Deep learning-based shaft sleeve quality traceability method and system

Abstract

This invention discloses a deep learning-based method and system for tracing the quality of automotive axle bushings. The method includes: acquiring inspection images, production line data, and warehouse data of automotive axle bushings; constructing a hypergraph network based on the production line data; generating a Local Sensitive Hash Index (LSHI) based on the hypergraph network using hyperedge features and corresponding node environmental features; using warehouse partitions as query anchors; performing fast nearest neighbor matching based on the query anchors to obtain a set of candidate vulnerable nodes; generating vulnerable process constraints from the candidate vulnerable node set through correlation analysis; performing deep learning on the inspection images of rust spots on the axle bushings based on the vulnerable process constraints to obtain traceability hyperpaths from vulnerable nodes to rust defects; and searching for root cause nodes through the hypergraph network based on each traceability hyperpath to obtain the quality traceability results. This method utilizes a hypergraph network combined with an LHI for nearest neighbor matching, improving the real-time performance and accuracy of quality traceability under industrial data and exhibiting good interpretability.

Description

Technical Field

[0001] This invention relates to the field of quality traceability, and in particular to a method and system for quality traceability of bushings based on deep learning. Background Technology

[0002] As core safety components for automotive wheel hub bearings, drive motor bearings, and gearbox bearings, bearing bushings affect the overall vehicle's driving safety, NVH performance, and service life. The surface quality, dimensional accuracy, and service stability of bushings are the main focus of quality monitoring. Surface corrosion is the most easily detected quality defect in the entire production, warehousing, and logistics chain of automotive bushings. The quality traceability system in the automotive bushing industry is based on batch-level material flow traceability using barcodes, QR codes, and RFID to enable information queries on product production batches, operators, and equipment numbers. However, the root cause location of corrosion defects relies on manual inspection, SPC statistical process control, and orthogonal experiments.

[0003] However, the production and use of bushings involve multiple variables such as raw materials, process parameters, environmental factors, and logistics conditions throughout the entire warehousing process, resulting in low root cause localization accuracy, long traceability cycles, and low adaptability to batch risk investigation. Existing quality traceability processes mostly focus on visual inspection of bushing surface defects, and the few root cause localization solutions mostly use sample analysis. When the labeled sample data is insufficient, spurious correlation interference is serious. Therefore, it is crucial to shorten the quality traceability cycle and improve the accuracy of root cause localization. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for tracing the quality of bushings based on deep learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: The first aspect of this invention provides a deep learning-based method for tracing the quality of bushings, comprising: Acquire inspection images, production line data, and warehouse data of automotive axle sleeves, and use the raw material batches, production process units, and warehouse partitions of the production line data as super nodes to construct a hypergraph network based on the super nodes; Based on the hypergraph network, extract the hyperedge features of material flow path and process coupling relationship, and generate local sensitive hash index based on the hyperedge features and the environmental features of the corresponding nodes; Using the storage partition as the query anchor point, a set of candidate vulnerable nodes is obtained by fast nearest neighbor matching based on the query anchor point through local sensitive hash index. The candidate vulnerable node set is used to generate vulnerable process constraints through correlation analysis. Based on the vulnerable process constraints, deep learning is performed on the detected image of bushing rust to obtain the source superpath from vulnerable node to rust defect. Based on each traceability superpath, the root cause node is searched using the hypergraph network, and confidence is ranked using DS evidence. The quality traceability results are obtained based on the production process unit and detection image samples of the ranked nodes.

[0006] Furthermore, the method for obtaining the supernode includes: Based on production line data and warehouse data, entities are extracted according to timestamps. Material supernodes, process supernodes, and warehouse supernodes are obtained by identifying raw material batches, production process units, and warehouse partitions according to entity identifiers. Production line data, warehouse data, and inspection images with corresponding timestamps are associated as attribute sets for each supernode. The material supernode is uniquely identified by the batch number, the process supernode is jointly identified by the process unit number and the process time period, and the warehouse supernode is spatially identified by the warehouse partition coordinates. The production line data includes the raw material batches, process units, processing environment data, processing sequence, heat treatment temperature curves, timestamps, processing equipment numbers, and storage zones for automotive axle bushings. The inspection images include surface inspection images, internal defect inspection images, and 3D point cloud data of the automotive axle bushings. The storage data includes the warehousing time of the automotive axle bushings, storage zone coordinates, stacking layers, storage period, storage environment data, and cross-regional transfer records.

[0007] Furthermore, the method for obtaining the hypergraph network includes: The material flow superedge is constructed based on the flow direction of raw material batches between material supernodes and process supernodes. The weight of the material flow superedge is positively correlated with the proportion of raw material batches consumed in each process. Based on the process supernode, process flow superedges are constructed through the process connection relationship of production process units. The process connection relationship includes cross-process unit flow and sequential flow of adjacent process units. The weight of the process flow superedge is positively correlated with the frequency of material flow between process units and negatively correlated with the processing time delay. Based on the process supernode and the storage supernode, a storage-related superedge is constructed through cross-regional transfer records. The weight of the storage-related superedge is positively correlated with the cross-regional transfer frequency and storage duration. The weights of the material flow hyperedges, process flow hyperedges, and storage association hyperedges are normalized. A hypergraph adjacency tensor is generated based on the hypernodes and normalized hyperedges. The product of the normalized hyperedge weights and the feature similarity between the hypernodes is used as the element value of the hypergraph adjacency tensor to obtain the hypergraph network.

