Neodymium-iron-boron magnet quality tracing method and system based on big data analysis

Abstract

The application discloses a neodymium-iron-boron magnet quality tracing method and system based on big data analysis, relates to the technical field of quality tracing, and solves the technical problems of high data collection missing rate, inability to support process state restoration and root cause analysis, difficult non-causal simulation error debugging, influence on new product development efficiency, and serious industrial chain information island; through isolation forest online detection of deformation anomalies, process fluctuations can be self-adapted; the deformation index quantifies the deformation severity of single-piece magnets, the Mahalanobis distance evaluates the stability of the batch relative to the historical baseline, and the mold wear or pressing pressure anomaly can be early warned, the size deviation is converted into quantifiable data labels, and the root cause is convenient to trace. The magnetic properties are preliminarily analyzed, if the risk is found, an equation dependency graph is constructed, an abnormal sensor or process parameter is located by using a graph neural network from error variables, key performance fingerprints are stored by a consortium chain, and it is ensured that the data cannot be tampered with and can be audited.

Description

Technical Field

[0001] This invention belongs to the field of quality traceability, specifically a quality traceability method and system for automotive neodymium iron boron magnets based on big data analysis. Background Technology

[0002] Neodymium iron boron (NdFeB) magnets for automotive applications are widely used in critical safety components of new energy vehicles, such as drive motors, braking systems, and electric power steering, due to their high remanence, high coercivity, and excellent energy product. Their magnetic properties directly determine the motor's power density, efficiency, and thermal stability, thus affecting the vehicle's range, acceleration performance, and braking safety. However, the manufacturing process of NdFeB magnets is extremely complex, typically involving more than ten steps, including rare earth smelting, powder preparation, orientation molding, sintering and tempering, grain boundary diffusion, surface coating, and magnetization testing. Fluctuations in each step can significantly impact the final magnetic properties. With the continuous increase in the production and sales of new energy vehicles, OEMs have placed far higher demands on the quality consistency, traceability, and after-sales safety of automotive magnets than on ordinary parts. Traditional quality traceability methods primarily target general attributes such as geometric dimensions, weight, and appearance, using manual inspection tools, barcode scanning, or simple manufacturing execution systems to record batch information.

[0003] The current quality control of NdFeB magnets for automotive applications faces multiple technical bottlenecks: First, data acquisition relies on manual recording or barcode scanning, resulting in high omission rates, significant delays, and the inability to obtain continuous time-series data such as temperature and vacuum levels. This makes it difficult to reconstruct the process state when magnetic performance is abnormal, and root cause analysis lacks effective support. Second, deformation detection relies on manual inspection tools or fixed-threshold machine vision, which is affected by ambient light and mold wear. Minor deformations are easily missed, making it impossible to quantify and trace back to specific processes, leading to safety risks such as abnormal noises during vehicle assembly and brake failure. Third, debugging is difficult after non-causal simulation models such as Modelica report errors. The solver only outputs single-equation errors, making it impossible to locate upstream multivariate coupling problems, which seriously slows down the new product development cycle. Fourth, information silos exist in the upstream and downstream of the industry chain. There is a lack of collaboration between rare earth plants, magnetic material plants, and OEMs. After demagnetization failures occur, there is serious buck-passing, low efficiency of offline communication, and concerns about trade secrets prevent data sharing, making it difficult to form a closed-loop optimization. Fifth, the lack of credible evidence for end-to-end traceability means that enterprise records are at risk of being tampered with and cannot serve as valid evidence in the event of an accident or dispute. As a result, OEMs face high recall costs and brand risks. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a quality traceability method and system for automotive neodymium iron boron magnets based on big data analysis, which is used to solve the technical problems of high data acquisition omission rate, inability to support process state restoration and root cause analysis, difficulty in debugging non-causal simulation errors, impact on new product development efficiency, and serious information silos in the industrial chain.

[0005] To address the aforementioned problems, the first aspect of this invention provides a method and system for quality traceability of automotive neodymium iron boron magnets based on big data analysis, comprising: Each traceability unit is assigned a unique identifier across the entire domain, and this identifier is bound to the magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; process parameters, environmental variables, and intermediate quality data are collected in real time through distributed edge computing nodes, and the data is stored in the data lake according to the identifier; Based on real-time data associated with identifiers stored in the data lake, anomalies are detected online using isolated forests, and deformation indexes and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to identifiers in a data-driven form. A preliminary analysis of the magnetic properties of neodymium iron boron magnets for automobiles is conducted. If there are risks to the magnetic properties in the preliminary analysis, a directed graph of variable dependence of the equation is constructed. Based on the preliminary analysis results and intermediate calculation values, a pre-trained graph neural network GNN-ErrTrace is used to analyze abnormal sensors and record the diagnosed abnormal sensors or parameters as acquisition anomaly markers. The key performance fingerprints and anomaly markers of enterprises in each industry chain are hashed and stored through a consortium blockchain, and gradient parameters are exchanged using a federated learning framework to conduct cross-enterprise collaborative anomaly detection; cross-enterprise batch similarity calculation is achieved based on the identifiers to generate an immutable end-to-end traceability summary.

[0006] Optionally, in one example of the above aspects, each traceability unit is assigned a globally unique identifier, and this identifier is bound to the magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; process parameters, environmental variables, and intermediate quality data are collected in real time through distributed edge computing nodes, and the weighted fusion data is stored in the data lake according to the identifier association, including the following steps: The smallest traceability unit is a single piece of automotive neodymium iron boron magnet or the smallest traceability batch defined in production, which serves as the sole granularity benchmark for quality traceability. By using a composite coding rule that includes production batch code, production line station code, timestamp, serial number, and material specification code, a unique traceability ID is generated globally for each smallest traceability unit, ensuring that the ID is uniquely located across the entire factory area and the entire production cycle, and that each individual / batch of magnets is uniquely located. The traceability ID has inherent semantic attributes, directly linking to material model, production date, and production line information, providing a unique primary key identifier for subsequent full-process traceability and quality positioning; In response to the material characteristics of neodymium iron boron magnets, a unique traceability ID is bound to the individual magnet and the batch tooling pallet at each of the fully discrete process nodes, including raw material warehousing, powder making, molding, sintering, tempering, finishing, plating, testing, and finished product warehousing. Edge computing gateways are deployed at each process station in the discrete manufacturing workshop. The gateways collect core production data in real time, including process parameters, environmental parameters, and intermediate quality data. The compressed, structured, and time-series core production data is linked and bound to a primary key using a unique traceability ID and stored in the data lake.

