feature changes, while forming obvious feature differences in structural fractures or layer mixing regions, thus improving the accuracy and stability of layer mixing identification.

[0141] In this embodiment, the layer-by-layer penetration mechanism is specifically as follows:

[0142] For any layer sequence position i, select the connection modulation features corresponding to the i-th layer, the (i-1)-th layer and the (i+1)-th layer at the same spatial position;

[0143] When the (i-1)th or (i+1)th layer sequence position does not exist, the connection modulation feature of the i-th layer is used as the connection modulation feature of the (i-1)th or (i+1)th layer in the calculation. The feature values ​​of the connection modulation features of the three selected layer sequence positions are averaged to obtain the path average feature of the corresponding layer sequence position i.

[0144] Subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i-1)-th layer, and subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i+1)-th layer. Then, average the two results to obtain the path change features at the corresponding layer position i.

[0145] The path average feature and path variation feature are concatenated along the channel dimension to obtain the sequence enhancement feature at sequence position i. The sequence enhancement features of all sequence positions are arranged in sequence index order to obtain the sequence enhancement feature tensor.

[0146] In this invention, the sequence penetration mechanism jointly calculates the connectivity modulation features of adjacent sequence positions along the sequence direction, ensuring that the sequence enhancement features of each sequence position simultaneously contain structural information from its own layer, the layer above, and the layer below. This introduces a continuous description along the thickness direction into the feature representation. By performing an averaging operation on the connectivity modulation features to obtain path-averaged features, it can stably represent regions with gradual changes in the interlayer structure, reducing the impact of local random fluctuations on the results. By performing difference calculations on adjacent sequence positions to obtain path change features, it can highlight locations of abrupt changes in the interlayer structure, making density discontinuities more apparent in specific areas. A clear response is formed in the feature space; after the two are combined, a composite feature expression with both stability and sensitivity is formed at the same sequence position; further, at the boundary sequence position, the current layer feature is used to replace the missing sequence feature in the calculation, so that all sequence positions meet the unified calculation structure and ensure the consistency of feature distribution in the sequence direction; through the above processing, the sequence penetration mechanism can form smooth transition features in continuous structural regions and abrupt feature responses in mixed-layer or structurally anomalous regions, thereby improving the ability to characterize the continuous changes in interlayer structure and providing a stable and discriminative feature basis for layer-by-layer peeling and structural trajectory tracking.

[0147] In this embodiment, step seven specifically includes:

[0148] Based on the layered structure feature tensor, layer-by-layer peeling is performed according to the thickness direction of the slab. The layered structure features corresponding to each layer position are extracted sequentially according to the layer sequence index from the surface layer to the core layer. The layered structure features corresponding to each layer position are arranged according to the layer sequence index to generate a structure sequence from the surface layer to the core layer.

[0149] For any spatial location, extract the feature values ​​of the inter-layer structure features corresponding to each layer sequence position in the structural sequence, and arrange them according to the layer sequence index order to form the layer sequence feature sequence corresponding to the spatial location.

[0150] To continuously track structural changes at the same spatial location in different layers, the feature values ​​corresponding to the previous layer sequence position in the corresponding sequence feature sequence are subtracted from the feature values ​​corresponding to the next layer sequence position in the sequence feature sequence, and the absolute value of the subtraction result is taken to obtain the sequence of interlayer change values.

[0151] The numerical sequence of inter-layer changes is arranged in stratum index order, and the numerical sequence of inter-layer changes corresponding to each spatial location is taken as the structural change trajectory of the corresponding spatial location.

[0152] By summarizing the structural change trajectories of all spatial locations, we obtain the set of inter-layer structural change trajectories;

[0153] This invention, by performing layer-by-layer peeling of the feature tensor of the interlayer structure and constructing a structural sequence, enables the structural distribution from the surface layer to the core layer inside the slab to be expressed in a sequential form, which is beneficial for grasping the continuous change law of the interlayer structure as a whole. Furthermore, by constructing the sequence feature sequence at the same spatial location and performing continuous difference calculation, subtle changes between layers can be amplified layer by layer, thereby strengthening the response of structurally abrupt regions at the feature level. The structural change trajectory formed on this basis can intuitively reflect the transition relationship between different layers, making continuous regions appear as smooth change trajectories, while mixed-layer or abnormal regions appear as abrupt trajectories. By summarizing the structural change trajectories at all spatial locations, a set of interlayer structural change trajectories with spatial continuity can be formed, thereby improving the positioning accuracy of mixed-layer locations and enhancing the adaptability to complex internal structural changes.

