A deep learning-based obsolete data storage device detection method and system

Abstract

The application discloses a kind of based on deep learning's abandoned data storage device detection method and system, it is related to computer vision technical field, including the following steps: step one: acquisition multi-view and normal sample image, form image set;Step two: complete registration normalization, construct component structure record and local coordinate domain;Step three: propagation damage potential energy, generate damage propagation mask and mechanism constraint abnormal image;Step four: construct improved DestSeg model, extract teacher and student features;Step five: encode position distance and connection direction, form structure relationship tensor;Step six: fusion relationship difference and feature difference, form single-view abnormal image;Step seven: analyze position change and keep interval, generate observation track persistence;Step eight: re-calibrate abnormal image and segment damage area, output damage type, position and grade.The application improves damage detection accuracy and stability.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for detecting abandoned data storage devices based on deep learning. Background Technology

[0002] With the expansion of data centers, the upgrading of terminal equipment, and the increasing scale of electronic waste recycling, technologies for visual damage detection and condition assessment of discarded data storage devices have received widespread attention. Existing detection methods mainly rely on manual visual inspection, single-view image recognition, or surface defect segmentation using ordinary convolutional networks. However, these methods generally suffer from the following problems in practical applications: The shell, interface, and encapsulation areas of discarded data storage devices have complex structures, with occlusion, scale differences, and directional variations between different components. Single-view images cannot fully present cracks, deformations, and ablation areas, easily leading to missed damage detection and location deviations. Existing deep learning detection models typically rely on a large number of real damage samples for training. The damage morphology of discarded equipment is dispersed, and sample annotation is costly. Abnormal images generated by simple pasting or random perturbation lack component connection relationships and damage propagation patterns, resulting in inconsistencies between training data and real damage mechanisms. Traditional feature difference methods focus on corresponding pixel or local texture changes, making it difficult to describe changes in positional distance and connection direction between components. At the same time, they lack cross-view anomaly response persistence analysis, easily misidentifying reflections, stains, and local shadows as damage, affecting the accuracy of damage area segmentation, damage type determination, and damage level assessment.

[0003] Therefore, how to provide a method and system for detecting abandoned data storage devices based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a method and system for detecting abandoned data storage devices based on deep learning. This invention employs multi-view image analysis and an improved DestSeg model, combined with damage potential energy propagation, structural relationship encoding, and observed trajectory persistence to complete damage region segmentation, type identification, and level assessment. It has the advantages of high detection accuracy, low false detection rate, and strong adaptability.

[0005] A method for detecting obsolete data storage devices based on deep learning according to an embodiment of the present invention includes the following steps: Step 1: Collect multi-view images from abandoned data storage devices and normal sample images to generate an image set; Step 2: Perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains; Step 3: Determine the damage initiation point based on the component structure record, propagate the damage potential energy along the component connection direction, and generate a damage propagation mask and mechanism constraint anomaly image; Step 4: Construct an improved DestSeg model, inputting normal sample images and mechanism-constrained abnormal images into the teacher network and the denoised student network respectively to generate teacher features and student features; Step 5: Map teacher and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher and student structural relationship tensors; Step 6: Calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher characteristics and student characteristics; Step 7: Map the single-view anomaly map to the local coordinate domain, record the positional changes and persistence intervals of the anomaly response along the view sequence, and generate the observation trajectory persistence. Step 8: Recalibrate the single-view anomaly map based on the observed orbit persistence, segment the damaged area, and generate the damage type, corresponding location, and damage level.

[0006] Optionally, step one specifically includes: The waste data storage device is fixed on the rotating stage, and the imaging distance, focal length, illumination intensity and acquisition angle interval are set. The rotation axis of the rotating stage and the image acquisition coordinate system are calibrated. The rotating stage is driven to complete one revolution around the rotation axis, and images of the abandoned data storage device are acquired at acquisition angle intervals. The device identification and acquisition position are recorded, and multi-view images of the abandoned data storage device are generated. Images from the same model of non-destructive data storage device are acquired according to imaging distance, focal length, illumination intensity and acquisition orientation. Images corresponding to each acquisition orientation are collected according to device identification to generate normal sample images. An image set is generated by associating multi-view images of abandoned data storage devices with normal sample images based on device identification and acquisition location.

[0007] Optionally, step two specifically includes: Select a multi-view image of a discarded data storage device from the image set as the reference image, extract the corresponding feature points in the reference image and the remaining multi-view images of the discarded data storage devices, and calculate the spatial transformation parameters based on the coordinate deviation of the corresponding feature points. Based on spatial transformation parameters, multi-view images of abandoned data storage devices and normal sample images are mapped to the coordinate range of the reference image. The pixel spacing of the mapped images is adjusted according to a uniform pixel size to generate a registered image set. Edge response extraction and closed region segmentation are performed on the registered image set. The component boundaries are determined based on the closed regions. The contact length and interval distance between component boundaries are calculated. The connection relationship is determined based on the contact length threshold and the interval distance threshold. Configure component identifiers for component boundaries, write the component identifiers, component boundary coordinates, and connection relationships into the component structure record, and establish a local coordinate domain with the center of the component boundary as the origin and the main direction of the component boundary as the horizontal axis.

[0008] Optionally, step three specifically includes: Read the component identifier, component boundary coordinates, and connection relationships from the component structure record; calculate the boundary curvature based on the component boundary coordinates; calculate the connection distance and connection direction deviation based on the connection relationships; and generate a damage initiation score based on the boundary curvature, connection distance, and connection direction deviation. The component boundary coordinates are sorted according to the damage initiation score. The component boundary coordinates corresponding to the maximum damage initiation score are determined as the damage initiation point, and an initial damage potential energy is configured for the damage initiation point. Taking the damage initiation point as the propagation source, the damage potential energy propagates along the connection direction of the components corresponding to the connection relationship. The potential energy attenuation is calculated based on the connection distance and the deviation of the connection direction, and the damage potential energy corresponding to the component boundary coordinates is updated step by step. The component boundary coordinates where the damage potential energy reaches the propagation threshold are marked as the damage propagation location. A damage propagation mask is generated based on the damage propagation location and the corresponding damage potential energy. The damage propagation mask is mapped onto a normal sample image, and the pixel values ​​are adjusted according to the corresponding damage potential energy to generate a mechanism-constrained abnormal image.