[0008] Furthermore, the method for obtaining the hyperedge features includes: Based on the hypergraph network, the material flow hyperedge is used to extract the flow sequence of raw material batches through process units, the consumption ratio between each process, and the flow time span to obtain the hyperedge features of the material flow path. Based on the process flow hyperedge, the similarity of processing parameters of adjacent process units, the quality entropy of cross process units, and the processing time delay are extracted to obtain the hyperedge features of process coupling relationship. The quality entropy is obtained based on the information entropy of the process defect rate.

[0009] Furthermore, the method for obtaining the locality-sensitive hash index includes: Based on the hyperedge features, the processing environment data and warehousing environment data of the corresponding nodes are used as environmental features. Convolutional embedding is performed using the hyperedge feature vector, environmental feature vector, and hypergraph network through the hyperedge Laplacian algorithm to obtain the associated feature vector of the hypernode. Based on the detected images of the hypernodes, positive and negative sample pairs are divided using rust defect labels, and a rust similarity matrix is ​​calculated. The formula for calculating the rust similarity matrix is ​​as follows: ; in Let be the rust similarity matrix, and let be a symmetric positive definite matrix. Positive sample pairs gather, negative sample pairs gather, For sample pairs labeled with different rust defects, For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For transpose; The rust similarity matrix is ​​subjected to Cholliski decomposition to obtain the projection matrix. A locality-sensitive hash function is then generated based on the projection matrix. The locality-sensitive hash function is as follows: ; in For the first Locality-sensitive hash function, For associated feature vectors, Let be the projection matrix. For matrix multiplication, For inner product operations, For the first A random projection vector is obtained by independent sampling according to a standard Gaussian distribution. For the first A random offset, based on a uniform distribution Obtained through random sampling. This refers to the granularity of the hash bucket. This is the floor function; The output of the locality-sensitive hash function is combined into a hash signature to obtain a hash table. Based on the hash table, each supernode identifier and the associated feature vector are stored in the corresponding hash bucket to obtain the locality-sensitive hash index of the supernode identifier and the hash bucket position.

[0010] Furthermore, the method for obtaining the source tracing superpath includes: Correlation analysis is performed on the candidate vulnerable node set to obtain the correlation degree between each node attribute and the rust defect label. Vulnerability process constraints are generated based on these correlation degrees, and a spatial heat map of process defects is generated based on these constraints. The vulnerability process constraints include critical values ​​and actual detected values ​​of process parameters. The calculation formula for the spatial heat map of process defects is as follows: ; in This is a space thermal map of process defects. , For spatial coordinate indexes on the feature map, For rows in the height direction, , For columns in the width direction, , The height of the convolutional feature map. The width of the convolutional feature map is given by the image being detected. This represents the total number of vulnerable process constraints. This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. For the first The actual measured values ​​of each process parameter For the first Critical values ​​of each process parameter For the first Coordinates of high-incidence areas of rust defects corresponding to each vulnerable process constraint. For the first The spatial influence range of the Gaussian kernel corresponding to each process constraint in the horizontal and vertical directions is preset according to the process unit through the camera field of view and the workpiece size. The spatial heatmap of the process defect and the convolutional feature map of the detection image are weighted and fused to obtain the fused feature. The calculation formula of the fused feature is as follows: ; in This is the fused feature map after process attention weighting. , For convolutional feature maps, For the number of channels, Indicates that F and spliced ​​as The channel tensor, For activation function, This is element-wise multiplication; The fused feature map is subjected to global average pooling to obtain the defect region feature vector. The defect region feature vector is concatenated with the associated feature vector of the candidate vulnerable node to construct a node feature matrix. The node feature matrix is ​​input into the hypergraph convolutional neural network. The rust defect label of the detected image is used as a supervision signal to perform deep learning on the hypergraph convolutional neural network. The optimization objective is to minimize the weighted sum of the defect prediction loss and the vulnerability constraint regularization term. The connection probability from each candidate vulnerable node to the defect node is output. The connection probability is used as the hyperedge propagation weight. Based on the propagation weight, the source hyperpath from the defect node to the candidate vulnerable node is extracted inversely through beam search.

[0011] Furthermore, the method for obtaining the quality traceability results includes: Based on each traceability superpath, reverse tracing is performed through a hypergraph network. The set of supernodes on each traceability superpath is extracted. Based on the set of supernodes, image evidence is obtained by detecting defect labels in the defect region and node association region of the image. Process evidence is obtained based on the difference between the process parameters corresponding to the set of supernodes and the standard range. Structural evidence is obtained based on the centrality index of the set of supernodes in the hypergraph network. The image evidence, process evidence and structural evidence are synthesized using Dempster-Shafer evidence theory. The basic probability distribution of the synthesis of each node in the set of supernodes is calculated. The centrality index includes degree centrality or betweenness centrality. The confidence function and similarity function of each node are calculated based on the basic probability allocation. The nodes are sorted in descending order according to the confidence function. Target root cause nodes are screened based on a preset confidence threshold. The production process unit, raw material batch and detection image sample of the target root cause node are used as quality sources to obtain quality traceability results. The quality traceability results include root cause node identification, confidence score and defect propagation hyperedge path.

