Federal learning-based full-life-cycle management method and system for automobile parts

Abstract

The invention discloses an automobile part full life cycle management method and system based on federated learning, and relates to the technical field of federated learning, and the method comprises the steps: carrying out the real-time reasoning through a to-be-reasoned feature view and an effective model version, and calculating the uncertainty and drift distance to obtain a real-time prediction result and an evidence obtaining candidate queue; after an evidence obtaining budget is distributed according to the evidence obtaining candidate queue, an evidence obtaining packet is collected and rechecking is completed, label backflow and credibility weights are generated, the label backflow is written into the to-be-trained sample set, and an incremental training sample set is obtained; executing local incremental training by adopting the incremental training sample set and the credibility weight, and generating a local model updating package; and performing hierarchical federation aggregation on the local model update package, generating a verifiable aggregation proof, writing the verifiable aggregation proof into a model version account book, and releasing an effective model version and a registration record. According to the method, the prediction stability is improved, traceability and verifiability of model evolution are realized, and cross-subject full-life-cycle collaborative management is supported.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a method and system for full lifecycle management of automotive parts based on federated learning. Background Technology

[0002] Automotive parts lifecycle management encompasses the continuous chain of parts from R&D and design, manufacturing quality inspection, warehousing and logistics, assembly and delivery, operation and maintenance, to recycling and remanufacturing. Related technologies are evolving alongside the digitalization and intelligent connectivity of the automotive industry. On one hand, PLM, MES, WMS, after-sales work orders, and vehicle network data channels are gradually being integrated. Event-driven data governance and quality traceability methods are widely adopted in the industry, and parts identification coding, batch tracking, and process parameter recording are becoming increasingly standardized. On the other hand, machine learning-based fault prediction, remaining life estimation, and tiered recall handling are continuously being deepened in domestic and international research and engineering practices. Edge computing and multi-party collaborative training are used to reduce edge latency and improve the model's adaptability to complex operating conditions. Against the backdrop of simultaneously increasing requirements for privacy compliance and data element circulation, federated learning, as an important path for multi-agent collaborative modeling, is increasingly being used in cross-enterprise quality modeling, cross-regional maintenance prediction, and supply chain risk assessment scenarios. It is also developing in tandem with technologies such as encrypted computing, access control, and traceable evidence storage, making the intelligent lifecycle management of "data not leaving the domain but capabilities being shareable" a foundation for engineering implementation.

[0003] Several key shortcomings in existing technologies hinder the stable improvement of full lifecycle management capabilities. First, the distribution of multi-entity data exhibits strong heterogeneity and strong non-independent identical distribution characteristics. Significant differences exist between the production and after-sales ends in terms of field definitions, sampling frequencies, noise patterns, and missing data modes. Traditional centralized aggregation methods are difficult to sustain in the long term due to compliance and cost constraints. Existing federated learning practices often neglect event-level evidence correlation and permission semantics, resulting in insufficient availability and auditability of training data. Models are prone to definitional drift and attribution issues when migrating across entities. Second, after-sales fault and failure labels are naturally scarce and of varying quality. Relying solely on passively transmitted labels can lead to sample selection bias. Conceptual drift combined with label noise can cause predictive instability under new batches of materials, new process parameters, new regional climates, and new driving habits. Furthermore, the model iteration process lacks verifiable aggregation credentials and version evolution evidence chains, making it difficult to support demanding scenarios such as recall disputes, warranty arbitration, and regulatory verification. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a federated learning-based method for the full lifecycle management of automotive parts to address the problems of compliant, unified, and shared event flows from multiple entities, unstable predictions due to a lack of labels and concept drift, and a lack of traceable and auditable model iterations.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, this invention provides a method for managing the entire lifecycle of automotive parts based on federated learning. This method includes: collecting event stream data from the entire lifecycle of automotive parts and encapsulating it into event packages, generating evidence summaries and signature identifiers to obtain a queue of event packages to be processed and a set of event evidence summaries; performing compliance verification, unified desensitization, and permission labeling on the queue of event packages to be processed and the set of event evidence summaries, and extracting task fields to obtain a feature view to be inferred and a training sample set; performing real-time inference using the feature view to be inferred and the effective model version, and calculating uncertainty and drift to obtain real-time prediction results and a candidate queue for evidence collection; allocating an evidence collection budget based on the candidate queue for evidence collection, collecting evidence packages, and completing verification, generating label feedback and credibility weights, and writing the label feedback into the training sample set to obtain an incremental training sample set; performing local incremental training using the incremental training sample set and credibility weights, and generating a local model update package; performing hierarchical federated aggregation on the local model update package, generating verifiable aggregation proofs, writing them into the model version ledger, and then publishing the effective model version and registration records.

[0007] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for collecting automotive parts lifecycle event stream data and encapsulating it into event packets are as follows. Collect event stream data of the entire life cycle of automotive parts, and read the newly added record payload according to the incremental event triggering collection rules, and write it into the collection temporary storage area; The newly added recorded loads in the data collection and storage area are processed to standardize component identification, unify timestamps and main body identification, resulting in a standardized set of event fields. The event type is determined based on the standardized event field set and the event type mapping table, and a set of encapsulation preparation fields is generated. An event package is constructed by encapsulating a set of preparation fields, and the event package header and body information are written to obtain a complete event package.

[0008] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for obtaining the queue of event packets to be processed and the set of event evidence summaries are as follows: Perform normalized serialization on the complete event package, calculate the evidence digest, establish the correspondence between the evidence digest and the event sequence number, and obtain the event evidence digest set; Perform private key signing on the evidence digest, generate a signature identifier, and write it into the event packet evidence field. Sort the event packets and write them into the pending event packet queue to obtain the pending event packet queue.