[0009] Optionally, step four specifically includes: An improved DestSeg model is established, which includes a teacher network and a denoised student network. The teacher network is set as a feature encoding structure with frozen parameters, and the denoised student network is set as an encoding and decoding structure with updated parameters. The teacher network and the denoised student network have the same feature output level and number of feature channels. Based on the device identification and acquisition location, normal sample images, mechanism constraint abnormal images generated based on normal sample images, and corresponding damage propagation masks are associated, and the pixel size of normal sample images, mechanism constraint abnormal images, and damage propagation masks is unified to generate training samples. Normal sample images are input into the teacher network for hierarchical feature encoding, while abnormal images constrained by mechanisms are input into the denoising student network for hierarchical feature encoding and feature decoding. Intermediate features of teachers and students are extracted from the same feature output level. The damage propagation mask is scaled according to the spatial dimensions of the teacher's intermediate features and the student's intermediate features. The damage propagation location is determined based on the scaled damage propagation mask. The denoising recovery deviation of the teacher's intermediate features and the student's intermediate features within the damage propagation location is calculated. The consistency deviation of the teacher's intermediate features and the student's intermediate features outside the damage propagation location is also calculated. The parameters of the denoised student network are updated based on the denoising recovery bias and consistency bias. The training samples are repeatedly input until the change in denoising recovery bias in adjacent training rounds is no greater than the convergence threshold. The parameters of the denoised student network are fixed, and the intermediate features of teachers are aggregated into teacher features and the intermediate features of students are aggregated into student features.

[0010] Optionally, step five specifically includes: The teacher and student features are mapped to the local coordinate domain according to the feature output level. The component identifiers are configured for the teacher and student feature positions based on the component boundary coordinates, and the teacher local feature table and the student local feature table are generated. Read the connection relationships in the component structure record, filter the component identifiers with connection relationships, extract feature positions from the teacher local feature table and the student local feature table according to the filtering results, and establish teacher feature position pairs and student feature position pairs; Calculate the coordinate distance and direction angle relative to the horizontal axis of the teacher feature position pairs in the local coordinate domain, and generate the teacher position distance matrix and the teacher connection direction matrix. Calculate the coordinate distance and direction angle relative to the horizontal axis of the student feature position pairs in the local coordinate domain, and generate the student position distance matrix and the student connection direction matrix. Based on the unified matrix index of component identification and feature output hierarchy, the teacher position distance matrix and teacher connection direction matrix are written into the teacher structure relation tensor, and the student position distance matrix and student connection direction matrix are written into the student structure relation tensor.

[0011] Optionally, step six specifically includes: Align the teacher structure relation tensor and the student structure relation tensor according to the component identifier and feature output level, calculate the position distance difference and connection direction difference of the corresponding tensor positions respectively, normalize the position distance difference and connection direction difference, and fuse the normalization results according to the set weight to generate the relation difference tensor. Based on the component identifier and feature output level corresponding to the relation difference tensor, the relation difference values ​​in the relation difference tensor are mapped to the local coordinate domain, and the relation difference values ​​are aggregated according to the feature position to generate a relation difference map; Align teacher features and student features according to the feature output hierarchy, calculate the channel feature difference at the corresponding feature position, average the absolute value of the channel feature difference, map the average result to the local coordinate domain, and generate a feature difference map. Adjust the relationship difference map and feature difference map to a uniform spatial size, normalize the numerical values ​​of the relationship difference map and feature difference map, calculate the abnormal response at the corresponding feature position according to the set fusion weight, collect the abnormal response according to the collection direction, and generate a single-view abnormal map.

[0012] Optionally, step seven specifically includes: Based on the acquisition location, read the single-view anomaly map, determine the component to which the anomaly response belongs according to the component identifier, convert the anomaly response position from image coordinates to local coordinates, arrange the anomaly responses according to the local coordinates, and generate a local anomaly map; Arrange the local anomaly maps in ascending order of acquisition location to form a view sequence. Calculate the local coordinate distance between anomaly response locations with consistent component identification in adjacent acquisition locations. Connect the anomaly responses with local coordinate distances not greater than the tracking threshold in sequence to generate anomaly response trajectories. The difference between the abscissa and ordinate of adjacent abnormal response positions in the abnormal response trajectory is calculated along the view sequence. The position change is generated based on the difference between the abscissa and ordinate. The starting and ending acquisition positions of the continuous existence of the abnormal response trajectory are recorded to generate the holding interval. The average position change is calculated based on the position changes of the abnormal response trajectory. The length of the holding interval is calculated based on the starting and ending acquisition azimuths. The average position change is reverse normalized, and the length of the holding interval is forward normalized. The normalized results are fused according to the set weights to generate the observation trajectory persistence.

[0013] Optionally, step eight specifically includes: Based on the abnormal response trajectory, the observation trajectory persistence is mapped to the single-view anomaly map corresponding to the acquisition azimuth. The abnormal response is recalibrated using the observation trajectory persistence, and a recalibrated single-view anomaly map is generated. Based on the segmentation threshold, the damage locations in the single-view anomaly map are screened and recalibrated. Damage locations with adjacent local coordinates and consistent component identification are connected to generate damage regions. Extract the area of ​​the damaged region, the aspect ratio of the boundary, the extension angle of the main direction relative to the component boundary, and the recalibration anomaly response. Determine the damage type based on the aspect ratio and extension angle, and generate the damage degree value based on the area and the recalibration anomaly response. The corresponding location is determined based on the component identification and center coordinates of the damaged area, and the damage level is determined based on the level range corresponding to the damage severity value, thus generating the damage type, corresponding location, and damage level.