[0154] In this embodiment, step eight specifically includes:

[0155] Based on the set of inter-layer structural change trajectories, the structural change trajectory corresponding to each spatial location is read point by point, and the inter-layer change values ​​corresponding to each layer sequence position in the structural change trajectory are extracted.

[0156] In the structural change trajectory, the inter-layer change value corresponding to the next sequence position is compared with the inter-layer change value corresponding to the previous sequence position. When the inter-layer change value corresponding to the next sequence position is greater than the inter-layer change value corresponding to the previous sequence position, the corresponding spatial position is marked as an abnormal change position.

[0157] For spatial locations marked as abnormal changes, a continuous search is performed in the stratum sequence direction according to the stratum index order. The locations of adjacent stratum sequence positions that satisfy the inter-stratum change value corresponding to the next stratum sequence position is greater than the inter-stratum change value corresponding to the previous stratum sequence position are connected to form a structural intrusion path.

[0158] All structural intrusion paths are summarized according to their spatial coordinates, and the spatial locations where structural intrusion paths exist are marked as mixed-layer locations;

[0159] Arrange all mixed-layer locations according to spatial coordinates to generate the mixed-layer spatial distribution result;

[0160] The interlayer variation value is calculated from the difference in the interlayer structural characteristics of adjacent sequence positions. Its physical meaning corresponds to the intensity of the change in the internal structure of the slab in the thickness direction. When the structure is in a normal paving state, the density and particle distribution between each layer have a gradual relationship, and the interlayer variation value remains relatively stable or fluctuates slowly in the sequence direction. However, when mixed layering occurs, the core layer or surface structure invades the adjacent layer. After the local structural difference appears in the initial layer, it will gradually expand to the subsequent layers as the invasion process occurs, so that the interlayer variation value shows an increasing trend from small to large. Therefore, using the interlayer variation value corresponding to the next sequence position being greater than the interlayer variation value corresponding to the previous sequence position as the criterion can effectively screen out areas where the structural change is continuously enhanced, thereby eliminating the interference of random noise or isolated abrupt change points.

[0161] Furthermore, by continuously connecting positions that satisfy the incremental condition along the stratigraphic sequence, discrete anomalous change points can be transformed into directional structural intrusion paths, allowing the process of structural expansion from one layer to another to be expressed in path form. This processing method can simultaneously reflect the intensity and direction of structural change, making the mixed-layer region spatially continuous rather than scattered, thereby improving the stability and positioning accuracy of mixed-layer identification and enhancing its adaptability to complex internal structural changes.

[0162] Example 1: To verify the feasibility of this invention in practice, it was applied to the paving and forming section of a continuous flat-press particleboard production line with an annual output of 220,000 m³. The production line has a slab width of 1.83 m, a designed thickness of 18 mm, and a stable conveying speed of 1.0 m / s. It employs a typical three-layer paving structure, with fine particleboard on the top and bottom surfaces and coarse particleboard in the core layer. During actual production, due to airflow disturbances and uneven material distribution during paving, the coarse particleboard in the core layer tends to float while the fine particleboard on the surface sinks, resulting in a mixed-layer structure. This leads to fluctuations in the internal bonding strength of the board within ±16%, accompanied by defects such as localized delamination and blistering. Existing detection methods mainly rely on manual sampling or surface visual identification, which cannot continuously monitor structural changes in the thickness direction of the slab. The mixed-layer problem is usually only discovered after hot pressing, causing significant economic losses.