[0009] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the following steps are taken: The queue of event packets to be processed and the set of event evidence summaries undergo unified desensitization and permission labeling based on compliance verification standards. Perform consistency checks between the event packet evidence fields in the pending event packet queue and the event evidence digest set, and perform signature validity checks to obtain the compliance check event packet queue; Perform unified field processing on the compliance verification event package queue according to the event type mapping table, and obtain a unified event package queue by unifying the unit format and time format; According to the desensitization rule table, replace the plaintext fields of the sensitive fields in the unified event packet queue, and retain the desensitization mark to obtain the desensitized event packet queue; Write the de-identified event package queue into the permission label field according to the permission policy table, and bind the permission label field to the event type and subject identifier to obtain the permission label event package queue.

[0010] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for obtaining the feature view to be inferred and the training sample set are as follows: Based on the permission-labeled event packet queue and task field mapping table, extract task fields and generate a feature view to be reasoned that can be used for real-time reasoning; The training fields are extracted from the permission labeling event packet queue according to the task field mapping table, and then aggregated into a training sample set according to the time window.

[0011] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for obtaining the real-time prediction results and the evidence candidate queue are as follows: Load the effective model version that matches the task identifier of the feature view to be inferred, forming the real-time inference input; Real-time inference is performed on the real-time inference input to obtain risk scores and risk categories, which are then written into the event sequence number association field to obtain real-time prediction results; The uncertainty is calculated based on the real-time prediction results, and a one-to-one correspondence is established between the uncertainty and the event sequence number to obtain the uncertainty result set; The drift amount is updated within the sliding time window based on the feature view to be inferred, and the drift amount is associated with the event type to obtain the drift amount result set; Based on the uncertainty result set, drift result set and evidence collection triggering rules, the event sequence number that needs supplementary evidence is selected, and an evidence collection request identifier is generated to obtain the evidence collection candidate queue.

[0012] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for obtaining the incremental training sample set are as follows: Based on the evidence collection candidate queue and combined with the evidence collection budget allocation rules, the evidence collection budget amount corresponding to each evidence collection candidate is calculated, and the evidence collection budget allocation result set is obtained. Based on the evidence collection budget allocation result set, generate an evidence collection request and send it to the evidence collection target entity to obtain evidence collection materials that match the event sequence number and obtain the evidence collection package; The integrity and evidence digest consistency checks are performed on the evidence package, and the signature validity check is performed to obtain the approved evidence package; Based on the verified evidence package, a tag return stream is generated and the tag return stream is bound to the event sequence number to form a tag return stream record; The credibility weight of the tag return record is calculated according to the credibility assessment rules, and the credibility weight is bound to the tag return record; The label backflow records are written into the training sample set, and the credibility weight field bound to the label backflow records is written simultaneously to form an incremental training sample set.

[0013] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for generating the local model update package are as follows: Perform field completeness and label consistency checks on the incremental training sample set to obtain the training-ready sample set. Load the effective model version that matches the task identifier of the training-ready sample set, and establish the trainable parameter set and training round identifier to obtain the incremental training initialization state. Weighted sampling is performed on the training-ready sample set according to the confidence weight, and incremental training batches are constructed to obtain the incremental training batch set; Local incremental training is performed on the incremental training initialization state using an incremental training batch set, and loss convergence information is recorded in each training round to obtain the updated local model parameters; Perform differential calculation on the updated local model parameters and the effective model version to generate an update summary. Bind the update summary with the task identifier to obtain the local model update difference and update summary. The local model update difference and update summary are encapsulated into a local model update package.

[0014] As a preferred embodiment of the federated learning-based automotive parts lifecycle management method described in this invention, the specific steps for releasing and implementing the model version and registration records are as follows: Receive local model update packages uploaded by each participant, and aggregate and group them according to the supply chain topology mapping rules and align them with the time window to obtain an aggregated group update set; The first-level aggregation is performed on the aggregation group update set at the edge aggregation node to generate the regional aggregation results and regional aggregation summary, which are then combined into a regional model update package; The regional model update package is uploaded to the consortium aggregation node, and a second-level aggregation is performed to generate candidate global model versions and a global aggregation summary. Verifiable aggregation proofs are generated for the first-level and second-level aggregation processes, and the verifiable aggregation proofs are bound to the global aggregation digest to obtain verifiable aggregation proofs. Write the version identifier, global aggregation summary, and verifiable aggregation proof of the candidate global model version into the model version ledger to generate a queryable effective model version registration record; Based on the effective model version registration record, the candidate global model version is published as the effective model version, and the effective model version and the effective model version registration record are distributed to each participating party to obtain the effective model version and the effective model version registration record.

[0015] Secondly, this invention provides a federated learning-based automotive parts lifecycle management system, comprising: an event encapsulation module for collecting automotive parts lifecycle event stream data and encapsulating it into event packages, generating evidence summaries and signature identifiers, and obtaining a queue of event packages to be processed and a set of event evidence summaries; a governance construction module for performing compliance verification, unified desensitization, and permission labeling on the queue of event packages to be processed and the set of event evidence summaries, and extracting task fields to obtain a feature view to be inferred and a training sample set; and an inference triggering module for performing real-time inference using the feature view to be inferred and the effective model version. The system calculates uncertainty and drift to obtain real-time prediction results and a candidate queue for evidence collection; the evidence collection feedback module allocates the evidence collection budget based on the candidate queue, collects evidence collection packages, completes verification, generates label feedback and credibility weights, and writes the label feedback into the training sample set to obtain the incremental training sample set; the incremental training module performs local incremental training using the incremental training sample set and credibility weights, and generates a local model update package; the aggregation and evidence storage module performs hierarchical federated aggregation on the local model update package, generates verifiable aggregation proofs, writes them into the model version ledger, and then publishes the effective model version and registration record.

[0016] The beneficial effects of this invention are as follows: By using the evidence collection and feedback step, the evidence collection candidate queue obtained by real-time inference is used to allocate the evidence collection budget and collect evidence collection packages for review. After verification, the label feedback and credibility weight are generated to form an incremental training sample set, which realizes targeted evidence supplementation and noise reduction training in scenarios with few labels, thereby improving prediction stability. By using the aggregation and evidence storage step, the local model update package is aggregated in a hierarchical federated manner and verifiable aggregation proof is generated and written into the model version ledger to publish the effective model version and registration record, thereby realizing traceable and verifiable model evolution and supporting cross-entity full life cycle collaborative management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a federated learning-based approach to the full lifecycle management of automotive parts.