[0014] A deep learning-based system for detecting obsolete data storage devices according to an embodiment of the present invention includes: The image acquisition module is used to acquire multi-view images of discarded data storage devices and normal sample images to generate an image set; The registration and modeling module is used to perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains. Anomaly construction module is used to determine the damage initiation point and propagate the damage potential energy along the component connection direction, generating a damage propagation mask and mechanism constraint anomaly image; The feature extraction module is used to build the improved DestSeg model. Normal sample images and mechanism-constrained abnormal images are input into the teacher network and the denoised student network, respectively, to generate teacher features and student features. The relation encoding module is used to map teacher features and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher structural relation tensors and student structural relation tensors; The difference fusion module is used to calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher features and student features. The orbit analysis module is used to map the single-view anomaly map to the local coordinate domain, record the position changes and persistence intervals of the anomaly response along the view sequence, and generate the observation orbit persistence. The damage output module is used to recalibrate the single-view anomaly map based on the observed orbital persistence, segment the damage area, and generate the damage type, corresponding location, and damage level.

[0015] The beneficial effects of this invention are: This invention acquires multi-view images of discarded data storage devices and normal sample images, performs spatial registration and scale normalization, and establishes component structure records and local coordinate domains based on component boundaries and connection relationships. This allows device images from different acquisition locations to be compared within a unified structural reference. Damage initiation points are determined based on boundary curvature, connection distance, and connection direction deviation, and damage potential energy is propagated along the component connection direction. This generates damage propagation masks and mechanism-constrained anomaly images that conform to the device structure and damage propagation laws, reducing reliance on a large number of real damage samples and improving the consistency between training samples and actual damage morphology.

[0016] This invention extracts teacher and student features using an improved DestSeg model and encodes positional distances and connection directions in the local coordinate domain. It fuses structural relationship differences with feature differences to generate a single-view anomaly map, enhancing the ability to identify component deformation, cracks, and ablation areas. By analyzing the positional changes and persistence intervals of anomaly responses along the viewpoint sequence and recalibrating the single-view anomaly map using observation trajectory persistence, it can suppress short-term anomaly responses caused by reflections, stains, and localized shadows, improving the accuracy and stability of damage area segmentation, damage type determination, corresponding location identification, and damage level assessment. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for detecting abandoned data storage devices based on deep learning, as proposed in this invention. Figure 2 This is a schematic diagram of the improved DestSeg model processing of a deep learning-based method for detecting abandoned data storage devices proposed in this invention. Figure 3 This is a schematic diagram illustrating the generation of observation orbit persistence in a deep learning-based method for detecting abandoned data storage devices proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for detecting obsolete data storage devices based on deep learning, comprising the following steps: Step 1: Collect multi-view images from abandoned data storage devices and normal sample images to generate an image set; Step 2: Perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains; Step 3: Determine the damage initiation point based on the component structure record, propagate the damage potential energy along the component connection direction, and generate a damage propagation mask and mechanism constraint anomaly image; Step 4: Construct an improved DestSeg model, inputting normal sample images and mechanism-constrained abnormal images into the teacher network and the denoised student network respectively to generate teacher features and student features; Step 5: Map teacher and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher and student structural relationship tensors; Step 6: Calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher characteristics and student characteristics; Step 7: Map the single-view anomaly map to the local coordinate domain, record the positional changes and persistence intervals of the anomaly response along the view sequence, and generate the observation trajectory persistence. Step 8: Recalibrate the single-view anomaly map based on the observed orbit persistence, segment the damaged area, and generate the damage type, corresponding location, and damage level.

[0020] In this embodiment, step one specifically includes: The waste data storage device is fixed at the center of the rotating stage, with the imaging surface of the device facing the image acquisition end. The imaging distance and focal length are set according to the device size. The lighting intensity is adjusted so that the brightness of the device surface is within the image acquisition range. The acquisition angle interval is set according to the ratio of the rotation angle of the rotating stage to the number of images acquired. An image acquisition coordinate system is established with the center of the rotating stage as the origin, and the position of the rotation axis in the image acquisition coordinate system is determined. The rotating platform is controlled to rotate around the rotation axis from the initial acquisition position. The rotation stops when it reaches a certain acquisition angle interval and the image of the waste data storage device is acquired. The rotation angle of the rotating platform relative to the initial acquisition position is used as the acquisition position. The device identifier and acquisition position are written into the acquired waste data storage device image. The acquired waste data storage device images are arranged in ascending order of acquisition position to generate multi-view images of the waste data storage device. Select a data storage device with the same model as the discarded data storage device and no surface damage. Keep the imaging distance, focal length and illumination intensity unchanged. Control the rotating stage to rotate at the same acquisition angle interval. Acquire images at the acquisition positions corresponding to the multi-view images of the discarded data storage device. Write the device identifier and acquisition position to the acquired images. Collect the images corresponding to each acquisition position according to the device identifier to generate normal sample images. Read the device identifier and acquisition location carried in multi-view images of abandoned data storage devices and normal sample images, establish a correspondence between abandoned data storage device images and normal sample images with the same acquisition location, and collect the corresponding relationship and associated images according to the device identifier to generate an image set.

