A method and system for foreign object detection and handling in intelligent paper shredders that integrates visual recognition

Abstract

This invention provides a method and system for foreign object detection and handling in intelligent paper shredders that integrates visual recognition. It relates to the field of intelligent paper shredder technology. The method includes acquiring a sequence of target object images, performing spatiotemporal joint analysis to extract texture motion trajectories and obtain material flexibility descriptions to form spatiotemporal fusion features, constructing a multi-scale response map to locate the foreign object, calculating the force distribution based on the blade motion trajectory, determining the cumulative fatigue damage and crack density prediction values ​​of functional blocks, assessing the lifespan decay, and generating processing instructions. This invention improves the accuracy of foreign object detection and blade life prediction in paper shredders, reducing the risk of equipment damage.

Description

Technical Field

[0001] This invention relates to the field of intelligent paper shredder technology, and in particular to an intelligent paper shredder foreign object detection and handling method and system that integrates visual recognition. Background Technology

[0002] Paper shredders are common document destruction devices in office and home environments, used to shred paper documents to protect information security. Traditional paper shredders mainly rely on mechanical structures to cut and shred documents. With the development of intelligent technology, modern paper shredders have gradually incorporated sensors, intelligent control and other functions, enabling intelligent operation such as automatic paper feeding and automatic start and stop.

[0003] Foreign object detection is a crucial step in ensuring the safe operation of shredders during the shredding process. Foreign objects such as metal clips, paper clips, and plastic cards entering the shredder can damage the blades or cause machine malfunctions. Currently, foreign object detection in shredders primarily relies on current sensing, pressure sensing, or simple photoelectric sensors, monitoring changes in motor load or blockages in the paper feed path to determine the presence of foreign objects. Some high-end shredders are also equipped with metal detectors specifically designed to detect metallic foreign objects.

[0004] With the increasing prevalence of office automation, the frequency of use and the types of documents processed by paper shredders are becoming more diverse, placing higher demands on the accuracy and intelligence of foreign object detection. Existing technologies still have shortcomings in foreign object detection and handling. Current paper shredder foreign object detection mainly relies on a single sensor method, lacking the ability to comprehensively analyze the material characteristics of foreign objects. It is difficult to accurately distinguish different types of foreign objects, especially for non-metallic foreign objects such as plastic cards and tape, which has limited recognition capabilities and is prone to misjudgment or missed detection. Traditional foreign object detection methods cannot assess the potential damage to the shredder blades in real time, nor can they intelligently adjust processing strategies according to the characteristics of different foreign objects. They often rely on simple shutdown protection measures, reducing equipment efficiency. Existing technologies lack a precise assessment mechanism for the lifespan of the shredder blades, and cannot predict blade damage and plan maintenance based on the actual impact of foreign objects on the blades. This leads to premature blade failure or unnecessary increased maintenance costs, affecting the long-term stable operation of the equipment. Summary of the Invention

[0005] This invention provides a method and system for foreign object detection and handling in an intelligent paper shredder that integrates visual recognition, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for foreign object detection and handling in an intelligent paper shredder that integrates visual recognition, comprising:

[0007] Acquire a multi-frame image sequence of the target object in the paper feed channel;

[0008] Spatiotemporal joint analysis is performed on a multi-frame image sequence. The surface texture motion trajectory of the target object is extracted by optical flow field calculation. The material flexibility description is obtained by deformation tensor decomposition of the surface texture motion trajectory. The material flexibility description is then tensor-stitched with the image spatial gradient distribution to form a spatiotemporal fusion feature.

[0009] A multi-scale response map is constructed by pyramid decomposition of spatiotemporal fusion features. Anomaly responses are iteratively transmitted between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence position and intensity peak of the anomaly responses.

[0010] Geometric contours are extracted based on the location results of the foreign object area. A virtual contact scene is constructed in the spatial coordinate system by combining the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection.

[0011] The paper shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. The current crack density prediction value is determined by combining historical damage data, and the life decay of each functional block is evaluated.

[0012] Processing instructions are generated based on the amount of paper life decay, and the paper feeding mechanism and shredder assembly are driven to perform corresponding actions based on the processing instructions.

[0013] In one optional embodiment, spatiotemporal joint analysis of a multi-frame image sequence is performed, and the surface texture motion trajectory of the target object is extracted by optical flow field calculation, including:

[0014] Pixel displacement field calculation is performed between adjacent frames in a multi-frame image sequence to obtain the initial optical flow field;

[0015] The initial optical flow field is decomposed into spatial scales, and motion components are calculated at multiple preset spatial scales. Motion components with a coverage area greater than a preset first spatial scale threshold are identified as overall motion components, and motion components with a coverage area less than a preset second spatial scale threshold are identified as local deformation motion components. Local deformation motion components are extracted to form a local deformation motion field.

[0016] A temporal correlation relationship is established for each pixel in the local deformation motion field. A motion chain graph is constructed by tracking the position evolution of each pixel between consecutive frames. The neighborhood cooperation index of each pixel is calculated on the motion chain graph. The neighborhood cooperation index is obtained by measuring the angle between the motion vector of the pixel and the motion amplitude difference between the pixel and the pixels in its spatial neighborhood.

[0017] The motion chain graph is divided into regions based on the neighborhood collaboration index, and pixels that meet the preset collaboration conditions are merged into texture motion regions.

[0018] Extract the motion path of each texture motion region and output the motion path as the surface texture motion trajectory of the target object.

[0019] In one optional embodiment, a material flexibility description is obtained through deformation tensor decomposition of the surface texture motion trajectory. The material flexibility description is then tensor-concatenated with the image spatial gradient distribution to form a spatiotemporal fusion feature, including:

[0020] Based on the motion vectors of the surface texture motion trajectory at each pixel position in the spatial domain, a second-order deformation tensor is constructed at each pixel position.

[0021] Eigenvalue decomposition is performed on the second-order deformation tensor to obtain the eigenvalue set of each pixel position. The ratio of the maximum eigenvalue to the minimum eigenvalue in the eigenvalue set is calculated and used as the deformation direction distribution parameter of each pixel position.

[0022] For the deformation direction distribution parameters of each pixel position, calculate the parameter difference value in the spatial neighborhood, and group the adjacent pixel positions with parameter difference values ​​less than the preset difference threshold into homogeneous deformation regions. Calculate the Euclidean distance between the deformation vector of each pixel position in each homogeneous deformation region and the average deformation vector of the region. Obtain the mean and variance of the Euclidean distance to form a material flexibility description.

[0023] Spatial gradient calculation is performed on each frame of a multi-frame image sequence to obtain the gradient magnitude distribution and gradient direction distribution;

[0024] The material flexibility description is aligned with the gradient magnitude and gradient direction distribution in the spatial domain. At each alignment position, the material flexibility parameters in the material flexibility description are tensor-concatenated with the gradient magnitude and gradient direction at the corresponding position to form a spatiotemporal fusion feature.

[0025] In one optional embodiment, a multi-scale response map is constructed by pyramid decomposition of the spatiotemporal fusion features. Abnormal responses are iteratively propagated between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence location and intensity peak of the abnormal responses, including:

[0026] The spatiotemporal fusion features are decomposed into a pyramid according to spatial resolution, and the response values ​​of pixel positions are calculated at each decomposition level to construct a multi-scale response map.

[0027] Based on the highest resolution level of the multi-scale response map, the pixel positions where the response value exceeds the preset seed threshold are selected to determine the abnormal response seed points. According to the response value and spatial gradient direction of each abnormal response seed point, the diffusion direction and diffusion weight to the adjacent pixel positions are calculated, and the response value is transferred to the adjacent pixel positions to obtain the updated response distribution of the current level.