[0007] Optionally, in one example of the above aspects, based on real-time data associated with identifiers stored in the data lake, anomalies are detected online using isolated forests, and deformation indices and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to the identifiers in a data-driven form, including the following steps: Based on process parameters and intermediate quality data from real-time core production data associated with identifiers stored in the data lake; Each magnet generates a deformation feature vector and an image feature vector; By isolating the anomalies in the online deformation feature vector and image feature vector of the forest, magnets that detect anomalies are marked as suspicious magnets; For the point cloud of a single magnet, the normal deviation is calculated point by point, and after being divided by the design tolerance, it is normalized to obtain the deformation index. For a batch of magnets using the same mold and the same process parameters, calculate the average deformation characteristic value. Covariance For the new batch of magnets, calculate the deformation index and the mean / variance of the magnet area deviation for each piece to form a batch-level deformation feature vector. Calculate the Mahalanobis distance between the new batch and the historical baseline. ; If the new batch is far from the historical baseline Mahalanobis... If the result is greater than the threshold, the process of the new batch is stable; otherwise, the process of the new batch is unstable. The process stability analysis result of the new batch is used as an extended field, associated with the unique traceability ID of the magnet, and stored in the industrial data lake.

[0008] Optionally, in one example of the above aspects, a preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets is performed. If the preliminary analysis reveals risks to the magnetic properties, a directed graph of variable dependencies for the equations is constructed, including the following steps: A preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets was conducted by verifying the tolerance of magnetic property deviations. Let the nominal remanence value of this batch in the inventory be... The measured remanence value is The remanence value after temperature compensation is analyzed using the temperature compensation formula. By calculating the residual magnetism deviation rate, the verification conditions for the preliminary analysis are set. If the initial analysis fails to meet the verification conditions, it is determined that there is a risk to the magnetic properties. Based on all variables and equation constraints in the remanence calculation and verification, a directed graph of variable dependencies of the equations is constructed. Based on all variables and equation constraints in the calculation and verification of remanence, a directed graph G=(V,E) is constructed, with each variable corresponding to a node; Optionally, in one example of the above aspects, the pre-trained graph neural network GNN-ErrTrace is used to analyze abnormal sensors based on preliminary analysis results and intermediate calculated values, and the diagnosed abnormal sensors or parameters are recorded as acquisition anomaly markers, including the following steps: Mark the variables that trigger anomalies during magnetic property verification as error nodes. The intermediate calculated values ​​of all variables during the verification process are used as node feature vectors and input into the pre-trained root cause analysis model, which adopts the GNN-ErrTrace model. By inputting a pre-trained root variable analysis model with a variable dependency directed graph as its structure, message passing and feature aggregation are performed on the graph, starting from the error reporting node. Starting from the beginning, the error is propagated back along the added directed edges; the probability of each node becoming a root cause anomaly is calculated, and finally, the set of root cause variables most likely to cause the error is output. } and their corresponding probabilities; The set of source variables output by the model { }, mapping back to the corresponding physical acquisition source, including: If the root variable in the set is If the root cause variable is..., then the problem is identified as a temperature sensor malfunction; if the root cause variable is... If the root cause is parameter α, the problem is identified as a faulty residual magnetism measurement sensor / equipment; if the root cause is parameter α, the problem is identified as an incorrect temperature coefficient parameter configuration; otherwise, the root cause is faulty. If the root source variable has a corresponding physical acquisition source, obtain the sensor or parameter record corresponding to the physical acquisition source, mark the corresponding sensor as an abnormal sensor, and add an acquisition abnormality mark to the parameter record of the abnormal sensor; if the root source variable itself is abnormal, then the root source variable is marked as abnormal.

[0009] Optionally, in one example of the above aspects, the trained root cause variable analysis model includes the following steps: By verifying magnetic properties under normal operating conditions, simulating processes, and modeling equipment control solvers, known errors are injected into the solver to generate failure cases. For each known error injected, record: the sequence of intermediate values ​​for all variables during the solution process; and the variable that ultimately reports the error. Known root cause variables of the errors; obtain simulation failure case data; Obtain real-world anomaly cases from actual production where the root cause has been identified, and label them as: error variables; intermediate values ​​recorded during the solution process; confirmed root cause variables; and obtain real failure case data. For both simulation failure case data and real failure case data, node features are set to intermediate values ​​to construct a variable dependency directed graph. The GNN-ErrTrace model is trained, with the input being the variable dependency graph and node features, and the output being the probability that each node is the root cause. The node with the highest probability of being the root cause is selected as the root variable node.

[0010] Optionally, in one example of the above aspects, the key attribute fingerprints and anomaly markers of each participant in the upstream and downstream of the industry chain are stored in a distributed ledger, and a federated learning mechanism is used to perform privacy-preserving anomaly detection locally in each enterprise, exchanging only model gradient parameters, including the following steps: Based on the upstream and downstream participants in the automotive rare earth magnet industry chain, a traceability alliance chain for automotive rare earth magnets will be established, with industry chain nodes including: rare earth refinery node, magnetic material plant node, motor plant node, OEM node and third-party testing agency node. Extract key attribute fingerprints from the original data based on the nodes of the industrial chain; Each enterprise completes the statistical and hash calculation of key attribute fingerprints and anomaly markers locally to generate batch fingerprints for this step without uploading any original data; the fingerprint hash value and access control pointer are written into the consortium blockchain to form a batch on-chain record; other nodes verify data integrity through on-chain hash and apply for authorized access through access control pointer; A federated learning mechanism is adopted to perform privacy-preserving anomaly detection locally in each enterprise, and a federated regression model is established, in which different nodes in the same industrial chain only exchange gradient data. By adding noise to the gradient data using differential privacy, each enterprise uploads the noisy gradient data to the consortium blockchain or federated coordination node for gradient descent training. This process includes: the coordination node aggregating all local gradients to obtain the global gradient and updating the model parameters using gradient descent; the updated global model parameters being distributed to all enterprise nodes for the next round of local training of the federated regression model; and the new training data being uploaded with the noisy gradient to repeat the gradient descent training process until the model converges. By using the trained federated regression model, companies can analyze potential anomaly markers based on the key attribute fingerprints of each participant in the upstream and downstream of the industrial chain.