[0012] A second aspect of the present invention provides a deep learning-based bushing quality traceability system, comprising: Data acquisition module: used to acquire inspection images, production line data and warehouse data of automobile axle sleeves, and to construct a hypergraph network based on the raw material batches, production process units and warehouse partitions of the production line data as super nodes; Index generation module: used to extract hyperedge features of material flow paths and process coupling relationships based on the hypergraph network, and generate local sensitive hash indexes based on the hyperedge features and the environmental features of the corresponding nodes; Candidate vulnerable node screening module: Used to use the storage partition as the query anchor point, and perform fast nearest neighbor matching based on the query anchor point through local sensitive hash index to obtain a set of candidate vulnerable nodes; Source tracing path generation module: used to generate vulnerable process constraints from the candidate vulnerable node set through correlation analysis, and perform deep learning on the detected image of bushing rust based on the vulnerable process constraints to obtain the source tracing hyperpath from vulnerable node to rust defect; The traceability module is used to search for root cause nodes based on each traceability superpath using the hypergraph network, and to rank them by confidence using DS evidence. Based on the production process unit and detection image samples of the ranked nodes, the module obtains the quality traceability results.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention constructs a hypergraph network containing three types of supernodes: materials, processes, and storage. Based on the relationships between material flow, process coupling, and storage, it extracts hyperedge features, thereby improving the coverage of quality traceability. Through correlation analysis, it generates vulnerable process constraints, constructs a spatial heatmap of process defects, and fuses it with the convolutional features of the detection images. It integrates industrial process priors into the deep learning model, guiding the model to focus on defect-sensitive areas, thus improving the accuracy and interpretability of the traceability path. By fusing image evidence, process evidence, and structural evidence through DS evidence theory, it calculates and ranks the confidence of candidate root cause nodes, improving the multi-source nature of root cause judgment, and outputting structured traceability results for production line quality. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of the deep learning-based bushing quality traceability method in this embodiment of the invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, this invention provides a deep learning-based method for tracing the quality of bushings, including: S1 acquires the inspection images, production line data, and warehouse data of the automotive axle sleeve, and uses the raw material batches, production process units, and warehouse partitions of the production line data as super nodes to construct a hypergraph network based on the super nodes; In the actual assessment, inspection images, production line data, and warehousing data of the crankshaft bushings of the 4G63S4T engine were obtained. Quality traceability was performed on 12 bushings with rust defects detected in storage area C05. The bushing production process included raw material preparation, forging, heat treatment, machining, surface protection, semi-finished product storage, and assembly. Production line data was managed by the MES system, and warehousing data was managed by the WMS system. 500 bushing blanks of raw material batch BL20250301, 40CrMo alloy steel, produced from March 1st to March 5th, were selected. This involved 6 production process units: G01 preparation, G02 forging, G03 heat treatment, G04 machining, G05 surface protection, and G06 semi-finished product transfer. There were 3 storage areas: C03 raw material area, C04 semi-finished product transition area, and C05 semi-finished product storage area. Twelve bushings with rust defects were subsequently detected in zone C05, representing a defect rate of 2.4%. Three types of inspection images of the bushings were collected using a vision inspection system, with a uniform resolution of 1024×1024. Rust defect areas were collected and labeled using an industrial area array camera, including rust on the outer and end faces of the bushings in zone C05, obtaining surface inspection images. Internal defect inspection images were also obtained using industrial CT scanning; no internal defects were found in this batch. Three-dimensional point cloud data was collected using a structured light 3D scanner to assist in locating the spatial position of the rust. Production line data was extracted using the MES system, with key fields including: process unit number G01, process name "material cutting," raw material batch BL20250301, processing environment data (temperature / humidity) of 25℃ / 45%RH, processing sequence 1, and heat treatment temperature curve 2025-03-01. 08:00, timestamp T01, processing equipment number C05, target storage zone -", based on the WMS system, the storage data is extracted, the core fields are "storage zone coordinates C03 (X10, Y20), inbound time 2025-03-01 07:30, stacking layer number 3, storage period 15h, storage environment data (temperature / humidity / corrosive gas) 25℃ / 45%RH / 0.001%SO2, cross-regional transfer record - none"; In the actual evaluation, entity extraction was performed based on the timestamps of production line data and warehouse data. Three types of supernodes were categorized by entity type, and unique identifiers and attribute associations were established. The material supernode, denoted as N_M1, uses the raw material batch number BL20250301 as its unique identifier. Its associated attributes include production line data related to the raw material batch in the production line data, C03 partition data related to the warehouse data, and batch inspection images. The process supernodes use a combined identifier of "process unit number and process time period," totaling six: N_P1(G01, 08:00-09:30), N_P2(G02, 09:30-11:00), N_P3(G03, 1... The three types of hypernodes—N_P4(G04, 14:00-16:00), N_P5(G05, 16:00-17:00), and N_P6(G06, 17:00-17:10)—are associated with production line data and inspection images for their respective processes. Three warehouse hypernodes, N_W1(C03, X10, Y20), N_W2(C04, X15, Y20), and N_W3(C05, X20, Y20), are spatially identified by warehouse partition coordinates. Each hypernode is associated with warehouse data and inspection images for its corresponding partition. Based on the business associations of these three types of hypernodes, three types of hyperedges are constructed, and a hypergraph is generated. The adjacency tensor constructs material flow hyperedges E_M1 (N_M1→N_P1), E_M2 (N_M1→N_P2), E_M3 (N_M1→N_P3), E_M4 (N_M1→N_P4), E_M5 (N_M1→N_P5), and E_M6 (N_M1→N_P6) based on the flow direction of raw material batches. The hyperedge weights are positively correlated with the consumption ratio of each process. In the current batch of 500 blanks, no effective loss was recorded in any process, and the consumption ratio is 1 for each process. After normalization, the weights of all such hyperedges are 1. Process flow hyperedges are also constructed based on process connection relationships. The sequential flow hyperedge of adjacent processes is E_P1 (N_P1→N_P2). E_P1, E_P2 (N_P2→N_P3), E_P3 (N_P3→N_P4), E_P4 (N_P4→N_P5), E_P5 (N_P5→N_P6); Since there is no cross-process rework in the current batch, the cross-process unit flow excess edge is zero. Its excess edge weight is positively correlated with the material flow frequency and negatively correlated with the processing time delay. The flow frequency of each process is 500 times, 1 time / piece, and the processing time delays are 0h (E_P1), 0h (E_P2), 0h (E_P3), 0h (E_P4), and 0.17h (E_P5), respectively. After normalization calculation, the weight of E_P1-E_P4 is 1, and the weight of E_P5 is 0.85; Based on cross-regional transfer records, warehouse-related hyperedges are constructed as E_W1 (N_P1→N_W1), E_W2 (N_P2→N_W2), E_W3 (N_P5→N_W3), and