[0028] The updated response distribution is downsampled to the next level. At the next level, the diffusion direction and diffusion weight of each pixel position are recalculated based on the response value passed from the previous level. Abnormal response is passed on and repeated until the lowest resolution level is reached. During the passing process, the trajectory of the response value change of each pixel position at each level is recorded.

[0029] Calculate the spatial offset of the response value of each pixel position between adjacent layers. When it is continuously less than the preset convergence threshold, mark the pixel position as the convergence state and extract the corresponding response value as the intensity peak.

[0030] Pixels with intensity peaks exceeding a preset peak threshold are spatially connected and their locations are marked to determine the foreign object region location.

[0031] In one optional embodiment, the geometric contour is extracted based on the foreign object area location result, and a virtual contact scene is constructed in a spatial coordinate system in combination with the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection position, including:

[0032] Boundary tracking is performed on the foreign object region location results to extract the geometric contour of the foreign object region;

[0033] Extract the spatial coordinates of each boundary point from the geometric contour, calculate the tangent and normal directions for each boundary point, construct the local coordinate system for each boundary point, and calculate the curvature distribution and surface tilt distribution of the foreign object region under the local coordinate system.

[0034] The motion trajectory of the paper shredder blade is obtained, and the motion trajectory is aligned with the geometric contour in a spatial coordinate system. In the aligned spatial coordinate system, the intersection time and intersection position of the motion trajectory and the geometric contour are determined to construct a virtual contact scene.

[0035] In the virtual contact scenario, the curvature distribution and surface tilt angle distribution corresponding to the intersection positions are extracted. Based on the speed and direction of movement of the shredder blade at the intersection moment, the relative contact speed between the shredder blade and the foreign object area is calculated. The impact force amplitude at each intersection position is calculated based on the relative contact speed and curvature distribution. The direction of the impact force is determined based on the surface tilt angle distribution. The impact force amplitude and direction at each intersection position are mapped along the blade surface of the shredder blade to obtain the force distribution of the blade.

[0036] In one optional embodiment, calculating the impact force amplitude at each intersection point based on the relative contact velocity and curvature distribution includes:

[0037] Local extreme points of curvature distribution are extracted at the intersection points, and these local extreme points are used as the stress concentration cores.

[0038] The curvature gradient is calculated from each stress concentration core to the surrounding area, and the stress diffusion radius is determined based on the magnitude of the curvature gradient. Within the stress diffusion radius, the peak stress of the stress concentration core is calculated based on the relative contact velocity, and the stress distribution at the surrounding location is calculated based on the distance from the stress concentration core and the attenuation rate of the curvature gradient.

[0039] When there are multiple stress concentration cores at the intersection, the stress fields of each stress concentration core are superimposed, the location of the stress peak after superposition is extracted, and the stress peak is converted into the impact force amplitude.

[0040] In one optional embodiment, the shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. The current crack density prediction value is determined by combining historical damage data. The assessment of the lifespan decay of each functional block includes:

[0041] The paper shredder blade is divided into multiple functional blocks along the blade length and blade width directions;

[0042] Extract the stress time sequence of each functional block from the stress distribution of the blade, perform peak detection, and count the amplitude and frequency of the stress peak points;

[0043] A stress cycle spectrum is established based on the amplitude and frequency of the peak stress point. The number of cycles in different amplitude ranges is counted. Based on the preset material fatigue characteristic curve, the number of cycles in each amplitude range is converted into damage contribution and accumulated to obtain the fatigue damage accumulation index.

[0044] Acquire historical damage data for each functional block, including the cumulative fatigue damage index and measured crack density within historical working cycles;

[0045] Historical damage data of each functional block is used to construct a damage evolution chain in the order of working cycles. The first mutation point of fatigue damage accumulation index and the second mutation point of crack density are extracted. The damage propagation delay coefficient is calculated based on the time lag of the first and second mutation points. The fatigue damage accumulation index of the current cycle is combined with the damage propagation delay coefficient and mapped to the crack density response time. The corresponding measured crack density is extracted, the current crack density prediction value is determined, and the remaining life is calculated based on the difference between the preset failure threshold and the current crack density prediction value. The life decay of each functional block is evaluated.

[0046] A second aspect of this invention provides an intelligent paper shredder foreign object detection and handling system integrating visual recognition, comprising:

[0047] The first unit is used to acquire a multi-frame image sequence of the target object in the paper feed channel;

[0048] The second unit is used to perform spatiotemporal joint analysis on multi-frame image sequences. It extracts the surface texture motion trajectory of the target object through optical flow field calculation, obtains the material flexibility description through deformation tensor decomposition of the surface texture motion trajectory, and spatiotemporally fused features by tensor splicing the material flexibility description with the image spatial gradient distribution.

[0049] The third unit is used to construct a multi-scale response map by pyramid decomposing the spatiotemporal fusion features, iteratively transmitting the abnormal response between the different levels of the multi-scale response map, and determining the location result of the foreign object region based on the convergence position and intensity peak of the abnormal response.

[0050] The fourth unit is used to extract the geometric contour based on the location result of the foreign object area, and to construct a virtual contact scene in the spatial coordinate system in combination with the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection position.

[0051] The fifth unit is used to divide the shredder blade into multiple functional blocks, calculate the fatigue damage accumulation index of each functional block based on the stress distribution of the blade, determine the current crack density prediction value by combining historical damage data, and assess the life decay of each functional block.

[0052] The sixth unit is used to generate processing instructions based on the amount of life decay, and drive the paper feeding mechanism and shredder assembly to perform corresponding actions based on the processing instructions.

[0053] A third aspect of the present invention provides an electronic device, comprising:

[0054] processor;

[0055] Memory used to store processor-executable instructions;

[0056] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0058] In this embodiment of the invention, by performing spatiotemporal joint analysis on multiple image sequences, the surface texture motion trajectory is extracted and the material flexibility description is obtained. Then, tensor splicing with the image spatial gradient distribution is performed to form spatiotemporal fusion features, which realizes the accurate differentiation of the material properties of different objects and effectively solves the technical problem that traditional methods cannot accurately identify foreign objects of different materials. The technical solution of constructing multi-scale response maps by pyramid decomposition and iteratively transmitting abnormal responses between each level can adapt to the detection of foreign objects of different sizes and shapes, improving the accuracy and robustness of detection, especially significantly improving the detection effect for small foreign objects. The method of extracting geometric contours based on the foreign object area positioning results and constructing a virtual contact scene by combining the shredder blade motion trajectory realizes the accurate calculation of the blade force distribution, providing a scientific basis for subsequent blade life assessment. By dividing the shredder blade into multiple functional blocks and calculating the fatigue damage accumulation index of each block, and predicting crack density by combining historical damage data, the accurate assessment of the shredder blade life is realized, which greatly improves the service life and safety of the equipment. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the foreign object detection and handling method for an intelligent paper shredder that integrates visual recognition, according to an embodiment of the present invention.

[0060] Figure 2 This is a flowchart for analyzing the contact between a foreign object and the blade. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a flowchart illustrating the foreign object detection and handling method for an intelligent paper shredder integrating visual recognition, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] Acquire a multi-frame image sequence of the target object in the paper feed channel;

[0065] Spatiotemporal joint analysis is performed on a multi-frame image sequence. The surface texture motion trajectory of the target object is extracted by optical flow field calculation. The material flexibility description is obtained by deformation tensor decomposition of the surface texture motion trajectory. The material flexibility description is then tensor-stitched with the image spatial gradient distribution to form a spatiotemporal fusion feature.

[0066] A multi-scale response map is constructed by pyramid decomposition of spatiotemporal fusion features. Anomaly responses are iteratively transmitted between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence position and intensity peak of the anomaly responses.

[0067] Geometric contours are extracted based on the location results of the foreign object area. A virtual contact scene is constructed in the spatial coordinate system by combining the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection.

