An automobile electronic component marking code checking and supply chain tracing method

Abstract

The application discloses a kind of automobile electronic parts marking code check and supply chain traceability method, comprising: forming original marking image data, obtaining illumination calibration marking image;Mapping to global uniform world coordinate system, generating normalized physical position coding features;Generate time series radio frequency sensing dataset, filter the radio frequency noise caused by multipath effect, generate confidence weighted radio frequency feature vector;Get marking code label check features;Construct warehouse in and out event space-time feature tensor, generate multi-node logistics track matching sequence;Complete asset authenticity check and output asset anti-fake verification label;Generate supply chain traceability closed-loop management instructions containing spatial grid location, time slice state, flow legality proof and tamper-proof hash features.The application improves the ability of abnormal flow identification and path authenticity verification in supply chain traceability process.

Description

Technical Field

[0001] This invention relates to the field of automotive electronic component technology, and in particular to a method for verifying the marking codes of automotive electronic components and tracing their supply chain. Background Technology

[0002] With the development of intelligent manufacturing and digital supply chain management for automotive electronic components, the demand for identity verification, status tracking, and full lifecycle traceability of automotive electronic components during production, warehousing, transportation, and circulation is constantly increasing. Existing automotive electronic component warehousing management systems typically rely on manual data entry, barcode scanning, or a single RFID method to record and manage component information, making it difficult to simultaneously ensure comprehensive verification of the integrity of component surface markings, changes in spatial location, warehousing circulation status, and the authenticity of the supply chain.

[0003] In the process of verifying the entry and exit of automotive electronic components, traditional visual inspection-based marking code recognition methods are often affected by factors such as high reflectivity of metal component surfaces, defects in laser-etched characters, blurred character boundaries, and local texture variations. This leads to inaccurate feature extraction of the marked characters, making it difficult to reliably verify the marking information on the component surfaces. Existing warehouse positioning methods typically rely on fixed-position coding or single-sensor positioning, which is insufficient to accurately describe the spatial position changes and dynamic flow status of components between different warehouse nodes.

[0004] Existing RFID-based warehouse tracking methods typically rely on single tag reads to determine component locations. However, factors such as metal shelf reflections, environmental obstructions, and multipath propagation of radio frequency signals can lead to signal fluctuations, read anomalies, and location inaccuracies, resulting in insufficient reliability of RFID sensing data during warehouse operations. Similarly, existing supply chain traceability methods often establish flow relationships based on single ledger records or simple time-series associations. These methods struggle to integrate visual verification information, spatial location information, RFID sensing information, and blockchain-based evidence, making it difficult to effectively verify the authenticity of supply chain flow paths and the consistency of asset status.

[0005] Existing technologies for warehousing and outbound verification and supply chain traceability of automotive electronic components suffer from several problems, including insufficient accuracy in verifying component surface markings, inadequate ability to correlate warehouse space status, susceptibility of radio frequency sensing data to environmental influences, difficulty in verifying the authenticity of multi-node flow in the supply chain, and insufficient reliability in matching logistics trajectories throughout the entire lifecycle. Summary of the Invention

[0006] One objective of this invention is to propose a method for verifying the marking codes of automotive electronic components and for supply chain traceability. This invention improves the ability to identify abnormal flow and verify the authenticity of the path during the supply chain traceability process.

[0007] A method for verifying the marking codes of automotive electronic components and for supply chain traceability according to an embodiment of the present invention includes:

[0008] During the circulation of automotive electronic components, the original DPM code image of the surface of the automotive electronic components to be circulated is acquired by the high-definition vision sensor of the production line, forming the original marking image data. Nonlinear grayscale anisotropic diffusion filtering is performed on the original marking image data to suppress the mirror artifacts caused by high reflectivity on the metal surface, and the illumination-calibrated marking image is obtained.

[0009] Using a binocular depth camera, the spatial pose of the mechanical gripper arm containing automotive electronic components is acquired synchronously to obtain the initial physical gripping pose. The initial physical gripping pose is then input into a spatial homogeneous transformation matrix and mapped to a global unified world coordinate system to generate normalized physical position coding features.

[0010] By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, the time-series signal strength index stream within a continuous time window is captured, generating a time-series RF sensing dataset. For the time-series RF sensing dataset, a sliding time window is used to perform short-time energy averaging and Gaussian kernel density estimation, filtering out RF noise caused by multipath effects, and generating a confidence-weighted RF feature vector.

[0011] The illumination-calibrated marking image is input into a pre-trained marking code defect detection network to delineate the geometric topological boundary of the marking characters and extract the character distortion components to obtain the marking code label verification features.

[0012] Normalized physical location coding features and confidence-weighted radio frequency feature vectors are concatenated at the feature level, and a strong spatiotemporal correlation is established based on the component identity ID. A spatiotemporal feature tensor of warehouse entry and exit events is constructed. The on-chain flow trajectory path in the distributed local ledger of each warehouse network is retrieved. The multi-node spatiotemporal state convergence of the spatiotemporal feature tensor of warehouse entry and exit events is calculated by the least squares optimal alignment algorithm, and a multi-node logistics trajectory matching sequence is generated.

[0013] Obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger, perform asymmetric encrypted digital signature verification based on the label verification features, complete the asset authenticity verification and output the asset anti-counterfeiting verification label;

[0014] Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a supply chain trust score distribution is constructed. The maximum flow allocation path of the entire life cycle is extracted according to the flow network flow balance rule, and a supply chain traceability closed-loop management instruction containing spatial grid location, time slice status, flow legality proof and tamper-proof hash features is generated.

[0015] Optionally, during the flow of automotive electronic components, the acquisition of original DPM code images of the laser-etched surfaces of the automotive electronic components to be flowed through high-definition vision sensors on the production line includes:

[0016] When automotive electronic components enter the transfer and acquisition station, fixed-focus imaging is performed to acquire the original DPM code image and form the original marking image data.

[0017] The original marking image data is normalized in terms of gray dimensions to obtain a normalized original marking grayscale image.

[0018] Based on the normalized original labeled grayscale image, a weighted image of the mirror pseudo-effect of the metal surface is constructed.

[0019] Using the normalized original labeled grayscale image as the initial input for nonlinear grayscale anisotropic diffusion filtering, the directional grayscale difference is calculated to generate nonlinear diffusion control coefficients.

[0020] Based on the directional gray-level difference, nonlinear diffusion control coefficient and metal surface mirror pseudo-effect weight map, nonlinear gray-level anisotropic diffusion filtering is performed on the normalized original marking gray-level image to obtain the filtered marking gray-level image.

[0021] Perform grayscale boundary truncation on the filtered grayscale image to obtain the illumination-calibrated grayscale image.

[0022] Optionally, the step of simultaneously acquiring the spatial pose of the mechanical gripper arm where the automotive electronic components are located using a binocular depth camera to obtain the initial physical gripping pose includes:

[0023] The initial physical grasping pose is obtained by simultaneously acquiring the spatial pose of the mechanical gripper arm where the automotive electronic components are located using a binocular depth camera.

[0024] Based on the installation and calibration data of the binocular depth camera and the correspondence between the coordinates of the calibration reference point, a spatial homogeneous transformation matrix from the binocular depth camera coordinate system to the global unified world coordinate system is constructed.

[0025] The initial physical grasp pose is input into the homogeneous transformation matrix of the space and mapped to the global unified world coordinate system to obtain the physical grasp pose in the world coordinate system.

[0026] The physical grasping pose in the world coordinate system is normalized to obtain normalized spatial position components and normalized attitude components.

[0027] Following a fixed order of the X-axis, Y-axis, Z-axis, roll attitude angle, pitch attitude angle, and yaw attitude angle in the global unified world coordinate system, the normalized spatial position components and normalized attitude components are vectorized and encoded to generate normalized physical position coding features.

[0028] Optionally, the step of periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes to capture the time-series signal strength index stream within a continuous time window includes:

[0029] By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, the temporal signal strength index stream within a continuous time window is captured.

[0030] A time-series radio frequency sensing dataset is generated based on the time-series signal strength index stream.

[0031] A sliding time window is constructed for the time-series radio frequency sensing dataset to obtain multiple window radio frequency sensing sub-sequences;

[0032] Perform short-time energy averaging on each window radio frequency sensing subsequence to obtain the window short-time energy average;

[0033] Gaussian kernel density estimation is performed on each window radio frequency sensing subsequence to obtain the window radio frequency density estimate;

[0034] Window RF confidence weights are generated based on the estimated window RF density.

[0035] A confidence-weighted radio frequency feature vector is generated based on the window's short-time energy average and the window's radio frequency confidence weight.

[0036] Optionally, the step of inputting the illumination-calibrated marking image into a pre-trained marking code defect detection network, outlining the geometric topological boundaries of the marking characters, and extracting the character distortion components includes:

[0037] The illumination calibration and marking image is input into the training anisotropic heterogeneous multi-branch network to obtain the horizontal stroke response feature map, the vertical stroke response feature map and the local texture response feature map, and generate a set of multi-branch heterogeneous feature matrices.

[0038] Perform channel normalization and spatial alignment on the set of multi-branch heterogeneous feature matrices to generate channel normalized feature tensor parameters;

[0039] Based on the channel-normalized feature tensor parameters, the convolution parameters of each branch are equivalently transformed to generate equivalent transformation weights and equivalent translation biases with aligned dimensions.