[0021] In this embodiment, step two specifically includes: Multi-view images of abandoned data storage devices with the initial acquisition orientation are selected from the image set as reference images. Feature points with stable grayscale changes are extracted from the reference image and other multi-view images of abandoned data storage devices. The correspondence between feature points is established based on the distance of the feature description vector. Corresponding feature points with coordinate transformation errors exceeding the error threshold are removed. The rotation component, translation component, and scaling component are solved based on the coordinates of the retained corresponding feature points to generate spatial transformation parameters. Based on the spatial transformation parameters, coordinate transformation is performed on the multi-view images of the abandoned data storage device. Normal sample images are mapped to the coordinate range of the reference image according to the spatial transformation parameters corresponding to the same acquisition orientation. The scale coefficient is calculated based on the ratio between the mapped image size and the uniform pixel size. The pixel spacing of the mapped image is adjusted using the scale coefficient to generate a registered image set. The grayscale change amplitude of adjacent pixels is calculated for the registered image set. Edge response pixels are selected based on the grayscale change amplitude. Spatially continuous edge response pixels are connected and edge gaps are closed. Closed regions are divided according to the area enclosed by the closed edges. The contour of the closed region is determined as the component boundary. The overlapping pixel length of adjacent component boundaries is counted as the contact length. The minimum pixel distance between non-contact component boundaries is calculated as the interval distance. The connection relationship is determined based on the contact length threshold and the interval distance threshold. Configure component identifiers according to the arrangement of component boundaries within the coordinate range of the reference image. Write the component identifiers, component boundary coordinates, and connection relationships into the component structure record according to the corresponding relationship. Calculate the geometric center of the component boundary coordinates as the origin of the coordinate system. Determine the main direction of the component boundary based on the extension direction of the component boundary coordinates. Set the main direction of the component boundary as the horizontal axis and set the direction that passes through the origin of the coordinate system and is perpendicular to the horizontal axis as the vertical axis to establish a local coordinate domain.

[0022] In this embodiment, step three specifically includes: Read the component identifier, component boundary coordinates, and connection relationship from the component structure record. Extract the corresponding component boundary coordinates according to the component identifier. Calculate the boundary curvature based on the directional change formed by adjacent boundary coordinates. Calculate the minimum coordinate distance and determine the connection distance based on the boundary coordinates of two sets of connected components. Calculate the angle between the main direction of the component boundary and the direction of the connecting line and determine the connection direction deviation. Normalize the boundary curvature, connection distance, and connection direction deviation. Sum the normalized results according to the set weights to generate the damage initiation score corresponding to the boundary coordinates of each component. Arrange the component boundary coordinates in descending order of damage initiation score, read the maximum damage initiation score and the corresponding component boundary coordinates, determine the corresponding component boundary coordinates as the damage initiation point, write the preset potential energy value at the damage initiation point as the initial damage potential energy, and make the damage initiation point the starting position for the propagation of damage potential energy. Using the damage initiation point as the propagation source, the boundary coordinates of the components connected to the damage initiation point are read, and the initial damage potential energy is transmitted along the connection direction of the components corresponding to the connection relationship. The potential energy attenuation is calculated based on the propagation loss caused by the increase in connection distance and the direction loss caused by the increase in connection direction deviation. The potential energy attenuation is subtracted from the transmitted damage potential energy to obtain the updated damage potential energy, and the damage potential energy corresponding to the component boundary coordinates is updated step by step according to the connection relationship. By comparing the damage potential energy corresponding to the component boundary coordinates with the propagation threshold, the component boundary coordinates that reach the propagation threshold are marked as damage propagation locations. A damage propagation mask is generated according to the coordinate range of the damage propagation location in the normal sample image. The damage propagation mask is then applied to the normal sample image. The pixel adjustment range is determined based on the damage potential energy corresponding to each damage propagation location, and the pixel values ​​within the damage propagation location are adjusted to generate a mechanism-constrained abnormal image.

[0023] In this embodiment, step four specifically includes: An improved DestSeg model is established, which includes a teacher network and a denoising student network. The teacher network adopts a layer-by-layer downsampling feature encoding structure, loads normal image feature extraction parameters and stops parameter updates. The denoising student network adopts a feature encoding structure and a layer-by-layer upsampling feature decoding structure corresponding to the teacher network. The output positions with the same spatial size in the teacher network and the denoising student network are set as the same feature output level, and the output dimension of each feature output level is adjusted so that the teacher network and the denoising student network have the same number of feature channels at the same feature output level. Read the device identifier and acquisition location carried by the normal sample image, read the device identifier and acquisition location corresponding to the mechanism constraint abnormal image, establish the association between the normal sample image, the mechanism constraint abnormal image and the damage propagation mask with the same device identifier and acquisition location, adjust the size of the normal sample image and the mechanism constraint abnormal image according to the set number of image rows and image columns, and adjust the damage propagation mask according to the same row and column scaling ratio to generate training samples; Normal sample images from the training samples are input into the teacher network, and feature encoding is performed sequentially according to the feature output level. The feature data output by each feature output level is recorded as the teacher intermediate features. Mechanism constraint abnormal images from the training samples are input into the denoising student network, and feature encoding and feature decoding are performed sequentially. The output data that is consistent with the teacher intermediate feature space size and the number of feature channels is recorded as the student intermediate features. The damage propagation mask is scaled according to the number of rows and columns of the teacher's intermediate features. The coordinates of the scaled numerical indicators of damage are determined as the damage propagation location. The channel difference between the teacher's intermediate features and the student's intermediate features within the damage propagation location is calculated coordinate by coordinate. The absolute values ​​of the channel differences are averaged to generate the denoising recovery bias. The coordinates of the scaled numerical indicators of no damage are determined as the area outside the damage propagation location. The consistency bias is generated according to the same calculation method. The denoising recovery bias and the consistency bias are added together to form the basis for parameter updates. The parameters of the denoising student network are adjusted in the direction of reducing the parameter update basis. The training samples are repeatedly input and the denoising recovery bias of each training round is recorded. The absolute difference of the denoising recovery bias between adjacent training rounds is calculated until the absolute difference is not greater than the convergence threshold. The parameter updates are stopped and the parameters of the denoising student network are fixed. Teacher features are generated by collecting teacher intermediate features according to the feature output hierarchy, and student features are generated by collecting student intermediate features according to the feature output hierarchy.

