AOI detection image and result data association system under industrial big data

Abstract

The present application relates to the technical field of industrial big data, and specifically relates to an AOI detection image and result data association system under industrial big data. The system obtains the optical feature vector and spatial coordinates of the automatic optical detection image, maps the spatial coordinates to a preset grid space, reduces the dimension of the optical feature vector, and splices the grid code to generate a space-time feature composite hash key; a hash index table with the composite hash key as the table header is constructed in the memory layer, and a balanced multi-path lookup tree index with the optical feature vector as the key and the composite hash key as the value is constructed in the solid state disk layer; when receiving an association request, the target composite hash key is queried in the memory layer hash index table, the defect category is converted into an auxiliary search vector when a miss occurs, and a range query is performed through the solid state disk layer index to output an image storage address. The present application shortens the response time of cross-node association search, reduces the disk input / output load of the distributed storage node, and improves the association accuracy of unstructured images and structured result data.

Description

Technical Field

[0001] This invention relates to the field of industrial big data technology, specifically to a system for associating AOI detection images and results data under industrial big data. Background Technology

[0002] In industrial big data scenarios, automated optical inspection equipment generates massive amounts of high-resolution inspection images and corresponding structured defect result data. Current conventional correlation solutions use relational databases for table partitioning, recording the image storage path in the image table and the defect category, coordinates, and other results in the result table. The two tables are linked by a foreign key using the product serial number. When the original image corresponding to a specific defect result needs to be retrieved, the system extracts the foreign key from the result table based on the query command and uses this foreign key to perform a cross-table join query in the image table. Because the image paths and result data are scattered across different physical storage nodes, the join query process requires multiple data packet exchanges at the network layer and random read / write operations at the disk layer to locate the actual physical address of the image.

[0003] When faced with massive amounts of high-resolution image data, the aforementioned conventional solutions generate a large number of disk input / output operations through cross-node joint queries, resulting in high response latency and high system resource consumption when distributed storage nodes perform image-result data association retrieval. Summary of the Invention

[0004] The purpose of this invention is to provide a system for associating AOI detection images and result data under industrial big data, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An industrial big data-based AOI inspection image and result data association system includes an optical feature extraction component, a coordinate mapping component, a composite hash key generation component, a memory layer index construction component, a solid-state drive layer index construction component, and an association retrieval component; The optical feature extraction component acquires the optical feature vector of the AOI detection image; The coordinate mapping component maps the spatial coordinates of the printed circuit board to a preset grid space; The composite hash key generation component performs dimensionality reduction on the optical feature vector and concatenates it with the grid code of the grid space to generate a spatiotemporal feature composite hash key. The memory layer index building component builds a hash index table in the memory layer with the spatiotemporal feature composite hash key as the header; The solid-state drive layer index building component constructs a B+ tree index in the solid-state drive layer with the optical feature vector as the key and the spatiotemporal feature composite hash key as the value. When the associated retrieval component receives a result data association request containing spatial coordinates and defect categories, it calculates the corresponding target spatiotemporal feature composite hash key, queries the hash index table, and converts the defect category into an auxiliary retrieval vector if no match is found. It then performs a range query on the solid-state drive layer using the B+ tree index and outputs the associated image data storage address system.

[0006] Preferably, the composite hash key generation component includes a locality-sensitive hash function call subcomponent and a bit concatenation subcomponent; The locality-sensitive hash function calls the sub-component to perform multiple random projections on the optical feature vector to obtain multiple projection hash codes; The bit splicing sub-component arranges the multiple projected hash codes in sequence to form an optical hash sequence, and converts the grid code of the grid space into a binary grid sequence; The bit splicing sub-component aligns and splices the high-order bits of the binary grid sequence with the low-order bits of the optical hash sequence to generate a fixed-length bit string, which is then used as the spatiotemporal feature composite hash key.

[0007] Preferably, the memory layer index building component includes a hash bucket allocation subcomponent and a linked list mounting subcomponent; The hash bucket allocation sub-component performs a modulo operation on the spatiotemporal feature composite hash key to determine the bucket number of the hash bucket; The linked list mounting sub-component creates a linked list head node in the hash bucket, encapsulates the spatiotemporal feature composite hash key and the corresponding image data storage address into a linked list node data structure, and mounts the linked list node data structure to the linked list head node. The linked list node data structure includes a timestamp field and a version number field.

[0008] Preferably, the solid-state drive layer index building component includes a vector space partitioning sub-component and a tree node insertion sub-component; The vector space partitioning sub-component uses a hierarchical clustering algorithm to cluster all optical feature vectors stored in the solid-state drive layer to obtain multiple cluster center points. The tree node insertion sub-component uses the cluster center point as the routing key for the root node and intermediate nodes of the B+ tree index, uses the optical feature vector as the leaf node key of the B+ tree index, and writes the spatiotemporal feature composite hash key as the value of the leaf node into the continuous physical pages of the solid-state drive layer.

[0009] Preferably, the association retrieval component includes a memory matching subcomponent and a hit determination subcomponent; The memory matching subcomponent performs a modulo operation based on the target spatiotemporal feature composite hash key to locate the target hash bucket and traverses the linked list node data structure in the target hash bucket. The hit determination sub-component extracts the spatiotemporal feature composite hash key from the linked list node data structure and performs an XOR operation with the target spatiotemporal feature composite hash key. When the XOR operation result is zero and the version number field in the linked list node data structure is in a valid state, it is determined as a hit. When the XOR operation result is non-zero or no matching item is found after traversal, it is determined as a miss.

[0010] Preferably, the association retrieval component further includes a category mapping subcomponent and a range retrieval subcomponent; The category mapping subcomponent pre-stores a mapping dictionary between defect categories and standard feature vectors. It obtains the standard feature vector corresponding to the defect category by looking up the mapping dictionary and uses the standard feature vector as the auxiliary retrieval vector. The range retrieval subcomponent calculates the spatial distance between the auxiliary retrieval vector and the routing key contained in the intermediate node of the B+ tree index. When the spatial distance is less than a preset radius, it enters the subtree corresponding to the intermediate node to continue the comparison until the leaf node is traversed and the image data storage address is output.

[0011] Preferably, the locality-sensitive hash function call sub-component includes a frequency domain transformation sub-unit and a projection matrix construction sub-unit; The frequency domain conversion subunit converts the optical feature vector from the spatial domain to the frequency domain to obtain a frequency domain feature sequence. The projection matrix construction sub-unit constructs a dimension-reduced projection matrix by extracting frequency components whose variance is greater than a preset energy threshold according to the energy distribution variance of the frequency domain feature sequence. The locality-sensitive hash function call sub-component uses the dimensionality-reduced projection matrix to perform matrix multiplication on the frequency domain feature sequence to obtain the multiple projected hash codes.