[0019] Figure 2 This is a schematic diagram of a federated learning-based automotive parts lifecycle management system.

[0020] Figure 3 A flowchart illustrating the data link structure built for event encapsulation and governance.

[0021] Figure 4 A flowchart for the proactive evidence collection, tag return, and hierarchical federated aggregation evidence storage chain. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4This is one embodiment of the present invention, which provides a method for full lifecycle management of automotive parts based on federated learning, including the following steps: S1: Collect event stream data of the entire life cycle of automotive parts and encapsulate it into event packages, and generate evidence digests and signature identifiers to obtain a queue of event packages to be processed and a set of event evidence digests; S1.1: Collect event stream data of the entire life cycle of automotive parts, and read the newly added record payload according to the incremental event triggering collection rules, and write it into the collection temporary storage area; Furthermore, incremental event triggering collection rules are configured for the business record sources in the production, quality inspection, warehousing and logistics, assembly, operation, maintenance, and recycling stages. The incremental event triggering collection rules use changes in the record update timestamp field as the trigger condition and the component identification field as the location condition. After the incremental event triggering collection rule is triggered, the newly added record payload and record source identifier are read from the corresponding business record source and written to the collection temporary storage area.

[0026] S1.2: Standardize the component identification, unify the timestamp and the main body identification of the newly added record load in the acquisition temporary storage area to obtain a standardized event field set; Furthermore, the newly added record payload is read from the acquisition buffer in the order of the written timestamps, and the component identifier field, timestamp field, and record source identifier field are parsed. The component identifier is standardized according to the component identifier field rules to ensure consistency of character set, length, and check bits, and a standardized component identifier is output. The timestamp is unified according to the timestamp field rules to convert timestamps with different precision and time zones to a unified time zone and precision, and a unified timestamp is output. The subject identifier is unified according to the subject identifier dictionary to map the record source identifier to the subject identifier and output the subject identifier. The standardized component identifier, unified timestamp, and subject identifier, together with the business fields in the newly added record payload, are written into the standardized event field set.

[0027] S1.3: Determine the event type based on the standardized event field set and the event type mapping table, and generate the encapsulation preparation field set; Furthermore, the standardized event field set is read one by one, and the event type is determined according to the event type mapping table. The event type mapping table uses the existence, value range, and semantic label of the field as the determination conditions and outputs the event type. The standardized event field set that has completed the event type determination is used to generate a packaging preparation field set. The packaging preparation field set includes the component identifier, event type, event timestamp, and subject identifier, and generates an event sequence number. The event sequence number is a combination of the subject identifier, component identifier, and event timestamp, with an appended sequence number to distinguish multiple records under the same event timestamp.

[0028] It should be noted that the event type mapping table is determined by the field dictionary of each business record source and the business event definition and is fixed in tabular form. The table content includes the event type name and judgment condition field set and the event package information field list. The component identification field rules and timestamp field rules and the subject identification dictionary are also fixed in tabular form and referenced when configuring incremental event triggering collection rules to ensure that the standardized event field set maintains a consistent standard across different business record sources.

[0029] S1.4: Construct an event package using the encapsulated preparation field set, and write the event package header information and event package body information to obtain a complete event package; Furthermore, the encapsulation preparation field set is read and an event package is created. The event package is written with event header information and event body information. The event header information is written with component identifier, event type, event timestamp, subject identifier, and event sequence number. The event body information is written with corresponding business fields extracted from the standardized event field set according to the field list defined in the event type mapping table. After the event header information and event body information are written, a complete event package is formed.

[0030] S1.5: Perform normalized serialization on the complete event package, calculate the evidence digest, establish the correspondence between the evidence digest and the event sequence number, and obtain the event evidence digest set; Furthermore, normalized serialization is performed on the complete event package. Normalized serialization uses a fixed field order and fixed character encoding to concatenate the event package header information and event package body information into a byte sequence to be digested. An evidence digest is calculated on the byte sequence to be digested, and a correspondence is established between the evidence digest and the event sequence number. The event evidence digest set is written in an append-only manner, and each record contains the event sequence number and the evidence digest.

[0031] S1.6: Perform private key signing on the evidence digest, generate a signature identifier, and write it into the event packet evidence field. Sort the event packets and write them into the pending event packet queue to obtain the pending event packet queue. Furthermore, the event evidence digest set is read and a private key signature is performed on the evidence digest to generate a signature identifier. The signature identifier is bound to the subject identifier and written into the event packet evidence field of the complete event packet. The complete event packets written into the event packet evidence field are sorted by event timestamp and event sequence number and then appended to the queue of event packets to be processed. The entries written into the queue of event packets to be processed contain the complete event packet and the queue enqueue timestamp.

[0032] S2: Perform compliance verification and desensitization and permission labeling on the event package queue and event evidence summary set to be processed, and extract task fields to obtain the feature view to be reasoned and the sample set to be trained; S2.1: Perform consistency verification between the event packet evidence field of the event packet queue to be processed and the event evidence digest set, and perform signature validity verification to obtain the compliance verification event packet queue; Furthermore, complete event packets are read one by one from the queue of event packets to be processed, and the evidence digest and signature identifier are extracted from the evidence field of the event packet. At the same time, the complete event packet is normalized and serialized, and a verification digest is calculated. The verification digest is then checked for consistency with the evidence digest in the evidence field of the event packet. After the consistency check passes, the corresponding evidence digest record is retrieved from the event evidence digest set according to the event sequence number, and the evidence digest matching check is completed. After the evidence digest matching check passes, the signature identifier is validated based on the public key associated with the subject identifier, and the complete event packet that passes the check is written into the compliance verification event packet queue.