[0024] In this embodiment, step five specifically includes: The teacher and student features are read according to the feature output level. Based on the ratio between the image size and the feature size, the teacher and student feature positions are converted into image coordinates. The origin, horizontal axis, and vertical axis of the local coordinate domain are read. The displacement of the image coordinates relative to the origin is projected onto the horizontal axis and vertical axis to obtain the local coordinates. The component identifier is configured according to whether the local coordinates fall within the range enclosed by the component boundary coordinates. The feature output level, component identifier, local coordinates, and feature values ​​are written into the teacher local feature table and the student local feature table, respectively. Read the connection relationship in the component structure record, form component identifier pairs by combining the component identifiers at both ends of the connection relationship, extract the teacher feature positions within the corresponding component boundary range from the teacher local feature table and extract the student feature positions within the corresponding component boundary range from the student local feature table, and combine the teacher feature positions and student feature positions according to the corresponding order of the local coordinates to establish teacher feature position pairs and student feature position pairs. For teacher feature position pairs, calculate the difference between the x-coordinates and y-coordinates of the two teacher feature positions. Take the square root of the sum of the squares of the x-coordinate difference and the squares of the y-coordinate difference to obtain the coordinate distance. Calculate the direction angle relative to the x-axis based on the difference between the y-coordinates and the x-coordinates. Arrange the coordinate distances and direction angles according to the component identification pairs and feature output hierarchy to generate a teacher position distance matrix and a teacher connection direction matrix. The same processing procedure is used for student feature position pairs to generate a student position distance matrix and a student connection direction matrix. Establish a unified matrix index according to the arrangement order of component identifier pairs and feature output level. Write the coordinate distance and direction angle that are consistent with the index in the teacher position distance matrix and the teacher connection direction matrix into the same tensor position to generate the teacher structure relationship tensor. Write the student position distance matrix and the student connection direction matrix into the corresponding tensor positions according to the unified matrix index to generate the student structure relationship tensor.

[0025] In this embodiment, step six specifically includes: Read the teacher and student structural relationship tensors according to the component identifier and feature output level. Establish a correspondence between the positions of tensors with the same component identifier and feature output level. Calculate the absolute difference between the distance between the teacher and student positions in the corresponding tensor positions to generate the position distance difference. Calculate the minimum angle between the teacher's connection direction and the student's connection direction to generate the connection direction difference. Perform numerical normalization based on the maximum value of the position distance difference and the maximum value of the connection direction difference. Multiply the normalization results by the set weights and add them together. Write the calculation results according to the corresponding tensor positions to generate the relationship difference tensor. Read the component identifier and feature output level corresponding to each relation difference value in the relation difference tensor, determine the component boundary range in the local coordinate domain based on the component identifier, determine the feature position based on the teacher feature position pair and student feature position pair corresponding to the relation difference value, write the relation difference value into the local coordinate corresponding to the feature position, calculate the average value of the relation difference value received for the same local coordinate, arrange the average value according to the local coordinate, and generate a relation difference map. Read teacher and student features with the same spatial size and number of feature channels according to the feature output level, establish a correspondence between the same feature positions and the same feature channels, calculate the difference between the teacher feature value and the student feature value in the corresponding feature channel, obtain the absolute value of the difference, and average the absolute values ​​of the differences contained in the same feature position. Write the average result into the local coordinate domain according to the component identifier corresponding to the feature position, arrange the average result according to the local coordinate, and generate a feature difference map. Read the number of rows and columns of the relationship difference map and the feature difference map, adjust the relationship difference map and the feature difference map to a uniform spatial size according to the ratio of the number of rows and columns, and perform numerical normalization according to the minimum and maximum values ​​in the map respectively. Multiply the normalized relationship difference value and feature difference value by the set fusion weight and add them together to generate the abnormal response at each feature location. Write the abnormal response to the corresponding abandoned data storage device multi-view image according to the acquisition direction to generate a single-view abnormal map.

[0026] In this embodiment, step seven specifically includes: Based on the acquisition location, read the single-view anomaly map, read the image coordinates corresponding to each anomaly response position in the single-view anomaly map, determine the component to which the anomaly response belongs based on the component boundary coordinates and configure the component identifier, project the displacement of the anomaly response position relative to the origin of the local coordinate domain to the horizontal and vertical axes to obtain the local coordinates of the anomaly response position, write the anomaly response value into the corresponding position according to the component identifier and local coordinates, and generate a local anomaly map. Local anomaly maps are arranged in ascending order of acquisition location to form a view sequence. Two local anomaly maps corresponding to adjacent acquisition locations are selected in sequence. Within the range where the component identification is consistent, the difference between the horizontal and vertical coordinates of the anomaly response locations is calculated. The square root of the sum of the squares of the horizontal and vertical coordinate differences is taken to obtain the local coordinate distance. Anomaly responses with a local coordinate distance not greater than the tracking threshold are connected, and the connection relationship is extended along the view sequence to generate anomaly response trajectories. The abnormal response positions in the abnormal response trajectory are read along the view sequence. The difference between the horizontal and vertical coordinates of adjacent abnormal response positions is calculated. The square root of the sum of the squares of the horizontal and vertical coordinate differences is taken to generate the positional change between adjacent acquisition positions. The acquisition position corresponding to the first occurrence of the abnormal response trajectory is read as the starting acquisition position. The last acquisition position of the continuous existence of the abnormal response trajectory is read as the ending acquisition position. The holding interval is generated based on the starting and ending acquisition positions. The positional changes in the abnormal response trajectory are summed and divided by the number of positional changes to obtain the average positional change. The azimuth difference between the termination and start of acquisition is calculated. The azimuth difference is converted into the hold interval length based on the acquisition angle interval. The maximum and minimum values ​​of the average positional change in all abnormal response trajectories are used to perform reverse normalization, so that the normalization result increases when the average positional change decreases. The maximum and minimum values ​​of the hold interval length in all abnormal response trajectories are used to perform forward normalization, so that the normalization result increases when the hold interval length increases. The two normalization results are multiplied by a set weight and then summed to generate the observation trajectory persistence.