[0012] Preferably, the linked list mounting sub-component includes a node conflict detection sub-unit and a tail replacement sub-unit; Before mounting, the node conflict detection subunit traverses the existing linked list nodes following the head node of the linked list and compares the spatiotemporal feature composite hash key in the existing linked list nodes with the spatiotemporal feature composite hash key to be mounted. When there are identical values, the tail replacement subunit extracts the timestamp field corresponding to the spatiotemporal feature composite hash key to be mounted. When the timestamp field is later than the timestamp field in the existing linked list node, the image data storage address and version number field in the existing linked list node are updated to the corresponding data in the linked list node data structure to be mounted.

[0013] Preferably, the range retrieval sub-component includes a Gaussian kernel mapping sub-unit and a distance calculation sub-unit; The Gaussian kernel mapping subunit performs Gaussian kernel function mapping on the auxiliary retrieval vector and the routing key respectively to obtain the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector; The distance calculation subunit calculates the Mahalanobis distance between the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector, inputs the Mahalanobis distance as the spatial distance into the range retrieval subcomponent and compares it with the preset radius, and decides whether to enter the subtree based on the comparison result.

[0014] Preferably, the range retrieval subcomponent further includes a radius dynamic adjustment subunit and a concurrent lock control subunit; The radius dynamic adjustment subunit calculates the current read / write queue depth of the solid-state drive layer, calculates the adjustment coefficient according to the read / write queue depth using a linearly decreasing function, and multiplies the preset radius by the adjustment coefficient to obtain the dynamic query radius. The concurrent lock control subunit applies a spin lock to the physical page corresponding to the subtree before entering the subtree, and releases the spin lock after traversing the leaf nodes corresponding to the subtree and outputting the image data storage address.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention maps spatial coordinates to a preset grid space, reduces the dimensionality of optical feature vectors, and concatenates them with grid codes to generate a spatiotemporal feature composite hash key. A hash index table with the composite hash key as the header is constructed in memory, and a balanced multi-way search tree index with the optical feature vector as the key and the composite hash key as the value is constructed in the solid-state drive layer. When a correlation request is received, the hash index table in memory is queried first. If no match is found, the defect category is converted into an auxiliary retrieval vector, and a range query is performed through the solid-state drive layer index. This scheme converts spatial location and optical features into fixed-length hash keys for direct memory-level mapping, avoiding the cross-table join query operation logic of relational databases, shortening the response time of cross-node correlation retrieval, and reducing the disk I / O load of distributed storage nodes.

[0016] 2. This invention improves the stability of the hash key construction process by performing random projection on optical feature vectors using a locality-sensitive hash function to obtain projected hash codes, and then aligning and splicing them with a binary grid sequence. It enhances the controllability of memory-level data updates by creating a linked list head node in the hash bucket and attaching linked list nodes containing timestamp and version number fields. It overcomes the limitations of single-feature matching by mapping defect categories to standard feature vectors as auxiliary retrieval vectors and calculating Mahalanobis distance using a Gaussian kernel function in range queries. Furthermore, it improves the robustness of the system operation in high-concurrency retrieval scenarios by dynamically adjusting the query radius based on the solid-state drive layer read / write queue depth and combining spinlocks for concurrent control of physical pages. Attached Figure Description

[0017] Figure 1 This is a flowchart of the optical feature extraction and spatiotemporal feature composite hash key generation process of the present invention; Figure 2 This is a flowchart of the locality-sensitive hash function call and bit concatenation process of the present invention; Figure 3 This is a flowchart of the memory layer hash index construction and linked list mounting process of the present invention; Figure 4 This is a flowchart of the solid-state drive layer B+ tree index construction process of the present invention; Figure 5 This is a flowchart of the memory layer hash index matching and hit determination process of the present invention; Figure 6 This is a flowchart of the solid-state drive layer range retrieval and dynamic adjustment process of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides an AOI inspection image and result data association system under industrial big data, deployed in a distributed storage and inspection cluster of a PCB production line. After the AOI inspection equipment completes the scanning of the PCB board surface, it generates a high-resolution grayscale inspection image and corresponding inspection metadata. The inspection metadata includes the physical dimension parameters of the PCB board surface, the lens distortion correction parameters of the inspection equipment, the scanning start coordinates, and the scanning resolution. The optical feature extraction component receives the high-resolution grayscale inspection image output by the AOI inspection equipment, performs optical feature extraction operations, and generates the corresponding optical feature vector of the inspection image.

[0020] Specifically, the optical feature extraction component first performs preprocessing operations on the input high-resolution grayscale detection image. These preprocessing operations include grayscale value normalization, Gaussian filtering for noise reduction, and region of interest (ROI) segmentation. The grayscale value normalization operation maps the pixel grayscale values ​​of the detection image to the [0,1] interval, following the following operational rules:

[0021] in, To detect the coordinates in the image The original grayscale value of the pixel. To detect the minimum gray value of all pixels in an image, To detect the maximum gray value of all pixels in an image, The normalized coordinates are The grayscale value of the pixel.

[0022] The Gaussian filtering denoising operation uses a 5×5 Gaussian kernel to perform convolution operations on the normalized image, filtering out Gaussian noise generated during image acquisition. The region of interest (ROI) segmentation operation, based on the PCB physical size parameters in the detection metadata, removes invalid non-board areas at the image edges, retaining the valid image area corresponding to the PCB board surface as input for subsequent feature extraction. The optical feature extraction component performs multi-dimensional optical feature extraction on the valid image area, extracting features including pixel gray-level statistical features, edge gradient features, and texture features. Pixel gray-level statistical features include the gray-level mean, gray-level variance, gray-level skewness, and gray-level kurtosis of the valid image area; edge gradient features use the Sobel operator to calculate the gradient magnitude in the horizontal and vertical directions, generating a statistical histogram feature of the gradient magnitude; texture features use the gray-level co-occurrence matrix to calculate the contrast, correlation, energy, and inverse moment features of the valid image area. The optical feature extraction component concatenates all the extracted features in a preset order to generate a one-dimensional optical feature vector, with a fixed dimension denoted as D.

[0023] Table 1. Comparison of Optical Feature Vector Dimensions and Physical Meanings of AOI Detection Images 1 1-4 Pixel grayscale statistical features, including grayscale mean, variance, skewness, and kurtosis. The real number field, normalized interval [0,1] 2 5-20 Marginal gradient statistical histogram features, amplitude percentage of 16 gradient intervals. The real number field, the interval [0,1], has a sum of 1. 3 21-24 Gray-level co-occurrence matrix texture features, including contrast, correlation, energy, and inverse difference moment. The real number field, normalized interval [0,1]

[0024] The dimensions and physical meanings of the optical feature vectors defined in Table 1 remain fixed during system operation, ensuring that the optical feature vectors generated from different detected images have the same dimensions and aligned feature semantics, providing a unified input format for subsequent dimensionality reduction processing and hash key generation.