[0068] The paper shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. The current crack density prediction value is determined by combining historical damage data, and the life decay of each functional block is evaluated.

[0069] Processing instructions are generated based on the amount of paper life decay, and the paper feeding mechanism and shredder assembly are driven to perform corresponding actions based on the processing instructions.

[0070] In one optional implementation, spatiotemporal joint analysis of a multi-frame image sequence is performed, and the surface texture motion trajectory of the target object is extracted through optical flow field calculation, including:

[0071] Pixel displacement field calculation is performed between adjacent frames in a multi-frame image sequence to obtain the initial optical flow field;

[0072] The initial optical flow field is decomposed into spatial scales, and motion components are calculated at multiple preset spatial scales. Motion components with a coverage area greater than a preset first spatial scale threshold are identified as overall motion components, and motion components with a coverage area less than a preset second spatial scale threshold are identified as local deformation motion components. Local deformation motion components are extracted to form a local deformation motion field.

[0073] A temporal correlation relationship is established for each pixel in the local deformation motion field. A motion chain graph is constructed by tracking the position evolution of each pixel between consecutive frames. The neighborhood cooperation index of each pixel is calculated on the motion chain graph. The neighborhood cooperation index is obtained by measuring the angle between the motion vector of the pixel and the motion amplitude difference between the pixel and the pixels in its spatial neighborhood.

[0074] The motion chain graph is divided into regions based on the neighborhood collaboration index, and pixels that meet the preset collaboration conditions are merged into texture motion regions.

[0075] Extract the motion path of each texture motion region and output the motion path as the surface texture motion trajectory of the target object.

[0076] In one specific implementation, pixel displacement field calculation is performed between adjacent frames in a multi-frame image sequence to obtain an initial optical flow field. Specifically, gradient-based optical flow estimation methods, such as the Horn-Schunck algorithm or the Lucas-Kanade algorithm, can be used to calculate the pixel displacement between two adjacent frames in the image sequence. For example, for the t-th frame and the (t+1)-th frame in the image sequence, the horizontal displacement u(x, y) and vertical displacement v(x, y) of each pixel (x, y) between the two frames are calculated to form an optical flow vector field (u, v). This optical flow vector field contains the motion information of all objects in the scene, but it needs to be further decomposed to extract the surface texture motion of the target object.

[0077] The initial optical flow field is spatially decomposed, and motion components are calculated at multiple preset spatial scales. Specifically, multi-scale analysis methods such as wavelet transform or Laplacian pyramid decomposition are used to decompose the optical flow field into motion components at different spatial scales. For example, a five-level Laplacian pyramid is constructed, corresponding to the motion components at the pixel level, local region level, medium region level, large region level, and global level, respectively. At each scale level, the corresponding optical flow component and its spatial coverage are calculated.

[0078] Based on preset spatial scale thresholds, motion components with a coverage area greater than a preset first spatial scale threshold are identified as global motion components, while motion components with a coverage area less than a preset second spatial scale threshold are identified as local deformation motion components. For example, motion components covering more than 50% of the image area are identified as global motion components (such as global displacement caused by camera movement), while motion components covering less than 5% of the image area are identified as local deformation motion components (such as local texture deformation of an object's surface). By extracting local deformation motion components and constructing a local deformation motion field, this step effectively separates the interference between the surface texture motion of the target object and other motions.

[0079] A temporal correlation is established for each pixel in the local deformation motion field, and a motion chain graph is constructed by tracking the positional evolution of each pixel across consecutive frames. Specifically, for each pixel p(x, y) in the local deformation motion field, its corresponding position p'(x', y') in frame t+1 is calculated based on its position in frame t and the optical flow vector, where x' = x + u(x, y) and y' = y + v(x, y). The corresponding position p' in frame t+2 is calculated in the same way, and so on, forming a motion trajectory chain of pixel p throughout the entire sequence. If a pixel cannot find a valid corresponding position in a subsequent frame (e.g., due to occlusion or removal from the field of view), the motion chain of that point terminates.

[0080] The neighborhood coordination index is calculated for each pixel on the motion chain graph. The neighborhood coordination index is obtained by measuring the angle between the motion vectors of a pixel and the difference in motion amplitude between the pixel and its spatial neighbors. For example, for a pixel p in a local deformation motion field, n pixels in its spatial neighborhood are selected to form a set N(p). The angle θi between the motion vector of p and each point qi in N(p) and the difference in motion amplitude Δmi are calculated. The angle between the motion vectors measures directional consistency, and the difference in motion amplitude measures velocity consistency. The neighborhood coordination index C(p) can be expressed as a weighted combination of these two differences, reflecting the degree of motion coordination between pixel p and its neighbors. Regions with high coordination typically correspond to texture motion on the same object surface.

[0081] The motion chain graph is divided into regions based on the neighborhood cooperation index. Pixels whose neighborhood cooperation index meets the preset cooperation conditions are grouped into texture motion regions. A threshold τ is set; when C(p)≥τ, pixel p is determined to have high cooperation. Neighboring high-cooperation pixels are connected into regions using segmentation algorithms such as region growing or graph cut, forming multiple texture motion regions. Pixels within each region have similar motion characteristics and belong to the texture of the same object surface.

[0082] The motion paths of each texture motion region are extracted and output as the surface texture motion trajectory of the target object. For each identified texture motion region R, the center position of the region in each frame is calculated, and these center positions are connected to form the overall motion path of the region. Simultaneously, representative feature points (such as corner points or texture-rich points) within the region can be extracted, and the motion trajectories of these feature points in the sequence are tracked as local motion features of the surface texture. The overall motion path of the region reflects the overall displacement and rotation of the object's surface, while the local trajectories of the feature points reflect the subtle deformations of the surface texture.

[0083] The above technical solution enables accurate extraction of the motion trajectory of the surface texture of the target object, effectively separates global motion and local deformation, and improves the robustness of trajectory extraction through neighborhood co-operation analysis.

[0084] In one optional implementation, a material flexibility description is obtained through deformation tensor decomposition of the surface texture motion trajectory. This material flexibility description is then tensor-concatenated with the image spatial gradient distribution to form a spatiotemporal fusion feature, including:

[0085] Based on the motion vectors of the surface texture motion trajectory at each pixel position in the spatial domain, a second-order deformation tensor is constructed at each pixel position.

[0086] Eigenvalue decomposition is performed on the second-order deformation tensor to obtain the eigenvalue set of each pixel position. The ratio of the maximum eigenvalue to the minimum eigenvalue in the eigenvalue set is calculated and used as the deformation direction distribution parameter of each pixel position.

[0087] For the deformation direction distribution parameters of each pixel position, calculate the parameter difference value in the spatial neighborhood, and group the adjacent pixel positions with parameter difference values ​​less than the preset difference threshold into homogeneous deformation regions. Calculate the Euclidean distance between the deformation vector of each pixel position in each homogeneous deformation region and the average deformation vector of the region. Obtain the mean and variance of the Euclidean distance to form a material flexibility description.

[0088] Spatial gradient calculation is performed on each frame of a multi-frame image sequence to obtain the gradient magnitude distribution and gradient direction distribution;

[0089] The material flexibility description is aligned with the gradient magnitude and gradient direction distribution in the spatial domain. At each alignment position, the material flexibility parameters in the material flexibility description are tensor-concatenated with the gradient magnitude and gradient direction at the corresponding position to form a spatiotemporal fusion feature.

[0090] In one specific implementation, based on the acquired surface texture motion trajectory, a second-order deformation tensor is constructed for each pixel location in the spatial domain. The deformation tensor is constructed based on the motion vectors of the pixel location and its eight neighboring pixels. For example, for a pixel at location (100, 150), the acquired motion vector of this point is (2.3, -1.5) pixels / frame. By analyzing the differences in motion vectors between this point and its eight surrounding pixels, a 2×2 deformation tensor is constructed. This deformation tensor describes the deformation characteristics of the local region at this location, recording deformation information such as stretching, shearing, and rotation of the object's surface in this region.