[0163] In this embodiment, a continuous transmission scan of the slab is performed using an online X-ray flaw detection device to obtain a transmission image sequence. The transmission image sequence is then subjected to energy spectrum response decoupling and fusion processing to construct a fused transmission image sequence. Subsequently, the images are stacked according to spatial correspondence to form a three-dimensional data volume inside the slab. Structural nodes and connections are established within the three-dimensional data volume to construct an interlayer structural continuity topological field. A sequence unfolding is then performed along the thickness direction to form a sequence unfolding tensor set. This sequence unfolding tensor set is input into an improved RepLKNet model. Through the sequence penetration mechanism in the sequence enhancement feature construction module, cross-layer feature modeling is achieved, outputting a trans-layer structural feature tensor. Further, a layer-by-layer peeling process is performed on the trans-layer structural feature tensor, and continuous tracking is performed at the same spatial location to generate a set of interlayer structural change trajectories. Finally, based on the structural change trajectory set, abnormal change locations are identified, and the spatial distribution results of the mixed-layer area are generated through structural intrusion path unfolding analysis, achieving online positioning of the mixed-layer region.

[0164] To verify the actual effect of the present invention, three comparison schemes were set up. Comparison scheme one is surface vision + manual sampling; comparison scheme two is single-energy X-ray imaging + two-dimensional convolution recognition; comparison scheme three is interlayer structural continuity topological field + traditional RepLKNet model scheme. The method of the present invention uses interlayer structural continuity topological field and improved RepLKNet model. The comparison results are shown in Table 1.

[0165] Table 1. Statistical Comparison of Mixed-Layer Recognition Performance

[0166] index Comparison Option 1 Comparison Option 2 Comparison Option 3 Method of the present invention Mixed-layer recognition accuracy / % 77.8 86.5 91.2 97.6 False negative rate / % 20.5 12.4 7.8 2.6 False positive rate / % 10.8 8.1 6.2 3.3 Positioning error / mm 15.6 9.7 5.4 2.5 Consistency rate with cut sample / % 76.2 85.3 90.8 97.1

[0167] The indicators in Table 1 above quantify the mixed-layer identification capability from different perspectives. The mixed-layer identification accuracy rate reflects the overall correct judgment rate and is a core indicator for measuring the comprehensive performance of the system. The present invention achieves 97.6%, which is significantly higher than the comparison scheme, indicating that it can more accurately distinguish between normal areas and mixed-layer areas under complex structural conditions. The false negative rate represents the proportion of mixed-layer areas that actually exist but are not identified. The lower the indicator, the stronger the ability to capture anomalies. The false negative rate of the present invention is 2.6%, indicating that it can effectively identify the vast majority of mixed-layer situations and reduce the risk of hidden dangers entering subsequent processes. The false positive rate represents the proportion of normal areas that are mistakenly identified as mixed-layer areas. This indicator directly affects the effectiveness of production intervention. The present invention controls the rate at 3.3%, indicating fewer false alarms and helping to avoid unnecessary process adjustments.

[0168] Positioning error measures the spatial deviation between the identified mixed-layer position and the actual position. This invention reduces the error to 2.5mm, indicating high positioning accuracy in both the thickness and planar directions, providing a precise basis for subsequent quality control. The consistency rate with the cut sample reflects the degree of consistency between the online detection results and the actual detection results after hot pressing. A higher rate indicates that the online monitoring results are closer to the true structural state. This invention achieves 97.1%, indicating high reliability of the detection results. Overall, these indicators show that this invention not only has a significant advantage in overall identification capability but also demonstrates a high level of performance in anomaly detection, misjudgment control, and spatial positioning, thereby achieving stable identification and accurate positioning of the mixed-layer structure inside particleboard.

[0169] This embodiment constructs a continuous topological field of interlayer structure and combines it with an improved RepLKNet model to achieve continuous modeling and dynamic analysis of the internal structure in the thickness direction during particleboard forming. This effectively expresses the previously difficult-to-observe interlayer change process. Through layer-by-layer peeling and continuous tracking of structural changes, complex three-dimensional structural changes are transformed into analyzable trajectory forms. Furthermore, analysis of structural intrusion paths enables accurate identification and spatial positioning of mixed-layer regions. This invention not only detects mixed-layer problems in advance but also clarifies their expansion direction and change trend, thus providing a reliable basis for real-time control during the production process and improving the stability and consistency of particleboard forming quality.