[0025] The coordinate mapping component receives the physical coordinate system parameters of the PCB board surface and the spatial coordinates corresponding to the defect detection results from the inspection metadata, and maps the physical spatial coordinates of the PCB board surface to a preset grid space. Specifically, the preset grid space is a two-dimensional discrete grid system that corresponds one-to-one with the physical plane of the PCB board surface. The upper left corner of the physical plane of the PCB board surface is the origin, the horizontal direction to the right is the positive X-axis direction, and the vertical direction downwards is the positive Y-axis direction. The X-axis value range is [0, L], and the Y-axis value range is [0, W], where L is the physical length of the PCB board surface and W is the physical width of the PCB board surface, both in millimeters. The preset grid space divides the X-axis value range into M consecutive intervals and the Y-axis value range into N consecutive intervals, forming a two-dimensional grid matrix of M columns and N rows. Each grid cell corresponds to a fixed physical region of the PCB board surface, and each grid cell has a unique grid code. The coordinate mapping process follows the following operation rules:

[0026]

[0027] in, This refers to the X-axis coordinate value within the physical plane of the PCB board. This refers to the Y-axis coordinate value within the physical plane of the PCB board. The total number of columns in the preset grid space. The total number of rows in the preset grid space. The physical length of the PCB board surface. The physical width of the PCB board surface. For floor operations, The mapped grid column number, The grid row number is the mapped grid row number. The grid column number ranges from [0, M-1], and the grid row number ranges from [0, N-1].

[0028] The coordinate mapping component generates the corresponding grid code based on the calculated grid column number and grid row number. The grid code is an integer value obtained by concatenating the grid row number and grid column number into binary code with a preset bit width. The bit width of the grid code is a fixed value to ensure that the grid codes corresponding to all grid cells have the same binary length.

[0029] The composite hash key generation component receives the optical feature vector output from the optical feature extraction component and the grid code output from the coordinate mapping component. It performs dimensionality reduction on the optical feature vector, concatenates the dimensionality-reduced feature data with the grid code, and generates a spatiotemporal feature composite hash key. Specifically, the dimensionality reduction process uses principal component analysis (PCA) to perform a linear transformation on the D-dimensional optical feature vector, mapping the original high-dimensional feature vector to a low-dimensional feature space, resulting in a d-dimensional feature vector, where d is much smaller than D. The transformation matrix of the PCA algorithm is obtained by performing covariance matrix decomposition on the historical optical feature vector dataset of AOI detection images. The transformation matrix is ​​calculated and fixed during system initialization, and the fixed transformation matrix is ​​directly called during system operation to perform the dimensionality reduction operation. The dimensionality reduction operation follows these rules:

[0030] in, The input is an optical feature vector with a dimension of 1×D. For the pre-generated principal component analysis transformation matrix of dimension D×d, This is the dimensionality-reduced feature vector with a dimension of 1×d obtained after dimensionality reduction.

[0031] The composite hash key generation component converts the values ​​of each dimension of the dimensionality-reduced feature vector into fixed-length binary codes according to a preset quantization rule, generating an optical feature binary sequence. Simultaneously, it converts the grid code into a fixed-length binary grid sequence. The optical feature binary sequence and the binary grid sequence are then aligned and concatenated in a preset order to generate a fixed-length bit string, which serves as the spatiotemporal feature composite hash key. The length of the spatiotemporal feature composite hash key remains fixed during system operation, ensuring that all generated hash keys have the same bit length, providing a unified key-value format for subsequent hash index construction.

[0032] The memory-layer index construction component receives the spatiotemporal feature composite hash key output by the composite hash key generation component, along with the corresponding image data storage address of the AOI detection image. It then constructs a hash index table in the memory layer, with the spatiotemporal feature composite hash key as the head. Specifically, the memory layer is the random access memory space of the server node. The hash index table uses an open-chaining method to handle hash collisions. The hash index table contains multiple consecutive hash buckets, each with a unique bucket number. The storage addresses of each hash bucket are allocated contiguously in the memory space. The memory-layer index construction component performs a hash operation on the spatiotemporal feature composite hash key to obtain the corresponding hash bucket number. A linked list structure is created in the corresponding hash bucket, and the spatiotemporal feature composite hash key and the corresponding image data storage address are encapsulated as linked list nodes and attached to this linked list structure, completing the construction of the hash index table. The hash index table resides permanently in the memory layer. During system startup, the memory space is pre-allocated and initialized. During system operation, it continuously receives new hash keys and storage addresses, performing linked list node insertion and update operations.

[0033] The solid-state drive (SSD) layer index building component receives the optical feature vector output by the optical feature extraction component, the spatiotemporal feature composite hash key output by the composite hash key generation component, and the image data storage address of the corresponding AOI detection image. It then constructs a B+ tree index on the SSD layer, using the optical feature vector as the key and the spatiotemporal feature composite hash key as the value. Specifically, the SSD layer is the solid-state storage medium space of the server node. The B+ tree index is a balanced multi-way search tree structure, containing a root node, intermediate nodes, and leaf nodes. The root and intermediate nodes store routing keys for index path routing. The leaf nodes store complete key-value pairs and their corresponding data storage addresses. All leaf nodes are connected by a doubly linked list for easy range query operations. The SSD layer index building component uses the optical feature vector as the key of the B+ tree index and the combination of the spatiotemporal feature composite hash key and the corresponding image data storage address as the value. It performs a node insertion operation on the B+ tree, writing the key-value pairs into consecutive physical pages of the SSD layer, thus completing the construction of the B+ tree index. The metadata of the B+ tree index is stored at a fixed physical address on the solid-state drive layer. The metadata is loaded into the memory layer when the system starts up, and the index query and update operations are performed based on the metadata during system operation.

[0034] The association retrieval component receives result data association requests from external systems. These requests include the PCB board surface spatial coordinates and defect category corresponding to the defect to be queried. The association retrieval component calculates the corresponding target spatiotemporal feature composite hash key based on the spatial coordinates in the association request, prioritizing a query of the hash index table in the memory layer. Specifically, the association retrieval component calls the coordinate mapping component to map the spatial coordinates in the association request to a preset grid space, obtaining the corresponding grid code. Simultaneously, it calls the composite hash key generation component to perform dimensionality reduction processing based on the optical feature vector of the AOI detection image corresponding to the defect. The dimensionality-reduced feature data is then concatenated with the grid code to generate the target spatiotemporal feature composite hash key. The association retrieval component queries the hash index table in the memory layer based on the target spatiotemporal feature composite hash key. When the query is successful, it directly outputs the image data storage address stored in the corresponding linked list node in the hash index table. When the query is unsuccessful, the association retrieval component converts the defect category in the association request into an auxiliary retrieval vector, performs a range query operation through the B+ tree index in the solid-state drive layer, and outputs the associated image data storage address.