[0091] The constructed second-order deformation tensor is decomposed into eigenvalues, yielding two eigenvalues ​​for each pixel location. Taking the location (100, 150) as an example, the two eigenvalues ​​are 0.85 and 0.32, respectively. The ratio of these two eigenvalues, 0.85 / 0.32 = 2.66, is calculated and used as the deformation direction distribution parameter for that pixel location. This parameter describes the anisotropy of the deformation at that location; a larger value indicates more non-uniform deformation, and a smaller value indicates more uniform deformation. By performing the same processing on every pixel across the entire image region, a complete deformation direction distribution parameter map is obtained.

[0092] The parameter difference of the deformation direction distribution parameters of each pixel position within its spatial neighborhood is calculated. Using a 5×5 neighborhood window, the difference in deformation direction distribution parameters between the center pixel and other pixels in the neighborhood is calculated. A preset difference threshold of 0.5 is set, and adjacent pixel positions with parameter difference values ​​less than this threshold are grouped into homogeneous deformation regions. For example, within a 5×5 region (100-104, 150-154), a homogeneous deformation region containing 18 pixels is formed. After processing the entire image, multiple different homogeneous deformation regions are formed, and the pixels within each region have similar deformation characteristics.

[0093] For each homogeneous deformation region, the Euclidean distance between the deformation vector of each pixel within the region and the average deformation vector of the region is calculated. Taking the homogeneous deformation region consisting of the aforementioned 18 pixels as an example, the average deformation vector of the pixels within this region is (2.1, -1.3), and the deformation vector of the first pixel within the region is (2.3, -1.5), with a Euclidean distance of 0.28 between them. The Euclidean distance between all pixels within the region and the average deformation vector is calculated, resulting in a set of distance values. The mean of this set of distance values ​​is 0.28, and the variance is 0.06. These two parameters constitute the material flexibility description of the region; the mean reflects the consistency level of deformation within the region, and the variance reflects the stability of deformation. The material flexibility description ultimately forms a feature map of the same size as the original image, where each pixel position is associated with the material flexibility parameters of its corresponding homogeneous deformation region.

[0094] Simultaneously, spatial gradient calculations are performed on each frame of the multi-frame image sequence using the Sobel operator. The horizontal and vertical gradients are calculated for each frame, yielding the gradient magnitude and direction distributions. For the pixel at position (100, 150) in the nth frame, the calculated gradient magnitude is 85, and the gradient direction is 67 degrees. The gradient magnitude distribution reflects the intensity variation of texture edges in the image, while the gradient direction distribution reflects the orientation of the texture edges.

[0095] After gradient calculation, the material flexibility description is aligned with the gradient magnitude and direction distributions in the spatial domain. At each aligned pixel location, the mean and variance of the material flexibility description are combined with the corresponding gradient magnitude and direction to form a four-dimensional feature vector. For example, the feature vector at location (100, 150) is [0.28, 0.06, 85, 67]. This combination forms a spatiotemporal fusion feature that incorporates the deformation characteristics and texture features of the object's material.

[0096] For example, for ordinary A4 paper, the homogeneous deformation area is usually large, covering 25%-40% of the image, with a low mean (typically 0.15-0.30) and low variance (typically 0.03-0.10) for the material flexibility description. For metallic foreign objects, the homogeneous deformation area is small, typically less than 5% of the image, with a high mean (typically 0.60-0.85) and high variance (typically 0.20-0.35) for the material flexibility description. For semi-rigid objects such as plastic cards, the homogeneous deformation area typically covers 10%-20%, with a mean of approximately 0.40-0.55 and a variance of approximately 0.12-0.18 for the material flexibility description. Printed text areas on the paper surface, due to their high gradient amplitude but similar material flexibility to ordinary paper, exhibit a unique combination of features, with the gradient amplitude component in the spatiotemporal fusion features significantly higher than that of pure paper areas.

[0097] This spatiotemporal fusion feature fully combines the material deformation characteristics and surface texture features of an object, effectively distinguishing different materials such as paper, plastic, metal, and fabric. It provides a reliable feature basis for intelligent paper shredders to identify foreign objects and significantly improves the accuracy of foreign object detection.

[0098] In one optional implementation, a pyramid decomposition is performed on the spatiotemporal fusion features to construct a multi-scale response map. Abnormal responses are iteratively propagated between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence location and intensity peak of the abnormal responses, including:

[0099] The spatiotemporal fusion features are decomposed into a pyramid according to spatial resolution, and the response values ​​of pixel positions are calculated at each decomposition level to construct a multi-scale response map.

[0100] Based on the highest resolution level of the multi-scale response map, the pixel positions where the response value exceeds the preset seed threshold are selected to determine the abnormal response seed points. According to the response value and spatial gradient direction of each abnormal response seed point, the diffusion direction and diffusion weight to the adjacent pixel positions are calculated, and the response value is transferred to the adjacent pixel positions to obtain the updated response distribution of the current level.

[0101] The updated response distribution is downsampled to the next level. At the next level, the diffusion direction and diffusion weight of each pixel position are recalculated based on the response value passed from the previous level. Abnormal response is passed on and repeated until the lowest resolution level is reached. During the passing process, the trajectory of the response value change of each pixel position at each level is recorded.

[0102] Calculate the spatial offset of the response value of each pixel position between adjacent layers. When it is continuously less than the preset convergence threshold, mark the pixel position as the convergence state and extract the corresponding response value as the intensity peak.

[0103] Pixels with intensity peaks exceeding a preset peak threshold are spatially connected and their locations are marked to determine the foreign object region location.

[0104] # Detailed Implementation

[0105] In one specific implementation, the spatiotemporal fusion features are processed through pyramid decomposition to construct a multi-level structure according to different spatial resolutions. For example, the feature map at the original resolution is used as the first level L1, and then progressively downsampled to generate lower-resolution levels L2, L3, and L4, where the resolution of each level is half that of the previous level. At each level, the response value at each pixel location is calculated to obtain a multi-scale response map representing the degree of anomaly. The response value calculation method can employ a metric function based on the consistency of local region feature distribution, that is, for each pixel location, the difference between the feature and the feature distribution of the surrounding region is calculated.

[0106] After the multi-scale response map is constructed, seed points are selected at the highest resolution level L1. The selection criteria are pixel locations whose response values ​​exceed a preset seed threshold Ts (e.g., Ts=0.7); these locations are considered the starting points of potential anomaly regions. For each selected seed point, the spatial gradients in eight surrounding directions are calculated to determine the diffusion direction of the response value. The diffusion weights are calculated based on the feature similarity between the center point and its neighboring points and the difference in response values.

[0107] For a seed point (x, y), calculate its diffusion weights to neighboring pixels (x+i, y+j), where i, j ∈ {-1, 0, 1} and are not simultaneously 0. The diffusion weights are directly proportional to feature similarity and inversely proportional to response value difference. Based on the calculated diffusion direction and weights, the response values ​​of the seed point are proportionally distributed to neighboring pixels, updating their response values. This process is iterated multiple times across the entire layer until the response distribution stabilizes, thus obtaining the updated L1 layer response distribution.

[0108] The updated response distribution at level L1 is downsampled to level L2, maintaining consistency of response values ​​across their corresponding spatial locations. At level L2, based on the response values ​​passed down from L1, the diffusion direction and diffusion weight at each pixel location are recalculated, and a similar response propagation process is performed. This cross-level propagation mechanism allows anomalous responses to be integrated and optimized at different scales, enhancing adaptability to foreign objects of varying sizes.