E_W4 (N_P6→N_W3). The weights of these hyperedges are positively correlated with the frequency of cross-regional transfers and the storage duration. Specifically, E_W1 has a transfer frequency of 0 and a storage duration of 0.5 hours, with a weight of 0.1; E_W2 has a transfer frequency of 1 and a storage duration of 5.2 hours, with a weight of 0.6; and E_W3 has a transfer frequency of 1 and a storage duration of 4.8 days, with a weight of 0.95. _W4 has a transfer frequency of 1, a storage duration of 4.8 days, and a weight of 0.95. After normalizing the weights of the three types of hyperedges, the hypergraph adjacency tensor element values ​​are calculated based on the hypernode feature similarity. The tensor element value is the product of the normalized hyperedge weight and the feature similarity. The cosine similarity between material and process hypernodes is 0.9, and the cosine similarity between process and warehousing hypernodes is 0.7. A 9×9 hypergraph adjacency tensor is generated. The element values ​​of N_M1 and N_P1 are 0.9, and the element values ​​of N_P5 and N_W3 are 0.665, thus obtaining the hypergraph network. S2 extracts the hyperedge features of material flow path and process coupling relationship based on the hypergraph network, and generates a local sensitive hash index based on the hyperedge features and the environmental features of the corresponding nodes; In the actual evaluation, the material flow sequence E_M1-E_M6 was extracted, from 1 to 6, with each sequence having a consumption ratio of 1. The flow time spans were E_M1: 1.5h, E_M2: 2.5h, E_M3: 5h, E_M4: 8h, E_M5: 10h, and E_M6: 10.17h, resulting in a 6-dimensional material flow feature vector. The processing parameter similarity of E_P1-E_P5 was extracted, with temperature and humidity similarities of 0.92, 0.88, 0.95, and 0.98 for adjacent processes, based on the process defect rate. The quality entropy is calculated, with a defect rate of 0 for each process and a quality entropy of 0. Processing time delays are 0, 0, 0, 0.17 hours, forming a 5-dimensional process coupling feature vector. Hyperedge feature vectors and environmental feature vectors are used as inputs, where the environmental feature vectors represent the temperature / humidity of the processing environment and the temperature / humidity / corrosive gas concentration of the storage environment. Convolutional embedding is performed using the hyperedge Laplacian algorithm, with two hypergraph convolutional layers and ReLU activation function, generating 256-dimensional associated feature vectors X1-X9 for 9 supernodes, corresponding to N_M1. Based on the rust defect labels in the detected images, positive sample pairs P{(N_W3,N_P5), (N_W3,N_P6), (N_P5,N_P6)} and negative sample pairs N{(N_W3,N_M1), (N_W3,N_P1), (N_W3,N_W1)} are divided into positive sample pairs and negative sample pairs N{(N_W3,N_M1), (N_W3,N_P1), (N_W3,N_W1)}. The rust similarity matrix M is calculated, and the Cholliski decomposition is performed on M to obtain a 256×256 dimensional lower triangular projection matrix L, constructing 10 local... A sensitive hash function is used, where the random projection vector follows a standard Gaussian distribution, the random offset follows a uniform distribution in [0,10], the hash bucket granularity is 10, the LSH function is obtained, 5 independent hash tables are constructed, the associated feature vectors of 9 supernodes are mapped to 10-bit hash signatures and stored in the corresponding hash buckets, and a local sensitive hash index is established for the supernode identifier and hash bucket position. The hash signature of N_W3 is [3,5,7,2,9,1,4,6,8,0], and it is stored in bucket 3 of hash table 1 and bucket 5 of hash table 2. S3 uses the storage partition as the query anchor point, and performs fast nearest neighbor matching through the local sensitive hash index based on the query anchor point to obtain a set of candidate vulnerable nodes; In the actual assessment, based on the storage supernode N_W3 detected by rust defects, the C05 partition was used as the query anchor point. Its 256-dimensional associated feature vector was extracted and substituted into the LSH function to generate a query hash signature. Candidate hash buckets were located in 5 hash tables, and supernodes within the buckets were extracted and cosine similarity was calculated. Among them, the similarity between candidate supernodes N_P5 and N_W3 was 0.92, so they were selected; the similarity between N_P6 and N_W3 was 0.9, so they were selected; the similarity between N_W2 and N_W3 was 0.75, so they were selected; and the similarity between N_P3 and N_W3 was 0.60, so they were removed. The candidate vulnerable node set S{N_P5 (surface protection), N_P6 (semi-finished product transfer), N_W2 (C04 storage)} was determined. S4 generates vulnerable process constraints for the candidate vulnerable node set through correlation analysis, and performs deep learning on the detected image of bushing rust based on the vulnerable process constraints to obtain the source superpath from vulnerable nodes to rust defects; In the actual assessment, Pearson correlation coefficient analysis was performed on the candidate vulnerable node set to calculate the correlation between node attributes and rust defect labels. Key vulnerable factors with a correlation greater than 0.7 were screened, generating three vulnerable process constraints. Among them, vulnerable process constraint y1 is "Surface protection process rust-preventive oil spraying thickness: 0.008mm / 0.02mm", with a correlation of 0.95; y2 is "C04 storage zone storage duration: 5.2h / 4h", with a correlation of 0.82; and y3 is "C05 storage zone corrosive gas concentration: 0.003%SO2 / 0.002%SO2", with a correlation of 0.98. The shafts corresponding to the candidate vulnerable nodes were then analyzed. Convolutional feature maps are extracted from the detected images. These feature maps are extracted using ResNet50 with dimensions H×W×2048 (H=32, W=32). The generated process defect spatial heatmap has dimensions 2×32×1. Indicator functions y1, y2, and y3 are set to 1 if the actual detected value is greater than the critical value of the process parameters. The coordinates of the high-rust-affected area (u1, v1) = (16, 16) are located on the outer surface of the bushing. (u2, v2) = (16, 16) and (u3, v1) = (16, 16) have Gaussian kernel spatial bandwidth of 2. Based on the camera field of view of 50mm×50mm and the bushing outer diameter of 30mm, the convolutional feature map is... The feature map is concatenated with the channel-wise thermal map of process defects, and the number of channels is reduced to 2048 through 1×1 convolution. After Sigmoid activation, attention weights are generated to obtain fused features. Based on the fused features, global average pooling is performed to obtain a 2048-dimensional vector, which is then concatenated with the 256-dimensional feature vector associated with candidate vulnerable nodes to obtain a 96-dimensional node feature matrix. The node feature matrix is ​​then input into a hypergraph convolutional neural network (HGNN), where the model structure of HGNN is "input layer → 2 hypergraph convolutional layers → attention fusion layer → output layer Softmax". The hypergraph convolutional layers are the hypergraph adjacency tensors generated in step S1, and the rust defect labels of the detected images are used as... The supervisory signal is set with a learning rate of 0.001 and 70 training rounds. The connection probabilities from the candidate vulnerable node to the defective node N_W3 are output as follows: N_P5→N_W3 (0.98), N_P6→N_W3 (0.85), and N_W2→N_P5 (0.90). Based on the connection probabilities, the source tracing hyperpath is extracted in reverse through beam search, resulting in two source tracing hyperpaths. The beam width of the beam search is 2. The source tracing hyperpaths P1 are N_P5 (surface protection) → E_W3 → N_W3 (C05 warehouse) and P2 are N_W2 (C04 warehouse) → N_P5 (surface protection) → E_W3 → N_W3 (C05 warehouse). S5 searches for root cause nodes based on each traceability superpath using a hypergraph network, and ranks them by confidence using DS evidence. Based on the production process units and detection image samples of the ranked nodes, it obtains the quality traceability results.