[0035] This embodiment fully realizes the associated storage and retrieval of AOI detection images and result data. By fusing spatial coordinates and optical features to generate composite hash keys, a hash index is built in the memory layer to achieve fast retrieval, and a B+ tree index is built in the solid-state drive layer to achieve fallback range retrieval. This avoids the operation logic of cross-table join queries in relational databases and shortens the response path of associated retrieval.

[0036] In a preferred embodiment, reference Figure 2 The composite hash key generation component includes a locality-sensitive hash function call subcomponent and a bit-concatenation subcomponent. The locality-sensitive hash function call subcomponent receives the optical feature vector output by the optical feature extraction component, performs multiple random projections on the optical feature vector, and obtains multiple projected hash codes. The bit-concatenation subcomponent receives the multiple projected hash codes output by the locality-sensitive hash function call subcomponent and the grid code output by the coordinate mapping component, and performs a bit-concatenation operation to generate a spatiotemporal feature composite hash key.

[0037] Specifically, the locality-sensitive hash function call sub-component includes a frequency domain transformation sub-unit and a projection matrix construction sub-unit. The frequency domain transformation sub-unit receives optical feature vectors and transforms them from the spatial domain to the frequency domain to obtain a frequency domain feature sequence. The frequency domain transformation is implemented using the Discrete Fourier Transform, and the transformation process follows these operational rules:

[0038] in, Let be the value of the nth dimension of the spatial domain optical eigenvectors, and D be the total dimension of the optical eigenvectors. The imaginary unit, It is the complex value of the k-th frequency component of the frequency domain characteristic sequence.

[0039] The frequency domain transformation subunit performs a modulo operation on the frequency domain feature sequence to obtain the amplitude of each frequency component, generating a frequency domain amplitude sequence. The projection matrix construction subunit receives the frequency domain amplitude sequence and calculates the energy distribution variance corresponding to each frequency component. The energy distribution variance is obtained by statistically calculating the frequency domain amplitude sequence dataset of historical AOI detection images. The variance of each frequency component characterizes its discriminative power across different detection images. The projection matrix construction subunit presets an energy threshold, truncates frequency components with variances greater than the preset energy threshold, retains the dimension of the corresponding frequency domain feature sequence, and constructs a dimension-reduced projection matrix. The number of rows in the dimension-reduced projection matrix is ​​consistent with the dimension D of the optical feature vector, and the number of columns is consistent with the number of truncated frequency components. In the dimension-reduced projection matrix, the diagonal position of the column corresponding to the truncated frequency component has a value of 1, and the other positions have values ​​of 0.

[0040] The locality-sensitive hash function call subcomponent uses a dimension-reduced projection matrix to perform matrix multiplication on the frequency domain feature sequence, obtaining a dimension-reduced frequency domain feature vector. Then, it performs K random projection operations on the dimension-reduced frequency domain feature vector using K preset random projection vectors. Each projection operation generates a 1-bit hash code, and the K projection operations generate a K-bit projected hash code sequence. The generation of the random projected hash code follows these operational rules:

[0041] in, The frequency domain eigenvectors after dimensionality reduction. For the k-th random projection vector, For vector dot product operation, Let K be the output value of the k-th locality-sensitive hash function, where K is the total number of random projections. The output values ​​of the K hash functions are arranged in order to form a K-bit projected hash code.

[0042] The locality-sensitive hash function call subcomponent performs multiple sets of random projection operations. Each set of random projection operations generates a K-bit projected hash code, ultimately yielding multiple projected hash codes. The bit-concatenation subcomponent arranges these multiple projected hash codes in a preset order to form an optical hash sequence. The total bit length of the optical hash sequence is the sum of the bit lengths of all sets of projected hash codes. The bit-concatenation subcomponent converts the grid code in the grid space into a fixed-length binary grid sequence. The bit length of the binary grid sequence matches the value range of the grid code, ensuring that all binary grid sequences corresponding to grid codes have the same bit length.

[0043] The bit-joining sub-component aligns and joins the high-order bits of the binary lattice sequence with the low-order bits of the optical hash sequence to generate a fixed-length bit string, which is then used as the spatiotemporal feature composite hash key. Specifically, in the alignment and joining operation, the binary lattice sequence serves as the high-order bit segment of the spatiotemporal feature composite hash key, and the optical hash sequence serves as the low-order bit segment. There are no gap bits between the high-order and low-order bit segments; they are directly joined to form a continuous fixed-length bit string. The total bit length of the fixed-length bit string is the sum of the bit lengths of the binary lattice sequence and the optical hash sequence, and this total bit length remains fixed during system operation.

[0044] Table 2 Comparison of bit concatenation rules between projected hash codes and binary grid sequences 1 Binary grid sequence 16 bits 0-15 digits Mesh encoding output by the coordinate mapping component 2 Optical hash sequence 48 bits 16-63 bits Locality-sensitive hash function calls multiple sets of projected hash codes output by the child component total Spatiotemporal feature composite hash key 64 bits 0-63 bits High-level segments and low-level segments are directly spliced ​​together

[0045] The bit splicing rules defined in Table 2 remain fixed during system operation, ensuring that the high-order bits of the generated spatiotemporal feature composite hash key carry the spatial location information of the PCB board surface, and the low-order bits carry the optical feature information of the detection image, thereby realizing the fusion encoding of spatial features and optical features. At the same time, it ensures that all hash keys have a fixed length of 64 bits, which is suitable for the operation requirements of the memory layer hash index table.

[0046] Preferably, the Hamming distance of the projected hash code generated by the Locality Sensitive Hash Function (LSH) call subcomponent is positively correlated with the Euclidean distance of the corresponding optical feature vectors. That is, the smaller the Euclidean distance between two optical feature vectors, the smaller the Hamming distance of the corresponding projected hash code. This ensures that during hash index queries, hash keys corresponding to detection images with similar optical features have a higher collision probability, improving the matching efficiency of range queries. During the bit-joining process, the bit-joining subcomponent performs cyclic redundancy check (CRC) calculations on both the binary grid sequence and the optical hash sequence, appending the calculated checksum to the end of the fixed-length bit string for data integrity verification during subsequent index queries.

[0047] This embodiment performs frequency domain transformation and random projection on optical feature vectors using a locality-sensitive hash function to generate a projected hash code with feature preservation capability. This code is then aligned and concatenated with a binary grid sequence carrying spatial location information to generate a spatiotemporal feature composite hash key. This improves the stability and feature distinguishability of the hash key construction process, ensuring that hash keys corresponding to similar optical features and detection images at the same spatial location have consistent encoding logic, thus providing a foundation for fast matching of memory-level hash indexes.