[0170] 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 monitoring particleboard forming based on machine vision, characterized in that, Includes the following steps: Step 1: Install an online X-ray flaw detection device along the particleboard laying and conveying path to continuously scan the slab in the forming stage and obtain a sequence of transmission images. Step 2: Perform energy spectrum response decoupling and fusion processing sequentially on the transmission image sequence to construct a fused transmission image sequence; Step 3: Stack the fused transmission image sequence according to spatial correspondence to form a three-dimensional data volume inside the slab; Step 4: In the three-dimensional data volume inside the slab, establish structural nodes based on voxels as basic units, establish connection relationships, and construct the interlayer structural continuity topological field; Step 5: Perform a sequence unfolding of the interlayer structural continuity topological field along the slab thickness direction to generate a structural slice group that maintains the interlayer correspondence, forming a sequence unfolding tensor group; Step 6: Input the layer-order unfolded tensor group into the improved RepLKNet model. The improved RepLKNet model includes an input embedding module, a layer-order enhanced feature construction module, a layer-order skip compression module, and an output encoding module. The layer-order enhanced feature construction module embeds a layer-order penetration mechanism and outputs a layer-by-layer structure feature tensor. Step 7: Based on the layered structure feature tensor, perform layer-by-layer peeling process according to the thickness direction of the slab to generate a structural sequence from the surface layer to the core layer, and continuously track the structural changes in different layers at the same spatial location to generate a set of interlayer structural change trajectories. Step 8: Identify the locations of abnormal changes based on the set of interlayer structural change trajectories, and perform structural intrusion path expansion analysis to generate mixed-layer spatial distribution results.

2. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step one specifically involves: An online X-ray flaw detection device is arranged below the particleboard laying and conveying path along the conveying direction. The online X-ray flaw detection device includes an X-ray emission source and a linear array detector, with the linear array detector located on the opposite side of the X-ray emission source. During the process of the slab moving along the conveying path in the forming stage, the X-ray emission source continuously emits a beam of X-rays, which is received by the linear array detector after passing through the slab. The linear array detector samples the transmitted rays line by line to obtain transmission signal data corresponding to the internal structure of the slab. A rotary encoder is installed on the conveyor belt drive shaft. Based on the displacement signal output by the rotary encoder and the sampling frequency of the linear array detector, the spatial positions corresponding to adjacent sampling rows are matched. Subtract the reference transmission signal value at the corresponding position in the blank-free state from the transmission signal value at each sampling point, and divide the subtraction result by the reference transmission signal value to obtain the normalized transmission intensity value. The normalized transmission intensity values ​​are arranged into a transmission image according to the sampling row order, and stored sequentially by number according to the transport direction to generate a transmission image sequence.

3. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step two specifically involves: The transmitted image sequence is subjected to energy spectrum response decoupling processing. Using the linear array detector in the online X-ray flaw detection device, the transmitted rays are divided into three energy ranges according to their energy magnitude: low energy range, medium energy range, and high energy range. The pixels of each sampling row in the transmission image sequence are processed point by point according to their spatial location. The normalized transmission intensity values ​​corresponding to the same spatial location in the low-energy region, medium-energy region and high-energy region are extracted respectively and classified according to the energy range. The normalized transmission intensity values ​​belonging to the low-energy region are arranged according to the original sampling row order and pixel position to generate a low-energy region transmission image. The normalized transmission intensity values ​​belonging to the medium-energy region are arranged according to the original sampling row order and pixel position to generate a medium-energy region transmission image. The normalized transmission intensity values ​​belonging to the high-energy region are arranged according to the original sampling row order and pixel position to generate a high-energy region transmission image. The low-energy region transmission image, the medium-energy region transmission image, and the high-energy region transmission image are fused. The normalized transmission intensity value of the low-energy region is added to the normalized transmission intensity value of the medium-energy region and then divided by two to obtain a first intermediate value. The first intermediate value is added to the normalized transmission intensity value of the high-energy region and then divided by two to obtain the fused transmission intensity value. The fused transmission intensity values ​​are arranged according to the original sampling row order and pixel position to generate a fused transmission image. Each fused transmission image is then continuously numbered and stored sequentially according to the sampling order corresponding to the transport direction to construct a fused transmission image sequence.

4. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step three specifically involves: The fused transmission image sequence is read in chronological order according to the transport direction, and the time interval between adjacent fused transmission images is converted into the corresponding spatial interval; Based on the correspondence between the conveyor belt speed and the sampling frequency, each fused transmission image is assigned a corresponding spatial position coordinate in the conveying direction, and the fused transmission images are arranged in sequence according to the spatial position coordinates. Using the pixels in each fused transmission image as two-dimensional plane coordinates and the spatial position coordinates of the corresponding fused transmission image as the third-dimensional coordinates, a three-dimensional coordinate mapping is performed on each pixel. According to the three-dimensional coordinate mapping relationship, the fused transmission images are arranged sequentially according to their spatial positions, and the distance relationship between adjacent fused transmission images is determined based on the spatial interval. The superimposed three-dimensional coordinate points and their corresponding fused transmission intensity values ​​are organized in a unified manner. The fused transmission intensity value corresponding to each three-dimensional coordinate position is used as a voxel value to construct a three-dimensional data volume inside the slab composed of multiple three-dimensional coordinate points.

5. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step four specifically involves: In the three-dimensional data volume inside the slab, each voxel corresponding to a three-dimensional coordinate position is used as a structural node, and the fused transmission intensity value is used as the attribute value of the structural node. Within the 3D data body, for each structural node, spatially adjacent voxels are selected as adjacent nodes, and the adjacency range is limited to directly adjacent voxels in the X, Y and Z directions. Subtract the attribute values ​​between each structural node and its corresponding adjacent node, and take the absolute value of the subtraction result to obtain the voxel intensity difference between adjacent voxels. The voxel intensity differences between all adjacent voxels are summarized and statistically analyzed. The voxel intensity differences are sorted from smallest to largest, and the median of the sorting results is selected as the criterion for structural similarity judgment. When the difference in voxel intensity between a structural node and its adjacent node is less than the structural similarity criterion, a connection relationship is established between the structural node and its adjacent node. The process of establishing connection relationships for all structural nodes is repeated to form a spatial network structure composed of multiple structural nodes and connection relationships, forming a connection structure distributed along the thickness direction of the slab, and constructing a continuous topological field of interlayer structure.

6. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step five specifically involves: The interlayer structural continuity topological field is processed in layers along the thickness direction of the slab. The set of structural nodes corresponding to each thickness position is extracted in ascending order of Z coordinate, and a structural slice is formed by the set of structural nodes at each thickness position. Arrange the structural slices sequentially according to the thickness direction of the slab to generate a structural slice group; For two adjacent structural slices, for structural nodes at the same two-dimensional plane coordinate position, it is determined whether there is a connection relationship in the Z direction. The position with a connection relationship is marked with a value of 1, and the position without a connection relationship is marked with a value of 0, forming an inter-layer connection mark matrix with the same size as the structural slice. Arrange the structural node attribute values ​​in each layer of structural slice into a two-dimensional matrix according to the two-dimensional plane coordinate order, and arrange the corresponding inter-layer connection mark matrix according to the same two-dimensional plane coordinate order. Following the order of the slab thickness, the structural slice matrix of each layer and the corresponding interlayer connection mark matrix are arranged alternately. After the structural slice matrix of the i-th layer, the interlayer connection mark matrix between the i-th layer and the (i+1)-th layer is arranged to form a layer sequence expansion tensor group.

7. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step six specifically involves: The hierarchical unfolded tensor group is input into the input embedding module. The structural slice matrix and the inter-layer connection label matrix are mapped to the same spatial position. The value of each position in the structural slice matrix is ​​multiplied with the value of the corresponding position in the inter-layer connection label matrix. The product result is added to the value of the corresponding position in the structural slice matrix to obtain the connection modulation feature. The feature is then arranged according to the hierarchical index order to generate the hierarchical embedding feature tensor. The hierarchical embedding feature tensor is input into the hierarchical enhancement feature construction module, which introduces a hierarchical penetration mechanism to obtain the hierarchical enhancement feature tensor. The sequence augmentation feature tensor is input into the sequence skip compression module, which performs cross-sequence difference calculation on the sequence augmentation features at each sequence position along the sequence index direction. For any sequence position i, select the sequence enhancement features corresponding to the i-th and i+2-th layers at the same spatial position. When the i+2-th sequence position does not exist, use the sequence enhancement features of the i-th sequence position as the substitute value of the i+2-th sequence position in the calculation. Subtract the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer and take the absolute value to obtain the inter-layer difference features at the corresponding sequence position i. Add the feature values ​​of the sequence enhancement features of the i-th layer and the (i+2)-th layer, and then divide by two to obtain the inter-layer average feature at the corresponding sequence position i. The inter-layer difference features and the inter-layer average features are concatenated along the channel dimension to form the compressed sequence feature at sequence position i. Compressed stratigraphic features are generated by arranging all stratigraphic positions in stratigraphic index order; The compressed stratum feature tensor is input into the output encoding module, and nonlinear transformation is performed point by point on the feature values ​​of the compressed stratum feature corresponding to each stratum position in the compressed stratum feature tensor. Perform an exponential operation on the feature value of each compressed sequence feature, subtract 1 from the exponential operation result, and divide by the sum of the exponential operation result and 1 to obtain the corresponding layered structure feature. Arrange all the inter-layer structure features corresponding to the layer sequence positions in order of the layer sequence index, and output the inter-layer structure feature tensor.

