CN-121706993-B - Industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method

CN121706993BCN 121706993 BCN121706993 BCN 121706993BCN-121706993-B

Abstract

The embodiment of the invention provides an industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method, and belongs to the technical field of intelligent industry. The method forms light representation of the abrupt change period on the edge side in a structured log and window statistics mode, only returns a key sample and a key context when a trigger condition is met, so that return bandwidth and return cost are reduced, meanwhile, weak supervision and enhancement instructions are provided for the abrupt change sample through cloud cognitive write-back, the light edge model has low-cost self-training feasibility, further, damage to iteration by noise samples is restrained through a layered conflict buffer and dynamic gate mechanism, stability after the online is ensured through conflict suppression and stability keeping constraint, and finally reusable continuous optimization is formed through the effect association of a new reasoning log and a return packet, so that edge reasoning misjudgment rate is reduced, reasoning stability is improved, and instantaneity and light deployment constraint are considered.

Inventors

ZENG MING
ZHONG XIAOQING
HUANG XIAOYU
Zhou Yefen
LIU JUNJIE
Shui Yuanfei
ZHANG HUAN
LI BO
LI WEI

Assignees

工业云制造(四川)创新中心有限公司

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (9)

1. An industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method is characterized by comprising the following steps of: acquiring continuous input data of industrial edge nodes, performing edge reasoning processing one by one to generate a reasoning result, and writing the reasoning result into a reasoning log; performing window statistics processing according to the time sliding window based on the reasoning log to generate window statistics results of each time window and mutation digests representing short-time mutation degrees; when the mutation abstract meets a preset triggering condition, executing key sample extraction processing, key log cutting processing and conflict sample judgment processing from continuous input data and reasoning logs in a corresponding mutation period, and packaging the key sample extraction processing, the key log cutting processing and the conflict sample judgment processing with the mutation abstract to form a mutation reflux packet; Cloud cognitive write-back processing is carried out on the mutation reflow package to generate a weak tag corresponding to the key sample, a weak tag confidence coefficient and an enhanced instruction set, so as to form a write-back mutation reflow package; Constructing a patch training set based on the back-written mutation reflux packet, and executing incremental training processing to generate an updated edge reasoning model version, wherein the method specifically comprises the steps of executing hierarchical conflict buffer processing and sample score calculation processing based on the back-written mutation reflux packet, executing dynamic gating screening processing based on a sample score calculation result, executing risk constraint weighting processing based on the dynamic gating screening result, constructing the patch training set, executing incremental training on the patch training set, and issuing back-writing and multiplexing scheduling; And performing deployment packaging processing and gray level release processing on the updated edge reasoning model version to generate a new reasoning log, and performing effect write-back and multiplexing scheduling based on the time range association relationship between the new reasoning log and the abrupt change reflux packet.
2. The industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method according to claim 1, wherein obtaining continuous input data of industrial edge nodes, performing edge reasoning processing piece by piece to generate reasoning results, and writing the reasoning results into a reasoning log, comprises: acquiring continuous input data of each industrial edge node, generating an input number and a time stamp for associated tracing for each piece of input data, and forming an input record with an index; Performing edge reasoning processing on the indexed input records one by one to generate a prediction result, a prediction confidence coefficient and reasoning time consumption corresponding to each piece of input data; And combining the input number, the timestamp, the prediction result, the prediction confidence, the reasoning time consumption and the model version number into a single log record, and converging to obtain a reasoning log.
3. The industrial scenario-oriented lightweight model cloud edge collaborative self-training evolution method according to claim 1, wherein performing window statistics processing according to a temporal sliding window based on the inference log to generate window statistics results for each temporal window and mutation digests characterizing short-time mutation levels, comprises: performing sliding segmentation processing on the reasoning log according to a preset window length and a preset step length to obtain a plurality of time windows arranged in time sequence; For any time window, performing abnormal output density statistics, high confidence abnormal duty ratio statistics, confidence mean statistics and reasoning time-consuming mean statistics based on log records in the time window respectively to generate window statistics results of the time window; performing a stable baseline generation process based on window statistics corresponding to the time window determined to be stable, to obtain a stable baseline; for any time window, performing deviation calculation processing based on a window statistical result of the time window and the stable baseline to obtain a mutation score of the time window; and combining the mutation score with the window statistical result of the time window and the stable baseline to generate a mutation abstract.