[0011] Optionally, in one example of the above aspects, cross-enterprise batch similarity calculation based on identifiers to generate an immutable end-to-end traceability summary includes the following steps: Based on the unique traceability ID assigned to each magnet / batch, the gradient data of each node in the industrial chain are associated, and a unified gradient performance vector is constructed for each batch. Based on the magnet performance vector corresponding to the anomaly marker ; Historical batch magnet performance vectors from other nodes in the same industry chain The weighted cosine similarity is calculated within the trusted execution environment of the consortium blockchain. If the weighted cosine similarity of the performance vectors of two batches of magnets is greater than the threshold, it is determined that they can be traced back to the same rare earth source batch. Based on the similarity comparison results and combined with the key attribute fingerprints that have been uploaded to the blockchain at each stage, a full-chain traceability summary is generated for this traceability ID, including: a list of rare earth source batches that can be traced back to the same source and the corresponding key attribute fingerprints; The end-to-end traceability summary data is hashed to generate a summary hash value. The summary hash and access control pointer are written into the consortium blockchain and confirmed by consensus among all nodes.

[0012] According to another aspect of this disclosure, a quality traceability system for automotive NdFeB magnets based on big data analysis is provided. This system uses the aforementioned method for quality traceability of automotive NdFeB magnets based on big data analysis to achieve quality traceability of automotive NdFeB magnets based on big data analysis.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention uses an isolated forest method to detect deformation anomalies online, without relying on a fixed threshold, and can adapt to process fluctuations. The deformation index quantifies the severity of deformation of a single magnet, and the Mahalanobis distance assesses the stability of a batch relative to a historical baseline. It can provide early warning of mold wear or abnormal pressing pressure, transforming the vague "dimensional deviation" into a quantifiable data marker, which facilitates tracing the root cause.

[0014] This invention performs a preliminary analysis of magnetic properties. If a risk is identified, an equation dependency graph is constructed, and a graph neural network is used to backpropagate from the error variable to locate the abnormal sensor or process parameter. This solves the problem of "single equation error but no causal chain" in non-causal simulation models, reducing debugging time from several days to hours and minimizing blind guessing.

[0015] This invention uses a consortium blockchain to store key performance fingerprints, ensuring data immutability and auditability. The federated learning framework allows companies to exchange only gradient parameters, not raw data, breaking down information barriers while protecting trade secrets. Cross-enterprise batch similarity calculation can quickly identify batches from the same source, enabling virtual traceability and solving the problem of data silos in the supply chain. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

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

[0019] Please see Figures 1-2 The first aspect of this invention provides a method and system for quality traceability of neodymium iron boron magnets for automobiles based on big data analysis, including: Each traceability unit is assigned a unique identifier across the entire domain, and this identifier is bound to the magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; process parameters, environmental variables, and intermediate quality data are collected in real time through distributed edge computing nodes, and the data is stored in the data lake according to the identifier; Based on real-time data associated with identifiers stored in the data lake, anomalies are detected online using isolated forests, and deformation indexes and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to identifiers in a data-driven form. A preliminary analysis of the magnetic properties of neodymium iron boron magnets for automobiles is conducted. If there are risks to the magnetic properties in the preliminary analysis, a directed graph of variable dependence of the equation is constructed. Based on the preliminary analysis results and intermediate calculation values, a pre-trained graph neural network GNN-ErrTrace is used to analyze abnormal sensors and record the diagnosed abnormal sensors or parameters as acquisition anomaly markers. The key performance fingerprints and anomaly markers of enterprises in each industry chain are hashed and stored through a consortium blockchain, and gradient parameters are exchanged using a federated learning framework to conduct cross-enterprise collaborative anomaly detection; cross-enterprise batch similarity calculation is achieved based on the identifiers to generate an immutable end-to-end traceability summary.

[0020] Specifically, in this embodiment, a globally unique identifier is bound to the magnet entity through an environmentally resistant coupling method, ensuring accurate identification of the neodymium iron boron magnet and achieving end-to-end traceability from raw materials to loading and after-sales service, avoiding chain breaks caused by identifier detachment or damage. Distributed edge nodes collect process parameters, environmental variables, and intermediate quality data close to the equipment, reducing network bandwidth pressure. At the same time, the data lake can store massive amounts of heterogeneous data, providing complete and timely data support for subsequent analysis, solving the problems of data fragmentation and offline lag in traditional manufacturing.

[0021] Using isolated forests for online detection of deformation anomalies does not rely on fixed thresholds and can adapt to process fluctuations; the deformation index quantifies the severity of deformation of individual magnets, and the Mahalanobis distance assesses the stability of batches relative to historical baselines. It can provide early warnings of mold wear or abnormal pressing pressure, transforming vague "dimensional deviations" into quantifiable data markers, making it easier to trace the root cause.

[0022] A preliminary analysis of the magnetic properties is performed. If a risk is identified, an equation dependency graph is constructed, and a graph neural network is used to backpropagate from the error variable to locate the abnormal sensor or process parameter. This approach solves the problem of "single equation error but no causal chain" in non-causal simulation models, reducing debugging time from several days to hours and minimizing guesswork.

[0023] By storing key performance fingerprints through a consortium blockchain, data immutability and auditability are ensured. The federated learning framework allows companies to exchange only gradient parameters, not raw data, breaking down information barriers while protecting trade secrets. Cross-enterprise batch similarity calculations can quickly identify batches from the same source, enabling virtual traceability and solving the problem of data silos in the supply chain. The traceability summary generated based on identifiers contains key fingerprints and anomaly markers from each stage, from rare earth elements to the complete vehicle. Consensus through the consortium blockchain allows it to serve as evidence in quality disputes and meets automotive-grade regulations for the traceability of critical safety components.