[0040] The equivalent transformation weights with dimensional alignment and the equivalent translation bias are structurally reparameterized and fused to generate a single-branch equivalent 3x3 main convolution kernel parameter set;

[0041] Based on the parameter set of the single-branch equivalent 3x3 main convolution kernel, forward inference is performed on the illumination calibration and marking image to generate an anisotropic marking candidate response field.

[0042] Local variational detection is performed on the anisotropic marking candidate response field along the topological skeleton of the marked character to generate a stroke distortion feature differential field;

[0043] Adaptive threshold segmentation and connected component filtering are performed on the differential field of stroke distortion features to generate a damage and defect mask for component marking codes.

[0044] Based on the damage and defect mask of the marking code and the marking character area of ​​the parts, a set of characteristic indicators for the quality compliance of the parts is constructed.

[0045] Based on the component quality compliance feature index set and the component marking code damage defect mask, mark code label verification features are generated.

[0046] Optionally, the step of concatenating the normalized physical location coding features and the confidence-weighted radio frequency feature vector at the feature level, and establishing a strong spatiotemporal association based on the component identity ID, includes:

[0047] The normalized physical location coding features and the confidence-weighted radio frequency feature vector are concatenated at the feature level to obtain the fusion perception feature vector for parts entering and leaving the warehouse.

[0048] Establish a strong spatiotemporal correlation of component entry and exit fusion perception feature vectors based on component identity ID, and construct a spatiotemporal feature tensor for warehouse entry and exit events;

[0049] Retrieve the on-chain flow path from the distributed local ledger of each warehouse outlet;

[0050] Based on the spatiotemporal feature tensor of warehouse inbound and outbound events and the on-chain flow trajectory path, multi-node spatiotemporal aligned sample pairs are constructed.

[0051] Calculate the spatiotemporal alignment residual based on multi-node spatiotemporal aligned sample pairs;

[0052] Based on the spatiotemporal alignment residual, calculate the spatiotemporal state convergence of multiple nodes;

[0053] Based on the convergence of the spatiotemporal states of multiple nodes and the on-chain flow trajectory path, a multi-node logistics trajectory matching sequence is generated.

[0054] Optionally, the step of obtaining the pre-stored supply chain source certificate in the blockchain distributed ledger and performing asymmetric encrypted digital signature verification based on the tag verification features includes:

[0055] Based on the component identity ID, obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger and generate source factory certificate parsing data;

[0056] Generate a current label verification summary based on the label verification features;

[0057] Asymmetric encrypted digital signature verification is performed based on the source factory-existing evidence parsing data to obtain the source signature verification result;

[0058] Based on the current marking verification summary and the source factory certificate analysis data, the marking consistency verification result is obtained;

[0059] Based on the source verification results and the labeling consistency verification results, the authenticity of the assets is verified and an asset anti-counterfeiting verification label is output.

[0060] Optionally, the step of constructing a supply chain trust score distribution based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences includes:

[0061] Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a set of supply chain circulation nodes is constructed;

[0062] Based on the set of supply chain flow nodes, spatial grid locations and time slice states are divided;

[0063] Based on the asset anti-counterfeiting verification label, the set of supply chain circulation nodes is assigned anti-counterfeiting credibility weights to obtain anti-counterfeiting credibility weights.

[0064] Based on the multi-node logistics trajectory matching sequence, the trajectory matching credibility of the supply chain flow node set is assigned, and a supply chain trust score distribution is constructed.

[0065] Based on the supply chain trust score distribution, construct a full life cycle circulation network;

[0066] Based on the full lifecycle flow network, the maximum flow allocation path for the entire lifecycle flow is extracted according to the flow network traffic balancing rules;

[0067] Based on the maximum flow allocation path throughout the entire lifecycle, generate closed-loop management instructions for supply chain traceability.

[0068] The beneficial effects of this invention are:

[0069] (1) This invention achieves multi-directional fine-grained defect perception for laser-etched characters on the metal surface of automotive electronic components. By constructing an anisotropic heterogeneous multi-branch network containing horizontal slice convolutional branches, vertical slice convolutional branches, and local texture response branches, the horizontal stroke continuity features, vertical stroke continuity features, and local defect texture features of the etched characters are extracted respectively. Through batch normalization spatial alignment and structural reparameterization fusion, the multi-branch feature extraction capability in the training stage is converted into a single-branch equivalent convolutional inference structure. While maintaining the multi-directional feature perception capability, the computational complexity of online detection is reduced, and the real-time processing capability of the warehouse edge detection node is improved.

[0070] (2) This invention constructs a multi-source heterogeneous spatiotemporal correlation model for the warehousing inbound and outbound process by fusing normalized physical location coding features with confidence-weighted radio frequency feature vectors, thereby improving the reliability of component status perception in complex warehousing environments. By performing coordinate system-one mapping and normalized coding on the spatial pose of the mechanical gripper arm acquired by the binocular depth camera, and constructing a sliding time window for the continuous sampling signal of the UHF RFID active tag, the radio frequency confidence weight of the window is generated by short-time energy averaging and Gaussian kernel density estimation, which suppresses the low-confidence radio frequency perception window affected by multipath effects. The spatial location features and radio frequency perception features are fused at the feature level, so that the spatial state and radio frequency state in the component warehousing flow process form a stable correlation, thereby improving the accuracy of inbound and outbound event identification and component status tracking.

[0071] (3) This invention achieves reliable verification of the entire lifecycle circulation path of automotive electronic components by combining blockchain source evidence storage and signature verification, multi-node logistics trajectory matching, and maximum flow optimization supply chain closed-loop analysis method. This invention generates the current tag verification summary based on the tag verification features, and realizes the authenticity verification of source factory evidence storage through asymmetric encrypted digital signature verification. It constructs the supply chain trust score distribution by combining asset anti-counterfeiting verification tags and multi-node logistics trajectory matching sequences. It uniformly describes the supply chain circulation nodes, node connection relationships, and node credibility through flow network modeling. It extracts the reliable circulation path of the entire lifecycle using flow conservation constraints and maximum flow allocation path solving methods, and generates a supply chain traceability closed-loop management instruction containing spatial grid position, time slice status, circulation legality proof, and tamper-proof hash features, thereby improving the ability to identify abnormal circulation and verify the authenticity of the path during the supply chain traceability process. Attached Figure Description

[0072] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0073] Figure 1 This is a flowchart of a method for verifying the marking codes of automotive electronic components and tracing the supply chain, as proposed in this invention.

[0074] Figure 2 This is a comparison of the processing effects of the original DPM code image to the illumination-calibrated marking image in Embodiment 2 of the method for verifying and tracing the marking codes of automotive electronic components proposed in this invention. Detailed Implementation

[0075] Example 1: Reference Figure 1 A method for verifying the marking codes of automotive electronic components and for supply chain traceability, comprising:

[0076] During the circulation of automotive electronic components, the original DPM code image of the surface of the automotive electronic components to be circulated is acquired by the high-definition vision sensor of the production line, forming the original marking image data. Nonlinear grayscale anisotropic diffusion filtering is performed on the original marking image data to suppress the mirror artifacts caused by high reflectivity on the metal surface, and the illumination-calibrated marking image is obtained.

[0077] In this embodiment, during the flow of automotive electronic components, the original DPM code image of the laser-etched surface of the automotive electronic components to be flowed is acquired by a high-definition vision sensor on the production line, including:

[0078] When automotive electronic components enter the transfer and acquisition station, fixed-focus imaging is performed to acquire the original DPM code image and form the original marking image data.

[0079] In Example 1, a high-definition vision sensor on the production line is used to perform fixed-focus imaging of the laser-etched area on the surface of the automotive electronic parts to be processed. The pixel area where the DPM code is located in the laser-etched area is determined as the marking acquisition area to obtain the original DPM code image. The original DPM code image is then represented by grayscale matrix according to two-dimensional pixel coordinates to form the original marking image data.

[0080] The original marking image data is normalized in terms of gray dimensions to obtain a normalized original marking grayscale image.

[0081] In Example 1, the dark field grayscale difference is obtained by subtracting the dark field grayscale reference value corresponding to the production line high-definition vision sensor in the original marking image acquisition scene from the original grayscale response value at the two-dimensional pixel coordinates. The grayscale normalization scale reference value is obtained by subtracting the dark field grayscale reference value corresponding to the production line high-definition vision sensor in the original marking image acquisition scene from the bright field grayscale reference value corresponding to the production line high-definition vision sensor in the original marking image acquisition scene and adding a grayscale normalization stability coefficient. The grayscale normalization scale reference value is obtained by dividing the dark field grayscale difference by the grayscale normalization scale reference value. The normalized original marking grayscale image is obtained.

[0082] The minimum resolvable grayscale change of the high-definition vision sensor on the production line in the original marking image acquisition scenario is obtained, and the larger value is taken after comparing it with the preset lower limit stability value to obtain the grayscale normalization stability coefficient.

[0083] Based on the normalized original labeled grayscale image, a weighted image of the mirror pseudo-effect of the metal surface is constructed.

[0084] In Example 1, the gray-level surge response value is obtained by calculating the difference between the normalized original labeled gray value at each two-dimensional pixel coordinate in the normalized original labeled gray image and the normalized gray mean in the local neighborhood of the labeled image centered on the two-dimensional pixel coordinate.

[0085] The mirror artifact deviation response value is obtained by subtracting the normalized gray standard deviation in the local neighborhood of the marked image from the gray-level sudden increase response value.