[0033] S2.2: Perform unified field processing on the compliance verification event package queue according to the event type mapping table, and obtain the unified event package queue by unifying the unit format and time format; Furthermore, complete event packages are read one by one from the compliance verification event package queue, and the field scope of the event package body information fields is uniformly processed according to the event type mapping table. The uniform field scope processing includes field name mapping, field value enumeration uniformity, and field unit conversion. At the same time, the event timestamp in the event package header information is uniformly processed for time format and written back to the event package header information. After the field scope uniformity processing is completed, the updated complete event package is written into the uniform field event package queue.

[0034] S2.3: Replace the plaintext fields of sensitive fields in the unified event packet queue according to the desensitization rule table, and retain the desensitization mark to obtain the desensitized event packet queue; Furthermore, complete event packets are read one by one from the unified event packet queue, and sensitive fields and plaintext fields are located according to the desensitization rule table. Sensitive fields and plaintext fields are replaced using the replacement method defined in the desensitization rule table, and a desensitization mark field is written into the event packet body information. The desensitization mark field records the status of the sensitive fields that have been replaced and the replacement type identifier. After the desensitization process is completed, the updated complete event packet is written into the desensitized event packet queue.

[0035] S2.4: Write the de-identified event package queue into the permission label field according to the permission policy table, and bind the permission label field to the event type and subject identifier to obtain the permission label event package queue; Furthermore, complete event packages are read one by one from the de-identified event package queue, and the values ​​of the permission labeling fields are calculated according to the permission policy table. The permission policy table outputs the permission level and the range of accessible fields using the event type and the subject identifier as indexes. The permission labeling fields are written into the event package header information and the range of accessible fields are written into the event package body information. After the permission labeling fields are bound to the event type and the subject identifier, the complete event package is written into the permission labeling event package queue.

[0036] S2.5: Based on the permission-labeled event packet queue and task field mapping table, extract task fields and generate a feature view to be reasoned that can be used for real-time reasoning; Furthermore, complete event packets are read one by one from the permission-annotated event packet queue, and task fields are extracted according to the task field mapping table. The task field mapping table outputs a list of fields to be inferred and field position mapping rules, indexed by task identifier and event type. Fields to be inferred are extracted from the event packet header and event packet body information according to the field position mapping rules, and records to be inferred are formed according to the event sequence number. At the same time, the event timestamp is retained for time-series sorting. After the records to be inferred are summarized, they are sorted by event timestamp and written into the feature view to be inferred.

[0037] S2.6: Extract training fields from the permission labeling event packet queue according to the task field mapping table, and aggregate them into a training sample set according to the time window; Furthermore, complete event packages are read one by one from the permission-annotated event package queue, and training fields are extracted according to the task field mapping table. The training field mapping rules output a list of training fields and a list of sample key fields. The sample key fields include component identifiers, event types, and time window identifiers. Time window aggregation is performed on the training fields using the sample key fields to generate training sample records. The training sample records are written to the training sample set and retain the association index with the event sequence number for traceability.

[0038] It should be noted that the event type mapping table is fixed in tabular form and includes event type names and field definition consistency rules, field unit conversion rules, and time format consistency rules. The desensitization rule table is fixed in tabular form and includes sensitive field names, replacement method identifiers, and desensitization mark writing rules. The permission policy table is fixed in tabular form and includes the mapping relationship between event type and subject identifier to permission level and accessible field range identifier. The task field mapping table is fixed in tabular form and includes the mapping relationship between task identifier and event type to the list of fields to be inferred and the list of training fields.

[0039] S3: Perform real-time inference using the feature view to be inferred and the effective model version, and calculate the uncertainty and drift to obtain real-time prediction results and evidence candidate queue; S3.1: Load the effective model version that matches the task identifier of the feature view to be inferred, forming the real-time inference input; Furthermore, the task identifier is read from the feature view to be inferred and the effective model version registration record is accessed using the task identifier as the retrieval key. The effective model version registration record returns the effective model version identifier that matches the task identifier and the index of the effective model version storage location. The effective model version is loaded according to the index of the effective model version storage location and the record to be inferred in the feature view to be inferred is read. The record to be inferred and the event sequence number are bound in the same data structure to form the real-time inference input.

[0040] S3.2: Perform real-time inference on the real-time inference input to obtain the risk score and risk category, and write them into the event sequence number association field to obtain the real-time prediction result; Furthermore, the real-time inference input is fed into the effective model version to perform real-time inference and output risk scores and risk categories. The risk scores and risk categories are bound to the event sequence numbers in the output stage and written into the event sequence number association field. The risk scores and risk categories containing the event sequence number association field are summarized and written into the real-time prediction results.

[0041] S3.3: Calculate the uncertainty based on the real-time prediction results, and establish a one-to-one correspondence between the uncertainty and the event sequence number to obtain the uncertainty result set; Furthermore, the confidence level output corresponding to the risk category is read from the real-time prediction results, and the uncertainty is calculated using the confidence level dispersion. The confidence level dispersion is measured by the concentration of the confidence level distribution of the risk category, and the uncertainty value is output. A one-to-one correspondence is established between the uncertainty value and the event sequence number and written into the uncertainty result set.

[0042] S3.4: Update the drift amount within the sliding time window based on the feature view to be inferred, and associate the drift amount with the event type to obtain the drift amount result set; Furthermore, after reading the records to be inferred from the feature view and grouping them by event type, the sliding time window update process is entered. The sliding time window update process measures the difference between the key field distribution of the records to be inferred in the current time window and the key field distribution in the reference time window and outputs the drift value. The drift value is then associated with the event type and written into the drift result set.