8. The method for monitoring particleboard forming based on machine vision according to claim 7, characterized in that, The layer-by-layer penetration mechanism is specifically as follows: For any layer sequence position i, select the connection modulation features corresponding to the i-th layer, the (i-1)-th layer and the (i+1)-th layer at the same spatial position; When the (i-1)th or (i+1)th layer sequence position does not exist, the connection modulation feature of the i-th layer is used as the connection modulation feature of the (i-1)th or (i+1)th layer in the calculation. The feature values ​​of the connection modulation features of the three selected layer sequence positions are averaged to obtain the path average feature of the corresponding layer sequence position i. Subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i-1)-th layer, and subtract the absolute values ​​of the modulation features connecting the i-th layer and the (i+1)-th layer. Then, average the two results to obtain the path change features at the corresponding layer position i. The path average feature and the path variation feature are concatenated along the channel dimension to form the sequence enhancement feature at sequence position i. Arrange the sequence augmentation features of all sequence positions in sequence index order to obtain the sequence augmentation feature tensor.

9. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step seven specifically involves: Based on the tensor of the interlayer structure features, a layer-by-layer peeling process is performed in the thickness direction of the slab. According to the order of the layer sequence index from the surface layer to the core layer, the interlayer structure features corresponding to each layer sequence position are extracted sequentially, and the interlayer structure features corresponding to each layer sequence position are arranged in the order of the layer sequence index to generate a structural sequence from the surface layer to the core layer. For any spatial location, extract the feature values ​​of the inter-layer structure features corresponding to each layer sequence position in the structural sequence, and arrange them according to the layer sequence index order to form the layer sequence feature sequence corresponding to the spatial location. To continuously track structural changes at the same spatial location in different layers, the feature values ​​corresponding to the previous layer sequence position in the corresponding sequence feature sequence are subtracted from the feature values ​​corresponding to the next layer sequence position in the sequence feature sequence, and the absolute value of the subtraction result is taken to obtain the sequence of interlayer change values. The numerical sequence of inter-layer changes is arranged in stratum index order, and the numerical sequence of inter-layer changes corresponding to each spatial location is taken as the structural change trajectory of the corresponding spatial location. By summarizing the structural change trajectories of all spatial locations, we obtain the set of inter-layer structural change trajectories.

10. The method for monitoring particleboard forming based on machine vision according to claim 1, characterized in that, Step eight specifically involves: Based on the set of interlayer structural change trajectories, the structural change trajectory corresponding to each spatial location is read point by point, and the interlayer change values ​​corresponding to each layer sequence position in the structural change trajectory are extracted. In the structural change trajectory, the inter-layer change value corresponding to the next sequence position is compared with the inter-layer change value corresponding to the previous sequence position. When the inter-layer change value corresponding to the next sequence position is greater than the inter-layer change value corresponding to the previous sequence position, the corresponding spatial position is marked as an abnormal change position. For spatial locations marked as abnormal changes, a continuous search is performed in the stratum sequence direction according to the stratum index order. The locations of adjacent stratum sequence positions that satisfy the inter-stratum change value corresponding to the next stratum sequence position is greater than the inter-stratum change value corresponding to the previous stratum sequence position are connected to form a structural intrusion path. All structural intrusion paths are summarized according to their spatial coordinates, and the spatial locations where structural intrusion paths exist are marked as mixed-layer locations; Arrange all mixed-layer locations according to spatial coordinates to generate the mixed-layer spatial distribution result.

Images