[0024] In one embodiment of the present invention, each traceability unit is assigned a globally unique identifier, and this identifier is bound to a magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; process parameters, environmental variables, and intermediate quality data are collected in real time through distributed edge computing nodes, and the weighted and fused data is stored in a data lake according to the identifier association, including the following steps: The smallest traceability unit is a single piece of automotive neodymium iron boron magnet or the smallest traceability batch defined in production, which serves as the sole granularity benchmark for quality traceability. By using a composite coding rule that includes production batch code, production line station code, timestamp, serial number, and material specification code, a unique traceability ID is generated globally for each smallest traceability unit, ensuring that the ID is uniquely located across the entire factory area and the entire production cycle, and that each individual / batch of magnets is uniquely located. The traceability ID has inherent semantic attributes, directly linking to material model, production date, and production line information, providing a unique primary key identifier for subsequent full-process traceability and quality positioning; In response to the material characteristics of neodymium iron boron magnets, a unique traceability ID is bound to the individual magnet and the batch tooling pallet at each of the fully discrete process nodes, including raw material warehousing, powder making, molding, sintering, tempering, finishing, plating, testing, and finished product warehousing. Edge computing gateways are deployed at each process station in the discrete manufacturing workshop to connect to sintering furnaces, presses, grinding equipment, environmental sensors, and online quality inspection equipment nearby. The gateways collect core production data in real time, including process parameters, environmental parameters, and intermediate quality data. Process parameters: sintering temperature, holding time, pressing pressure, rotation speed, coating process parameters, etc.; Environmental parameters: workshop temperature and humidity, dust concentration, oxygen content, etc.; Intermediate quality data: dimensional tolerances of each process, preliminary magnetic property test, appearance defects, intermediate metallographic test results, etc. The compressed, structured, and time-series core production data is linked and bound to a primary key using a unique traceability ID and stored in the data lake.

[0025] In this embodiment, the edge gateway performs lossless / lossy adaptive compression on the entire process, environment, and intermediate quality time-series data, reducing the overhead of storing and transmitting massive amounts of time-series data. The compressed, structured, time-series production quality data is then linked and bound using a unique traceability ID as a primary key, forming a complete traceability data chain from traceability ID to the entire process environment and intermediate quality. After being categorized and organized by batch, time sequence, and process, the data is uniformly integrated into the enterprise's industrial big data lake, achieving permanent storage, retrieval, and traceability of data throughout its entire lifecycle. This provides a complete data foundation for subsequent quality big data analysis, defect root cause localization, process optimization, and full-chain quality traceability.

[0026] In one embodiment of the present invention, based on real-time data associated with identifiers stored in a data lake, anomalies are detected online using isolated forests, and deformation indices and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to identifiers in a data-driven form, including the following steps: Based on process parameters and intermediate quality data from real-time core production data associated with identifiers stored in the data lake; each magnet generates a deformation feature vector and an image feature vector; the deformation feature vector... Image feature vector The deformation feature vector and image feature vector are bound to the traceability ID and stored in the data lake; the detection device in this embodiment adopts a line laser profilometer (to acquire 3D point cloud) + high-resolution industrial camera (to acquire surface image) combined with force sensor (to detect assembly simulation force).

[0027] By isolating the anomalies in the online deformation feature vector and image feature vector of the forest, magnets that detect anomalies are marked as suspicious magnets; For the point cloud of a single magnet, the normal deviation is calculated point by point, and after being divided by the design tolerance, it is normalized to obtain the deformation index. For a batch of magnets using the same mold and the same process parameters, calculate the average deformation characteristic value. Covariance For the new batch of magnets, calculate the deformation index and the mean / variance of the magnet area deviation for each piece to form a batch-level deformation feature vector. Calculate the Mahalanobis distance between the new batch and the historical baseline. The formula is: in, The deformation characteristic vector of the new batch of magnets can be composed of the deformation index of a single piece, the average deformation of each key region, the maximum deviation, etc. If the new batch is far from the historical baseline Mahalanobis... If the result is greater than the threshold, the process of the new batch is stable; otherwise, the process of the new batch is unstable. The process stability analysis result of the new batch is used as an extended field, associated with the unique traceability ID of the magnet, and stored in the industrial data lake.

[0028] In this embodiment, This indicates a significant anomaly in the deformation pattern of this batch, automatically triggering a "mold wear check" or "pressing pressure calibration" recommendation. This value is stored in the supply chain traceability card for downstream automakers to assess assembly risks.

[0029] In one embodiment of the present invention, a preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets is performed. If the preliminary analysis indicates a risk to the magnetic properties, a variable-dependent directed graph of the equation is constructed, including the following steps: A preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets was conducted by verifying the tolerance of magnetic property deviations. Let the nominal remanence value of this batch in the inventory be... The measured remanence value is The remanence value after temperature compensation is analyzed using a temperature compensation formula. Due to the influence of temperature, temperature compensation is introduced, and the formula is as follows: in, The ambient temperature at the time of measurement; The remanence temperature coefficient of the magnet is given in this embodiment. Typical value: -0.12% / ℃, meaning that for every 1℃ increase in temperature, the remanence decreases by 0.12%; This is the remanence value after temperature compensation, which is equivalent to converting the measured remanence to the equivalent remanence under the standard condition of 20℃, and is used for comparison with the nominal value. By calculating the residual magnetism deviation rate, the verification conditions for the preliminary analysis are set as follows: in, The residual magnetism deviation rate, This refers to the tolerance for deviation of automotive magnets; in this embodiment, for automotive magnets, (±3%), stricter than ordinary parts (±5%). If it fails, the automatic verification will fail, the batch will be suspended from production and the magnetic material factory will be notified.

[0030] If the initial analysis fails to meet the verification conditions, it is determined that there is a risk to the magnetic properties. Based on all variables and equation constraints in the remanence calculation and verification, a directed graph of variable dependencies of the equations is constructed. Among them, all variables and equation constraints for remanence calculation and verification include: The variable set V includes: batch-level deformable feature vectors Temperature T, measured remanence Nominal residual magnetism Temperature coefficient α, and residual magnetism after compensation Remanence deviation rate ; The equation constraints E include: the Mahalanobis distance between the new batch and the historical baseline, the temperature compensation formula, and the verification conditions for the preliminary analysis; extracting the deformation characteristics of magnet batches under the same mold and the same process parameters within several stable production cycles in the past (e.g., 50 consecutive batches without abnormalities), and calculating their mean and covariance.

[0031] Based on all variables and equation constraints from remanence calculation and verification, a directed graph G=(V,E) is constructed, with each variable corresponding to a node; such as Tmeas, Brmeas, α, Brcomp, and deviation rate. Directed edges are added according to the dependencies in the equations, including: from arrive Side, from arrive Side, from arrive Side, from arrive Side, from To the residual magnetism deviation rate edge, from To the edge of the residual magnetism deviation rate; In this embodiment, the directed edges include: → Temperature affects the residual magnetism after compensation; → The measured residual magnetism directly affects the compensation value; → Temperature coefficient affects compensation calculation; → Deviation rate, compensation value affects the final deviation; → Deviation rate, nominal value affects the calculation of deviation rate.