[0109] This process continues sequentially, with the abnormal response propagating from L2 to L3, then from L3 to L4, until it reaches the lowest resolution layer. Throughout this propagation, the response value and spatial location trajectory of each pixel at each layer are recorded; this trajectory information will be used for subsequent convergence analysis.

[0110] After completing multi-level response propagation, the spatial offset of each pixel's response value between adjacent levels is calculated. The offset reflects the stability of the response center at different resolutions. When the offset of a pixel's position between two or more consecutive levels is less than a preset convergence threshold Tc (e.g., Tc = 2 pixels), the pixel is marked as having reached a convergence state. Convergence indicates that this location is the central region of the real foreign object.

[0111] For converged pixel locations, their corresponding response values ​​are extracted as intensity peaks. Intensity peaks reflect the strength of the anomaly. A preset peak threshold Tp is set (e.g., Tp=0.8), and pixel locations with intensity peaks exceeding Tp are considered valid anomalies. These anomalies are spatially connected component labeled; adjacent anomalies are grouped into the same connected component, and each connected component represents a potential foreign object region.

[0112] Based on the spatial extent and shape characteristics of connected components, the final foreign object location result is determined. The center point, bounding box, or precise contour of each connected component can be calculated as the output of the foreign object location. For multiple adjacent connected components, it can also be determined whether they should be merged into the same foreign object region based on their distance and response value distribution characteristics, improving the accuracy and practicality of the location results.

[0113] By employing the aforementioned pyramid decomposition and multi-scale response propagation mechanism, this method effectively handles the detection of foreign objects of varying sizes and shapes, exhibiting strong robustness and accuracy. In practical applications, the seed threshold, convergence threshold, and peak threshold can be adjusted according to the specific scenario to achieve optimal detection results. Particularly in foreign object detection against complex backgrounds, this method, through multi-scale information fusion, effectively suppresses background interference, highlights the true foreign object target, and improves detection accuracy and reliability.

[0114] In one optional implementation, the geometric contour is extracted based on the location result of the foreign object area, and a virtual contact scene is constructed in a spatial coordinate system in combination with the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection position, including:

[0115] Boundary tracking is performed on the foreign object region location results to extract the geometric contour of the foreign object region;

[0116] Extract the spatial coordinates of each boundary point from the geometric contour, calculate the tangent and normal directions for each boundary point, construct the local coordinate system for each boundary point, and calculate the curvature distribution and surface tilt distribution of the foreign object region under the local coordinate system.

[0117] The motion trajectory of the paper shredder blade is obtained, and the motion trajectory is aligned with the geometric contour in a spatial coordinate system. In the aligned spatial coordinate system, the intersection time and intersection position of the motion trajectory and the geometric contour are determined to construct a virtual contact scene.

[0118] In the virtual contact scenario, the curvature distribution and surface tilt angle distribution corresponding to the intersection positions are extracted. Based on the speed and direction of movement of the shredder blade at the intersection moment, the relative contact speed between the shredder blade and the foreign object area is calculated. The impact force amplitude at each intersection position is calculated based on the relative contact speed and curvature distribution. The direction of the impact force is determined based on the surface tilt angle distribution. The impact force amplitude and direction at each intersection position are mapped along the blade surface of the shredder blade to obtain the force distribution of the blade.

[0119] In one specific implementation, boundary tracking is performed on the foreign object region localization result to extract the geometric contour of the foreign object region. Boundary tracking employs an improved contour tracking algorithm, starting scanning from the upper left corner of the foreign object region localization result image. When the first boundary point is encountered, it is used as the starting point, and the eight adjacent directions are detected sequentially in a clockwise direction to determine the next boundary point. After determining the next boundary point, contour tracking continues according to the same rules until it returns to the starting point, completing the extraction of a closed contour. To improve the accuracy of boundary tracking, image gradient information is incorporated during the tracking process. When multiple boundary directions are encountered, the direction with the largest gradient value is selected as the position of the next boundary point.

[0120] After extracting the geometric contour of the foreign object region, it is necessary to calculate the spatial coordinates of each boundary point on the contour. The boundary points in the image coordinate system are then transformed to the world coordinate system using camera calibration parameters to obtain the three-dimensional spatial coordinates (x, y, z) of the boundary points. For planar foreign objects, it can be assumed that the z-coordinate of all boundary points is the same; for foreign objects with thickness, depth information needs to be obtained through methods such as binocular vision or structured light.

[0121] After obtaining the spatial coordinates of the boundary points, it is necessary to calculate the tangent and normal directions of each boundary point and construct a local coordinate system. For a boundary point with spatial coordinates (xi, yi, zi), its tangent direction vector T = (tx, ty, tz) can be calculated through the coordinate difference between adjacent points. Specifically, three points before and after this point are selected, and the tangent direction is calculated using the central difference method:

[0122] For a boundary point Pi, its tangent direction vector T is calculated based on its six neighboring points. After the tangent direction is calculated, the normal direction N is obtained by solving for the perpendicular vector between the tangent direction vector and the surface. When calculating the normal, the weighted average of the normal vectors of the triangular plane formed by the adjacent points can be used as the normal direction of that point, and the weights can be set to the area of ​​the triangle.

[0123] In the local coordinate system, the curvature distribution and surface tilt angle distribution of the foreign object region are calculated. The curvature K can be calculated using the second derivative of the boundary points in the local coordinate system, while the surface tilt angle θ is calculated using the angle between the normal direction and the z-axis. To improve calculation accuracy, B-spline curves can be used to fit the boundary points, and then the curvature and surface tilt angle can be calculated using the fitted curves.

[0124] The motion trajectory of the shredder blade can be obtained through the blade tip pose data recorded by the robot control system or through a vision tracking system. The motion trajectory is represented as a series of position and orientation data at discrete time points {(p1, R1, t1), (p2, R2, t2), ..., (pn, Rn, tn)}, where pi represents the position coordinates, Ri represents the rotation matrix, and ti represents the timestamp.

[0125] Align the motion trajectory of the shredder blade with the geometric contour of the foreign object in a spatial coordinate system. Coordinate alignment includes: determining the transformation relationship between the two coordinate systems, namely the rotation matrix R and the translation vector T; and transforming the points of the foreign object contour or the shredder trajectory to the same coordinate system. Feature point matching and iterative nearest-point algorithms are used to achieve coordinate alignment.

[0126] In the aligned spatial coordinate system, the intersection time and position of the motion trajectory and the geometric contour are determined to construct a virtual contact scene. Specifically, the blade is simplified into a line segment or a plane, and the foreign object contour is simplified into a set of line segments or a curved surface; then, the minimum distance between the simplified models is calculated at each time point, and contact is determined to have occurred when the distance is less than a preset threshold. To improve computational efficiency, spatial partitioning techniques such as octrees can be used for acceleration.

[0127] In the virtual contact scenario, the corresponding curvature distribution and surface tilt angle distribution are extracted based on the intersection location. For each intersection point, the previously calculated curvature K and surface tilt angle θ are retrieved based on its position index on the foreign object contour. If the intersection point is not a boundary point, its curvature and surface tilt angle are calculated using an interpolation method.

[0128] Based on the velocity v and direction d of the shredder blades at the moment of intersection, calculate the relative contact velocity vr between the blades and the foreign object region. The relative contact velocity can be expressed as the projection of the blade velocity onto the normal direction of the foreign object surface: vr = v·n, where n is the normal direction of the foreign object surface.

[0129] The impact force amplitude F at each intersection point is calculated based on the relative contact velocity and curvature distribution. The impact force can be calculated using Hertz contact theory, where F is proportional to the square of the relative contact velocity vr and the square root of the curvature K at the contact point. For multiple intersection points, the impact force amplitude at each point needs to be calculated separately.