[0043] S3.5: Based on the uncertainty result set, drift result set and evidence collection triggering rules, filter the event sequence number that needs supplementary evidence, generate evidence collection request identifier, and obtain the evidence collection candidate queue; Furthermore, the uncertainty result set and drift result set are read and filtered according to the evidence collection triggering rules. The evidence collection triggering rules are indexed by event type and provide uncertainty triggering conditions and drift triggering conditions. When the uncertainty result set meets the uncertainty triggering condition or the drift result set meets the drift triggering condition, the corresponding event sequence number is written into the evidence collection candidate queue and an evidence collection request identifier is generated for each record. After the evidence collection request identifier is bound to the event sequence number, it is written into the evidence collection candidate queue.

[0044] It should be noted that the effective model version registration record is used to maintain the correspondence between the task identifier, the effective model version identifier, and the index of the effective model version storage location. The evidence collection triggering rule is used to maintain the correspondence between the event type, the uncertainty triggering condition, and the drift triggering condition. The difference measurement adopts a statistical calculation method based on distribution comparison and uses a table to solidify the range of key fields and the update method of the reference time window to ensure the consistency of drift calculation.

[0045] S4: After allocating the evidence collection budget and collecting evidence collection packages according to the evidence collection candidate queue and completing the review, generate the label backflow and credibility weight, and write the label backflow into the training sample set to obtain the incremental training sample set. S4.1: Based on the evidence collection candidate queue and combined with the evidence collection budget allocation rules, calculate the evidence collection budget amount corresponding to each evidence collection candidate and obtain the evidence collection budget allocation result set; Furthermore, the event sequence number, evidence request identifier, event type, and subject identifier are read one by one from the evidence candidate queue, and the evidence budget allocation rule is accessed. The evidence budget allocation rule uses the event type, uncertainty trigger state, and drift trigger state as input and output evidence budget amount calculation elements. The budget amount calculation is performed on the evidence budget amount calculation elements, and the budget amount is bound to the event sequence number and evidence request identifier and written into the evidence budget allocation result set.

[0046] S4.2: Based on the evidence collection budget allocation result set, generate an evidence collection request and send it to the evidence collection target entity, obtain the evidence collection materials matching the event sequence number, and obtain the evidence collection package; Furthermore, the event sequence number, evidence request identifier, and budget amount are read one by one from the evidence collection budget allocation result set, and the evidence material field list is located according to the event type mapping table. The evidence material field list limits the content scope and format requirements of the evidence materials. An evidence collection request is generated based on the evidence material field list and sent to the evidence collection target entity. The evidence collection target entity is determined by the entity identifier and the evidence materials matching the event sequence number are returned. The evidence materials, event sequence number, and evidence request identifier are encapsulated into an evidence collection package and written to the evidence collection package storage area.

[0047] S4.3: Perform integrity verification and evidence digest consistency verification on the evidence package, and perform signature validity verification to obtain the verified evidence package; Furthermore, the evidence package is read from the evidence package storage area and its integrity is verified. The integrity verification performs existence and format consistency checks on the fields corresponding to the field list of evidence materials and outputs the integrity verification results. For evidence packages that pass the integrity verification, the evidence digest consistency check is performed. The evidence digest consistency check performs a matching check on the evidence digest in the evidence package and the evidence digest corresponding to the event sequence number in the event evidence digest set. For evidence packages that pass the evidence digest consistency check, the signature validity check is performed. The signature validity check verifies the signature based on the public key associated with the subject identifier and outputs the verification results. The verified evidence packages are written into the set of verified evidence packages.

[0048] S4.4: Based on the verified evidence package, generate a tag return stream and bind the tag return stream with the event sequence number to form a tag return stream record; Furthermore, the review process involves reading each piece of evidence collected from the evidence package collection and extracting tag fields according to the tag generation rules. The tag generation rules are indexed by the event type and define the mapping relationship between the tag field source and the tag value. Tag return streams are generated based on the tag field source and tag value mapping relationship, and the tag return streams are bound to the event sequence number and written into the tag return stream record.

[0049] S4.5: Calculate the credibility weight for the tag return records according to the credibility assessment rules, and bind the credibility weight to the tag return records; Furthermore, the event sequence number and the verification results of tag return and evidence collection materials are read one by one from the tag return record, and the credibility assessment rules are accessed. The credibility assessment rules take the integrity verification results, evidence digest consistency verification results, and signature validity verification results as inputs and output credibility weights. The credibility weights are then bound to the tag return records and written back to the tag return records.

[0050] S4.6: Write the label backflow record into the training sample set, and simultaneously write the credibility weight field bound to the label backflow record to form an incremental training sample set; Furthermore, the training sample set is read and the training sample record is located based on the event sequence number and associated index. The label backflow record after binding the credibility weight is written into the label field of the training sample record and the credibility weight field is written simultaneously. The label backflow record that cannot match the event sequence number is appended to generate a training sample record according to the component identifier, event type and time window identifier and the same writing is completed. The training sample record set after writing is summarized into an incremental training sample set.

[0051] It should be noted that the evidence collection budget allocation rules are fixed in tabular form and include the mapping relationship between event types and budget calculation elements; the evidence collection material field list is fixed in tabular form and includes the mapping relationship between event types and the scope and format requirements of evidence collection material fields; the tag generation rules are fixed in tabular form and include the mapping relationship between event types and the source of tag fields and the tag values; and the credibility assessment rules are fixed in tabular form and include the mapping relationship between the combination of verification results and the credibility weight.

[0052] S5: Perform local incremental training using the incremental training sample set and confidence weights, and generate a local model update package; S5.1: Perform field completeness and label consistency checks on the incremental training sample set to obtain the training-ready sample set; Furthermore, training sample records are read one by one from the incremental training sample set, and the training field list is located according to the task field mapping table. Field completeness verification performs existence and format consistency checks on the corresponding fields in the training field list and outputs the field completeness verification results. Label consistency verification performs matching checks on the mapping relationship between the label fields of the training sample records and the label values ​​corresponding to the event types and outputs the label consistency verification results. Training sample records that pass both the field completeness verification and label consistency verification results are written into the training ready sample set.