4. The industrial scene oriented lightweight model cloud edge collaborative self-training evolution method according to claim 1, wherein when the abrupt digest meets a preset trigger condition, performing key sample extraction processing, key log clipping processing and conflict sample judgment processing from continuous input data and reasoning logs in a corresponding abrupt period, and packaging the same as the abrupt digest into an abrupt reflux packet, comprising: identifying a mutation time period meeting a preset triggering condition based on the mutation abstract, and determining a time window range corresponding to the mutation time period; In the mutation period, extracting high-confidence abnormal samples from the continuous input data based on an inference confidence sequencing rule, extracting a prediction turnover frequent sample based on an output instability sequencing rule, extracting time-consuming abnormal samples based on an inference time-consuming sequencing rule, and combining and removing the time-consuming abnormal samples to form a key sample; Cutting log records corresponding to the key samples from the reasoning logs in the abrupt change period to form key logs; Performing conflict rule judgment processing based on the key logs to form a conflict sample list, wherein the conflict rule judgment processing at least comprises edge cloud conclusion inconsistent triggering, short neighborhood output high-frequency flip triggering and abrupt period abnormal output flooding triggering; And packaging the key samples, the key logs, the conflict sample list and the mutation abstract into mutation reflux packages, and writing the mutation reflux packages into a cloud reflux queue.
5. The industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method of claim 1, wherein the performing cloud cognitive write-back processing on the abrupt change reflux packet to generate a weak tag, a weak tag confidence coefficient and an enhanced instruction set corresponding to the key sample, forming a written-back abrupt change reflux packet comprises: performing structured finishing on the key samples, the key logs, the conflict sample list and the mutation abstract in the mutation reflux packet to form a cloud write-back input packet; Performing weak tag generation processing on the key samples piece by piece based on the cloud write-back input packet, and generating corresponding weak tag confidence and reason abstracts for each weak tag; Executing enhancement instruction generation processing based on the cloud write-back input packet to form an enhancement instruction set consisting of directly executable data enhancement actions and parameter ranges thereof; Writing the weak tag, the weak tag confidence, the reason abstract and the enhanced instruction set into a write-back slot of the mutation reflux packet to form a mutation reflux packet after write-back.
6. The industrial scene oriented lightweight model cloud edge collaborative self-training evolution method according to claim 5, wherein the hierarchical conflict buffer processing and sample score computing processing are executed based on the back-written mutation reflux packet, and the method comprises the following steps: Establishing a unified sample index record for each key sample in the re-written mutation reflux packet and converging the unified sample index record into an index table; performing hierarchical buffer splitting processing based on the conflict sample list and conflict or unstable marks in the unified sample index record to form a trusted pool, a pending pool and a conflict pool; And executing edge cloud consistency score calculation processing, time consistency score calculation processing, burst consistency score calculation processing and conflict strength score calculation processing on the samples in the index table, and writing the score result back to the index table.
7. The industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method of claim 6, wherein performing dynamic gating screening processing based on the sample score computation results comprises: Performing a confidence calibration parameter solving process based on the stationary phase samples to obtain confidence calibration parameters, and converting the edge prediction confidence level into a calibrated confidence level; calculating a dynamic gating threshold based on the mutation score, performing a gating screening process, performing a joint gating decision process on the samples to generate a final trusted set, a final pending set, and a final conflict pool; The samples entering the final trusted set meet the conditions that the conflict strength score does not exceed a dynamic conflict threshold, the edge cloud consistency is non-negative, the time consistency is not lower than a preset threshold, and the confidence after calibration and the weak tag confidence are not lower than corresponding dynamic gating thresholds.
8. The industrial scene oriented lightweight model cloud edge collaborative self-training evolution method of claim 1, wherein performing risk constraint weighting processing and constructing a patch training set based on the dynamic gating screening result comprises: calculating training weights for samples in the final trusted set; performing enhanced sample generation processing on the final trusted set based on the enhanced instruction set, generating an enhanced sample set, and inheriting or attenuating the training weight for the enhanced sample according to a preset rule; Merging the final trusted set and the enhanced sample set and performing de-duplication processing to form a patch training set; Constructing conflict suppression constraint terms for samples in the final conflict pool, and constructing stability retention constraint terms for a stable reference sample set.
9. The industrial scene oriented lightweight model cloud edge collaborative self-training evolution method of claim 8, wherein performing incremental training on a patch training set and performing release write-back and reuse scheduling comprises: performing incremental training processing based on the patch training set, the conflict suppression constraint term and the stability maintenance constraint term to generate an updated edge reasoning model version; performing deployment packaging processing on the updated edge reasoning model version, and performing gray level release processing on the edge node to generate a new reasoning log; Performing window statistics and mutation score calculation processing according to the statistics caliber consistent with the mutation abstract based on the new reasoning log, and associating calculation results to corresponding mutation reflow packages based on a time range and a model version number; Writing the associated calculation result into the effect field of the mutation reflux packet to form a mutation reflux packet library with effect marks, and executing reflux priority generation processing and training scheduling strategy generation processing for construction and training scheduling of the subsequent mutation reflux packet.