[0086] By subtracting the normalized grayscale mean value in the local neighborhood of the marked image centered on the two-dimensional pixel coordinates from one, and then adding a grayscale normalization stability coefficient, the normalized reference value of the mirror artifact is obtained.

[0087] The initial mirror artifact weight value is obtained by dividing the mirror artifact deviation response value by the mirror artifact normalized reference value.

[0088] By limiting the initial mirror artifact weight values ​​to a continuous weight range of 0 to 1, a mirror artifact response weight map of the metal surface is obtained.

[0089] Using the normalized original labeled grayscale image as the initial input for nonlinear grayscale anisotropic diffusion filtering, the directional grayscale difference is calculated to generate nonlinear diffusion control coefficients.

[0090] In Example 1, the normalized original grayscale image of the marking is used as the initial input image for nonlinear grayscale anisotropic diffusion filtering. The right, left, down and up directions form a set of four neighborhood directions of the marking image. For each two-dimensional pixel coordinate, the grayscale value of the marking image at the current two-dimensional pixel coordinate is subtracted from the grayscale value of the marking image at the adjacent two-dimensional pixel coordinates in the corresponding neighborhood direction to obtain the directional grayscale difference.

[0091] ;

[0092] in, Represents the nonlinear diffusion control coefficient; Represents pixel coordinates in two-dimensional space; Indicates the current iteration round; Indicates the direction identifier of the four neighboring regions, with a value range of 1. (That is, the four directions: up, down, left, and right); Indicates directional grayscale difference; This indicates that the laser etching edge maintains the diffusion threshold; This represents the weight of the mirror artifact effect on the metal surface, with a value range of [value range missing]. ; This represents the gray-level normalization stability coefficient.

[0093] The diffusion threshold for laser etching edge preservation is obtained by statistically normalizing the absolute value of the directional gray-level difference in the edge region of the marked character in the original marked gray-level image and taking the median level. The value ranges from 0.05 to 0.3. The selection criteria are that the value is higher than the gray-level difference amplitude corresponding to the random gray-level noise on the metal surface and lower than the gray-level transition amplitude corresponding to the true edge of the laser-etched character. This allows the nonlinear gray-level anisotropic diffusion filter to preserve the edge of the DPM code character while suppressing the mirror artifacts on the metal surface.

[0094] Based on the directional gray-level difference, nonlinear diffusion control coefficient and metal surface mirror pseudo-effect weight map, nonlinear gray-level anisotropic diffusion filtering is performed on the normalized original marking gray-level image to obtain the filtered marking gray-level image.

[0095] In Example 1, the cumulative diffusion value of the four-neighbor direction is obtained by multiplying and summing the nonlinear diffusion control coefficients corresponding to each neighborhood direction in the four-neighbor direction set of the marked image with the directional gray-level difference.

[0096] The metal surface mirror artifact suppression term is obtained by multiplying the metal surface mirror artifact suppression weight by the gray difference between the diffusion gray value of the current iteration of the marking image and the gray mean of the local neighborhood of the marking image centered on the current two-dimensional pixel coordinates.

[0097] The diffusion grayscale value of the marking image in the next iteration is obtained by jointly updating the diffusion grayscale value of the current iteration's marked image, the cumulative diffusion value of the four-neighbor direction after being weighted by the diffusion iteration step size, and the metal surface mirror artifact suppression term after being weighted by the mirror artifact suppression weight. The value of the diffusion iteration step size is greater than 0 and less than or equal to 0.25.

[0098] When the number of iterations reaches the preset diffusion iteration number, and the change in the diffusion grayscale value of the marking image obtained from two adjacent iterations is less than the preset diffusion convergence threshold, the nonlinear grayscale anisotropic diffusion filtering is stopped, and the filtered marking grayscale image is obtained.

[0099] Perform grayscale boundary truncation on the filtered grayscale image to obtain the illumination-calibrated grayscale image.

[0100] In Example 1, the diffusion grayscale value of the marking image at each two-dimensional pixel coordinate in the filtered marking grayscale image is restricted to a uniform dimensionless grayscale range to obtain an illumination-calibrated marking image.

[0101] Using a binocular depth camera, the spatial pose of the mechanical gripper arm containing automotive electronic components is acquired synchronously to obtain the initial physical gripping pose. The initial physical gripping pose is then input into a spatial homogeneous transformation matrix and mapped to a global unified world coordinate system to generate normalized physical position coding features.

[0102] In this embodiment, a binocular depth camera is used to simultaneously acquire the spatial pose of the mechanical gripper arm where the automotive electronic components are located, to obtain the initial physical gripping pose, including:

[0103] The initial physical grasping pose is obtained by simultaneously acquiring the spatial pose of the mechanical gripper arm where the automotive electronic components are located using a binocular depth camera.

[0104] In Example 1, the spatial position components and attitude angle components of the mechanical gripper arm clamping the automotive electronic components are acquired at the same acquisition time using a binocular depth camera. The components are then vectorized and combined in the order of spatial position first and attitude angle last to obtain the initial physical gripping pose in the camera coordinate system.

[0105] Based on the installation and calibration data of the binocular depth camera and the correspondence between the coordinates of the calibration reference point, a spatial homogeneous transformation matrix from the binocular depth camera coordinate system to the global unified world coordinate system is constructed.

[0106] In Example 1, multiple calibration reference points are set up at the automotive electronic component transfer and acquisition station. The camera coordinate position of each calibration reference point in the binocular depth camera coordinate system and the world coordinate position of the corresponding calibration reference point in the global unified world coordinate system are obtained to establish the coordinate correspondence of the same point.

[0107] By performing rigid body registration on the corresponding coordinate relationships of the same points, the rotation and translation mapping relationship of the binocular depth camera coordinate system relative to the world coordinate system is obtained. Based on the rotation and translation mapping relationship, and combined with homogeneous zero-filling terms and homogeneous scale terms according to the spatial homogeneous transformation structure, a spatial homogeneous transformation matrix from the binocular depth camera coordinate system to the global unified world coordinate system is constructed.

[0108] The initial physical grasp pose is input into the homogeneous transformation matrix of the space and mapped to the global unified world coordinate system to obtain the physical grasp pose in the world coordinate system.

[0109] In Example 1, the spatial position components in the initial physical grasping pose are converted into homogeneous position vectors, and the spatial homogeneous transformation matrix from the binocular depth camera coordinate system to the global unified world coordinate system is input to obtain the spatial position components of the mechanical grasping arm in the global unified world coordinate system.

[0110] By converting the attitude angular components in the initial physical grasp pose into the grasp pose rotation relationship in the binocular depth camera coordinate system, and using the rotation mapping relationship from the binocular depth camera coordinate system to the global unified world coordinate system, the grasp pose rotation relationship in the binocular depth camera coordinate system is mapped to obtain the grasp pose rotation relationship in the global unified world coordinate system.

[0111] By rotating in the ZYX axis according to the preset order, the roll attitude angle component, pitch attitude angle component and yaw attitude angle component are calculated and extracted from the gripping attitude rotation relationship in the global unified world coordinate system. These components are then vectorized and combined with the spatial position component in the global unified world coordinate system in the order of spatial position first and attitude angle second to obtain the physical gripping pose in the world coordinate system.

[0112] The physical grasping pose in the world coordinate system is normalized to obtain normalized spatial position components and normalized attitude components.

[0113] In Example 1, the position offset is obtained by subtracting the lower limit of the world coordinates of the corresponding coordinate axis from each spatial position component, and then divided by the range of the world coordinates of the corresponding coordinate axis after adding the position normalization stabilization value to generate the normalized spatial position component.

[0114] Add the attitude angle offset reference value to each attitude angle component to obtain the attitude angle offset, and divide it by the attitude angle full-cycle scale value that has been added to the attitude normalization stable value to generate the normalized attitude component.

[0115] Following a fixed order of the X-axis, Y-axis, Z-axis, roll attitude angle, pitch attitude angle, and yaw attitude angle in the global unified world coordinate system, the normalized spatial position components and normalized attitude components are vectorized and encoded to generate normalized physical position coding features.

[0116] By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, the time-series signal strength index stream within a continuous time window is captured, generating a time-series RF sensing dataset. For the time-series RF sensing dataset, a sliding time window is used to perform short-time energy averaging and Gaussian kernel density estimation, filtering out RF noise caused by multipath effects, and generating a confidence-weighted RF feature vector.

[0117] In this embodiment, the UHF RFID active tags attached to automotive electronic components are periodically scanned by IoT warehouse edge detection nodes to capture the time-series signal strength index stream within a continuous time window, including:

[0118] By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, the temporal signal strength index stream within a continuous time window is captured.

[0119] In Example 1, the IoT warehouse edge detection node continuously scans the UHF RFID active tags attached to automotive electronic components according to a preset radio frequency scanning cycle, collects multiple radio frequency sampling times corresponding to the same component identity ID and the signal strength index value corresponding to each radio frequency sampling time, and continuously records them according to the radio frequency sampling times to form a time-series signal strength index stream.

[0120] A time-series radio frequency sensing dataset is generated based on the time-series signal strength index stream.

[0121] In Example 1, multiple radio frequency sampling times corresponding to the same component ID and the signal strength index value corresponding to each radio frequency sampling time are bound together and arranged in order from first to last radio frequency sampling times to generate a time-series radio frequency sensing dataset.