[0048] In a preferred embodiment, reference Figure 3The memory-level index construction component includes a hash bucket allocation subcomponent and a linked list mounting subcomponent. The hash bucket allocation subcomponent receives the spatiotemporal feature composite hash key output by the composite hash key generation component, performs a modulo operation on the spatiotemporal feature composite hash key, and determines the corresponding hash bucket number. The linked list mounting subcomponent receives the bucket number output by the hash bucket allocation subcomponent, along with the spatiotemporal feature composite hash key and the corresponding image data storage address, creates a linked list structure in the corresponding hash bucket, and performs the linked list node mounting operation.

[0049] Specifically, the total number of hash buckets in the memory-level hash index table is a preset fixed value, denoted as B. The total number of hash buckets B is a power of 2 to facilitate fast execution of the modulo operation. The hash bucket allocation sub-component converts the spatiotemporal feature composite hash key into the corresponding unsigned long integer value, performs a modulo operation on this value, and obtains the corresponding hash bucket number. The modulo operation follows these rules:

[0050] in, Let B be the unsigned long integer value corresponding to the spatiotemporal composite hash key, and let B be the total number of hash buckets in the hash index table. For modulo operation, The calculated hash bucket number has a value range of [0, B-1].

[0051] The hash bucket allocation subcomponent locates the storage address of the corresponding hash bucket in memory space based on the calculated bucket number and outputs this storage address to the linked list mounting subcomponent. The linked list mounting subcomponent creates a linked list head node in the corresponding hash bucket. The linked list head node stores the bucket number corresponding to that hash bucket and a pointer to the first node of the linked list. The linked list head node is created during hash bucket initialization and remains resident during system operation. The linked list mounting subcomponent encapsulates the spatiotemporal feature composite hash key, the corresponding image data storage address, a timestamp field, and a version number field into a linked list node data structure. The timestamp field stores the acquisition time of the AOI detection image corresponding to the linked list node, and the version number field stores the data version identifier of the linked list node. The initial value of the version number field is 1, and the version number value increments by 1 each time the node data is updated.

[0052] Table 3. Field Definitions and Value Rules for Linked List Node Data Structure 1 Spatiotemporal feature composite hash key Unsigned long integer 8 bytes Fixed 64-bit length The spatiotemporal composite hash key value corresponding to the node 2 Image data storage address character array 32 bytes Solid State Drive Layer Physical Address Format Storage path of the corresponding AOI detection image 3 Timestamp field Unsigned long integer 8 bytes Unix timestamp format AOI image acquisition time 4 Version number field Unsigned integer 4 bytes Positive integer, initial value 1 Version identifier of node data 5 Successor node pointer Unsigned long integer 8 bytes Memory address format Storage address of the next linked list node

[0053] The linked list node data structure defined in Table 3 remains fixed during system operation. All linked list nodes are encapsulated using a unified data structure to ensure consistent data reading format during linked list traversal. The timestamp field and version number field provide a basis for judging node data updates and conflict handling.

[0054] The linked list mounting subcomponent mounts the encapsulated linked list node data structure to the head node of the corresponding hash bucket. Specifically, when there is no existing node in the linked list corresponding to the hash bucket, the linked list mounting subcomponent sets the first node pointer of the head node to the storage address of the newly encapsulated linked list node, completing the mounting operation. When there is already a node in the linked list corresponding to the hash bucket, the linked list mounting subcomponent sets the successor node pointer of the newly encapsulated linked list node to the current head node of the head node, and then sets the first node pointer of the head node to the storage address of the newly encapsulated linked list node, using the head insertion method to complete the node mounting operation.

[0055] Preferably, the linked list mounting subcomponent includes a node conflict detection subunit and a tail replacement subunit. Before performing the linked list node mounting operation, the node conflict detection subunit traverses the existing linked list nodes following the head node in the corresponding hash bucket, extracts the spatiotemporal feature composite hash key from each existing linked list node, and compares it with the spatiotemporal feature composite hash key in the linked list node to be mounted. When the comparison result shows that the two hash keys are exactly the same, the node conflict detection subunit determines that there is a hash key conflict and triggers the tail replacement subunit to perform a node update operation; when no matching hash key is found after the traversal, the node conflict detection subunit determines that there is no hash key conflict and triggers the linked list mounting subcomponent to perform a head insertion mounting operation.

[0056] The tail replacement subunit receives the conflict trigger signal from the node conflict detection subunit, extracts the timestamp field value from the linked list node to be attached, and compares it with the timestamp field value from the conflicting existing linked list node. When the timestamp field value of the linked list node to be attached is later than the timestamp field value of the existing linked list node, the tail replacement subunit updates the image data storage address, timestamp field, and version number field of the existing linked list node to the corresponding data in the linked list node to be attached, and increments the version number field value of the existing linked list node by 1; when the timestamp field value of the linked list node to be attached is earlier than or equal to the timestamp field value of the existing linked list node, the tail replacement subunit abandons the write operation of the linked list node to be attached and retains the original data of the existing linked list node. The version number update follows the following calculation rules:

[0057] in, This is the original value of the version number field of an existing linked list node. This is the value of the version number field after the node update.

[0058] Furthermore, the linked list mounting subcomponent presets a maximum linked list length threshold. When the number of nodes in the linked list corresponding to the hash bucket reaches the preset maximum length threshold, the linked list mounting subcomponent traverses all nodes in the linked list, extracts the timestamp field value of each node, and deletes the node with the earliest timestamp, ensuring that the number of nodes in the linked list does not exceed the maximum length threshold and avoiding excessive memory space consumption. During system operation, the memory-level index building component periodically performs an expiration cleanup operation on all linked list nodes in the hash index table, deleting nodes whose timestamp field values ​​are earlier than the preset expiration time threshold, thus releasing the corresponding memory space.

[0059] This embodiment achieves fast allocation of hash buckets through modulo operations, constructs a linked list structure for the hash index using the open chaining method, adds timestamp and version number fields to the linked list nodes, and handles hash key collisions and updates node data through collision detection and tail replacement mechanisms, thereby enhancing the controllability of memory layer data updates. At the same time, the execution efficiency of node mounting is improved through head insertion, ensuring low-latency execution of write and query operations for the memory layer hash index.