[0053] S5.2: Load the effective model version that matches the task identifier of the training-ready sample set, and establish the trainable parameter set and training round identifier to obtain the incremental training initialization state; Furthermore, the task identifier is read from the training-ready sample set and used as the retrieval key to access the effective model version registration record. The effective model version registration record returns the effective model version identifier that matches the task identifier and the index of the effective model version storage location. Based on the index of the effective model version storage location, the effective model version is loaded and the list of trainable parameter identifiers of the effective model version is read to establish the trainable parameter set. At the same time, the training round identifier is generated and written to the training process record area to associate with loss convergence information. The effective model version, the trainable parameter set, and the training round identifier are summarized to form the incremental training initialization state.

[0054] S5.3: Perform weighted sampling on the training-ready sample set according to the confidence weight, and construct incremental training batches to obtain the incremental training batch set; Furthermore, training sample records are read from the training-ready sample set and the corresponding confidence weight field is read. The confidence weight is used as the sampling weight to perform weighted sampling to form a training sample selection sequence. The training sample records are allocated to multiple training batches according to the training sample selection sequence, and a batch identifier, task identifier, and training round identifier association field are written to each training batch. After the training batches are written, they are summarized to obtain an incremental training batch set.

[0055] S5.4: Use the incremental training batch set to perform local incremental training in the incremental training initialization state, and record the loss convergence information in each training round to obtain the updated local model parameters; Furthermore, the incremental training batch set is input into the incremental training initialization state in the order of batch identifiers and local incremental training is executed. Local incremental training updates the parameters for each training batch and records the loss convergence information corresponding to the training round identifier in the training process recording area. When all training batches in the incremental training batch set have been processed, the updated local model parameters are output.

[0056] S5.5: Perform differential calculation on the updated local model parameters and the effective model version to generate an update summary. Bind the update summary with the task identifier to obtain the local model update difference and update summary. Furthermore, the updated local model parameters are read, and the baseline parameters corresponding to the trainable parameter set in the effective model version are read. The updated local model parameters and the baseline parameters are calculated to obtain the local model update difference. An update summary is generated from the local model update difference, and the update summary is bound to the task identifier and written into the update summary record area. After the update summary record area is written, the local model update difference and update summary are obtained.

[0057] S5.6: Encapsulate the local model update difference and update summary into a local model update package; Furthermore, the local model update difference and update summary are read and packaged into a local model update package. The local model update package is written with the task identifier, effective model version identifier, training round identifier, local model update difference and update summary, and after consistency verification, it is written to the update package storage area.

[0058] It should be noted that the field completeness check is performed based on the list of training fields defined in the task field mapping table, the label consistency check is performed based on the label value mapping relationship defined in the label generation rules, the weighted sampling is performed using a random sampling method according to the weight ratio, the local incremental training is performed using a gradient descent-type optimization method and the trainable parameter set is determined by the trainable parameter identifier list of the effective model version, and the update summary includes the training sample quantity statistics, training round identifier and loss convergence information summary and is used for the audit record in the subsequent aggregation stage.

[0059] S6: Perform hierarchical federated aggregation on the local model update package, generate verifiable aggregation proofs, write them to the model version ledger, and publish the effective model version and registration records. S6.1: Receive the local model update packages uploaded by each participant, and aggregate and group them according to the supply chain topology mapping rules and align them with the time window to obtain the aggregated group update set; Furthermore, the system receives local model update packages uploaded by each participant and reads the task identifier, effective model version identifier, training round identifier, and update summary within the local model update package. Simultaneously, it performs an integrity check on the local model update package to confirm the availability of the local model update difference and update summary fields. Based on the supply chain topology mapping rules, it maps the subject identifier to the edge aggregation node identifier and completes the aggregation grouping according to the edge aggregation node identifier. At the same time, it performs time window alignment based on the training round identifier, removes unaligned entries, and writes them into the aggregation group update set.

[0060] S6.2: Perform the first-level aggregation on the edge aggregation node for the aggregation group update set, generate the regional aggregation results and regional aggregation summary, and combine them into a regional model update package; Furthermore, the local model update package of the same group is read from the aggregation group update set according to the edge aggregation node identifier, and the local model update difference and update summary are extracted. The first-level aggregation performs aggregation operation on the local model update difference of the same group to form a regional aggregation result, and generates a regional aggregation summary for the participating subject identifier set and the update summary of the participating local model update package. The regional aggregation result and the regional aggregation summary are written into the regional model update package and attached with the edge aggregation node identifier, training round identifier and task identifier.

[0061] S6.3: Upload the regional model update package to the consortium aggregation node and perform second-level aggregation to form candidate global model versions and a global aggregation summary; Furthermore, the regional model update package is uploaded to the consortium aggregation node, and a consistency check is performed on the uploaded regional model update package to confirm that the task identifier and the training round identifier are aligned. The second-level aggregation performs aggregation operation on the regional aggregation results corresponding to the same task identifier and the same training round identifier to form candidate global model versions, and generates a global aggregation summary for the regional aggregation summaries of the regional model update packages.

[0062] Perform aggregation operations to generate candidate global model versions, expressed as: ; in, The candidate global model version is identified as follows in the training rounds. The corresponding parameter set, The candidate global model version is identified as follows in the training rounds. The corresponding parameter set, To update the package index set for the region models participating in the second-level aggregation, Update the package index for the region model. Aggregate weights, Update the package index for the region model. The parameter difference set corresponding to the region aggregation result.

[0063] S6.4: Generate verifiable aggregation proofs for the first-level aggregation process and the second-level aggregation process, and establish a binding relationship between the verifiable aggregation proofs and the global aggregation summary to obtain verifiable aggregation proofs; Furthermore, in the first-level aggregation stage, the summary of the aggregated group update set entries, the summary of the participating entity identifier set and the summary of the regional aggregation results are recorded, and a first-level aggregation proof record is generated. In the second-level aggregation stage, the summary of the participating regional model update package and the global aggregation summary are recorded, and a second-level aggregation proof record is generated. The first-level aggregation proof record and the second-level aggregation proof record are merged to form a verifiable aggregation proof, and the verifiable aggregation proof and the global aggregation summary are bound and written into the proof storage area.