[0130] The direction of the impact force is determined by the distribution of the surface tilt angle. The direction of the impact force is primarily determined by the normal direction of the foreign object's surface at the contact point, but is also influenced by the direction of the blade's movement. Specifically, the direction of the impact force can be decomposed into a normal component and a tangential component. The normal component is along the surface normal direction, while the tangential component is related to the surface friction coefficient and the normal force.

[0131] The impact force amplitude and direction at each intersection point are mapped along the blade surface of the shredder to obtain the force distribution on the blade. The mapping process needs to consider the relative positions of the intersection points on the blade; typically, an influence function can be used to expand the point force into a distributed force. For example, for position x on the blade, the force can be expressed as the weighted sum of the impact forces at all intersection points, with the weight decreasing as the distance between x and the intersection points increases.

[0132] In practical applications, the blade force distribution calculated using the above method can be used to predict the shredder's cutting effect on foreign objects, assess the risk of damage to the blade from foreign objects, or optimize the shredder's cutting strategy. For example, when a high curvature point or a region with drastic changes in surface tilt angle is detected in the foreign object area, the feed speed or cutting angle of the shredder blade can be adjusted to reduce the peak impact force and extend the blade's lifespan.

[0133] like Figure 2 The diagram shows the flowchart for analyzing the contact between the foreign object and the blade.

[0134] In one optional implementation, calculating the impact force amplitude at each intersection point based on the relative contact velocity and curvature distribution includes:

[0135] Local extreme points of curvature distribution are extracted at the intersection points, and these local extreme points are used as the stress concentration cores.

[0136] The curvature gradient is calculated from each stress concentration core to the surrounding area, and the stress diffusion radius is determined based on the magnitude of the curvature gradient. Within the stress diffusion radius, the peak stress of the stress concentration core is calculated based on the relative contact velocity, and the stress distribution at the surrounding location is calculated based on the distance from the stress concentration core and the attenuation rate of the curvature gradient.

[0137] When there are multiple stress concentration cores at the intersection, the stress fields of each stress concentration core are superimposed, the location of the stress peak after superposition is extracted, and the stress peak is converted into the impact force amplitude.

[0138] In one specific implementation, upon collision, geometric data of the surfaces of the two objects are acquired, including surface expressions or discrete point cloud data. The geometric features of the intersection point are determined by calculating the relative positional relationship between the two objects. The intersection point is typically represented as a region rather than a single point; therefore, it is necessary to extract the curvature distribution features within this region.

[0139] The intersecting regions are meshed to obtain sufficiently fine curvature distribution data. Differential geometry methods are used to calculate the principal curvature values ​​at each mesh point, including the maximum principal curvature κ1 and the minimum principal curvature κ2. Curvature calculations can employ curvature estimation methods for discrete surfaces, such as fitting a quadratic surface or using eigenvalue analysis of the covariance matrix of a neighborhood point set.

[0140] After obtaining the curvature distribution, the curvature distribution map is analyzed to extract local extrema. Local extrema are points with the largest curvature values ​​in their neighborhood, identified by comparing the curvature values ​​of each grid point with its neighbors. These local extrema are considered stress concentration cores, representing the locations where impact forces are most concentrated. For example, when a sphere collides with an uneven surface, the vertices of the protruding parts on the surface typically exhibit local extrema of curvature, and these locations will bear the greatest impact stress.

[0141] The curvature gradient is calculated from each stress concentration core outwards to the surrounding region. The curvature gradient, denoted as ∇κ, describes the rate of change of curvature in space. In practical calculations, the gradient value can be approximated using the finite difference method. For each stress concentration core point P, the magnitude of the curvature gradient |∇κ| at its surrounding points is calculated.

[0142] The stress diffusion radius R is determined based on the magnitude of the curvature gradient. The stress diffusion radius is inversely proportional to the curvature gradient magnitude, i.e., R = α / |∇κ|, where α is a proportionality coefficient determined based on material properties. When the curvature gradient is large, the stress concentration effect is more pronounced, and the diffusion radius is smaller; conversely, when the curvature gradient is small, the stress distribution is more uniform, and the diffusion radius is larger.

[0143] After determining the stress diffusion radius, the peak stress σmax at the stress concentration core needs to be calculated based on the relative contact velocity v. The relative contact velocity refers to the velocity component of the two objects along the normal direction at the point of contact at the moment of collision. The peak stress calculation considers the process of kinetic energy being converted into strain energy and can be expressed as σmax = β·ρ·v 2 ·κ, where β is the material response coefficient, ρ is the material density, and κ is the principal curvature at the stress concentration core.

[0144] For points around the stress concentration core, the stress value decreases with increasing distance. Let the distance from point Q to the stress concentration core P be d. Then the stress value σQ at point Q can be expressed as: σQ = σmax·exp(-d / λ), where λ is the attenuation parameter, which is related to the stress transfer characteristics of the material and can be obtained by fitting experimental data. When d exceeds the stress diffusion radius R, the stress is considered to have attenuated to a negligible level.

[0145] In practical applications, when multiple stress concentration cores exist at an intersection, it is necessary to calculate the superposition effect of the stress fields generated by each stress concentration core. Suppose point X is affected by n stress concentration cores, then the total stress at point X is: σX = σ1 + σ2 + ... + σn, which is the linear superposition of the stress values ​​generated by each core.

[0146] After the stress fields are superimposed, the location of the stress peak and its corresponding stress value σpeak are found by scanning the entire intersection region. Finally, the stress peak is converted into the impact force amplitude F, and the calculation formula is: F = σpeak·A, where A is the contact area under force, which can be estimated by the geometric characteristics of the intersection region.

[0147] The above method accurately simulates the impact force distribution between geometrically shaped objects, providing theoretical support for structural strength design and collision safety analysis. This method comprehensively considers the geometric features and dynamic parameters of the object's surface, and compared to the traditional spring-damped model, it can more accurately reflect the mechanical behavior during actual collisions.

[0148] In one optional implementation, the shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. The current crack density prediction value is determined by combining historical damage data. The lifespan decay of each functional block is evaluated, including:

[0149] The paper shredder blade is divided into multiple functional blocks along the blade length and blade width directions;

[0150] Extract the stress time sequence of each functional block from the stress distribution of the blade, perform peak detection, and count the amplitude and frequency of the stress peak points;

[0151] A stress cycle spectrum is established based on the amplitude and frequency of the peak stress point. The number of cycles in different amplitude ranges is counted. Based on the preset material fatigue characteristic curve, the number of cycles in each amplitude range is converted into damage contribution and accumulated to obtain the fatigue damage accumulation index.

[0152] Acquire historical damage data for each functional block, including the cumulative fatigue damage index and measured crack density within historical working cycles;

[0153] Historical damage data of each functional block is used to construct a damage evolution chain in the order of working cycles. The first mutation point of fatigue damage accumulation index and the second mutation point of crack density are extracted. The damage propagation delay coefficient is calculated based on the time lag of the first and second mutation points. The fatigue damage accumulation index of the current cycle is combined with the damage propagation delay coefficient and mapped to the crack density response time. The corresponding measured crack density is extracted, the current crack density prediction value is determined, and the remaining life is calculated based on the difference between the preset failure threshold and the current crack density prediction value. The life decay of each functional block is evaluated.

[0154] In one specific implementation, during the shredder blade life assessment process, the shredder blade is divided into multiple functional blocks. The division method can be based on the blade's structural characteristics and operating mode, typically dividing it along its length into a front section, middle section, and rear section, and along its width into an edge section, transition section, and back section. For example, for a standard industrial shredder blade, a block can be divided every 50 millimeters along its length, and along its width, it can be divided into 10-millimeter, 20-millimeter, and 30-millimeter areas based on the edge, transition zone, and back structure characteristics, forming a grid-like functional block matrix.