[0122] A sliding time window is constructed for the time-series radio frequency sensing dataset to obtain multiple window radio frequency sensing sub-sequences;

[0123] In Example 1, according to the preset window length and preset window sliding step size, continuous signal strength index values ​​are sequentially extracted from the time-series radio frequency sensing dataset along the direction from the first to the last radio frequency sampling time to obtain multiple window radio frequency sensing subsequences, wherein each window radio frequency sensing subsequence is composed of continuous signal strength index values ​​within the same sliding time window.

[0124] The preset window length is obtained by statistically analyzing the minimum number of consecutive sampling points required for a stable radio frequency reading to be generated within a continuous time window for the same component ID.

[0125] By subtracting the number of overlapping sampling points from the preset window length according to the principle of maintaining continuous overlapping coverage between adjacent sliding time windows, the preset window sliding step length is obtained.

[0126] Perform short-time energy averaging on each window radio frequency sensing subsequence to obtain the window short-time energy average;

[0127] In Example 1, for each window radio frequency sensing subsequence, the signal strength index value in each window radio frequency sensing subsequence is squared to obtain the squared signal strength value. All squared signal strength values ​​are summed to obtain the total window signal strength energy value. The total window signal strength energy value is divided by the number of signal strength index values ​​in the window radio frequency sensing subsequence to obtain the short-time energy average value of the window.

[0128] Gaussian kernel density estimation is performed on each window radio frequency sensing subsequence to obtain the window radio frequency density estimate;

[0129] In Example 1, for each window RF sensing subsequence, the mean of all signal strength index values ​​in the window RF sensing subsequence is calculated to obtain the mean signal strength of the window; the degree of difference between each signal strength index value in the window RF sensing subsequence and the mean signal strength of the window is calculated, and the degree of difference is mapped by a Gaussian kernel based on the Gaussian kernel bandwidth to obtain multiple signal strength kernel density contribution values; the average of all signal strength kernel density contribution values ​​in the window RF sensing subsequence is performed to obtain the estimated value of the window RF density.

[0130] For each window RF sensing subsequence, the mean of all signal strength index values ​​in the window RF sensing subsequence is calculated to obtain the mean signal strength of the window. The degree of difference between each signal strength index value in the window RF sensing subsequence and the mean signal strength of the window is calculated, and Gaussian kernel mapping is performed on the degree of difference based on the Gaussian kernel bandwidth to obtain multiple signal strength kernel density contribution values. The average of all signal strength kernel density contribution values ​​in the window RF sensing subsequence is performed to obtain the estimated value of the window RF density.

[0131] Window RF confidence weights are generated based on the estimated window RF density.

[0132] In Example 1, for each window radio frequency sensing subsequence, the window radio frequency density estimate corresponding to the window radio frequency sensing subsequence is normalized and compared with the maximum value among the window radio frequency density estimates corresponding to all window radio frequency sensing subsequences to obtain a normalized radio frequency density estimate.

[0133] The normalized radio frequency density estimate is compared with the multipath noise density threshold. Window radio frequency sensing subsequences below the multipath noise density threshold are identified as low-confidence radio frequency sensing windows affected by multipath effects, while window radio frequency sensing subsequences above or equal to the multipath noise density threshold are identified as stable radio frequency sensing windows. Window radio frequency confidence weights are generated based on the comparison results, such that the window radio frequency confidence weights corresponding to low-confidence radio frequency sensing windows are less than those corresponding to stable radio frequency sensing windows.

[0134] By collecting normalized radio frequency density estimates corresponding to low-confidence radio frequency sensing windows and stable radio frequency sensing windows affected by multipath effects at the automotive electronic component transfer station, and taking the median value between the upper quartile level of the normalized radio frequency density estimate corresponding to the low-confidence radio frequency sensing window and the lower quartile level of the normalized radio frequency density estimate corresponding to the stable radio frequency sensing window, the multipath noise density threshold is obtained. The value range is 0.45 to 0.75. The selection criteria are that it is higher than the upper bound of discrete radio frequency density fluctuation when affected by metal shelf reflection, obstruction and multipath superposition and lower than the lower bound of continuous radio frequency density under stable reading state.

[0135] A confidence-weighted radio frequency feature vector is generated based on the window's short-time energy average and the window's radio frequency confidence weight.

[0136] In Example 1, for each window radio frequency sensing subsequence, the average short-time energy of the window corresponding to the window radio frequency sensing subsequence is multiplied by the window radio frequency confidence weight to obtain the confidence weighted radio frequency energy component; according to the sliding time window number in ascending order, the confidence weighted radio frequency energy components corresponding to all window radio frequency sensing subsequences are vectorized to generate the confidence weighted radio frequency feature vector.

[0137] The illumination-calibrated marking image is input into a pre-trained marking code defect detection network to delineate the geometric topological boundary of the marking characters and extract the character distortion components to obtain the marking code label verification features.

[0138] In this embodiment, the illumination-calibrated marking image is input into a pre-trained marking code defect detection network to delineate the geometric topological boundaries of the marking characters and extract the character distortion components, including:

[0139] The illumination calibration and marking image is input into the training anisotropic heterogeneous multi-branch network to obtain the horizontal stroke response feature map, the vertical stroke response feature map and the local texture response feature map, and generate a set of multi-branch heterogeneous feature matrices.

[0140] In Example 1, during the forward feature extraction process, zero-padding of spatial boundaries matching the corresponding convolution kernel size is performed on the 1x3 horizontal slice convolution branch and the 3x1 vertical slice convolution branch respectively. The local defect texture response, horizontal stroke continuity response and vertical stroke continuity response feature maps of the tagged characters are extracted by each branch respectively, and matrix splicing and arrangement are performed according to the channel dimension to generate a multi-branch heterogeneous feature matrix set.

[0141] Perform channel normalization and spatial alignment on the set of multi-branch heterogeneous feature matrices to generate channel normalized feature tensor parameters;

[0142] Based on the channel-normalized feature tensor parameters, the convolution parameters of each branch are equivalently transformed to generate equivalent transformation weights and equivalent translation biases with aligned dimensions.

[0143] In Example 1, backpropagation training is performed using illumination to calibrate the marking image and its marking code defect annotations to obtain convolution weights, with values ​​ranging from -1 to 1. The selection criteria are to enable the convolution response to distinguish marking character edges, scratches, and stroke defects.

[0144] By learning to scale the normalized branch feature response amplitude during training, a batch normalization scaling factor is obtained, with a value ranging from 0.1 to 3. The selection criterion is to avoid excessive suppression or amplification of the channel response.

[0145] The convolution bias is obtained by learning to shift the baseline response of the convolution output during training.

[0146] By performing a learnable translation on the normalized branch feature response center during the training process, batch normalized translation coefficients are obtained, with values ​​ranging from -2 to 2. The selection criteria are to ensure that the local defect texture response, horizontal stroke continuity response, and vertical stroke continuity response are at a unified feature center scale.

[0147] The convolution weights of each branch convolutional layer are fused with the corresponding batch normalized scaling factor to obtain the equivalent transformation weights. The convolution biases of each branch convolutional layer are fused with the corresponding batch normalized translation factor to obtain the equivalent translation biases.

[0148] The equivalent transformation weights with dimensional alignment and the equivalent translation bias are structurally reparameterized and fused to generate a single-branch equivalent 3x3 main convolution kernel parameter set;

[0149] In Example 1, the equivalent transformation weights corresponding to the 1x3 horizontal slice convolution are zero-padded and extended to the 3x3 convolution kernel space, and the equivalent transformation weights corresponding to the 3x1 vertical slice convolution are zero-padded and extended to the 3x3 convolution kernel space. They are then superimposed with the equivalent transformation weights corresponding to the 3x3 multi-path concurrent convolution in the same spatial position to obtain the single-branch equivalent 3x3 main convolution kernel weights.

[0150] The equivalent translation biases corresponding to the 3x3 multi-channel concurrent convolution, the 1x3 horizontal slice convolution, and the 3x1 vertical slice convolution are added channel by channel according to the output channel to obtain the single-branch equivalent 3x3 main convolution kernel bias.

[0151] The single-branch equivalent 3x3 main convolution kernel weights and single-branch equivalent 3x3 main convolution kernel biases are combined to generate the single-branch equivalent 3x3 main convolution kernel parameter set.

[0152] Based on the parameter set of the single-branch equivalent 3x3 main convolution kernel, forward inference is performed on the illumination calibration and marking image to generate an anisotropic marking candidate response field.

[0153] In Example 1, the illumination calibration marking image is input into an inference network consisting of a single-branch equivalent 3x3 main convolution kernel parameter set. For each marked character pixel region, a local image block centered on the current two-dimensional pixel coordinates is extracted. The local image block is then fused and convolved once using the single-branch equivalent 3x3 main convolution kernel parameter set to obtain the fused abnormal response value of the marked character.

[0154] When the abnormal response value of the fused tagged character is higher than the preset tagged candidate response threshold, the corresponding tagged character pixel area is determined as the tagged candidate response area. Spatial location mapping is performed on all tagged candidate response areas to generate an anisotropic tagged candidate response field.

[0155] The preset marking candidate response threshold is obtained by acquiring the median value between the upper quartile level of the fused marking character abnormal response value and the lower quartile level of the defective marking area.

[0156] Local variational detection is performed on the anisotropic marking candidate response field along the topological skeleton of the marked character to generate a stroke distortion feature differential field;

[0157] In Example 1, the topological skeleton of the marked character is extracted based on the main connected stroke region of the marked character in the anisotropic marked candidate response field, and a local sliding window is constructed with each skeleton sampling point on the topological skeleton of the marked character as the center.