Description

Industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method Technical Field The invention relates to the technical field of intelligent industry, in particular to an industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method. Background The task of sensing and discriminating an industrial site is usually deployed at an edge-side terminal to perform, for example, real-time identification of state changes, abnormal signs, or quality risks in a production process. The scene has the common characteristics that firstly, edge side computing power and storage resources are limited, a model needs to be kept light to meet the requirements of reasoning time delay and stable operation, secondly, field data distribution drifts along with factors such as working conditions, material batches, equipment states, environmental conditions and the like, so that the problems of rising of misjudgment rate, confidence degree distortion or abnormal fluctuation of reasoning time consumption and the like occur after the model operates for a period of time, thirdly, the industrial field is highly sensitive to misjudged business cost, the misjudgment cost difference of different categories is obvious, and the actual requirement is difficult to meet by simply taking the whole accuracy as a unique index. In the prior art, an offline centralized training and periodic manual updating mode is generally adopted, wherein data are collected in a cloud end in a centralized mode and are subjected to annotation training, and an updated model is sent to an edge node again. However, in industrial scenarios, data distribution drift tends to be bursty and periodic, and there are a number of special cases in the field where early coverage is difficult. If the device still relies on manual recovery, manual cleaning, manual labeling and centralized training in a fixed period, the device has the disadvantages of long updating period, high cost and incapability of timely restraining sudden erroneous judgment. On the other hand, some existing methods attempt to directly perform self-training or incremental learning on the edge side, but due to the lack of reliable label sources on the edge side and the lack of effective identification and data organization capability for the abrupt period, false self-training, catastrophic forgetting or false reinforcement on conflicting samples are easy to occur, so that the false judgment rate cannot be reduced under certain abrupt conditions, and new risks are introduced. In addition, the data of the industrial field mutation period often presents short-time flooding characteristics, namely abnormal output intensively bursts in a short window, a large number of logs and samples are generated by edge equipment, bandwidth and storage pressure are brought if the total quantity of logs is directly returned, and key conflict samples are easily missed if the samples are simply sampled, so that effective write-back and guidance cannot be formed by a cloud. Part of the existing schemes can be relieved by adding more complex edge models or introducing multi-source data fusion, so that the burden of end side deployment is further increased, and the lightweight and real-time constraints are not facilitated. Therefore, how to construct a self-training evolution method capable of forming low-cost sustainable iteration between cloud edges and maintaining stability under abrupt change conditions on the premise of not remarkably increasing the complexity of an edge side model and the data return cost and aiming at the complex problem of concentrated burst of misjudgment and lack of reliable labels caused by abrupt change period is still a technical problem to be solved in intelligent industrial edge landing. Disclosure of Invention The embodiment of the invention aims to provide a cloud edge collaborative self-training evolution method of a lightweight model for an industrial scene, which at least solves the problems of uncontrollable self-training caused by misjudgment concentrated burst, lack of reliable labels on the edge side and excessive total feedback cost of the model in an abrupt change period in the industrial scene, so that the model can be continuously optimized under the condition of low cost, and the real-time performance and stability of end side inference can be maintained. In order to achieve the aim, the first aspect of the invention provides an industrial scene-oriented lightweight model cloud edge collaborative self-training evolution method, which comprises the steps of obtaining continuous input data of an industrial edge node, executing edge reasoning processing one by one to generate reasoning results, writing the reasoning results into a reasoning log, executing window statistics processing according to time sliding windows based on the reasoning log to generate window statistics results of each time window and mutation abstracts