[0064] S6.5: Write the version identifier, global aggregation summary, and verifiable aggregation proof of the candidate global model version into the model version ledger to generate a queryable effective model version registration record; Furthermore, a version identifier is generated for the candidate global model version and bound to the task identifier and training round identifier. At the same time, the global aggregation summary and verifiable aggregation proof are read and written into the model version ledger. After the model version ledger is successfully written, an effective model version registration record is generated and written into the registration record storage area.

[0065] S6.6: Based on the effective model version registration record, publish the candidate global model version as the effective model version, and distribute the effective model version and the effective model version registration record to each participant to obtain the effective model version and the effective model version registration record; Furthermore, the effective model version registration record is read, and the candidate global model version pointed to by the version identifier in the effective model version registration record is marked as the effective model version. At the same time, the effective model version and the effective model version registration record are distributed to each participant, and the local cache of the effective model version registration record is updated in each participant to support subsequent loading.

[0066] It should be noted that the supply chain topology mapping rule is used to maintain the mapping relationship from the main body identifier to the edge aggregation node identifier and solidify it in tabular form. The time window alignment uses the training round identifier as the alignment key and solidifies the allowed alignment window strategy in tabular form. The first-level aggregation proof record and the second-level aggregation proof record use a summary chain method to record the summary of the aggregation input and aggregation output for verification. The global aggregation summary includes the summary set of the model update package of the participating region, the aggregation operation identifier, and the output candidate global model version identifier to support model version ledger query.

[0067] This embodiment also provides a federated learning-based automotive parts lifecycle management system, including: an event encapsulation module, used to collect automotive parts lifecycle event stream data and encapsulate it into event packages, and generate evidence summaries and signature identifiers to obtain a queue of event packages to be processed and a set of event evidence summaries; a governance construction module, used to perform compliance verification, unified desensitization, and permission labeling on the queue of event packages to be processed and the set of event evidence summaries, and extract task fields to obtain a feature view to be inferred and a training sample set; and an inference triggering module, used to perform real-time inference using the feature view to be inferred and the effective model version, and calculate... The system calculates uncertainty and drift to obtain real-time prediction results and a candidate queue for evidence collection; the evidence collection feedback module allocates the evidence collection budget based on the candidate queue, collects evidence collection packages, completes verification, generates label feedback and credibility weights, and writes the label feedback into the training sample set to obtain the incremental training sample set; the incremental training module performs local incremental training using the incremental training sample set and credibility weights, and generates a local model update package; the aggregation and evidence storage module performs hierarchical federated aggregation on the local model update package, generates verifiable aggregation proofs, writes them into the model version ledger, and then publishes the effective model version and registration records.

[0068] In summary, this invention achieves targeted evidence supplementation and noise reduction training in scenarios with few labels by: using the evidence collection and feedback step to allocate evidence collection budgets from the evidence candidate queue obtained through real-time inference and collecting evidence packages for review, generating label feedback and credibility weights to form an incremental training sample set; and through the aggregation and evidence storage step to perform hierarchical federated aggregation of local model update packages and generate verifiable aggregation proofs, which are written into the model version ledger to publish and activate the model version and registration records, thereby enabling traceable and verifiable model evolution and supporting cross-entity full lifecycle collaborative management.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for managing the entire lifecycle of automotive parts based on federated learning, characterized by: include, Collect event stream data of the entire life cycle of automotive parts and encapsulate it into event packages, and generate evidence digests and signature identifiers to obtain a queue of event packages to be processed and a set of event evidence digests; The event package queue and event evidence summary set to be processed are subjected to compliance verification, unified desensitization and permission labeling, and task fields are extracted to obtain the feature view to be reasoned and the sample set to be trained; Real-time inference is performed using the feature view to be inferred and the effective model version, and the uncertainty and drift are calculated to obtain real-time prediction results and evidence candidate queue; After allocating the evidence collection budget and collecting evidence collection packages according to the evidence collection candidate queue and completing the review, the label backflow and credibility weight are generated, and the label backflow is written into the training sample set to obtain the incremental training sample set. Local incremental training is performed using an incremental training sample set and confidence weights, and a local model update package is generated. Perform hierarchical federated aggregation on the local model update package, generate verifiable aggregation proofs, write them into the model version ledger, and then publish the effective model version and registration records.

2. The method for full lifecycle management of automotive parts based on federated learning as described in claim 1, characterized in that: The specific steps for collecting and encapsulating event stream data throughout the entire lifecycle of automotive components into event packages are as follows. Collect event stream data of the entire life cycle of automotive parts, and read the newly added record payload according to the incremental event triggering collection rules, and write it into the collection temporary storage area; The newly added recorded loads in the data collection and storage area are processed to standardize component identification, unify timestamps and main body identification, resulting in a standardized set of event fields. The event type is determined based on the standardized event field set and the event type mapping table, and a set of encapsulation preparation fields is generated. An event package is constructed by encapsulating a set of preparation fields, and the event package header and body information are written to obtain a complete event package.

3. The method for full lifecycle management of automotive parts based on federated learning as described in claim 2, characterized in that: The specific steps for obtaining the queue of event packets to be processed and the set of event evidence digests are as follows. Perform normalized serialization on the complete event package, calculate the evidence digest, establish the correspondence between the evidence digest and the event sequence number, and obtain the event evidence digest set; Perform private key signing on the evidence digest, generate a signature identifier, and write it into the event packet evidence field. Sort the event packets and write them into the pending event packet queue to obtain the pending event packet queue.