[0158] For each local sliding window, extract all candidate response values ​​within the coverage area of ​​the local sliding window, and subtract the minimum response value from the maximum response value to obtain the response value change range.

[0159] For each local sliding window, the average of all candidate response values ​​within the coverage area of ​​the local sliding window is first calculated to obtain the local average response value. Then, the local average response value is subtracted from each candidate response value within the local sliding window to obtain the response deviation value of the corresponding candidate response value. All response deviation values ​​are squared and averaged to obtain the local variance of the response value.

[0160] For each local sliding window, the skeleton sampling points within the local sliding window are arranged sequentially according to the extension direction of the topological skeleton of the labeled character. The difference between candidate response values ​​between adjacent skeleton sampling points is calculated to obtain the first-order response gradient. Then, the gradient difference between adjacent first-order response gradients is calculated to obtain the trend of response gradient change.

[0161] The response value variation amplitude, local variance of the response value, and trend of response gradient variation are jointly characterized to obtain character distortion components. Then, the character distortion components corresponding to each local sliding window are mapped back to the anisotropic labeling candidate response field according to their spatial position in the topological skeleton of the labeled character, generating a stroke distortion feature differential field.

[0162] Adaptive threshold segmentation and connected component filtering are performed on the differential field of stroke distortion features to generate a damage and defect mask for component marking codes.

[0163] In Example 1, gray-level statistics are performed on the distortion response values ​​corresponding to all pixel positions in the differential field of stroke distortion features to obtain a gray-level histogram of distortion response values. Based on the Otsu algorithm, candidate segmentation gray levels are traversed in the gray-level histogram of distortion response values, and the candidate segmentation gray level corresponding to the largest inter-class variance is determined as the adaptive segmentation threshold.

[0164] Pixel regions with areas above the adaptive segmentation threshold are identified as candidate defect regions. Two-dimensional connected component filtering is performed on the candidate defect regions, and connected regions with areas smaller than the preset isolated noise area threshold are removed to obtain the component marking code damage defect mask.

[0165] Based on the damage and defect mask of the marking code and the marking character area of ​​the parts, a set of characteristic indicators for the quality compliance of the parts is constructed.

[0166] In Example 1, the number of defect pixels that fall within the marking character area of ​​the component marking code damage defect mask is counted, and the ratio of this number to the total number of character pixels in the marking character area is used to calculate the defect area ratio.

[0167] For defect pixels in the marking character area that fall within the mask of the part marking code damage defect, perform connected component marking, determine each independent defect region formed by connecting adjacent defect pixels as a defect connected component, count the total number of defect connected components, and obtain the number of defect connected components.

[0168] Extract the outer contour boundary of each defect connected domain, and count the total number of boundary points whose change amplitude along the boundary direction exceeds the preset boundary curvature change threshold to obtain the defect boundary curvature change frequency.

[0169] The defect mask of the marking code on the component is spatially overlapped with the topological skeleton of the marking character to obtain the defect skeleton segment. The ratio of the cumulative length of the defect skeleton segment to the total length of the topological skeleton of the marking character is calculated to obtain the stroke defect length ratio.

[0170] The proportion of defect area, the number of defect connected domains, the frequency of defect boundary curvature abrupt changes, and the proportion of stroke defect length are vectorized and concatenated in a fixed order to construct a set of component quality compliance characteristic indicators.

[0171] Based on the component quality compliance feature index set and the component marking code damage defect mask, mark code label verification features are generated.

[0172] In Example 1, the component quality compliance characteristic index set is arranged and normalized in a fixed order of defect area ratio, number of defect connected domains, frequency of defect boundary curvature abrupt change, and proportion of stroke defect length to generate a quality compliance numerical vector.

[0173] By extracting the defect pixel positions, defect connected domain boundaries, defect connected domain center points, and defect connected domain bounding rectangle dimensions from the defect mask of the part marking code damage, defect spatial description data is constructed. Then, the defect spatial description data is vectorized and encoded according to the spatial arrangement order of the marking character areas to obtain the defect spatial distribution vector.

[0174] By concatenating the quality compliance numerical vector and the defect spatial distribution vector at the feature level, with quality indicators first and spatial distribution second, a fused feature vector for marking defects is obtained.

[0175] The dimension of the marking defect fusion feature vector is compressed by fully connected dimensionality reduction mapping, retaining the effective feature components that represent the marking boundary contour, stroke loss rate and character distortion components, and generating marking code label verification features.

[0176] Normalized physical location coding features and confidence-weighted radio frequency feature vectors are concatenated at the feature level, and a strong spatiotemporal correlation is established based on the component identity ID. A spatiotemporal feature tensor of warehouse entry and exit events is constructed. The on-chain flow trajectory path in the distributed local ledger of each warehouse network is retrieved. The multi-node spatiotemporal state convergence of the spatiotemporal feature tensor of warehouse entry and exit events is calculated by the least squares optimal alignment algorithm, and a multi-node logistics trajectory matching sequence is generated.

[0177] In this embodiment, the normalized physical location coding features and the confidence-weighted radio frequency feature vector are concatenated at the feature level, and a strong spatiotemporal association is established based on the component identity ID, including:

[0178] The normalized physical location coding features and the confidence-weighted radio frequency feature vector are concatenated at the feature level to obtain the fusion perception feature vector for parts entering and leaving the warehouse.

[0179] Establish a strong spatiotemporal correlation of component entry and exit fusion perception feature vectors based on component identity ID, and construct a spatiotemporal feature tensor for warehouse entry and exit events;

[0180] In Example 1, the component entry and exit fusion perception feature vector corresponding to the same component identity ID, the time of occurrence of the warehouse entry and exit event, and the warehouse outlet identifier are bound together to form a single warehouse entry and exit event feature unit.

[0181] Based on the order of the occurrence of warehouse entry and exit events, the feature units of all single warehouse entry and exit events corresponding to the same component ID are arranged in chronological order to construct the spatiotemporal feature tensor of warehouse entry and exit events.

[0182] Retrieve the on-chain flow path from the distributed local ledger of each warehouse outlet;

[0183] In Example 1, based on the component identity ID, the on-chain transfer record corresponding to the same component identity ID is read from the distributed local ledger of each warehouse network.

[0184] Arrange the on-chain circulation records corresponding to the same component identity ID in chronological order according to the on-chain record timestamps to obtain the on-chain circulation trajectory path. Each on-chain circulation record includes the component identity ID, on-chain storage network identifier, on-chain record timestamp, and on-chain hash identifier.

[0185] Based on the spatiotemporal feature tensor of warehouse inbound and outbound events and the on-chain flow trajectory path, multi-node spatiotemporal aligned sample pairs are constructed.

[0186] In Example 1, for each single warehouse entry / exit event feature unit in the spatiotemporal feature tensor of the warehouse entry / exit event, in the on-chain flow trajectory path corresponding to the same component identity ID, the on-chain flow record with the smallest absolute difference between the on-chain record timestamp and the time when the warehouse entry / exit event occurs is retrieved, and the on-chain flow record is determined as a candidate on-chain flow record.

[0187] Align and bind the feature units of a single warehouse entry / exit event with the flow records on the candidate chain to construct multi-node spatiotemporal aligned sample pairs. The alignment and binding does not require the warehouse outlet identification to be consistent as a prerequisite.

[0188] Calculate the spatiotemporal alignment residual based on multi-node spatiotemporal aligned sample pairs;

[0189] In Example 1, for each multi-node spatiotemporal alignment sample pair, the time difference between the time of the warehouse entry / exit event and the timestamp of the record on the chain is used as the time alignment residual.

[0190] The node consistency difference between the warehouse outlet identifier corresponding to the warehouse inbound / outbound event and the on-chain warehouse outlet identifier corresponding to the on-chain transfer record is used as the node alignment residual.

[0191] Using the global time offset and candidate path matching relationship as the alignment objects to be optimized, the joint least squares solution is performed on the corresponding time alignment residuals and node alignment residuals of all multi-node spatiotemporal candidate alignment samples to minimize the sum of the squares of all time alignment residuals and the weighted sum of all node alignment residuals, thus obtaining the spatiotemporal alignment residuals.

[0192] Based on the spatiotemporal alignment residual, calculate the spatiotemporal state convergence of multiple nodes;

[0193] In Example 1, the spatiotemporal alignment residuals are subjected to residual scale normalization to obtain normalized spatiotemporal alignment residuals.

[0194] The spatiotemporal state convergence of multiple nodes is obtained by dividing one by the sum of one and the normalized spatiotemporal alignment residual. The smaller the spatiotemporal alignment residual between the spatiotemporal feature tensor of warehouse entry and exit events and the on-chain flow trajectory path, the higher the spatiotemporal state convergence of multiple nodes. Conversely, the larger the spatiotemporal alignment residual between the spatiotemporal feature tensor of warehouse entry and exit events and the on-chain flow trajectory path, the lower the spatiotemporal state convergence of multiple nodes.

[0195] Based on the convergence of the spatiotemporal states of multiple nodes and the on-chain flow trajectory path, a multi-node logistics trajectory matching sequence is generated.

[0196] In Example 1, the spatiotemporal state convergence of the multi-node is compared with the preset trajectory matching convergence threshold. When the spatiotemporal state convergence of the multi-node is greater than or equal to the preset trajectory matching convergence threshold, the warehouse network identifier, the time of occurrence of the warehouse entry and exit event, and the fusion perception feature vector of the parts entry and exit in the spatiotemporal feature tensor of the corresponding warehouse entry and exit event are arranged in the order of the on-chain record timestamps in the on-chain flow trajectory path to generate a multi-node logistics trajectory matching sequence.