[0017] In the actual assessment, the traceability superpaths P1 and P2 were traced backwards to extract the supernode set {N_P5, N_W2, N_W3}. Based on the supernode set, three types of evidence were constructed: image evidence, process evidence, and structural evidence. These three types of evidence were then synthesized using Dempster's combination rule, and the basic probability distribution (BPA) of the synthesized evidence was calculated. The BPA of N_P5 was 0.96, and that of N_W2 was 0.58. The confidence function Bel and the plausibility function Pl were calculated for each node. Based on a pre-set confidence threshold of 0.7, the target root cause node was selected as N_P5 (surface protection process, process unit number G05). Based on the target root cause node, a quality traceability result was generated, including the root cause node identifier "N_P5 (G05, surface protection, 2025-03-01)". (16:00-17:00), confidence score 0.96, defect propagation super-edge path N_P5→E_W3→N_W3 (C05 warehouse)", the core root cause is that the thickness of the anti-rust oil spraying in the surface protection process was 0.008mm, which did not reach the critical threshold of 0.02mm, resulting in the failure of the bushing surface protection and the rapid generation of rust in the C05 warehouse environment. Improvement suggestions were obtained: optimize the spraying equipment parameters of the surface protection process, raise the lower limit of the anti-rust oil spraying thickness control to 0.02mm, reduce the temperature and humidity of the C05 warehouse zone to below 28℃ and below 50%RH, and install a corrosive gas purification device.