[0060] In a preferred embodiment, reference Figure 4 The solid-state drive (SSD) layer index construction component includes a vector space partitioning subcomponent and a tree node insertion subcomponent. The vector space partitioning subcomponent receives all optical feature vectors stored in the SSD layer and performs clustering operations on all optical feature vectors using a hierarchical clustering algorithm to obtain multiple cluster centers. The tree node insertion subcomponent receives the cluster centers output by the vector space partitioning subcomponent, uses the cluster centers as the root node and routing keys of the intermediate nodes of the B+ tree index, uses the optical feature vectors as the leaf node keys of the B+ tree index, and uses the combination of the spatiotemporal feature composite hash key and the corresponding image data storage address as the value of the leaf node, writing it into the contiguous physical pages of the SSD layer.

[0061] Specifically, the hierarchical clustering algorithm employs a bottom-up aggregation strategy. Initially, each optical feature vector is treated as an independent cluster. The Euclidean distance between any two clusters is calculated sequentially, and the two clusters with the smallest distance are merged into a new cluster. This merging operation is repeated until the number of clusters reaches a preset threshold. The mean vector of each cluster is used as the cluster centroid. The Euclidean distance calculation between clusters follows these rules:

[0062] in, For the i-th cluster, For the j-th cluster, For clusters The mean vector, For clusters The mean vector, For L2 norm operations, This represents the Euclidean distance between two clusters.

[0063] The vector space partitioning subcomponent sorts all cluster centers in ascending order of their L2 norm, generating a routing key sequence. This sequence is used for routing selection of intermediate nodes in the B+ tree index. When constructing the B+ tree index, the tree node insertion subcomponent stores the complete routing key sequence at the root node, with each routing key corresponding to the storage address of a child node. Intermediate nodes store the routing key sequence within their respective subtree ranges. Leaf nodes store optical feature vectors, their corresponding spatiotemporal feature composite hash keys, and image data storage addresses. All leaf nodes are connected by a doubly linked list, with the node order matching the L2 norm of the leaf node keys. The tree node insertion subcomponent writes the data of each node in the B+ tree index into contiguous physical pages on the solid-state drive (SSD). Each physical page is aligned with the SSD sector size to ensure data continuity and reduce the overhead of random read / write operations.

[0064] Table 4. B+ Tree Index Node Structure and Storage Rules root node Route key array Variable length Solid State Drive Layer Fixed Physical Pages Cluster center vector sequence of top-level routes root node Array of child node pointers Variable length Solid State Drive Layer Fixed Physical Pages Each routing key corresponds to the physical address of its child node. root node Number of child nodes 4 bytes Solid State Drive Layer Fixed Physical Pages Total number of child nodes corresponding to the root node intermediate nodes Route key array Variable length Solid State Drive Layer Continuous Physical Pages Cluster center vector sequence within subtree range intermediate nodes Array of child node pointers Variable length Solid State Drive Layer Continuous Physical Pages Each routing key corresponds to the physical address of its child node. intermediate nodes Number of child nodes 4 bytes Solid State Drive Layer Continuous Physical Pages The total number of child nodes corresponding to the intermediate node intermediate nodes Parent node pointer 8 bytes Solid State Drive Layer Continuous Physical Pages Physical address of the parent node leaf node Optical feature vector array Variable length Solid State Drive Layer Continuous Physical Pages Leaf nodes store key-value sequences leaf node Composite hash key array Variable length Solid State Drive Layer Continuous Physical Pages Composite hash key sequence corresponding to optical feature vectors leaf node storage address array Variable length Solid State Drive Layer Continuous Physical Pages Storage path sequence of corresponding AOI detection images leaf node predecessor node pointer 8 bytes Solid State Drive Layer Continuous Physical Pages Physical address of the preceding leaf node leaf node Successor node pointer 8 bytes Solid State Drive Layer Continuous Physical Pages Physical address of the subsequent leaf node leaf node Number of key-value pairs 4 bytes Solid State Drive Layer Continuous Physical Pages The total number of key-value pairs stored in the leaf nodes

[0065] The B+ tree index node structure defined in Table 4 remains fixed during system operation. The metadata of the root node is loaded into the memory layer when the system starts, and the data of the intermediate nodes and leaf nodes are loaded into the cache space of the memory layer as needed during the query process, ensuring that the read and write operations of the solid-state drive are minimized during the index query process.

[0066] refer to Figure 5 The association retrieval component includes a memory matching subcomponent and a hit determination subcomponent. The memory matching subcomponent receives the target spatiotemporal feature composite hash key corresponding to the association request, performs a modulo operation on the target spatiotemporal feature composite hash key to locate the corresponding target hash bucket, and traverses all linked list nodes in the target hash bucket. Specifically, the memory matching subcomponent uses the same modulo operation rule as the hash bucket allocation subcomponent to calculate the target hash bucket number corresponding to the target spatiotemporal feature composite hash key, locates the target hash bucket in memory based on the bucket number, and traverses each linked list node sequentially starting from the head node, extracting the spatiotemporal feature composite hash key and version number field from the node, and outputting them to the hit determination subcomponent.

[0067] The hit determination subcomponent receives the spatiotemporal feature composite hash key from the linked list node output by the memory matching subcomponent, performs an XOR operation with the target spatiotemporal feature composite hash key, and determines whether the query has a hit based on the operation result and the status of the version number field. The hit determination follows the following operation rules:

[0068] in, This refers to the unsigned long integer value corresponding to the spatiotemporal composite hash key in the linked list node. The unsigned long integer value corresponding to the composite hash key of the target spatiotemporal features. For bitwise XOR operation, This refers to the value of the version number field in the linked list node. For a preset set of valid version numbers, For logical AND operation, As a hit marker, when If the result is true, it is considered a hit; if the result is false, it is considered a miss.

[0069] Specifically, when the result of the XOR operation is zero, it indicates that all bits of the two hash keys are exactly the same, and the version number field of the linked list node is in a valid state. The hit determination subcomponent determines that the query has been hit and extracts the image data storage address in the linked list node as the retrieval result output. When the result of the XOR operation is non-zero, or when no matching item is found after traversing all linked list nodes in the target hash bucket, the hit determination subcomponent determines that the query has not been hit and triggers the associated retrieval component to perform a range query operation on the solid-state drive layer.

[0070] Further, refer to Figure 6 The association retrieval component also includes a category mapping subcomponent and a range retrieval subcomponent. The category mapping subcomponent pre-stores a mapping dictionary between defect categories and standard feature vectors. This mapping dictionary is constructed during system initialization. By performing feature extraction and cluster analysis on a dataset of defect samples from historical AOI detection images, a corresponding standard feature vector is generated for each defect category. The dimension of the standard feature vector is completely consistent with the dimension of the optical feature vector output by the optical feature extraction component. The category mapping subcomponent receives the defect category from the association request, retrieves the corresponding standard feature vector by looking up the mapping dictionary, and outputs this standard feature vector as an auxiliary retrieval vector to the range retrieval subcomponent.