[0197] Specifically, by collecting the spatiotemporal convergence of multi-node states corresponding to historically confirmed normal flow trajectory samples and abnormal mismatch trajectory samples, and taking the median value between the upper quartile level of the spatiotemporal convergence of multi-node states corresponding to abnormal mismatch trajectory samples and the lower quartile level of the spatiotemporal convergence of multi-node states corresponding to normal flow trajectory samples, a preset trajectory matching convergence threshold is obtained. The indicated value is 0.85, and the value range is from 0.75 to 0.95. The selection criteria are that it is higher than the upper bound of the convergence formed by abnormal mismatch trajectories under the conditions of time offset, network point mismatch, and missing on-chain records, and lower than the lower bound of the convergence formed by normal flow trajectories under the condition of continuous matching of multiple warehouse network points.

[0198] Obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger, perform asymmetric encrypted digital signature verification based on the label verification features, complete the asset authenticity verification and output the asset anti-counterfeiting verification label;

[0199] In this embodiment, the process involves obtaining pre-stored supply chain origin factory certificates from the blockchain distributed ledger and performing asymmetric encrypted digital signature verification based on the tag verification features, including:

[0200] Based on the component identity ID, obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger and generate source factory certificate parsing data;

[0201] Generate a current label verification summary based on the label verification features;

[0202] Asymmetric encrypted digital signature verification is performed based on the source factory-existing evidence parsing data to obtain the source signature verification result;

[0203] In Example 1, the source digital signature and the factory public key certificate are extracted from the source factory-issued evidence parsing data. The factory public key certificate is parsed to obtain the factory verification public key. The source digital signature is decrypted and verified using the factory verification public key to restore the core plaintext on the source chain. The text digest in the core plaintext on the source chain is compared with the source-tagged text digest to obtain the source verification result.

[0204] If the text digest in the core plaintext on the source chain is completely consistent with the source-tagged text digest, the source signature verification result is obtained; otherwise, the source signature verification result is output.

[0205] Based on the current marking verification summary and the source factory certificate analysis data, the marking consistency verification result is obtained;

[0206] In Example 1, the source marking text digest is extracted from the source factory certificate parsing data, and the current marking verification digest is compared with the source marking text digest for lossless equality to obtain the marking consistency verification result.

[0207] When the current tagging verification summary is completely consistent with the source tagging text summary, the tagging consistency result is obtained; when they are inconsistent, the tagging consistency result is obtained.

[0208] Based on the source verification results and the labeling consistency verification results, the authenticity of the assets is verified and the asset anti-counterfeiting verification label is output.

[0209] In Example 1, when the source verification result is a successful source verification and the marking consistency verification result is a successful marking consistency verification, the automotive electronic components are identified as genuine assets; when the source verification result is a failed source verification or the marking consistency verification result is a failed marking consistency verification, the automotive electronic components are identified as suspected abnormal assets.

[0210] When automotive electronic components are identified as genuine assets, a genuine anti-counterfeiting verification label is output; when automotive electronic components are identified as suspected abnormal assets, an abnormal anti-counterfeiting verification label is output.

[0211] Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a supply chain trust score distribution is constructed. The maximum flow allocation path of the entire life cycle is extracted according to the flow network flow balance rule, and a supply chain traceability closed-loop management instruction containing spatial grid location, time slice status, flow legality proof and tamper-proof hash features is generated.

[0212] In this embodiment, a supply chain trust score distribution is constructed based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, including:

[0213] Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a set of supply chain circulation nodes is constructed;

[0214] Based on the set of supply chain flow nodes, spatial grid locations and time slice states are divided;

[0215] In Example 1, the warehouse network identifiers in the supply chain flow node set are mapped to the spatial grid positions of the corresponding warehouse network points, and the times of warehouse entry and exit events are mapped to the corresponding time slice states, so as to obtain the spatial grid position and time slice state corresponding to each supply chain flow node.

[0216] Based on the asset anti-counterfeiting verification label, the set of supply chain circulation nodes is assigned anti-counterfeiting credibility weights to obtain anti-counterfeiting credibility weights.

[0217] In Example 1, when the asset anti-counterfeiting verification label is a genuine product anti-counterfeiting verification label, the corresponding supply chain circulation node is assigned a genuine product credibility weight, which is then used as the anti-counterfeiting credibility weight corresponding to the supply chain circulation node.

[0218] When the asset anti-counterfeiting verification label is an abnormal anti-counterfeiting verification label, the corresponding supply chain circulation node will be assigned an abnormal credibility weight, which will be used as the anti-counterfeiting credibility weight corresponding to the supply chain circulation node, and the credibility weight of genuine products will be greater than the credibility weight of abnormal products.

[0219] Based on the multi-node logistics trajectory matching sequence, the trajectory matching credibility of the supply chain flow node set is assigned, and a supply chain trust score distribution is constructed.

[0220] Based on the supply chain trust score distribution, construct a full life cycle circulation network;

[0221] Based on the full lifecycle flow network, the maximum flow allocation path for the entire lifecycle flow is extracted according to the flow network traffic balancing rules;

[0222] The flow network flow balancing rule is as follows: In the flow network throughout the entire life cycle, the factory exit node at the source of the supply chain is determined as the source point, and the current warehouse inbound / outbound verification node is determined as the sink point. A flow conservation balance constraint is applied to each intermediate flow network node, where the total inflow is equal to the total outflow. A boundary capacity constraint is applied to each flow network edge, where the actual allocated flow is not greater than the corresponding flow capacity. The maximum feasible flow configuration state is calculated, and the single core backbone flow path with the highest flow contribution is extracted from the maximum feasible flow configuration state through reverse approximation, which is taken as the maximum flow allocation path for the entire life cycle flow.

[0223] Based on the maximum flow allocation path throughout the entire lifecycle, generate closed-loop management instructions for supply chain traceability.

[0224] Example 2: Reference Figure 2This embodiment uses the warehousing, in-warehouse transfer, verification and outbound processing, and on-chain traceability of automotive electronic control modules as application scenarios. The components have metal casings with DPM codes formed on their surfaces through laser etching, and are accompanied by UHF RFID active tags. In the warehouse, the components need to pass through the inbound inspection station, buffer shelf area, verification inspection station, and outbound transfer port. Because the metal casing is prone to mirror reflection, and the DPM code area has minor scratches, missing strokes, and worn character edges, and because RFID signals are easily obstructed, subjected to multipath reflections, and interfered with by nearby tags near metal shelves, traditional methods of single-scan, single-RFID reading, and manual path verification are insufficient to reliably verify component identity and supply chain authenticity.

[0225] A batch of 12,000 automotive electronic components was processed in the warehouse. Of these, 10,000 were normal components, 900 had minor surface scratches, 600 had missing DPM code strokes, 300 had strong metallic reflective interference, and 200 had abnormal movement trajectories. The system deployed high-definition vision sensors, binocular depth cameras, IoT warehouse edge detection nodes, and a blockchain distributed ledger access module within the warehouse inspection environment to perform visual marking verification, spatial pose acquisition, continuous RFID sensing, on-chain certificate storage and signature verification, and closed-loop supply chain traceability.

[0226] During a specific continuous operation period, an automotive electronic control module with the component identification ID "ECM-4F28-0736" enters the warehousing and inspection station. The production line's high-definition vision sensor performs fixed-focus imaging on the laser-etched area on the metal casing surface, obtaining a raw DPM code image with a resolution of 2448×2048. The system locates the 512×512 pixel DPM code marking acquisition area, detecting an average grayscale value of 138.6, a local maximum grayscale value of 248, and a local minimum grayscale value of 31. Approximately 7.8% of the pixels have a grayscale value higher than 230. Traditional grayscale thresholding methods misjudge the bright reflective spot on the upper right side as a character defect area and output a suspected marking anomaly.

[0227] The method in this embodiment performs grayscale matrixing on the original DPM code image to form the original marking image data. The production line's high-definition vision sensor has a dark-field grayscale reference value of 12 and a bright-field grayscale reference value of 247 in the original marking image acquisition scenario. The minimum resolvable grayscale change is 0.004, and the preset lower limit stability value is 0.006. Therefore, the grayscale normalization stability coefficient is set to 0.006. After grayscale dimension normalization, the grayscale transition amplitude of the true edges of the DPM code characters is concentrated between 0.21 and 0.38, and the random grayscale difference amplitude corresponding to the metallic reflective noise is concentrated between 0.04 and 0.13. The system constructs a metallic surface mirror artifact reflection weight map using a 7×7 pixel window as the local neighborhood of the marking image. The average metallic surface mirror artifact reflection weight in the strongly reflective area is 0.82, the average metallic surface mirror artifact reflection weight in the normal character edge area is 0.18, and the average metallic surface mirror artifact reflection weight in the background area is 0.09. Nonlinear grayscale anisotropic diffusion filtering was performed, with a diffusion threshold of 0.16 for laser etching edge preservation, a diffusion iteration step size of 0.18, a preset diffusion iteration count of 12, and a preset diffusion convergence threshold of 0.003. After filtering, the standard deviation of grayscale fluctuation in strongly reflective areas decreased from 0.092 to 0.027, and the average gradient preservation rate of DPM code character edges reached 94.6%, resulting in an illumination-calibrated marking image.