[0032] In one embodiment of the present invention, a pre-trained graph neural network GNN-ErrTrace is used to analyze abnormal sensors based on preliminary analysis results and intermediate calculation values, and the diagnosed abnormal sensors or parameters are recorded as acquisition anomaly markers, including the following steps: Mark the variables that trigger anomalies during magnetic property verification as error nodes. The intermediate calculated values ​​of all variables during the verification process, including batch-level deformable feature vectors, will be used. Ambient temperature during actual measurement The measured remanence value is Remanence value after temperature compensation Remanence deviation rate As the node feature vector, it is input into the pre-trained root cause analysis model, which adopts the GNN-ErrTrace model. By inputting a pre-trained root variable analysis model with a variable dependency directed graph as its structure, message passing and feature aggregation are performed on the graph, starting from the error reporting node. Starting from the beginning, the error is propagated back along the added directed edges; the probability of each node becoming a root cause anomaly is calculated, and finally, the set of root cause variables most likely to cause the error is output. } and their corresponding probabilities; For example, the model may identify "abnormal temperature sensor readings" or "deviation in residual magnetism measurement equipment data" as the core reasons for excessive deviation rates.

[0033] The set of source variables output by the model { }, mapping back to the corresponding physical acquisition source, including: If the root variable in the set is If the root cause variable is..., then the problem is identified as a temperature sensor malfunction; if the root cause variable is... If the root cause is parameter α, the problem is identified as a faulty residual magnetism measurement sensor / equipment; if the root cause is parameter α, the problem is identified as an incorrect temperature coefficient parameter configuration; otherwise, the root cause is faulty. If the root source variable has a corresponding physical acquisition source, obtain the sensor or parameter record corresponding to the physical acquisition source, mark the corresponding sensor as an abnormal sensor, and add an acquisition abnormality mark to the parameter record of the abnormal sensor; if the root source variable itself is abnormal, then the root source variable is marked as abnormal.

[0034] In this embodiment, the structured anomaly marker information is the structured acquisition anomaly marker generated for this verification batch, and its content includes: Unique traceability ID, bound to this batch / single magnet; anomaly type, such as "temperature sensor anomaly" or "residual magnetism measurement anomaly"; anomaly root cause variable name and model output probability; intermediate calculation anomaly values, such as temperature readings outside the normal range or invalid value markers; diagnostic timestamp and batch information.

[0035] The generated collection anomaly marker is used as an extended field and associated with the unique traceability ID of that batch / individual magnet. The marker can be stored in an industrial data lake or supply chain traceability card, along with process parameters, quality data, deformation analysis results, etc.

[0036] This serves as supplementary information regarding the failure of this magnetic performance verification, for reference by magnetic material manufacturers during their review; it provides a basis for sensor health status analysis in subsequent process stability analysis, supporting equipment maintenance decisions; and in downstream automotive assembly processes, it can be combined with anomaly markers to comprehensively assess the risk level of magnet batches.

[0037] In one embodiment of the present invention, the trained root cause variable analysis model includes the following steps: Through magnetic performance verification under normal operating conditions, process simulation, and equipment control solver models, these models have clear equation dependency graphs; known errors are injected into the solver to generate failure cases; For example: Sensor / measurement data errors: such as setting variables such as temperature, remanence, and pressure to NaN, infinity, or values ​​that are significantly deviated from the normal range; Parameter errors: such as setting physical parameters such as temperature coefficient, tolerance, and thermal conductivity to negative values ​​or values ​​that are outside the reasonable range; Calculation logic errors: such as incorrect formula input, unit mismatch, or missing dependencies.

[0038] For each known error injected, record: the sequence of intermediate values ​​for all variables during the solution process; and the variable that ultimately reports the error. The known root cause variable of the error, i.e., the variable you injected with the error; obtain simulation failure case data; Obtain real-world anomaly cases from actual production where the root cause has been identified, and label them as: error variables; intermediate values ​​recorded during the solution process; confirmed root cause variables; and obtain real failure case data. For both simulation failure case data and real failure case data, node features are set to intermediate values ​​to construct a variable dependency directed graph. The GNN-ErrTrace model is trained, with the input being the variable dependency graph and node features, and the output being the probability that each node is the root cause. The node with the highest probability of being the root cause is selected as the root variable node.

[0039] In this embodiment, when the probability of the same root cause occurring is at its maximum, multiple root cause variable nodes can be labeled.

[0040] The GNN-ErrTrace model uses a GraphSAGE (or GCN) architecture, which contains at least two graph convolutional layers. It uses the Mean Aggregator function to update node features. During model training, it uses the Cross-Entropy Loss function to calculate the error between the root cause prediction probability and the true label, and uses the Adam optimizer to update the parameters.

[0041] In one embodiment of the present invention, the key attribute fingerprints and anomaly markers of each participant in the upstream and downstream of the industry chain are stored in a distributed ledger, and a federated learning mechanism is used to perform privacy-preserving anomaly detection locally in each enterprise, exchanging only model gradient parameters, including the following steps: Based on the upstream and downstream participants in the automotive rare earth magnet industry chain, a traceability alliance chain for automotive rare earth magnets will be established, with industry chain nodes including: rare earth refinery node, magnetic material plant node, motor plant node, OEM node and third-party testing agency node. Based on the nodes of the industry chain, extract the key attribute fingerprints of the original data that cannot be reversed, including: Rare earth batch fingerprint: {Ore source ID, mean Dy content, mean Tb content, impurity distribution histogram hash}; Magnetic material batch fingerprint: {Br mean, Hcj mean, sintering curve feature vector hash}; OEM assembly fingerprint: {VIN range, motor model, assembly air gap value statistics}; Each enterprise completes the statistical and hash calculation of key attribute fingerprints and anomaly markers locally to generate batch fingerprints for this stage without uploading any original data; the fingerprint hash value and access control pointer are written into the consortium blockchain to form a batch on-chain record; other nodes verify the data integrity through on-chain hash and apply for authorized access through access control pointer; thus achieving the effect of "verifiable, tamper-proof, and privacy-controlled" evidence storage.