[0071] The range retrieval subcomponent receives the auxiliary retrieval vector output by the category mapping subcomponent, calculates the spatial distance between the auxiliary retrieval vector and the routing key contained in the intermediate node in the B+ tree index, and when the spatial distance is less than the preset radius, it enters the subtree corresponding to the intermediate node to continue the comparison until the leaf node is reached, and outputs the image data storage address stored in the leaf node.

[0072] Preferably, the range retrieval subcomponent includes a Gaussian kernel mapping subunit and a distance calculation subunit. The Gaussian kernel mapping subunit receives the auxiliary retrieval vector and the routing key of the intermediate node in the B+ tree, and performs Gaussian kernel function mapping on the auxiliary retrieval vector and the routing key respectively, mapping the vector in the original feature space to a high-dimensional feature space, obtaining the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector. The mapping process of the Gaussian kernel function follows the following operational rules:

[0073] in, The input vector is in the original feature space. Let be the center vector in the original feature space. For L2 norm operations, The bandwidth parameter of the Gaussian kernel. The output value of the Gaussian kernel function is the high-dimensional feature space coordinates of vector u under the Gaussian kernel mapping centered at v.

[0074] The Gaussian kernel mapping subunit pre-defines multiple Gaussian kernel center vectors. These vectors are obtained through cluster analysis of a historical optical feature vector dataset. Each Gaussian kernel center vector corresponds to one dimension of the high-dimensional feature space. After the auxiliary retrieval vector and routing key are mapped using the Gaussian kernel, the resulting high-dimensional mapping vector has the same dimension as the number of Gaussian kernel center vectors. The distance calculation subunit receives the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector output by the Gaussian kernel mapping subunit, calculates the Mahalanobis distance between the two high-dimensional mapping vectors, and inputs the calculated Mahalanobis distance as the spatial distance into the range retrieval subunit. It compares this distance with a preset radius and determines whether to enter the corresponding subtree based on the comparison result. The Mahalanobis distance calculation follows these rules:

[0075] in, To assist in high-dimensional mapping vectors, For routing high-dimensional mapping vectors, Let be the covariance matrix of all mapping vectors in the high-dimensional feature space. It is the inverse of the covariance matrix. This is the transpose of the vector difference. Let be the Mahalanobis distance between two high-dimensional mapping vectors.

[0076] The covariance matrix S is obtained during system initialization by performing statistical calculations on a high-dimensional mapping vector dataset of historical optical feature vectors. This matrix remains fixed during system operation. The distance calculation subunit directly calls the inverse of the pre-generated covariance matrix to calculate the Mahalanobis distance, avoiding the performance overhead of real-time matrix inversion. The Mahalanobis distance calculation considers the correlation between different dimensions in the high-dimensional feature space, eliminating the influence of dimensionality and improving the accuracy of similarity calculation between feature vectors.

[0077] Furthermore, the range retrieval subcomponent also includes a radius dynamic adjustment subunit and a concurrency lock control subunit. The radius dynamic adjustment subunit continuously monitors the read / write queue depth of the current solid-state drive layer, calculates an adjustment coefficient based on the read / write queue depth using a linearly decreasing function, and multiplies the preset radius by the adjustment coefficient to obtain the dynamic query radius. The calculation of the adjustment coefficient follows the following linear decreasing rule:

[0078] in, This represents the current read / write queue depth of the solid-state drive layer. The maximum read / write queue depth preset for the solid-state drive layer. This is the preset minimum adjustment coefficient. The calculated adjustment factor has a range of values. .

[0079] The calculation of the dynamic query radius follows these rules:

[0080] in, This is the preset initial query radius. To adjust the coefficient, This is the calculated dynamic query radius.

[0081] The radius dynamic adjustment subunit outputs the calculated dynamic query radius to the range retrieval subcomponent. The range retrieval subcomponent compares the spatial distance with the dynamic query radius. When the read / write queue depth of the solid-state drive layer is large, the adjustment coefficient decreases, the dynamic query radius shrinks, reducing the number of subtrees that need to be traversed during the range query and lowering the read / write load of the solid-state drive. When the read / write queue depth of the solid-state drive layer is small, the adjustment coefficient increases, the dynamic query radius expands, and the matching recall rate of the range query is improved.

[0082] Before the range retrieval subcomponent enters the subtree corresponding to an intermediate node, the concurrency lock control subunit applies a spinlock to the physical page corresponding to that subtree. After traversing all leaf nodes of the subtree and outputting the image data storage address, the spinlock is released. Specifically, the spinlock is a read-write mutex lock for the physical page of the solid-state drive layer. When a physical page is subject to a spinlock, other retrieval requests cannot perform write operations on that physical page, and can only perform read operations, avoiding inconsistencies in query results caused by modifications to the physical page data during the range query. The concurrency lock control subunit presets a maximum waiting time for the spinlock. When the spinlock waiting time exceeds the preset maximum waiting time, the traversal operation of that subtree is abandoned, and other subtrees that meet the conditions are traversed, avoiding long-term blocking of retrieval requests. The spinlock timeout determination follows the following rules:

[0083] in, This represents the current waiting time of the spinlock. The preset maximum waiting time, This is a timeout indicator, when If the result is true, it indicates a timeout and the traversal operation on that subtree is abandoned.

[0084] After completing the range query at the solid-state drive layer, the association retrieval component sorts all the obtained image data storage addresses according to their corresponding spatial distances from smallest to largest, and outputs a sorted list of storage addresses for external systems to read the corresponding AOI detection image data. The association retrieval component encapsulates the spatiotemporal feature composite hash keys and corresponding image data storage addresses obtained during the range query into new linked list nodes, inserts them into the hash bucket corresponding to the memory-level hash index table, providing memory-level fast retrieval capabilities for subsequent identical query requests and improving the response efficiency of repeated queries.

[0085] This embodiment uses a hierarchical clustering algorithm to spatially partition optical feature vectors, constructs a B+ tree index based on optical feature vectors, achieves fast hit determination of the memory layer hash index through XOR operation, maps defect categories to standard feature vectors as auxiliary retrieval vectors, achieves high-dimensional feature similarity matching through Gaussian kernel mapping and Mahalanobis distance calculation, dynamically adjusts the query radius according to the read / write load of the solid-state drive, and combines spinlocks to achieve concurrent access control. This overcomes the limitations of single feature matching, improves the robustness of system operation in high-concurrency retrieval scenarios, realizes fallback retrieval when the memory layer is not hit, and ensures full-scenario coverage of AOI detection image and result data association retrieval.