[0228] The system inputs illumination calibration marking images into an anisotropic heterogeneous multi-branch network during training. Before deployment, 8000 sets of illumination calibration marking images are used for training, including 5600 sets of normal marking samples, 900 sets of surface scratch samples, 750 sets of stroke defect samples, 500 sets of character edge anomaly samples, and 250 sets of strong reflective interference samples. Training annotations include defect pixel locations, defect connected region boundaries, defect connected region center points, and the size of the bounding rectangle of the defect connected region. In the samples, the average response value of horizontal continuous regions of characters in the horizontal stroke response feature map is 0.76, and the average response value of abnormal broken regions is 0.31; the average response value of vertical continuous regions of characters in the vertical stroke response feature map is 0.79, and the average response value of abnormal broken regions is 0.34; the average response value of scratch candidate regions in the local texture response feature map is 0.68, and the average response value of normal background regions is 0.17. After channel normalization, the variance of the local defect texture response channel was adjusted from 0.184 to 0.996, the variance of the horizontal stroke continuity response channel was adjusted from 0.211 to 1.004, and the variance of the vertical stroke continuity response channel was adjusted from 0.206 to 0.998. After structural reparameterization fusion, the multi-branch structure during training was converted into a single-branch equivalent 3×3 main convolution kernel parameter set. The average inference time per image decreased from 31.8ms to 18.3ms, and the throughput of edge node processing increased from 31 images per second to 54 images per second.

[0229] In the DPM code area of ​​the component, the system detected a slight scratch approximately 11 pixels long and 2 pixels wide. The abnormal response value of the fused marking character reached 0.82 in the scratch area, averaged 0.29 in the normal character area, and averaged 0.11 in the background area. The preset marking candidate response threshold was determined by the median value between the upper quartile level of the abnormal response value in the qualified marking area (0.37) and the lower quartile level of the abnormal response value in the defective marking area (0.64), and was set to 0.505. The system identified areas with abnormal response values ​​higher than 0.505 as marking candidate response areas. After constructing an 11×11 pixel local sliding window along the topological skeleton of the marking character, the response value change amplitude of the window containing the scratch was 0.54, the local variance of the response value was 0.071, and the trend of the response gradient change was 0.38; the average response value change amplitude of the normal character stroke area was 0.18, the average local variance of the response value was 0.014, and the average trend of the response gradient change was 0.07. After Otsu adaptive threshold segmentation, the system determined the segmentation threshold to be 0.46, resulting in four candidate defect connected regions. Three of these, with areas smaller than 6 pixels, were discarded, leaving one connected region with an area of ​​23 pixels, which was identified as the true scratch defect region. The system calculated the defect area percentage (0.37%), the number of defect connected regions (1), the frequency of abrupt changes in defect boundary curvature (5 times), and the percentage of stroke loss length (2.57%), and generated a label verification feature.

[0230] The character recognition sequence of the component's marking code is "ECM4F280736". The system extracts the character recognition sequence from the marking code label verification features and generates the current marking verification digest. Based on the component's identity ID, it reads the supply chain source factory certificate from the blockchain distributed ledger, obtaining the source marking verification digest, source digital signature, and factory public key certificate. The system parses the factory public key certificate to obtain the factory verification public key, and uses the factory verification public key to decrypt and verify the source digital signature, restoring the core plaintext on the source chain. In the sample, the text digest in the core plaintext on the source chain is completely consistent with the source marking text digest, and the current marking verification digest is completely consistent with the source marking text digest. Therefore, the system determines the component to be a genuine asset and outputs a genuine anti-counterfeiting verification label. Another component with the ID "ECM-7C91-1184" had an RFID tag that could be read normally, but its DPM code had missing strokes, and one character in the character recognition sequence was misread as B instead of 8. The current tag verification summary was inconsistent with the source tag text summary. The system will determine it as a suspected abnormal asset and output an abnormal anti-counterfeiting verification tag.

[0231] During spatial pose acquisition, the binocular depth camera simultaneously acquires the spatial pose of the mechanical gripper arm with the component ID "ECM-4F28-0736". The position components in the camera coordinate system are: X-axis 0.426m, Y-axis 0.318m, Z-axis 0.742m; roll angle 1.8°; pitch angle -2.4°; yaw angle 89.6°. The system establishes a coordinate correspondence between corresponding points using six calibration reference points. After rigid body registration, the average reprojection error is 2.6mm, and the maximum reprojection error is 4.1mm. After spatial homogeneous transformation matrix mapping, the position components of the component in the global unified world coordinate system are: X-axis 3.842m, Y-axis 1.276m, Z-axis 0.914m; roll angle 2.1°; pitch angle -1.9°; yaw angle 90.4°. After normalizing the storage space range to 0m to 5m on the X-axis, 0m to 3m on the Y-axis, and 0m to 2m on the Z-axis, the system obtains normalized spatial position components of 0.768, 0.425, and 0.457, and normalized attitude components of 0.506, 0.495, and 0.751, and generates normalized physical location coding features.

[0232] During continuous RFID sensing, the IoT warehouse edge detection node continuously sampled the component 64 times. During the stable reading phase, the signal strength index value was concentrated between -48dBm and -54dBm; when passing the edge of the metal shelf, the signal strength index value fluctuated between -41dBm and -67dBm in some windows. The system constructed 15 window RF sensing sub-sequences with 8 sampling points of a preset window length and 4 sampling points of a preset window sliding step size. The 8 signal strength index values ​​of the 6th window were -49, -50, -51, -50, -52, -49, -50, and -51, with a short-time average energy of 2525.5, a normalized RF density estimate of 0.96, and a window RF confidence weight of 0.94, thus it was determined to be a stable RF sensing window. The eight signal strength indicators for the ninth window are -44, -61, -42, -67, -45, -63, -41, and -66, with a short-time average energy of 3032.6. However, the normalized RF density estimate is only 0.40, lower than the multipath noise density threshold of 0.58, and the RF confidence weight for the window is reduced to 0.21. Traditional single-scan RFID reading methods misclassify components as adjacent shelf areas within the window, thus reducing the impact of abnormal windows on the final feature.

[0233] The system concatenates the normalized physical location coding features and the confidence-weighted radio frequency feature vector to obtain the fusion sensing feature vector for parts entering and leaving the warehouse. Parts form four single-event warehouse entry / exit feature units at the inbound inspection station, buffer shelf area, verification inspection station, and outbound transfer port, respectively, and a spatiotemporal feature tensor for these events is constructed. After reading the on-chain flow trajectory path, the system retrieves the on-chain flow record with the smallest absolute difference between the timestamp of the record and the time of the warehouse entry / exit event, constructing a multi-node spatiotemporal aligned sample pair. The four time alignment residuals of the normal sample are 1.2s, 1.9s, 1.4s, and 2.1s, respectively, with zero node consistency differences. The normalized spatiotemporal alignment residual is 0.061, and the multi-node spatiotemporal state convergence is 0.943, higher than the preset trajectory matching convergence threshold of 0.85, generating a multi-node logistics trajectory matching sequence. The abnormal sample "ECM-2B64-3095" was read by an RFID tag at the cache shelf node, but the corresponding node record in the chain flow trajectory path was missing. In addition, the warehouse network identification between the verification detection node and the cache shelf node was not continuous. Its normalized spatiotemporal alignment residual was 0.684, and the multi-node spatiotemporal state convergence was 0.594, which was lower than the preset trajectory matching convergence threshold. Therefore, it was not judged as normal flow.

[0234] The system constructs a set of supply chain circulation nodes based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences. For "ECM-4F28-0736", each supply chain circulation node is assigned a genuine product credibility weight of 0.98. The cosine similarity of the component inbound / outbound fusion perception feature vectors between adjacent matching nodes are 0.974, 0.961, and 0.982, respectively, all higher than the consistency threshold of 0.92. The trajectory matching credibility weight is set to 0.96. The preset anti-counterfeiting fusion weight is 0.6, and the preset trajectory fusion weight is 0.4. The system calculates a supply chain circulation node trust score of approximately 0.972. The system treats supply chain flow nodes as flow network nodes. Nodes with valid flow relationships between adjacent time-slice states, related to the verification of inbound and outbound operations for automotive electronic component warehousing and the sequence of supply chain traceability methods, are connected as flow network edges. The flow capacity from the source factory node to the inbound inspection node is 0.971, from the inbound inspection node to the cache shelf node is 0.971, from the cache shelf node to the verification inspection node is 0.968, and from the verification inspection node to the outbound transfer node is 0.970. The system extracts a single core backbone flow path using maximum flow solving and flow decomposition topology stripping algorithms, and generates a closed-loop supply chain traceability management instruction containing component identity ID, spatial grid location, time-slice state, proof of flow validity, and tamper-proof hash characteristics.

[0235] In a simulated batch verification of 12,000 parts, the traditional method achieved an accuracy of 94.1% in recognizing normal DPM codes, while the method in this embodiment achieved 99.0%; the traditional method achieved an accuracy of 82.7% in recognizing samples with strong reflective interference, while the method in this embodiment achieved 97.8%; the traditional method achieved a recall rate of 84.4% for surface scratch samples, while the method in this embodiment achieved 96.7%; the traditional method achieved a recall rate of 81.3% for stroke-defect samples, while the method in this embodiment achieved 96.0%; the traditional method had an average positioning deviation of 6.8 pixels for defect areas, while the method in this embodiment achieved 1.9 pixels; the traditional method had an average spatial position error of 18.6 mm, while the method in this embodiment achieved 4.2 mm; the traditional method had a filtering rate of 72.6% for RFID multipath interference, while the method in this embodiment achieved 95.8%; the traditional method had an abnormal mismatch trajectory recognition rate of 78.8%, while the method in this embodiment achieved 96.9%; and the traditional method took an average of 3.8 seconds to complete the complete traceability of a single part, while the method in this embodiment took 0.9 seconds. In the entire simulated batch, the traditional method produced 754 misclassified samples, while the method in this embodiment reduced the number of misclassified samples to 118, and the overall misclassification rate decreased from 6.28% to 0.98%.