[0155] For each functional block, a stress time sequence is extracted, and raw stress data is collected by strain sensors installed at key parts of the shredder. After filtering, the raw data is used to identify peak stress points using a peak detection algorithm. Peak detection employs a sliding window method; when the value at a point within the window is greater than all points on either side of the window and exceeds a preset threshold, it is marked as a peak point. For a typical shredder blade, the window width can be set to 0.5 seconds, and the threshold to 1.5 times the average stress. After peak point detection, the amplitude and occurrence time of each peak point are recorded, and the frequency distribution of different amplitude intervals is statistically analyzed.

[0156] A stress cycle spectrum is established based on peak value detection results, dividing the peak amplitude range into several intervals, such as 50-100MPa, 100-150MPa, etc., and the number of cycles within each interval is counted. Rainflow counting is used to convert irregular loads into equivalent stress cycles, more accurately reflecting the fatigue damage accumulation process. For example, for carbon steel shredder blades, their fatigue characteristics can be preset based on the material's SN curve (stress-cycle curve). The fatigue damage accumulation index D is calculated using linear cumulative damage theory (Miner's criterion).

[0157] The damage contribution of a single stress range is obtained by dividing the number of cycles in each stress range by the fatigue life of the material at that stress level. The damage contributions of all ranges are then summed to form the total fatigue damage accumulation index. When this index approaches or reaches 1, it indicates that the material is about to experience fatigue failure.

[0158] Acquire historical damage data for each functional block, including the cumulative fatigue damage index recorded during historical operating cycles and the measured crack density obtained through high-precision optical scanning or acoustic emission detection. Crack density is defined as the total length of cracks per unit area, and the unit can be mm / mm. 2 Historical data is arranged chronologically according to work cycles to form a time-series dataset, which is used to construct damage evolution chains.

[0159] When analyzing the damage evolution chain, a change point detection algorithm is used to identify the first abrupt change point in the fatigue damage accumulation index and the second abrupt change point in the crack density. The change point detection can be performed using the CUSUM algorithm or the Bayesian change point detection method. For a typical damage evolution process of a paper shredder blade, the first abrupt change point usually occurs when the fatigue damage accumulation index reaches the range of 0.4-0.6, while the second abrupt change point in the crack density appears in subsequent working cycles.

[0160] Calculate the time lag between the first and second abrupt change points, i.e., the number of working cycles between the two abrupt change points. This lag reflects the time delay from the accumulation of micro-damage to the formation of macro-cracks. Define a damage propagation delay coefficient based on the lag, which represents the time relationship between the transformation of micro-fatigue damage and macro-cracks. The delay coefficient can be determined by fitting an exponential or power function, with typical values ​​ranging from 1.2 to 2.5, depending on material properties and operating conditions.

[0161] By utilizing a damage propagation delay factor, the fatigue damage accumulation index of the current cycle is mapped to the future crack density response time. Specifically, using the current cycle as a baseline, the number of working cycles corresponding to the delay factor is advanced to determine the corresponding measured crack density as the current crack density prediction value. For example, if the delay factor is 2 working cycles and the current cycle is the 10th cycle, then the measured crack density of the 8th cycle is used as the current crack density prediction value.

[0162] The remaining life is calculated based on the difference between the preset failure threshold and the current predicted crack density. The failure threshold is determined based on safety standards and historical data, such as being set to 3.5 mm / mm. 2 The remaining number of working cycles can be estimated by dividing the difference between the current predicted crack density and the failure threshold by the average growth rate of the crack density. The average growth rate is obtained through regression analysis of historical damage data. For example, if the failure threshold is 3.5 mm / mm... 2 The current predicted crack density is 2.0 mm / mm. 2 The average growth rate is 0.1 mm / mm 2 If the remaining lifespan is approximately 15 working cycles, then the remaining lifespan is approximately 15 working cycles.

[0163] By combining the remaining life estimates of each functional block, the most vulnerable block is identified as the decisive factor in the overall blade life, thereby guiding the blade maintenance and replacement strategy, effectively reducing the maintenance cost of the shredder and improving its operational reliability.

[0164] The intelligent paper shredder foreign object detection and handling system integrating visual recognition according to embodiments of the present invention includes:

[0165] The first unit is used to acquire a multi-frame image sequence of the target object in the paper feed channel;

[0166] The second unit is used to perform spatiotemporal joint analysis on multi-frame image sequences. It extracts the surface texture motion trajectory of the target object through optical flow field calculation, obtains the material flexibility description through deformation tensor decomposition of the surface texture motion trajectory, and spatiotemporally fused features by tensor splicing the material flexibility description with the image spatial gradient distribution.

[0167] The third unit is used to construct a multi-scale response map by pyramid decomposing the spatiotemporal fusion features, iteratively transmitting the abnormal response between the different levels of the multi-scale response map, and determining the location result of the foreign object region based on the convergence position and intensity peak of the abnormal response.

[0168] The fourth unit is used to extract the geometric contour based on the location result of the foreign object area, and to construct a virtual contact scene in the spatial coordinate system in combination with the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection position.

[0169] The fifth unit is used to divide the shredder blade into multiple functional blocks, calculate the fatigue damage accumulation index of each functional block based on the stress distribution of the blade, determine the current crack density prediction value by combining historical damage data, and assess the life decay of each functional block.

[0170] The sixth unit is used to generate processing instructions based on the amount of life decay, and drive the paper feeding mechanism and shredder assembly to perform corresponding actions based on the processing instructions.

[0171] A third aspect of the present invention provides an electronic device, comprising:

[0172] processor;

[0173] Memory used to store processor-executable instructions;

[0174] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0175] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0176] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for foreign matter detection and processing of a smart paper shredder with visual recognition, characterized in that, include: Acquire a multi-frame image sequence of the target object in the paper feed channel; Spatiotemporal joint analysis is performed on a multi-frame image sequence. The surface texture motion trajectory of the target object is extracted by optical flow field calculation. The material flexibility description is obtained by deformation tensor decomposition of the surface texture motion trajectory. The material flexibility description is then tensor-stitched with the image spatial gradient distribution to form a spatiotemporal fusion feature. A multi-scale response map is constructed by pyramid decomposition of spatiotemporal fusion features. Anomaly responses are iteratively transmitted between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence position and intensity peak of the anomaly responses. Geometric contours are extracted based on the location results of the foreign object area. A virtual contact scene is constructed in the spatial coordinate system by combining the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection. The paper shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. The current crack density prediction value is determined by combining historical damage data, and the life decay of each functional block is evaluated. Processing instructions are generated based on the amount of paper life decay, and the paper feeding mechanism and shredder assembly are driven to perform corresponding actions based on the processing instructions.

2. The method according to claim 1, characterized in that, Spatiotemporal joint analysis of multiple image sequences is performed, and the surface texture motion trajectory of the target object is extracted through optical flow field calculation, including: Pixel displacement field calculation is performed between adjacent frames in a multi-frame image sequence to obtain the initial optical flow field; The initial optical flow field is decomposed into spatial scales, and motion components are calculated at multiple preset spatial scales. Motion components with a coverage area greater than a preset first spatial scale threshold are identified as overall motion components, and motion components with a coverage area less than a preset second spatial scale threshold are identified as local deformation motion components. Local deformation motion components are extracted to form a local deformation motion field. A temporal correlation relationship is established for each pixel in the local deformation motion field. A motion chain graph is constructed by tracking the position evolution of each pixel between consecutive frames. The neighborhood cooperation index of each pixel is calculated on the motion chain graph. The neighborhood cooperation index is obtained by measuring the angle between the motion vector of the pixel and the motion amplitude difference between the pixel and the pixels in its spatial neighborhood. The motion chain graph is divided into regions based on the neighborhood collaboration index, and pixels that meet the preset collaboration conditions are merged into texture motion regions. Extract the motion path of each texture motion region and output the motion path as the surface texture motion trajectory of the target object.