Claims

1. A system for associating AOI inspection images and results data under industrial big data, characterized in that, It includes optical feature extraction components, coordinate mapping components, composite hash key generation components, memory layer index building components, solid-state drive layer index building components, and association retrieval components; The optical feature extraction component acquires the optical feature vector of the AOI detection image; The coordinate mapping component maps the spatial coordinates of the printed circuit board to a preset grid space; The composite hash key generation component performs dimensionality reduction on the optical feature vector and concatenates it with the grid code of the grid space to generate a spatiotemporal feature composite hash key. The memory layer index building component builds a hash index table in the memory layer with the spatiotemporal feature composite hash key as the header; The solid-state drive layer index building component constructs a B+ tree index in the solid-state drive layer with the optical feature vector as the key and the spatiotemporal feature composite hash key as the value. When the associated retrieval component receives a result data association request containing spatial coordinates and defect categories, it calculates the corresponding target spatiotemporal feature composite hash key, queries the hash index table, and converts the defect category into an auxiliary retrieval vector if no match is found. It then performs a range query on the solid-state drive layer using the B+ tree index and outputs the associated image data storage address system.

2. The AOI detection image and result data association system under industrial big data as described in claim 1, characterized in that, The composite hash key generation component includes a locality-sensitive hash function call subcomponent and a bit concatenation subcomponent; The locality-sensitive hash function calls the sub-component to perform multiple random projections on the optical feature vector to obtain multiple projection hash codes; The bit splicing sub-component arranges the multiple projected hash codes in sequence to form an optical hash sequence, and converts the grid code of the grid space into a binary grid sequence; The bit splicing sub-component aligns and splices the high-order bits of the binary grid sequence with the low-order bits of the optical hash sequence to generate a fixed-length bit string, which is then used as the spatiotemporal feature composite hash key.

3. The system for associating AOI detection images and results data under industrial big data as described in claim 1, characterized in that, The memory layer index building component includes a hash bucket allocation subcomponent and a linked list mounting subcomponent; The hash bucket allocation sub-component performs a modulo operation on the spatiotemporal feature composite hash key to determine the bucket number of the hash bucket; The linked list mounting sub-component creates a linked list head node in the hash bucket, encapsulates the spatiotemporal feature composite hash key and the corresponding image data storage address into a linked list node data structure, and mounts the linked list node data structure to the linked list head node. The linked list node data structure includes a timestamp field and a version number field.

4. The system for associating AOI detection images and results data under industrial big data as described in claim 1, characterized in that, The solid-state drive layer index construction component includes a vector space partitioning sub-component and a tree node insertion sub-component; The vector space partitioning sub-component uses a hierarchical clustering algorithm to cluster all optical feature vectors stored in the solid-state drive layer to obtain multiple cluster center points. The tree node insertion sub-component uses the cluster center point as the routing key for the root node and intermediate nodes of the B+ tree index, uses the optical feature vector as the leaf node key of the B+ tree index, and writes the spatiotemporal feature composite hash key as the value of the leaf node into the continuous physical pages of the solid-state drive layer.

5. The system for associating AOI detection images and results data under industrial big data as described in claim 1, characterized in that, The association retrieval component includes a memory matching subcomponent and a hit determination subcomponent; The memory matching subcomponent performs a modulo operation based on the target spatiotemporal feature composite hash key to locate the target hash bucket and traverses the linked list node data structure in the target hash bucket. The hit determination sub-component extracts the spatiotemporal feature composite hash key from the linked list node data structure and performs an XOR operation with the target spatiotemporal feature composite hash key. When the XOR operation result is zero and the version number field in the linked list node data structure is in a valid state, it is determined as a hit. When the XOR operation result is non-zero or no matching item is found after traversal, it is determined as a miss.

6. The system for associating AOI detection images and results data under industrial big data as described in claim 1, characterized in that, The associated retrieval component also includes a category mapping sub-component and a range retrieval sub-component; The category mapping subcomponent pre-stores a mapping dictionary between defect categories and standard feature vectors. It obtains the standard feature vector corresponding to the defect category by looking up the mapping dictionary and uses the standard feature vector as the auxiliary retrieval vector. The range retrieval subcomponent calculates the spatial distance between the auxiliary retrieval vector and the routing key contained in the intermediate node of the B+ tree index. When the spatial distance is less than a preset radius, it enters the subtree corresponding to the intermediate node to continue the comparison until the leaf node is traversed and the image data storage address is output.

7. The AOI detection image and result data association system under industrial big data as described in claim 2, characterized in that, The local sensitive hash function call sub-component includes a frequency domain transformation sub-unit and a projection matrix construction sub-unit; The frequency domain conversion subunit converts the optical feature vector from the spatial domain to the frequency domain to obtain a frequency domain feature sequence. The projection matrix construction sub-unit constructs a dimension-reduced projection matrix by extracting frequency components whose variance is greater than a preset energy threshold according to the energy distribution variance of the frequency domain feature sequence. The locality-sensitive hash function call sub-component uses the dimensionality-reduced projection matrix to perform matrix multiplication on the frequency domain feature sequence to obtain the multiple projected hash codes.

8. The AOI detection image and result data association system under industrial big data as described in claim 3, characterized in that, The linked list mounting sub-component includes a node conflict detection sub-unit and a tail replacement sub-unit; Before mounting, the node conflict detection subunit traverses the existing linked list nodes following the head node of the linked list and compares the spatiotemporal feature composite hash key in the existing linked list nodes with the spatiotemporal feature composite hash key to be mounted. When there are identical values, the tail replacement subunit extracts the timestamp field corresponding to the spatiotemporal feature composite hash key to be mounted. When the timestamp field is later than the timestamp field in the existing linked list node, the image data storage address and version number field in the existing linked list node are updated to the corresponding data in the linked list node data structure to be mounted.

9. The industrial big data-based AOI detection image and result data association system according to claim 6, characterized in that, The range retrieval subcomponent includes a Gaussian kernel mapping subunit and a distance calculation subunit; The Gaussian kernel mapping subunit performs Gaussian kernel function mapping on the auxiliary retrieval vector and the routing key respectively to obtain the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector; The distance calculation subunit calculates the Mahalanobis distance between the auxiliary high-dimensional mapping vector and the routing high-dimensional mapping vector, inputs the Mahalanobis distance as the spatial distance into the range retrieval subcomponent and compares it with the preset radius, and decides whether to enter the subtree based on the comparison result.

10. The industrial big data-based AOI detection image and result data association system according to claim 9, characterized in that, The range retrieval subcomponent also includes a radius dynamic adjustment subunit and a concurrent lock control subunit; The radius dynamic adjustment subunit calculates the current read / write queue depth of the solid-state drive layer, calculates the adjustment coefficient according to the read / write queue depth using a linearly decreasing function, and multiplies the preset radius by the adjustment coefficient to obtain the dynamic query radius. The concurrent lock control subunit applies a spin lock to the physical page corresponding to the subtree before entering the subtree, and releases the spin lock after traversing the leaf nodes corresponding to the subtree and outputting the image data storage address.

Images