[0027] In this embodiment, step eight specifically includes: Read the observation trajectory persistence, abnormal response location and acquisition azimuth corresponding to the abnormal response trajectory. Map the observation trajectory persistence to the single-view anomaly map corresponding to the acquisition azimuth based on the abnormal response location. Use the observation trajectory persistence as the recalibration coefficient of the abnormal response. Adjust the abnormal response value using the recalibration coefficient to generate a recalibrated abnormal response. Arrange the recalibrated abnormal responses according to the abnormal response location in the single-view anomaly map to generate a recalibrated single-view anomaly map. The recalibration anomaly response in the recalibration single-view anomaly map is numerically normalized. The normalization result is compared with the segmentation threshold. Anomaly response locations not less than the segmentation threshold are marked as damage locations. The adjacency relationship between damage locations is determined based on local coordinates. Damage locations with adjacent local coordinates and consistent component identification are connected into continuous regions. Region division is completed according to the boundary of continuous regions to generate damage regions. Read the damage location contained in the damage area, calculate the area of ​​the area based on the boundary range enclosed by the damage location, generate the aspect ratio of the boundary based on the span of the boundary range in the horizontal and vertical directions of the local coordinates, generate the extension angle based on the angle between the extension direction of the damage area and the main direction of the component boundary, determine the damage type according to the judgment range corresponding to the boundary aspect ratio and the extension angle, collect the recalibration abnormal response in the damage area, and generate the damage degree value by combining the area area. Read the component identifier corresponding to the damaged area, determine the center coordinates based on the local coordinates of the damaged location within the damaged area, and use the component identifier and center coordinates together to determine the corresponding location of the damaged area. Match the damage severity value with the level range, determine the damage level based on the matching result, collect the damage type, corresponding location and damage level, and generate the damage type, corresponding location and damage level.

[0028] A deep learning-based system for detecting obsolete data storage devices, comprising: The image acquisition module is used to acquire multi-view images of discarded data storage devices and normal sample images to generate an image set; The registration and modeling module is used to perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains. Anomaly construction module is used to determine the damage initiation point and propagate the damage potential energy along the component connection direction, generating a damage propagation mask and mechanism constraint anomaly image; The feature extraction module is used to build the improved DestSeg model. Normal sample images and mechanism-constrained abnormal images are input into the teacher network and the denoised student network, respectively, to generate teacher features and student features. The relation encoding module is used to map teacher features and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher structural relation tensors and student structural relation tensors; The difference fusion module is used to calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher features and student features. The orbit analysis module is used to map the single-view anomaly map to the local coordinate domain, record the position changes and persistence intervals of the anomaly response along the view sequence, and generate the observation orbit persistence. The damage output module is used to recalibrate the single-view anomaly map based on the observed orbital persistence, segment the damage area, and generate the damage type, corresponding location, and damage level.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the appearance inspection of waste data storage devices in an electronic waste processing production line. The inspection targets included mechanical hard drives, solid-state drives, and mobile storage devices. Damage types included casing cracks, corner deformation, interface damage, surface ablation, and localized dents. A total of 1200 devices were tested, including 300 normal devices and 900 damaged devices. 1386 damaged areas were manually marked. During the inspection, each device was fixed on a rotating stage, with an imaging distance of 450 mm, a focal length of 35 mm, an illumination intensity of 900 lux, and an acquisition angle interval of 30 degrees. Each device acquired 12 multi-view images of the waste data storage device, and images of undamaged data storage devices of the same model were acquired under the same acquisition conditions as normal sample images. The test lasted for 30 working days, accumulating 14400 multi-view images of waste data storage devices and 3600 normal sample images.

[0030] After image acquisition, spatial registration and scale normalization were performed on multi-view images of the abandoned data storage device and normal sample images, uniformly adjusting the images to 512×512 pixels. By extracting component boundaries and connection relationships, component structure records and local coordinate domains corresponding to the shell, interface, label area, and corner areas were established. For each component boundary, the boundary curvature, connection distance, and connection direction deviation were calculated to determine the damage initiation point and configure the initial damage potential energy. The damage potential energy propagated step by step along the component connection direction, with a propagation threshold set to 0.42. The component boundary coordinates that reached the propagation threshold were marked as damage propagation positions, and a damage propagation mask was generated accordingly. After the damage propagation mask was mapped onto the normal sample images, the pixel values ​​were adjusted according to the damage potential energy, generating a total of 7200 mechanism-constrained anomaly images to address the problems of insufficient number of real damage samples and lack of structural constraints in randomly synthesized anomaly images.

[0031] The improved DestSeg model consists of a teacher network and a denoising student network. The teacher network parameters are kept frozen, while the denoising student network updates its parameters based on training samples. Training samples consist of normal sample images, mechanistic constraint anomaly images, and damage propagation masks. The training run is set to 120 epochs, with a batch size of 16. After mapping teacher and student features to the local coordinate domain, the model generates teacher and student structural relationship tensors based on the component identifier and feature output hierarchical encoding location distance and connection direction. The model simultaneously calculates structural relationship differences and channel feature differences, and generates a single-view anomaly map with fusion weights of 0.55 and 0.45, respectively. This ensures that structural changes caused by interface breaks, corner deformations, and shell cracks are prominently represented, reducing false negatives caused by relying solely on texture differences.

[0032] Single-view anomaly maps are arranged according to the acquisition orientation to form a view sequence, and the tracking threshold is set to 3% of the diagonal length of the local coordinate domain. Anomaly responses with consistent component identification and local coordinate distance not exceeding the tracking threshold in adjacent acquisition orientations are connected to form anomaly response trajectories. The average position change and hold interval length are calculated based on the anomaly response trajectories, and reverse normalization and forward normalization are performed respectively to generate observation track persistence. The observation track persistence is used to recalibrate the single-view anomaly maps, enhancing the real damage responses that appear continuously in adjacent acquisition orientations and suppressing the reflection, stain, and shadow responses that appear only in a single acquisition orientation. The segmentation threshold is set to 0.48. After the damage area is formed, the damage type, corresponding location, and damage level are determined based on the area, boundary aspect ratio, extension angle relative to the main direction of the component boundary, and recalibrated anomaly responses. Comparison schemes include a defect segmentation scheme based on a single-view convolutional neural network, an anomaly segmentation scheme using the original DestSeg model, and the detection scheme proposed in this invention, as shown in Table 1: Table 1 Comparison of Equipment Damage Detection Results