[0018] In this embodiment, the method for obtaining the supernode includes: Based on production line data and warehouse data, entities are extracted according to timestamps. Material supernodes, process supernodes, and warehouse supernodes are obtained by identifying raw material batches, production process units, and warehouse partitions according to entity identifiers. Production line data, warehouse data, and inspection images with corresponding timestamps are associated as attribute sets for each supernode. The material supernode is uniquely identified by the batch number, the process supernode is jointly identified by the process unit number and the process time period, and the warehouse supernode is spatially identified by the warehouse partition coordinates. The production line data includes the raw material batches, process units, processing environment data, processing sequence, heat treatment temperature curves, timestamps, processing equipment numbers, and storage zones for automotive axle bushings. The inspection images include surface inspection images, internal defect inspection images, and 3D point cloud data of the automotive axle bushings. The storage data includes the warehousing time of the automotive axle bushings, storage zone coordinates, stacking layers, storage period, storage environment data, and cross-regional transfer records.

[0019] In this embodiment, the method for obtaining the hypergraph network includes: The material flow superedge is constructed based on the flow direction of raw material batches between material supernodes and process supernodes. The weight of the material flow superedge is positively correlated with the proportion of raw material batches consumed in each process. Based on the process supernode, process flow superedges are constructed through the process connection relationship of production process units. The process connection relationship includes cross-process unit flow and sequential flow of adjacent process units. The weight of the process flow superedge is positively correlated with the frequency of material flow between process units and negatively correlated with the processing time delay. Based on the process supernode and the storage supernode, a storage-related superedge is constructed through cross-regional transfer records. The weight of the storage-related superedge is positively correlated with the cross-regional transfer frequency and storage duration. The weights of the material flow hyperedges, process flow hyperedges, and storage association hyperedges are normalized. A hypergraph adjacency tensor is generated based on the hypernodes and normalized hyperedges. The product of the normalized hyperedge weights and the feature similarity between the hypernodes is used as the element value of the hypergraph adjacency tensor to obtain the hypergraph network.

[0020] In this embodiment, the method for obtaining the hyperedge feature includes: Based on the hypergraph network, the material flow hyperedge is used to extract the flow sequence of raw material batches through process units, the consumption ratio between each process, and the flow time span to obtain the hyperedge features of the material flow path. Based on the process flow hyperedge, the similarity of processing parameters of adjacent process units, the quality entropy of cross process units, and the processing time delay are extracted to obtain the hyperedge features of process coupling relationship. The quality entropy is obtained based on the information entropy of the process defect rate.

[0021] In this embodiment, the method for obtaining the locality-sensitive hash index includes: Based on the hyperedge features, the processing environment data and warehousing environment data of the corresponding nodes are used as environmental features. Convolutional embedding is performed using the hyperedge feature vector, environmental feature vector, and hypergraph network through the hyperedge Laplacian algorithm to obtain the associated feature vector of the hypernode. Based on the detected images of the hypernodes, positive and negative sample pairs are divided using rust defect labels, and a rust similarity matrix is ​​calculated. The formula for calculating the rust similarity matrix is ​​as follows: ; in Let be the rust similarity matrix, and let be a symmetric positive definite matrix. Positive sample pairs gather, negative sample pairs gather, For sample pairs labeled with different rust defects, For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For transpose; The rust similarity matrix is ​​subjected to Cholliski decomposition to obtain the projection matrix. A locality-sensitive hash function is then generated based on the projection matrix. The locality-sensitive hash function is as follows: ; in For the first Locality-sensitive hash function, For associated feature vectors, Let be the projection matrix. For matrix multiplication, For inner product operations, For the first A random projection vector is obtained by independent sampling according to a standard Gaussian distribution. For the first A random offset, based on a uniform distribution Obtained through random sampling. This refers to the granularity of the hash bucket. This is the floor function; The output of the locality-sensitive hash function is combined into a hash signature to obtain a hash table. Based on the hash table, each supernode identifier and the associated feature vector are stored in the corresponding hash bucket to obtain the locality-sensitive hash index of the supernode identifier and the hash bucket position.

[0022] In this embodiment, the method for obtaining the source tracing superpath includes: Correlation analysis is performed on the candidate vulnerable node set to obtain the correlation degree between each node attribute and the rust defect label. Vulnerability process constraints are generated based on these correlation degrees, and a spatial heat map of process defects is generated based on these constraints. The vulnerability process constraints include critical values ​​and actual detected values ​​of process parameters. The calculation formula for the spatial heat map of process defects is as follows: ; in This is a space thermal map of process defects. , For spatial coordinate indexes on the feature map, For rows in the height direction, , For columns in the width direction, , The height of the convolutional feature map. The width of the convolutional feature map is given by the image being detected. This represents the total number of vulnerable process constraints. This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. For the first The actual measured values ​​of each process parameter For the first Critical values ​​of each process parameter For the first Coordinates of high-incidence areas of rust defects corresponding to each vulnerable process constraint. For the first The spatial influence range of the Gaussian kernel corresponding to each process constraint in the horizontal and vertical directions is preset according to the process unit through the camera field of view and the workpiece size. The spatial heatmap of the process defect and the convolutional feature map of the detection image are weighted and fused to obtain the fused feature. The calculation formula of the fused feature is as follows: ; in This is the fused feature map after process attention weighting. , For convolutional feature maps, For the number of channels, Indicates that F and spliced ​​as The channel tensor, For activation function, This is element-wise multiplication; The fused feature map is subjected to global average pooling to obtain the defect region feature vector. The defect region feature vector is concatenated with the associated feature vector of the candidate vulnerable node to construct a node feature matrix. The node feature matrix is ​​input into the hypergraph convolutional neural network. The rust defect label of the detected image is used as a supervision signal to perform deep learning on the hypergraph convolutional neural network. The optimization objective is to minimize the weighted sum of the defect prediction loss and the vulnerability constraint regularization term. The connection probability from each candidate vulnerable node to the defect node is output. The connection probability is used as the hyperedge propagation weight. Based on the propagation weight, the source hyperpath from the defect node to the candidate vulnerable node is extracted inversely through beam search.