3. The method according to claim 1, characterized in that, The material flexibility description is obtained by deformation tensor decomposition of surface texture motion trajectory. The material flexibility description is then tensor-concatenated with the image spatial gradient distribution to form spatiotemporal fusion features, including: Based on the motion vectors of the surface texture motion trajectory at each pixel position in the spatial domain, a second-order deformation tensor is constructed at each pixel position. Eigenvalue decomposition is performed on the second-order deformation tensor to obtain the eigenvalue set of each pixel position. The ratio of the maximum eigenvalue to the minimum eigenvalue in the eigenvalue set is calculated and used as the deformation direction distribution parameter of each pixel position. For the deformation direction distribution parameters of each pixel position, calculate the parameter difference value in the spatial neighborhood, and group the adjacent pixel positions with parameter difference values ​​less than the preset difference threshold into homogeneous deformation regions. Calculate the Euclidean distance between the deformation vector of each pixel position in each homogeneous deformation region and the average deformation vector of the region. Obtain the mean and variance of the Euclidean distance to form a material flexibility description. Spatial gradient calculation is performed on each frame of a multi-frame image sequence to obtain the gradient magnitude distribution and gradient direction distribution; The material flexibility description is aligned with the gradient magnitude and gradient direction distribution in the spatial domain. At each alignment position, the material flexibility parameters in the material flexibility description are tensor-concatenated with the gradient magnitude and gradient direction at the corresponding position to form a spatiotemporal fusion feature.

4. The method according to claim 1, characterized in that, A multi-scale response map is constructed by pyramid decomposition of spatiotemporal fusion features. Anomaly responses are iteratively propagated between different levels of the multi-scale response map. The location of the foreign object region is determined based on the convergence location and peak intensity of the anomaly responses. The results include: The spatiotemporal fusion features are decomposed into a pyramid according to spatial resolution, and the response values ​​of pixel positions are calculated at each decomposition level to construct a multi-scale response map. Based on the highest resolution level of the multi-scale response map, the pixel positions where the response value exceeds the preset seed threshold are selected to determine the abnormal response seed points. According to the response value and spatial gradient direction of each abnormal response seed point, the diffusion direction and diffusion weight to the adjacent pixel positions are calculated, and the response value is transferred to the adjacent pixel positions to obtain the updated response distribution of the current level. The updated response distribution is downsampled to the next level. At the next level, the diffusion direction and diffusion weight of each pixel position are recalculated based on the response value passed from the previous level. Abnormal response is passed on and repeated until the lowest resolution level is reached. During the passing process, the trajectory of the response value change of each pixel position at each level is recorded. Calculate the spatial offset of the response value of each pixel position between adjacent layers. When it is continuously less than the preset convergence threshold, mark the pixel position as the convergence state and extract the corresponding response value as the intensity peak. Pixels with intensity peaks exceeding a preset peak threshold are spatially connected and their locations are marked to determine the foreign object region location.

5. The method according to claim 1, characterized in that, Based on the location results of the foreign object area, the geometric contour is extracted. A virtual contact scene is constructed in a spatial coordinate system based on the motion trajectory of the shredder blade. The force distribution on the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection points, including: Boundary tracking is performed on the foreign object region location results to extract the geometric contour of the foreign object region; Extract the spatial coordinates of each boundary point from the geometric contour, calculate the tangent and normal directions for each boundary point, construct the local coordinate system for each boundary point, and calculate the curvature distribution and surface tilt distribution of the foreign object region under the local coordinate system. The motion trajectory of the paper shredder blade is obtained, and the motion trajectory is aligned with the geometric contour in a spatial coordinate system. In the aligned spatial coordinate system, the intersection time and intersection position of the motion trajectory and the geometric contour are determined to construct a virtual contact scene. In the virtual contact scenario, the curvature distribution and surface tilt angle distribution corresponding to the intersection positions are extracted. Based on the speed and direction of movement of the shredder blade at the intersection moment, the relative contact speed between the shredder blade and the foreign object area is calculated. The impact force amplitude at each intersection position is calculated based on the relative contact speed and curvature distribution. The direction of the impact force is determined based on the surface tilt angle distribution. The impact force amplitude and direction at each intersection position are mapped along the blade surface of the shredder blade to obtain the force distribution of the blade.

6. The method according to claim 5, characterized in that, The impact force amplitude at each intersection point is calculated based on the relative contact velocity and curvature distribution, including: Local extreme points of curvature distribution are extracted at the intersection points, and these local extreme points are used as the stress concentration cores. The curvature gradient is calculated from each stress concentration core to the surrounding area, and the stress diffusion radius is determined based on the magnitude of the curvature gradient. Within the stress diffusion radius, the peak stress of the stress concentration core is calculated based on the relative contact velocity, and the stress distribution at the surrounding location is calculated based on the distance from the stress concentration core and the attenuation rate of the curvature gradient. When there are multiple stress concentration cores at the intersection, the stress fields of each stress concentration core are superimposed, the location of the stress peak after superposition is extracted, and the stress peak is converted into the impact force amplitude.

7. The method according to claim 1, characterized in that, The paper shredder blade is divided into multiple functional blocks. The fatigue damage accumulation index of each functional block is calculated based on the stress distribution of the blade. Combined with historical damage data, the current crack density prediction value is determined, and the lifespan decay of each functional block is assessed, including: The paper shredder blade is divided into multiple functional blocks along the blade length and blade width directions; Extract the stress time sequence of each functional block from the stress distribution of the blade, perform peak detection, and count the amplitude and frequency of the stress peak points; A stress cycle spectrum is established based on the amplitude and frequency of the peak stress point. The number of cycles in different amplitude ranges is counted. Based on the preset material fatigue characteristic curve, the number of cycles in each amplitude range is converted into damage contribution and accumulated to obtain the fatigue damage accumulation index. Acquire historical damage data for each functional block, including the cumulative fatigue damage index and measured crack density within historical working cycles; Historical damage data of each functional block is used to construct a damage evolution chain in the order of working cycles. The first mutation point of fatigue damage accumulation index and the second mutation point of crack density are extracted. The damage propagation delay coefficient is calculated based on the time lag of the first and second mutation points. The fatigue damage accumulation index of the current cycle is combined with the damage propagation delay coefficient and mapped to the crack density response time. The corresponding measured crack density is extracted, the current crack density prediction value is determined, and the remaining life is calculated based on the difference between the preset failure threshold and the current crack density prediction value. The life decay of each functional block is evaluated.

8. A smart paper shredder foreign object detection and handling system integrating visual recognition, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire a multi-frame image sequence of the target object in the paper feed channel; The second unit is used to perform spatiotemporal joint analysis on multi-frame image sequences. It extracts the surface texture motion trajectory of the target object through optical flow field calculation, obtains the material flexibility description through deformation tensor decomposition of the surface texture motion trajectory, and spatiotemporally fused features by tensor splicing the material flexibility description with the image spatial gradient distribution. The third unit is used to construct a multi-scale response map by pyramid decomposing the spatiotemporal fusion features, iteratively transmitting the abnormal response between the different levels of the multi-scale response map, and determining the location result of the foreign object region based on the convergence position and intensity peak of the abnormal response. The fourth unit is used to extract the geometric contour based on the location result of the foreign object area, and to construct a virtual contact scene in the spatial coordinate system in combination with the motion trajectory of the shredder blade. The force distribution of the blade is calculated based on the curvature distribution and surface tilt angle distribution at the intersection position. The fifth unit is used to divide the shredder blade into multiple functional blocks, calculate the fatigue damage accumulation index of each functional block based on the stress distribution of the blade, determine the current crack density prediction value by combining historical damage data, and assess the life decay of each functional block. The sixth unit is used to generate processing instructions based on the amount of life decay, and drive the paper feeding mechanism and shredder assembly to perform corresponding actions based on the processing instructions.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images