[0033] As shown in Table 1, the damage detection rate of the detection scheme of this invention is 10.3 percentage points higher than that of the single-view convolutional neural network scheme and 5.4 percentage points higher than that of the original DestSeg scheme, while the damage false detection rate is reduced by 8.2 percentage points and 4.5 percentage points, respectively. The region intersection-union ratio reaches 0.836, indicating that the segmentation results have a high degree of overlap with the manually marked regions, and the damage type identification accuracy reaches 94.8%. Although the average detection time per unit increases to 1.58 seconds, it still meets the continuous detection requirements of the production line. In the test, for metal shell equipment with obvious reflectivity, the traditional scheme had 73 false alarms, the original DestSeg scheme had 49 false alarms, and the present invention had only 21 false alarms. For small cracks near the interface, the present invention detected 112, and 109 were manually verified and confirmed, indicating that the component structure relationship encoding and observation track persistence can effectively enhance the real damage response and improve the accuracy and stability of damage detection of waste data storage equipment.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting obsolete data storage devices based on deep learning, characterized in that, Includes the following steps: Step 1: Collect multi-view images from abandoned data storage devices and normal sample images to generate an image set; Step 2: Perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains; Step 3: Determine the damage initiation point based on the component structure record, propagate the damage potential energy along the component connection direction, and generate a damage propagation mask and mechanism constraint anomaly image; Step 4: Construct an improved DestSeg model, inputting normal sample images and mechanism-constrained abnormal images into the teacher network and the denoised student network respectively to generate teacher features and student features; Step 5: Map teacher and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher and student structural relationship tensors; Step 6: Calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher characteristics and student characteristics; Step 7: Map the single-view anomaly map to the local coordinate domain, record the positional changes and persistence intervals of the anomaly response along the view sequence, and generate the observation trajectory persistence. Step 8: Recalibrate the single-view anomaly map based on the observed orbit persistence, segment the damaged area, and generate the damage type, corresponding location, and damage level.

2. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step one specifically involves: The waste data storage device is fixed on the rotating stage, and the imaging distance, focal length, illumination intensity and acquisition angle interval are set. The rotation axis of the rotating stage and the image acquisition coordinate system are calibrated. The rotating stage is driven to complete one revolution around the rotation axis, and images of the abandoned data storage device are acquired at acquisition angle intervals. The device identification and acquisition position are recorded, and multi-view images of the abandoned data storage device are generated. Images from the same model of non-destructive data storage device are acquired according to imaging distance, focal length, illumination intensity and acquisition orientation. Images corresponding to each acquisition orientation are collected according to device identification to generate normal sample images. An image set is generated by associating multi-view images of abandoned data storage devices with normal sample images based on device identification and acquisition location.

3. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step two specifically involves: Select a multi-view image of a discarded data storage device from the image set as the reference image, extract the corresponding feature points in the reference image and the remaining multi-view images of the discarded data storage devices, and calculate the spatial transformation parameters based on the coordinate deviation of the corresponding feature points. Based on spatial transformation parameters, multi-view images of abandoned data storage devices and normal sample images are mapped to the coordinate range of the reference image. The pixel spacing of the mapped images is adjusted according to a uniform pixel size to generate a registered image set. Edge response extraction and closed region segmentation are performed on the registered image set. The component boundaries are determined based on the closed regions. The contact length and interval distance between component boundaries are calculated. The connection relationship is determined based on the contact length threshold and the interval distance threshold. Configure component identifiers for component boundaries, write the component identifiers, component boundary coordinates, and connection relationships into the component structure record, and establish a local coordinate domain with the center of the component boundary as the origin and the main direction of the component boundary as the horizontal axis.

4. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step three specifically involves: Read the component identifier, component boundary coordinates, and connection relationships from the component structure record; calculate the boundary curvature based on the component boundary coordinates; calculate the connection distance and connection direction deviation based on the connection relationships; and generate a damage initiation score based on the boundary curvature, connection distance, and connection direction deviation. The component boundary coordinates are sorted according to the damage initiation score. The component boundary coordinates corresponding to the maximum damage initiation score are determined as the damage initiation point, and an initial damage potential energy is configured for the damage initiation point. Taking the damage initiation point as the propagation source, the damage potential energy propagates along the connection direction of the components corresponding to the connection relationship. The potential energy attenuation is calculated based on the connection distance and the deviation of the connection direction, and the damage potential energy corresponding to the component boundary coordinates is updated step by step. The component boundary coordinates where the damage potential energy reaches the propagation threshold are marked as the damage propagation location. A damage propagation mask is generated based on the damage propagation location and the corresponding damage potential energy. The damage propagation mask is mapped onto a normal sample image, and the pixel values ​​are adjusted according to the corresponding damage potential energy to generate a mechanism-constrained abnormal image.

5. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step four specifically involves: An improved DestSeg model is established, which includes a teacher network and a denoised student network. The teacher network is set as a feature encoding structure with frozen parameters, and the denoised student network is set as an encoding and decoding structure with updated parameters. The teacher network and the denoised student network have the same feature output level and number of feature channels. Based on the device identification and acquisition location, normal sample images, mechanism constraint abnormal images generated based on normal sample images, and corresponding damage propagation masks are associated, and the pixel size of normal sample images, mechanism constraint abnormal images, and damage propagation masks is unified to generate training samples. Normal sample images are input into the teacher network for hierarchical feature encoding, while abnormal images constrained by mechanisms are input into the denoising student network for hierarchical feature encoding and feature decoding. Intermediate features of teachers and students are extracted from the same feature output level. The damage propagation mask is scaled according to the spatial dimensions of the teacher's intermediate features and the student's intermediate features. The damage propagation location is determined based on the scaled damage propagation mask. The denoising recovery deviation of the teacher's intermediate features and the student's intermediate features within the damage propagation location is calculated. The consistency deviation of the teacher's intermediate features and the student's intermediate features outside the damage propagation location is also calculated. The parameters of the denoised student network are updated based on the denoising recovery bias and consistency bias. The training samples are repeatedly input until the change in denoising recovery bias in adjacent training rounds is no greater than the convergence threshold. The parameters of the denoised student network are fixed, and the intermediate features of teachers are aggregated into teacher features and the intermediate features of students are aggregated into student features.

6. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step five specifically involves: The teacher and student features are mapped to the local coordinate domain according to the feature output level. The component identifiers are configured for the teacher and student feature positions based on the component boundary coordinates, and the teacher local feature table and the student local feature table are generated. Read the connection relationships in the component structure record, filter the component identifiers with connection relationships, extract feature positions from the teacher local feature table and the student local feature table according to the filtering results, and establish teacher feature position pairs and student feature position pairs; Calculate the coordinate distance and direction angle relative to the horizontal axis of the teacher feature position pairs in the local coordinate domain, and generate the teacher position distance matrix and the teacher connection direction matrix. Calculate the coordinate distance and direction angle relative to the horizontal axis of the student feature position pairs in the local coordinate domain, and generate the student position distance matrix and the student connection direction matrix. Based on the unified matrix index of component identification and feature output hierarchy, the teacher position distance matrix and teacher connection direction matrix are written into the teacher structure relation tensor, and the student position distance matrix and student connection direction matrix are written into the student structure relation tensor.

7. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step six specifically involves: Align the teacher structure relation tensor and the student structure relation tensor according to the component identifier and feature output level, calculate the position distance difference and connection direction difference of the corresponding tensor positions respectively, normalize the position distance difference and connection direction difference, and fuse the normalization results according to the set weight to generate the relation difference tensor. Based on the component identifier and feature output level corresponding to the relation difference tensor, the relation difference values ​​in the relation difference tensor are mapped to the local coordinate domain, and the relation difference values ​​are aggregated according to the feature position to generate a relation difference map; Align teacher features and student features according to the feature output hierarchy, calculate the channel feature difference at the corresponding feature position, average the absolute value of the channel feature difference, map the average result to the local coordinate domain, and generate a feature difference map. Adjust the relationship difference map and feature difference map to a uniform spatial size, normalize the numerical values ​​of the relationship difference map and feature difference map, calculate the abnormal response at the corresponding feature position according to the set fusion weight, collect the abnormal response according to the collection direction, and generate a single-view abnormal map.

8. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step seven specifically involves: Based on the acquisition location, read the single-view anomaly map, determine the component to which the anomaly response belongs according to the component identifier, convert the anomaly response position from image coordinates to local coordinates, arrange the anomaly responses according to the local coordinates, and generate a local anomaly map; Arrange the local anomaly maps in ascending order of acquisition location to form a view sequence. Calculate the local coordinate distance between anomaly response locations with consistent component identification in adjacent acquisition locations. Connect the anomaly responses with local coordinate distances not greater than the tracking threshold in sequence to generate anomaly response trajectories. The difference between the abscissa and ordinate of adjacent abnormal response positions in the abnormal response trajectory is calculated along the view sequence. The position change is generated based on the difference between the abscissa and ordinate. The starting and ending acquisition positions of the continuous existence of the abnormal response trajectory are recorded to generate the holding interval. The average position change is calculated based on the position changes of the abnormal response trajectory. The length of the holding interval is calculated based on the starting and ending acquisition azimuths. The average position change is reverse normalized, and the length of the holding interval is forward normalized. The normalized results are fused according to the set weights to generate the observation trajectory persistence.

9. The method for detecting abandoned data storage devices based on deep learning according to claim 1, characterized in that, Step eight specifically involves: Based on the abnormal response trajectory, the observation trajectory persistence is mapped to the single-view anomaly map corresponding to the acquisition azimuth. The abnormal response is recalibrated using the observation trajectory persistence, and a recalibrated single-view anomaly map is generated. Based on the segmentation threshold, the damage locations in the single-view anomaly map are screened and recalibrated. Damage locations with adjacent local coordinates and consistent component identification are connected to generate damage regions. Extract the area of ​​the damaged region, the aspect ratio of the boundary, the extension angle of the main direction relative to the component boundary, and the recalibration anomaly response. Determine the damage type based on the aspect ratio and extension angle, and generate the damage degree value based on the area and the recalibration anomaly response. The corresponding location is determined based on the component identification and center coordinates of the damaged area, and the damage level is determined based on the level range corresponding to the damage severity value, thus generating the damage type, corresponding location, and damage level.

10. A deep learning-based system for detecting obsolete data storage devices, comprising executing the deep learning-based method for detecting obsolete data storage devices according to any one of claims 1 to 9, characterized in that it includes: The image acquisition module is used to acquire multi-view images of discarded data storage devices and normal sample images to generate an image set; The registration and modeling module is used to perform spatial registration and scale normalization on the image set, extract component boundaries and connection relationships, and generate component structure records and local coordinate domains. Anomaly construction module is used to determine the damage initiation point and propagate the damage potential energy along the component connection direction, generating a damage propagation mask and mechanism constraint anomaly image; The feature extraction module is used to build the improved DestSeg model. Normal sample images and mechanism-constrained abnormal images are input into the teacher network and the denoised student network, respectively, to generate teacher features and student features. The relation encoding module is used to map teacher features and student features to the local coordinate domain, encode positional distance and connection direction, and generate teacher structural relation tensors and student structural relation tensors; The difference fusion module is used to calculate the relationship difference between the teacher structural relationship tensor and the student structural relationship tensor, and generate a single-view anomaly graph by combining the feature differences between teacher features and student features. The orbit analysis module is used to map a single-view anomaly map to a local coordinate domain, record the positional changes and persistence intervals of the anomaly response along the view sequence, and generate the observation orbit persistence. The damage output module is used to recalibrate the single-view anomaly map based on the observed orbital persistence, segment the damage area, and generate the damage type, corresponding location, and damage level.

Images