[0023] In this embodiment, the method for obtaining the quality traceability result includes: Based on each traceability superpath, reverse tracing is performed through a hypergraph network. The set of supernodes on each traceability superpath is extracted. Based on the set of supernodes, image evidence is obtained by detecting defect labels in the defect region and node association region of the image. Process evidence is obtained based on the difference between the process parameters corresponding to the set of supernodes and the standard range. Structural evidence is obtained based on the centrality index of the set of supernodes in the hypergraph network. The image evidence, process evidence and structural evidence are synthesized using Dempster-Shafer evidence theory. The basic probability distribution of the synthesis of each node in the set of supernodes is calculated. The centrality index includes degree centrality or betweenness centrality. The confidence function and similarity function of each node are calculated based on the basic probability allocation. The nodes are sorted in descending order according to the confidence function. Target root cause nodes are screened based on a preset confidence threshold. The production process unit, raw material batch and detection image sample of the target root cause node are used as quality sources to obtain quality traceability results. The quality traceability results include root cause node identification, confidence score and defect propagation hyperedge path.

[0024] A second aspect of the present invention also provides a deep learning-based bushing quality traceability system, comprising: Data acquisition module: used to acquire inspection images, production line data and warehouse data of automobile axle sleeves, and to construct a hypergraph network based on the raw material batches, production process units and warehouse partitions of the production line data as super nodes; Index generation module: used to extract hyperedge features of material flow paths and process coupling relationships based on the hypergraph network, and generate local sensitive hash indexes based on the hyperedge features and the environmental features of the corresponding nodes; Candidate vulnerable node screening module: Used to use the storage partition as the query anchor point, and perform fast nearest neighbor matching based on the query anchor point through local sensitive hash index to obtain a set of candidate vulnerable nodes; Source tracing path generation module: used to generate vulnerable process constraints from the candidate vulnerable node set through correlation analysis, and perform deep learning on the detected image of bushing rust based on the vulnerable process constraints to obtain the source tracing hyperpath from vulnerable node to rust defect; The traceability module is used to search for root cause nodes based on each traceability superpath using the hypergraph network, and to rank them by confidence using DS evidence. Based on the production process unit and detection image samples of the ranked nodes, the module obtains the quality traceability results.

[0025] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A deep learning-based bushing quality traceability method, characterized in that, Includes the following steps: Acquire inspection images, production line data, and warehouse data of automotive axle sleeves, and use the raw material batches, production process units, and warehouse partitions of the production line data as super nodes to construct a hypergraph network based on the super nodes; Based on the hypergraph network, extract the hyperedge features of material flow path and process coupling relationship, and generate local sensitive hash index based on the hyperedge features and the environmental features of the corresponding nodes; Using the storage partition as the query anchor point, a set of candidate vulnerable nodes is obtained by fast nearest neighbor matching based on the query anchor point through local sensitive hash index. The candidate vulnerable node set is used to generate vulnerable process constraints through correlation analysis. Based on the vulnerable process constraints, deep learning is performed on the detected image of bushing rust to obtain the source superpath from vulnerable node to rust defect. Based on each traceability superpath, the root cause node is searched using the hypergraph network, and confidence is ranked using DS evidence. The quality traceability results are obtained based on the production process unit and detection image samples of the ranked nodes.

2. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the supernode includes: Based on production line data and warehouse data, entities are extracted according to timestamps. Material supernodes, process supernodes, and warehouse supernodes are obtained by identifying raw material batches, production process units, and warehouse partitions according to entity identifiers. Production line data, warehouse data, and inspection images with corresponding timestamps are associated as attribute sets for each supernode. The material supernode is uniquely identified by the batch number, the process supernode is jointly identified by the process unit number and the process time period, and the warehouse supernode is spatially identified by the warehouse partition coordinates. The production line data includes the raw material batches, process units, processing environment data, processing sequence, heat treatment temperature curves, timestamps, processing equipment numbers, and storage zones for automotive axle bushings. The inspection images include surface inspection images, internal defect inspection images, and 3D point cloud data of the automotive axle bushings. The storage data includes the warehousing time of the automotive axle bushings, storage zone coordinates, stacking layers, storage period, storage environment data, and cross-regional transfer records.

3. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the hypergraph network includes: The material flow superedge is constructed based on the flow direction of raw material batches between material supernodes and process supernodes. The weight of the material flow superedge is positively correlated with the proportion of raw material batches consumed in each process. Based on the process supernode, process flow superedges are constructed through the process connection relationship of production process units. The process connection relationship includes cross-process unit flow and sequential flow of adjacent process units. The weight of the process flow superedge is positively correlated with the frequency of material flow between process units and negatively correlated with the processing time delay. Based on the process supernode and the storage supernode, a storage-related superedge is constructed through cross-regional transfer records. The weight of the storage-related superedge is positively correlated with the cross-regional transfer frequency and storage duration. The weights of the material flow hyperedges, process flow hyperedges, and storage association hyperedges are normalized. A hypergraph adjacency tensor is generated based on the hypernodes and normalized hyperedges. The product of the normalized hyperedge weights and the feature similarity between the hypernodes is used as the element value of the hypergraph adjacency tensor to obtain the hypergraph network.

4. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the hyperedge features includes: Based on the hypergraph network, the material flow hyperedge is used to extract the flow sequence of raw material batches through process units, the consumption ratio between each process, and the flow time span to obtain the hyperedge features of the material flow path. Based on the process flow hyperedge, the similarity of processing parameters of adjacent process units, the quality entropy of cross process units, and the processing time delay are extracted to obtain the hyperedge features of process coupling relationship. The quality entropy is obtained based on the information entropy of the process defect rate.

5. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the locality-sensitive hash index includes: Based on the hyperedge features, the processing environment data and warehousing environment data of the corresponding nodes are used as environmental features. Convolutional embedding is performed using the hyperedge feature vector, environmental feature vector, and hypergraph network through the hyperedge Laplacian algorithm to obtain the associated feature vector of the hypernode. Based on the detected images of the hypernodes, positive and negative sample pairs are divided using rust defect labels, and a rust similarity matrix is ​​calculated. The formula for calculating the rust similarity matrix is ​​as follows: ； in Let be the rust similarity matrix, and let be a symmetric positive definite matrix. Positive sample pairs gather, negative sample pairs gather, For sample pairs labeled with different rust defects, For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For the first The associated feature vectors of each supernode. For transpose; The rust similarity matrix is ​​subjected to Cholliski decomposition to obtain the projection matrix. A locality-sensitive hash function is then generated based on the projection matrix. The locality-sensitive hash function is as follows: ； in For the first Locality-sensitive hash function, For associated feature vectors, Let be the projection matrix. For matrix multiplication, For inner product operations, For the first A random projection vector is obtained by independent sampling according to a standard Gaussian distribution. For the first A random offset, based on a uniform distribution Obtained through random sampling. This refers to the granularity of the hash bucket. This is the floor function; The output of the locality-sensitive hash function is combined into a hash signature to obtain a hash table. Based on the hash table, each supernode identifier and the associated feature vector are stored in the corresponding hash bucket to obtain the locality-sensitive hash index of the supernode identifier and the hash bucket position.

6. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the source tracing superpath includes: Correlation analysis is performed on the candidate vulnerable node set to obtain the correlation degree between each node attribute and the rust defect label. Vulnerability process constraints are generated based on these correlation degrees, and a spatial heat map of process defects is generated based on these constraints. The vulnerability process constraints include critical values ​​and actual detected values ​​of process parameters. The calculation formula for the spatial heat map of process defects is as follows: ； in This is a space thermal map of process defects. , For spatial coordinate indexes on the feature map, For rows in the height direction, , For columns in the width direction, , The height of the convolutional feature map. The width of the convolutional feature map is given by the image being detected. This represents the total number of vulnerable process constraints. This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. For the first The actual measured values ​​of each process parameter For the first Critical values ​​of each process parameter For the first Coordinates of high-incidence areas of rust defects corresponding to each vulnerable process constraint. For the first The spatial influence range of the Gaussian kernel corresponding to each process constraint in the horizontal and vertical directions is preset according to the process unit through the camera field of view and the workpiece size. The spatial heatmap of the process defect and the convolutional feature map of the detection image are weighted and fused to obtain the fused feature. The calculation formula of the fused feature is as follows: ； in This is the fused feature map after process attention weighting. , For convolutional feature maps, For the number of channels, Indicates that F and spliced ​​as The channel tensor, For activation function, This is element-wise multiplication; The fused feature map is subjected to global average pooling to obtain the defect region feature vector. The defect region feature vector is concatenated with the associated feature vector of the candidate vulnerable node to construct a node feature matrix. The node feature matrix is ​​input into the hypergraph convolutional neural network. The rust defect label of the detected image is used as a supervision signal to perform deep learning on the hypergraph convolutional neural network. The optimization objective is to minimize the weighted sum of the defect prediction loss and the vulnerability constraint regularization term. The connection probability from each candidate vulnerable node to the defect node is output. The connection probability is used as the hyperedge propagation weight. Based on the propagation weight, the source hyperpath from the defect node to the candidate vulnerable node is extracted inversely through beam search.

7. The deep learning-based bushing quality traceability method according to claim 1, characterized in that, The method for obtaining the quality traceability results includes: Based on each traceability superpath, reverse tracing is performed through a hypergraph network. The set of supernodes on each traceability superpath is extracted. Based on the set of supernodes, image evidence is obtained by detecting defect labels in the defect region and node association region of the image. Process evidence is obtained based on the difference between the process parameters corresponding to the set of supernodes and the standard range. Structural evidence is obtained based on the centrality index of the set of supernodes in the hypergraph network. The image evidence, process evidence and structural evidence are synthesized using Dempster-Shafer evidence theory. The basic probability distribution of the synthesis of each node in the set of supernodes is calculated. The centrality index includes degree centrality or betweenness centrality. The confidence function and similarity function of each node are calculated based on the basic probability allocation. The nodes are sorted in descending order according to the confidence function. Target root cause nodes are screened based on a preset confidence threshold. The production process unit, raw material batch and detection image sample of the target root cause node are used as quality sources to obtain quality traceability results. The quality traceability results include root cause node identification, confidence score and defect propagation hyperedge path.

8. A deep learning-based bushing quality traceability system, used to execute the deep learning-based bushing quality traceability method according to any one of claims 1 to 7, characterized in that, The system includes: Data acquisition module: used to acquire inspection images, production line data and warehouse data of automobile axle sleeves, and to construct a hypergraph network based on the raw material batches, production process units and warehouse partitions of the production line data as super nodes; Index generation module: used to extract hyperedge features of material flow paths and process coupling relationships based on the hypergraph network, and generate local sensitive hash indexes based on the hyperedge features and the environmental features of the corresponding nodes; Candidate vulnerable node screening module: Used to use the storage partition as the query anchor point, and perform fast nearest neighbor matching based on the query anchor point through local sensitive hash index to obtain a set of candidate vulnerable nodes; Source tracing path generation module: used to generate vulnerable process constraints from the candidate vulnerable node set through correlation analysis, and perform deep learning on the detected image of bushing rust based on the vulnerable process constraints to obtain the source tracing hyperpath from vulnerable node to rust defect; The traceability module is used to search for root cause nodes based on each traceability superpath using the hypergraph network, and to rank them by confidence using DS evidence. Based on the production process unit and detection image samples of the ranked nodes, the module obtains the quality traceability results.

Images