4. The method for full lifecycle management of automotive parts based on federated learning as described in claim 3, characterized in that: The process of performing compliance verification and unified desensitization and permission labeling on the queue of event packets to be processed and the set of event evidence digests involves the following steps: Perform consistency checks between the event packet evidence fields in the pending event packet queue and the event evidence digest set, and perform signature validity checks to obtain the compliance check event packet queue; Perform unified field processing on the compliance verification event package queue according to the event type mapping table, and obtain a unified event package queue by unifying the unit format and time format; According to the desensitization rule table, replace the plaintext fields of the sensitive fields in the unified event packet queue, and retain the desensitization mark to obtain the desensitized event packet queue; Write the de-identified event package queue into the permission label field according to the permission policy table, and bind the permission label field to the event type and subject identifier to obtain the permission label event package queue.

5. The method for full lifecycle management of automotive parts based on federated learning as described in claim 4, characterized in that: The specific steps for obtaining the feature view to be inferred and the training sample set are as follows. Based on the permission-labeled event packet queue and task field mapping table, extract task fields and generate a feature view to be reasoned that can be used for real-time reasoning; The training fields are extracted from the permission labeling event packet queue according to the task field mapping table, and then aggregated into a training sample set according to the time window.

6. The method for full lifecycle management of automotive parts based on federated learning as described in claim 5, characterized in that: The specific steps for obtaining the real-time prediction results and the evidence candidate queue are as follows. Load the effective model version that matches the task identifier of the feature view to be inferred, forming the real-time inference input; Real-time inference is performed on the real-time inference input to obtain risk scores and risk categories, which are then written into the event sequence number association field to obtain real-time prediction results; The uncertainty is calculated based on the real-time prediction results, and a one-to-one correspondence is established between the uncertainty and the event sequence number to obtain the uncertainty result set; The drift amount is updated within the sliding time window based on the feature view to be inferred, and the drift amount is associated with the event type to obtain the drift amount result set; Based on the uncertainty result set, drift result set and evidence collection triggering rules, the event sequence number that needs supplementary evidence is selected, and an evidence collection request identifier is generated to obtain the evidence collection candidate queue.

7. The method for full lifecycle management of automotive parts based on federated learning as described in claim 6, characterized in that: The specific steps for obtaining the incremental training sample set are as follows: Based on the evidence collection candidate queue and combined with the evidence collection budget allocation rules, the evidence collection budget amount corresponding to each evidence collection candidate is calculated, and the evidence collection budget allocation result set is obtained. Based on the evidence collection budget allocation result set, generate an evidence collection request and send it to the evidence collection target entity to obtain evidence collection materials that match the event sequence number and obtain the evidence collection package; The integrity and evidence digest consistency checks are performed on the evidence package, and the signature validity check is performed to obtain the approved evidence package; Based on the verified evidence package, a tag return stream is generated and the tag return stream is bound to the event sequence number to form a tag return stream record; The credibility weight of the tag return record is calculated according to the credibility assessment rules, and the credibility weight is bound to the tag return record; The label backflow records are written into the training sample set, and the credibility weight field bound to the label backflow records is written simultaneously to form an incremental training sample set.

8. The method for full lifecycle management of automotive parts based on federated learning as described in claim 7, characterized in that: The specific steps for generating the local model update package are as follows. Perform field completeness and label consistency checks on the incremental training sample set to obtain the training-ready sample set. Load the effective model version that matches the task identifier of the training-ready sample set, and establish the trainable parameter set and training round identifier to obtain the incremental training initialization state. Weighted sampling is performed on the training-ready sample set according to the confidence weight, and incremental training batches are constructed to obtain the incremental training batch set; Local incremental training is performed on the incremental training initialization state using an incremental training batch set, and loss convergence information is recorded in each training round to obtain the updated local model parameters; Perform differential calculation on the updated local model parameters and the effective model version to generate an update summary. Bind the update summary with the task identifier to obtain the local model update difference and update summary. The local model update difference and update summary are encapsulated into a local model update package.

9. The method for full lifecycle management of automotive parts based on federated learning as described in claim 8, characterized in that: The specific steps for releasing and registering the effective model version and records are as follows. Receive local model update packages uploaded by each participant, and aggregate and group them according to the supply chain topology mapping rules and align them with the time window to obtain an aggregated group update set; The first-level aggregation is performed on the aggregation group update set at the edge aggregation node to generate the regional aggregation results and regional aggregation summary, which are then combined into a regional model update package; The regional model update package is uploaded to the consortium aggregation node, and a second-level aggregation is performed to generate candidate global model versions and a global aggregation summary. Verifiable aggregation proofs are generated for the first-level aggregation process and the second-level aggregation process, and the verifiable aggregation proofs are bound to the global aggregation summary to obtain verifiable aggregation proofs. Write the version identifier, global aggregation summary, and verifiable aggregation proof of the candidate global model version into the model version ledger to generate a queryable effective model version registration record; Based on the effective model version registration record, the candidate global model version is published as the effective model version, and the effective model version and the effective model version registration record are distributed to each participating party to obtain the effective model version and the effective model version registration record.

10. A federated learning-based automotive parts lifecycle management system, based on the federated learning-based automotive parts lifecycle management method according to any one of claims 1 to 9, characterized in that: include, The event encapsulation module is used to collect event stream data of the entire life cycle of automotive parts and encapsulate it into event packages, and generate evidence digests and signature identifiers to obtain a queue of event packages to be processed and a set of event evidence digests. The governance module is used to perform compliance verification, unify desensitization and permission labeling on the queue of event packages to be processed and the set of event evidence summaries, and extract task fields to obtain the feature view to be reasoned and the sample set to be trained. The inference triggering module is used to perform real-time inference using the feature view to be inferred and the effective model version, and to calculate the uncertainty and drift to obtain real-time prediction results and evidence candidate queue. The evidence feedback module is used to allocate the evidence budget according to the evidence candidate queue, collect evidence packets and complete the review, generate label feedback and credibility weights, and write the label feedback into the training sample set to obtain the incremental training sample set. The incremental training module is used to perform local incremental training using an incremental training sample set and confidence weights, and to generate a local model update package. The aggregation and evidence storage module is used to perform hierarchical federated aggregation on local model update packages, generate verifiable aggregation proofs, write them into the model version ledger, and then publish and register the effective model version and registration records.

Images