[0236] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for verifying marking codes and tracing the supply chain of automotive electronic components, characterized in that, include: Raw marking image data is collected during the circulation of automotive electronic components and preprocessed to obtain illumination-calibrated marking images; Using a binocular depth camera, the spatial pose of the mechanical gripper arm containing automotive electronic components is acquired simultaneously to obtain the initial physical gripping pose and generate normalized physical position coding features. By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, a time-series radio frequency sensing dataset is generated and preprocessed to generate a confidence-weighted radio frequency feature vector. The illumination-calibrated marking image is input into a pre-trained marking code defect detection network to delineate the geometric topological boundary of the marking characters and extract the character distortion components to obtain the marking code label verification features. Normalized physical location coding features and confidence-weighted radio frequency feature vectors are concatenated at the feature level, and strong spatiotemporal correlation is established based on component identity IDs to construct a spatiotemporal feature tensor for warehouse entry and exit events. The on-chain flow trajectory path in the distributed local ledger of each warehouse network is retrieved, the multi-node spatiotemporal state convergence of the spatiotemporal feature tensor for warehouse entry and exit events is calculated, and a multi-node logistics trajectory matching sequence is generated. Obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger, perform asymmetric encrypted digital signature verification based on the label verification features, complete the asset authenticity verification and output the asset anti-counterfeiting verification label; Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a supply chain trust score distribution is constructed, the maximum flow allocation path of the entire life cycle is extracted, and a supply chain traceability closed-loop management instruction is generated.

2. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, During the flow of automotive electronic components, the original DPM code images of the laser-etched surfaces of the components to be transferred are acquired using high-definition vision sensors on the production line, including: When automotive electronic components enter the transfer and acquisition station, fixed-focus imaging is performed to acquire the original DPM code image and form the original marking image data. The original marking image data is normalized in terms of gray dimensions to obtain a normalized original marking grayscale image. Based on the normalized original labeled grayscale image, a weighted image of the mirror pseudo-effect of the metal surface is constructed. Using the normalized original labeled grayscale image as the initial input for nonlinear grayscale anisotropic diffusion filtering, the directional grayscale difference is calculated to generate nonlinear diffusion control coefficients. Based on the directional gray-level difference, nonlinear diffusion control coefficient and metal surface mirror pseudo-effect weight map, nonlinear gray-level anisotropic diffusion filtering is performed on the normalized original marking gray-level image to obtain the filtered marking gray-level image. Perform grayscale boundary truncation on the filtered grayscale image to obtain the illumination-calibrated grayscale image.

3. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, The process of simultaneously acquiring the spatial pose of the mechanical gripper arm containing automotive electronic components using a binocular depth camera to obtain the initial physical gripping pose includes: The initial physical grasping pose is obtained by simultaneously acquiring the spatial pose of the mechanical gripper arm where the automotive electronic components are located using a binocular depth camera. Based on the installation and calibration data of the binocular depth camera and the correspondence between the coordinates of the calibration reference point, a spatial homogeneous transformation matrix from the binocular depth camera coordinate system to the global unified world coordinate system is constructed. The initial physical grasp pose is input into the homogeneous transformation matrix of the space and mapped to the global unified world coordinate system to obtain the physical grasp pose in the world coordinate system. The physical grasping pose in the world coordinate system is normalized to obtain normalized spatial position components and normalized attitude components. Following a fixed order of the X-axis, Y-axis, Z-axis, roll attitude angle, pitch attitude angle, and yaw attitude angle in the global unified world coordinate system, the normalized spatial position components and normalized attitude components are vectorized and encoded to generate normalized physical position coding features.

4. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, The method of periodically scanning UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes to capture the time-series signal strength index stream within a continuous time window includes: By periodically scanning the UHF RFID active tags attached to automotive electronic components through IoT warehouse edge detection nodes, the temporal signal strength index stream within a continuous time window is captured. A time-series radio frequency sensing dataset is generated based on the time-series signal strength index stream. A sliding time window is constructed for the time-series radio frequency sensing dataset to obtain multiple window radio frequency sensing sub-sequences; Perform short-time energy averaging on each window radio frequency sensing subsequence to obtain the window short-time energy average; Gaussian kernel density estimation is performed on each window radio frequency sensing subsequence to obtain the window radio frequency density estimate; Window RF confidence weights are generated based on the estimated window RF density. A confidence-weighted radio frequency feature vector is generated based on the window's short-time energy average and the window's radio frequency confidence weight.

5. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, The process of inputting the illumination-calibrated marking image into a pre-trained marking code defect detection network to delineate the geometric topological boundaries of the marking characters and extract the character distortion components includes: The illumination calibration and marking image is input into the training anisotropic heterogeneous multi-branch network to obtain the horizontal stroke response feature map, the vertical stroke response feature map and the local texture response feature map, and generate a set of multi-branch heterogeneous feature matrices. Perform channel normalization and spatial alignment on the set of multi-branch heterogeneous feature matrices to generate channel normalized feature tensor parameters; Based on the channel-normalized feature tensor parameters, the convolution parameters of each branch are equivalently transformed to generate equivalent transformation weights and equivalent translation biases with aligned dimensions. The equivalent transformation weights with dimensional alignment and the equivalent translation bias are structurally reparameterized and fused to generate a single-branch equivalent 3x3 main convolution kernel parameter set; Based on the parameter set of the single-branch equivalent 3x3 main convolution kernel, forward inference is performed on the illumination calibration and marking image to generate an anisotropic marking candidate response field. Local variational detection is performed on the anisotropic marking candidate response field along the topological skeleton of the marked character to generate a stroke distortion feature differential field; Adaptive threshold segmentation and connected component filtering are performed on the differential field of stroke distortion features to generate a damage and defect mask for component marking codes. Based on the damage and defect mask of the marking code and the marking character area of ​​the parts, a set of characteristic indicators for the quality compliance of the parts is constructed. Based on the component quality compliance feature index set and the component marking code damage defect mask, mark code label verification features are generated.

6. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, The step of concatenating the normalized physical location coding features and the confidence-weighted radio frequency feature vector at the feature level, and establishing a strong spatiotemporal association based on the component identity ID, includes: The normalized physical location coding features and the confidence-weighted radio frequency feature vector are concatenated at the feature level to obtain the fusion perception feature vector for parts entering and leaving the warehouse. Establish a strong spatiotemporal correlation of component entry and exit fusion perception feature vectors based on component identity ID, and construct a spatiotemporal feature tensor for warehouse entry and exit events; Retrieve the on-chain flow path from the distributed local ledger of each warehouse outlet; Based on the spatiotemporal feature tensor of warehouse inbound and outbound events and the on-chain flow trajectory path, multi-node spatiotemporal aligned sample pairs are constructed. Calculate the spatiotemporal alignment residual based on multi-node spatiotemporal aligned sample pairs; Based on the spatiotemporal alignment residual, calculate the spatiotemporal state convergence of multiple nodes; Based on the convergence of the spatiotemporal states of multiple nodes and the on-chain flow trajectory path, a multi-node logistics trajectory matching sequence is generated.

7. The method for verifying and tracing the marking codes of automotive electronic components according to claim 1, characterized in that, The step of obtaining pre-stored supply chain source factory certificates in the blockchain distributed ledger and performing asymmetric encrypted digital signature verification based on the tag verification features includes: Based on the component identity ID, obtain the pre-stored supply chain source factory certificate in the blockchain distributed ledger and generate source factory certificate parsing data; Generate a current label verification summary based on the label verification features; Asymmetric encrypted digital signature verification is performed based on the source factory-existing evidence parsing data to obtain the source signature verification result; Based on the current marking verification summary and the source factory certificate analysis data, the marking consistency verification result is obtained; Based on the source verification results and the labeling consistency verification results, the authenticity of the assets is verified and an asset anti-counterfeiting verification label is output.

8. The method for verifying marking codes and tracing the supply chain of automotive electronic components according to claim 1, characterized in that, The construction of a supply chain trust score distribution based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences includes: Based on asset anti-counterfeiting verification labels and multi-node logistics trajectory matching sequences, a set of supply chain circulation nodes is constructed; Based on the set of supply chain flow nodes, spatial grid locations and time slice states are divided; Based on the asset anti-counterfeiting verification label, the set of supply chain circulation nodes is assigned anti-counterfeiting credibility weights to obtain anti-counterfeiting credibility weights. Based on the multi-node logistics trajectory matching sequence, the trajectory matching credibility of the supply chain flow node set is assigned, and a supply chain trust score distribution is constructed. Based on the supply chain trust score distribution, construct a full life cycle circulation network; Based on the full lifecycle flow network, the maximum flow allocation path for the entire lifecycle flow is extracted according to the flow network traffic balancing rules; Based on the maximum flow allocation path throughout the entire lifecycle, generate closed-loop management instructions for supply chain traceability.

Images