[0042] A federated learning mechanism is adopted to perform privacy-preserving anomaly detection locally in each enterprise, and a federated regression model is established, in which different nodes in the same industrial chain only exchange gradient data. By adding noise to the gradient data using differential privacy, each enterprise uploads the noisy gradient data to the consortium blockchain or federated coordination node for gradient descent training. This process includes: the coordination node aggregating all local gradients to obtain the global gradient and updating the model parameters using gradient descent; the updated global model parameters being distributed to all enterprise nodes for the next round of local training of the federated regression model; and the new training data being uploaded with the noisy gradient to repeat the gradient descent training process until the model converges. By using the trained federated regression model, companies can analyze potential anomaly markers based on the key attribute fingerprints of each participant in the upstream and downstream of the industrial chain.

[0043] In one embodiment of the present invention, the formula for calculating gradient data is as follows: in, These are the key attribute fingerprints and anomaly marker parameters of the global model at the t-th iteration. η represents the updated global model key attribute fingerprint and anomaly label parameters after the (t+1)th iteration; η is the learning rate, a hyperparameter that controls the step size of each parameter update. The gradient operator represents the derivative of a function, pointing in the direction of the fastest growth of the function. In optimization, we usually update parameters in the opposite direction of the gradient. The local loss function is one of the nodes in the industry chain, corresponding to the local gradient mentioned in the paper. Calculations are performed at each node of the industrial chain. This is the local loss function for another node in the industrial chain, corresponding to the prediction of the impact of different rare earth ratios on the lifespan of the vehicle motor. Differential privacy adds noise to the gradient, as shown in the formula: in, The gradient after adding noise; The original gradient is represented by N(·); N(·) represents a normal distribution. σ represents the standard deviation of the noise, controlling the noise intensity. The larger the σ, the higher the degree of privacy protection (the more difficult it is to reverse-engineer), but the accuracy of the model may decrease. This typically represents the norm limit of the clipping, which determines how much noise needs to be added to meet the privacy budget. The identity matrix represents the isotropic nature of the noise, meaning that the noise added to each parameter dimension of the model is independent and identically distributed. In one embodiment of the present invention, cross-enterprise batch similarity calculation is performed based on identifiers to generate an immutable end-to-end traceability summary, including the following steps: Based on the unique traceability ID assigned to each magnet / batch, the gradient data of each node in the industrial chain are associated, and a unified gradient performance vector is constructed for each batch. Based on the magnet performance vector corresponding to the anomaly marker ; Historical batch magnet performance vectors from other nodes in the same industry chain The weighted cosine similarity is calculated in the trusted execution environment of the consortium blockchain, using the following formula: in, The specific value of the magnet performance vector for the batch corresponding to the anomaly marker at the kth index; This refers to the specific value of the magnet performance vector of other nodes in the same industrial chain for the historical batches at the k-th index. The weight of the k-th indicator; for and Weighted cosine similarity; If the weighted cosine similarity of the performance vectors of two batches of magnets is greater than the threshold, it is determined that they can be traced back to the same rare earth source batch. Based on the similarity comparison results and combined with the key attribute fingerprints that have been uploaded to the blockchain at each stage, a full-chain traceability summary is generated for this traceability ID, including: a list of rare earth source batches that can be traced back to the same source and the corresponding key attribute fingerprints; The end-to-end traceability summary data is hashed to generate a summary hash value. The summary hash and access control pointer are written into the consortium blockchain and confirmed by consensus among all nodes.

[0044] In this embodiment, when At that time, it was believed that the two batches had highly consistent performance and could be traced back to the same rare earth source batch.

[0045] In this embodiment, the weight of the kth indicator... Based on domain knowledge, for example, Br weight 0.5, Hcj weight 0.4, and temperature level 0.1. When At that time, it was assumed that the two batches had highly consistent performance and could be traced back to the same rare earth source batch. This calculation was performed on-chain through a Trusted Execution Environment (TEE), with public inputs and outputs but confidential raw data.

[0046] In this embodiment, through federated anomaly detection, each enterprise trains a local anomaly detection model locally, and only uploads the model parameter gradients to the consortium blockchain aggregation server. The server then calculates the global anomaly threshold. When any enterprise detects an abnormal batch, other enterprises' local models can be passively updated, enabling collaborative early warning without sharing data.

[0047] In another embodiment of the present invention, a quality traceability system for automotive NdFeB magnets based on big data analysis is provided. This system employs the aforementioned method for quality traceability of automotive NdFeB magnets based on big data analysis to achieve quality traceability of automotive NdFeB magnets based on big data analysis, including: The identification and association module assigns a unique identifier to each traceability unit and binds the identifier to the magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; it collects process parameters, environmental variables and intermediate quality data in real time through distributed edge computing nodes, and stores the weighted and fused data in the data lake according to the identifier association. Defect Analysis Module: Based on real-time data associated with identifiers stored in the data lake, it uses isolated forests to detect anomalies online and analyzes deformation index and cross-batch Mahalanobis distance, binding the analysis results to identifiers in a data-driven form; Anomaly labeling module: Performs preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets. If the preliminary analysis indicates a risk to the magnetic properties, it constructs a variable-dependent directed graph of the equation. Using a pre-trained graph neural network GNN-ErrTrace, it analyzes the abnormal sensors based on the preliminary analysis results and intermediate calculation values, and records the diagnosed abnormal sensors or parameters as acquisition anomaly labels. Link tracing module: The key performance fingerprints of enterprises in each industry chain are hashed and stored through a consortium blockchain, and gradient parameters are exchanged using a federated learning framework to achieve cross-enterprise collaborative detection of anomalies.

[0048] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for quality traceability of neodymium iron boron magnets for automobiles based on big data analysis, characterized in that, include: Each traceability unit is assigned a unique identifier across the entire domain, and this identifier is bound to a magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; Process parameters, environmental variables, and intermediate quality data are collected in real time through distributed edge computing nodes, and the data is stored in the data lake according to the association of the identifiers. Based on real-time data associated with identifiers stored in the data lake, anomalies are detected online using isolated forests, and deformation indexes and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to identifiers in a data-driven form. A preliminary analysis of the magnetic properties of neodymium iron boron magnets for automobiles is conducted. If there are risks to the magnetic properties in the preliminary analysis, a directed graph of variable dependence of the equation is constructed. Based on the preliminary analysis results and intermediate calculation values, a pre-trained graph neural network GNN-ErrTrace is used to analyze abnormal sensors and record the diagnosed abnormal sensors or parameters as acquisition anomaly markers. The key performance fingerprints and anomaly markers of enterprises in each industry chain are hashed and stored through a consortium blockchain, and gradient parameters are exchanged using a federated learning framework to conduct cross-enterprise collaborative anomaly detection. Based on the identifier, cross-enterprise batch similarity calculation is performed to generate an immutable end-to-end traceability summary.

2. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, Each traceability unit is assigned a unique identifier across the entire domain, and this identifier is bound to a magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; The data, including process parameters, environmental variables, and intermediate quality data, is collected in real time through distributed edge computing nodes. The weighted and fused data is then stored in a data lake according to the associated identifiers. The process includes the following steps: The smallest traceability unit is a single piece of automotive neodymium iron boron magnet or the smallest traceability batch defined in production, which serves as the sole granularity benchmark for quality traceability. By using a composite coding rule that includes production batch code, production line station code, timestamp, serial number, and material specification code, a unique traceability ID is generated globally for each smallest traceability unit, ensuring that the ID is uniquely located across the entire factory area and the entire production cycle, and that each individual / batch of magnets is uniquely located. The traceability ID has inherent semantic attributes, directly linking to material model, production date, and production line information, providing a unique primary key identifier for subsequent full-process traceability and quality positioning; In response to the material characteristics of neodymium iron boron magnets, a unique traceability ID is bound to the individual magnet and the batch tooling pallet at each of the fully discrete process nodes, including raw material warehousing, powder making, molding, sintering, tempering, finishing, plating, testing, and finished product warehousing. Edge computing gateways are deployed at each process station in the discrete manufacturing workshop. The gateways collect core production data in real time, including process parameters, environmental parameters, and intermediate quality data. The compressed, structured, and time-series core production data is linked and bound to a primary key using a unique traceability ID and stored in the data lake.

3. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, Based on real-time data associated with identifiers stored in the data lake, anomalies are detected online using isolated forests, and deformation indices and cross-batch Mahalanobis distances are analyzed. The analysis results are then bound to identifiers in a data-driven format, including the following steps: Based on process parameters and intermediate quality data from real-time core production data associated with identifiers stored in the data lake, each magnet generates a deformation feature vector and an image feature vector; the deformation feature vector... Image feature vector The deformable feature vector and image feature vector are bound to the traceability ID and stored in the data lake; By isolating the anomalies in the online deformation feature vector and image feature vector of the forest, magnets that detect anomalies are marked as suspicious magnets; For the point cloud of a single magnet, the normal deviation is calculated point by point, and after being divided by the design tolerance, it is normalized to obtain the deformation index. For a batch of magnets using the same mold and the same process parameters, calculate the average deformation characteristic value. Covariance For the new batch of magnets, calculate the deformation index and the mean / variance of the magnet area deviation for each piece to form a batch-level deformation feature vector. Calculate the Mahalanobis distance between the new batch and the historical baseline. ; If the new batch is far from the historical baseline Mahalanobis... If the result is greater than the threshold, the process of the new batch is stable; otherwise, the process of the new batch is unstable. The process stability analysis result of the new batch is used as an extended field, associated with the unique traceability ID of the magnet, and stored in the industrial data lake.

4. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, A preliminary analysis of the magnetic properties of neodymium iron boron magnets for automotive applications is conducted. If the preliminary analysis reveals risks related to magnetic properties, a directed graph of variable dependencies for the equations is constructed, including the following steps: A preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets was conducted by verifying the tolerance of magnetic property deviations. Let the nominal remanence value of this batch in the inventory be... The measured remanence value is The remanence value after temperature compensation is analyzed using the temperature compensation formula, which is: in, The ambient temperature at the time of measurement; is the temperature coefficient of remanence of the magnet; This is the remanence value after temperature compensation; By calculating the residual magnetism deviation rate, the verification conditions for the preliminary analysis are set as follows: in, The residual magnetism deviation rate, For the deviation tolerance of automotive magnets; If the initial analysis fails to meet the verification conditions, it is determined that there is a risk to the magnetic properties. Based on all variables and equation constraints in the remanence calculation and verification, a directed graph of variable dependencies of the equations is constructed. Among them, all variables and equation constraints for remanence calculation and verification include: The variable set V includes: batch-level deformable feature vectors Temperature T, measured remanence Nominal residual magnetism Temperature coefficient α, and residual magnetism after compensation Remanence deviation rate ; The equation constraints E include: the Mahalanobis distance between the new batch and the historical baseline, the temperature compensation formula, and the verification conditions for the preliminary analysis; Based on all variables and equation constraints from remanence calculation and verification, a directed graph G=(V,E) is constructed, with each variable corresponding to a node; directed edges are added according to the dependencies in the equations, including: from arrive Side, from arrive Side, from arrive Side, from arrive Side, from To the residual magnetism deviation rate edge, from To the residual magnetism deviation rate edge.

5. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, Using a pre-trained graph neural network GNN-ErrTrace, based on preliminary analysis results and intermediate calculated values, abnormal sensors are analyzed, and the diagnosed abnormal sensors or parameters are recorded as acquisition anomaly markers. This includes the following steps: Mark the variables that trigger anomalies during magnetic property verification as error nodes. The intermediate calculated values ​​of all variables during the verification process, including batch-level deformable feature vectors, will be used. Ambient temperature during actual measurement The measured remanence value is Remanence value after temperature compensation Remanence deviation rate As the node feature vector, it is input into the pre-trained root cause analysis model, which adopts the GNN-ErrTrace model. By inputting a pre-trained root variable analysis model with a variable dependency directed graph as its structure, message passing and feature aggregation are performed on the graph, starting from the error reporting node. Starting from the beginning, the error is propagated back along the added directed edges; the probability of each node becoming a root cause anomaly is calculated, and finally, the set of root cause variables most likely to cause the error is output. } and their corresponding probabilities; The set of source variables output by the model { }, mapping back to the corresponding physical acquisition source, including: If the root variable in the set is If the root cause variable is..., then the problem is identified as a temperature sensor malfunction; if the root cause variable is... If the root cause is parameter α, the problem is identified as a faulty residual magnetism measurement sensor / equipment; if the root cause is parameter α, the problem is identified as an incorrect temperature coefficient parameter configuration; otherwise, the root cause is faulty. If the root source variable has a corresponding physical acquisition source, obtain the sensor or parameter record corresponding to the physical acquisition source, mark the corresponding sensor as an abnormal sensor, and add an acquisition abnormality mark to the parameter record of the abnormal sensor; if the root source variable itself is abnormal, then the root source variable is marked as abnormal.

6. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, The root cause variable analysis model is trained, including the following steps: By verifying magnetic properties under normal operating conditions, simulating processes, and modeling equipment control solvers, known errors are injected into the solver to generate failure cases. For each known error injected, record: the sequence of intermediate values ​​for all variables during the solution process; and the variable that ultimately reports the error. Known root cause variables of the errors; obtain simulation failure case data; Obtain real-world anomaly cases from actual production where the root cause has been identified, and label them as: error variables; intermediate values ​​recorded during the solution process; confirmed root cause variables; and obtain real failure case data. For both simulation failure case data and real failure case data, node features are set to intermediate values ​​to construct a variable dependency directed graph. The GNN-ErrTrace model is trained, with the input being the variable dependency graph and node features, and the output being the probability that each node is the root cause. The node with the highest probability of being the root cause is selected as the root variable node.

7. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, The key attribute fingerprints and anomaly markers of all participants in the upstream and downstream of the industry chain are stored in a distributed ledger, and a federated learning mechanism is used to perform privacy-preserving anomaly detection locally in each enterprise, exchanging only model gradient parameters. This includes the following steps: Based on the upstream and downstream participants in the automotive rare earth magnet industry chain, a traceability alliance chain for automotive rare earth magnets will be established, with industry chain nodes including: rare earth refinery node, magnetic material plant node, motor plant node, OEM node and third-party testing agency node. Based on the nodes of the industry chain, extract the key attribute fingerprints of the original data that cannot be reversed, including: Rare earth batch fingerprint: {Ore source ID, mean Dy content, mean Tb content, impurity distribution histogram hash}; Magnetic material batch fingerprint: {Br mean, Hcj mean, sintering curve feature vector hash}; OEM assembly fingerprint: {VIN range, motor model, assembly air gap value statistics}; Each enterprise completes the statistical and hash calculation of key attribute fingerprints and anomaly markers locally to generate batch fingerprints for this step without uploading any original data; the fingerprint hash value and access control pointer are written into the consortium blockchain to form a batch on-chain record; other nodes verify data integrity through on-chain hash and apply for authorized access through access control pointer; A federated learning mechanism is adopted to perform privacy-preserving anomaly detection locally in each enterprise, and a federated regression model is established, in which different nodes in the same industrial chain only exchange gradient data. By adding noise to the gradient data using differential privacy, each enterprise uploads the noisy gradient data to the consortium blockchain or federated coordination node for gradient descent training. This process includes: the coordination node aggregating all local gradients to obtain the global gradient and updating the model parameters using gradient descent; the updated global model parameters being distributed to all enterprise nodes for the next round of local training of the federated regression model; and the new training data being uploaded with the noisy gradient to repeat the gradient descent training process until the model converges. By using the trained federated regression model, companies can analyze potential anomaly markers based on the key attribute fingerprints of each participant in the upstream and downstream of the industrial chain.

8. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 7, characterized in that, The formula for calculating gradient data is: in, These are the key attribute fingerprints and anomaly marker parameters of the global model at the t-th iteration. The key attribute fingerprint and anomaly labeling parameters of the updated global model after the (t+1)th iteration are given; η is the learning rate. For gradient operators; This is the local loss function for one of the nodes in the industry chain; For another node in the industry chain, there is a local loss function; Differential privacy adds noise to the gradient, as shown in the formula: in, The gradient after adding noise; The original gradient is represented by N(·); N(·) represents a normal distribution. The standard deviation of the noise; This usually represents the norm limit of the clipping; It is an identity matrix.

9. The method for quality traceability of automotive neodymium iron boron magnets based on big data analysis according to claim 1, characterized in that, Based on identifiers, cross-enterprise batch similarity calculation is performed to generate an immutable end-to-end traceability summary, including the following steps: Based on the unique traceability ID assigned to each magnet / batch, the gradient data of each node in the industrial chain are associated, and a unified gradient performance vector is constructed for each batch. Based on the magnet performance vector corresponding to the anomaly marker ; Historical batch magnet performance vectors from other nodes in the same industry chain The weighted cosine similarity is calculated in the trusted execution environment of the consortium blockchain, using the following formula: in, The specific value of the k-th index of the magnet performance vector corresponding to the abnormal marker; This refers to the specific value of the magnet performance vector of other nodes in the same industrial chain for the historical batches at the k-th index. The weight of the k-th indicator; for and Weighted cosine similarity; If the weighted cosine similarity of the performance vectors of two batches of magnets is greater than the threshold, it is determined that they can be traced back to the same rare earth source batch. Based on the similarity comparison results and combined with the key attribute fingerprints that have been uploaded to the blockchain at each stage, a full-chain traceability summary is generated for this traceability ID, including: a list of rare earth source batches that can be traced back to the same source and the corresponding key attribute fingerprints; The end-to-end traceability summary data is hashed to generate a summary hash value. The summary hash and access control pointer are written into the consortium blockchain and confirmed by consensus among all nodes.

10. A quality traceability system for automotive neodymium iron boron magnets based on big data analysis, characterized in that, This system employs the big data analysis-based quality traceability method for automotive NdFeB magnets as described in any one of claims 1-9 to achieve quality traceability of automotive NdFeB magnets based on big data analysis, comprising: The identification and association module assigns a unique identifier to each traceability unit and binds the identifier to the magnetic entity in an environmentally resistant coupling manner throughout the discrete manufacturing process; it collects process parameters, environmental variables and intermediate quality data in real time through distributed edge computing nodes, and stores the weighted and fused data in the data lake according to the identifier association. Defect Analysis Module: Based on real-time data associated with identifiers stored in the data lake, it uses isolated forests to detect anomalies online and analyzes deformation index and cross-batch Mahalanobis distance, binding the analysis results to identifiers in a data-driven form; Anomaly labeling module: Performs preliminary analysis of the magnetic properties of automotive neodymium iron boron magnets. If the preliminary analysis indicates a risk to the magnetic properties, it constructs a variable-dependent directed graph of the equation. Using a pre-trained graph neural network GNN-ErrTrace, it analyzes the abnormal sensors based on the preliminary analysis results and intermediate calculation values, and records the diagnosed abnormal sensors or parameters as acquisition anomaly labels. Link tracing module: The key performance fingerprints of enterprises in each industry chain are hashed and stored through a consortium blockchain, and gradient parameters are exchanged using a federated learning framework to achieve cross-enterprise collaborative detection of anomalies.

Images